Patent Application
20240060990
Combination therapies have the potential to overcome acquired resistance to drugs. While single-cell RNA-seq, flow cytometry, mass cytometry, and western blots have provided cell compositions and drug responses of signaling proteins to study acquired resistance, subcellular and in-situ cellular distribution of physical signaling events and interactions are inefficiently captured in single cells due to the lack of spatial context of individual cells. Rationally designed combination therapies would necessitate an emerging validated multiplexed signaling assays directly in biological samples under drug perturbations.
Signaling networks are emerging approaches to studying the cellular alterations and decision-making in cancers. While the RNA expression data have suggested indirect networks of signaling proteins in many blood and solid cancers, single-cell mass cytometry, and phospho-flow have revealed direct networks of major signaling pathways and cascades at the protein level. However, these approaches demonstrate the signaling activity in cell suspensions without the location information about where cells reside and what their neighbors are in tissues and organs. In the meantime, image-based cell reports have been barcoded to dissect the signaling networks in live cells. These signaling network concepts achieve multiplexing of 100 signaling events, but they do not apply to fixed cells and tissues clinically used in studying signaling responses. Thus, multiplexed imaging methods are needed to visualize signaling networks even in archival formalin-fixed paraffin-embedded (FFPE) and fresh frozen patient tissues.
Current methods to analyze signaling pathways and receptor distributions include RNA sequencing and proteomics assays. However, this trend generally averages cellular information, causing loss of granularity due to single-cell heterogeneity. Flow cytometry and single-cell mass cytometry have then provided alternative approaches for single-cell analysis of signaling states and surface receptors; however, each cell's spatial information is lost due to the dissociation of the cell during measurements. Most molecular imaging methods have been limited to detecting a few signaling markers due to devices' spectral limitations, making it a challenge to visualize subcellular structures but with a limited number of biomarkers or signaling proteins. Thus, single-cell analysis of signaling mechanisms would benefit from emerging highly multiplex RNA and protein profiling technologies to decipher multiple signaling pathways. Spatial genomics assays have utilized image-based RNA barcoding based on multiplexed sequential fluorescence in-situ hybridization (FISH), providing ligand-receptor interactions in spatially resolved data. However, multiplexed RNA signatures would still suffer from indirect signaling networks. Imaging mass cytometry (IMC) and multiplexed ion beam imaging (MIBI) quantify up to 36 proteins using isotope-labeled antibody libraries imaged by specialized equipment. Besides, high-dimensional fluorescence imaging maps up to 50 proteins using DNA-barcoded imaging (CODEX), multiplexed fluorescence microscopy (MxIF), cyclic and sequential immunofluorescence techniques. These methods typically require specialized reagents and time-consuming labeling procedures, making them limited to research labs for discovery purposes.
Spatially resolved multiplexed imaging methods (IMC, MIBI, CODEX, and CyCIF) provided surprising single-cell heterogeneity in tissues with low-resolution imaging of 0.5 to 1-μm. Even cyclic immunofluorescence (CyCIF) needs biological samples to be manually removed from the sample holders, followed by staining outside the microscope and then placing the specimen back to the microscope for imaging. While protein-protein interaction (PPI) networks have yielded bulk level signaling interactions, automated and rapid multiplexed imaging-based direct signaling protein mapping technologies are needed to visualize the signaling factor colocalizations and subcellular protein interactions in single cells at high-resolution optical resolution.
These needs and others are at least partially satisfied by the present disclosure.
Disclosed herein is a molecular screening platform for imaging signaling proteins of human cell cultures and tissues from physiologically relevant cancer models using multiplexed measurement techniques to record spatial patterning of subcellular and cellular distributions in patients' biopsies with gene mutations. The biological lung cancer model system demonstrates how subcellular wiring of signaling factors provides subset patients that can be used for personalized and precision signaling therapies in lung cancers. This provides an advanced multiplexed protein imaging method to provide rapid experiments, automated microfluidics-based re-labeling and removal of signal, and imaging on the same platform. Here, a rapid multiplexed immunofluorescence (RapMIF) method is developed to visualize many molecules (30 to potentially 100 parameters) at high optical resolution. Protein interaction (RapMPI) assays are also used to resolve signaling networks of protein-protein interactions (PPIs) in cell cultures and tissues.
In an aspect, provided is a method of detecting multiple protein interactions within a cell or tissue sample, the method comprising: a) applying to the cell or tissue sample at least a first targeting moiety and a second targeting moiety, wherein the first targeting moiety interacts with a first specific protein of interest and the second targeting moiety interacts with a second protein of interest in the cell or tissue sample, wherein an interaction between the first specific protein of interest and the second specific protein of interest causes the first targeting moiety and second targeting moiety to be in close proximity to each other, wherein when the first targeting moiety and the second targeting moiety are in close proximity to each other, at least one detectable signal is produced; b) imaging the detectable signal or signals in the cell or tissue sample; c) deactivating the detectable signal or signals; d) repeating steps a), b), and c), wherein the first targeting moiety and the second targeting moiety are different in each repetition.
In another aspect, provided is a method of detecting subcellular spatial protein networks in a cell or tissue sample, the method comprising: a) applying to the cell or tissue sample at least one imaging moiety which can interact with a specific protein of interest in the cell or tissue sample, wherein when the at least one imaging moiety interacts with a specific protein of interest, it produces a detectable signal; b) imaging the detectable signal or signals in the cell or tissue sample; c) deactivating the detectable signal or signals; d) repeating steps a), b), and c), wherein the at least one imaging moiety is different in each repetition.
In another aspect, provided is a method of determining a signaling response in a cell or tissue, the method comprising: a) obtaining data on multiple protein interactions and subcellular spatial protein networks from a plurality of cell and tissue samples; b) analyzing the data using a machine learning algorithm; and c) using a neural network to create a signaling response model to an event in the cell or tissue.
In another aspect, provided is a system for multiplexed imaging comprising: a computer program capable of receiving at least one parameter from a user; a support for a cell or tissue sample; at least one port for delivering fluids to the support for the cell or tissue sample; and at least one port for aspirating fluids from the cell or tissue sample.
Other systems, methods, features and/or advantages will be or may become apparent to one with skill in the art upon examination of the following drawings and detailed description. It is intended that all such additional systems, methods, features and/or advantages be included within this description and be protected by the accompanying claims.
It is appreciated that certain features of the disclosure, which are, for clarity, described in the context of separate aspects, can also be provided in combination with a single aspect. Conversely, various features of the disclosure, which are, for brevity, described in the context of a single aspect, can also be provided separately or in any suitable subcombination. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. Methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present disclosure.
In this specification and in the claims that follow, reference will be made to a number of terms, which shall be defined to have the following meanings:
Throughout the description and claims of this specification, the word “comprise” and other forms of the word, such as “comprising” and “comprises,” means including but not limited to, and are not intended to exclude, for example, other additives, segments, integers, or steps. Furthermore, it is to be understood that the terms comprise, comprising, and comprises as they relate to various aspects, elements, and features of the disclosed invention also include the more limited aspects of “consisting essentially of” and “consisting of.”
As used herein, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to a “polymer” includes aspects having two or more such polymers unless the context clearly indicates otherwise.
Ranges can be expressed herein as from “about” one particular value and/or to “about” another particular value. When such a range is expressed, another aspect includes from the one particular value and/or to the other particular value. Similarly, when values are expressed as approximations, by use of the antecedent “about,” it will be understood that the particular value forms another aspect. It should be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint, and independently of the other endpoint.
As used herein, the terms “optional” or “optionally” mean that the subsequently described event or circumstance may or may not occur, and that the description includes instances where said event or circumstance occurs and instances where it does not.
For the terms “for example” and “such as,” and grammatical equivalences thereof, the phrase “and without limitation” is understood to follow unless explicitly stated otherwise.
In an aspect, provided is a method of detecting multiple protein interactions within a cell or tissue sample, the method comprising: a) applying to the cell or tissue sample at least a first targeting moiety and a second targeting moiety, wherein the first targeting moiety interacts with a first specific protein of interest and the second targeting moiety interacts with a second protein of interest in the cell or tissue sample, wherein an interaction between the first specific protein of interest and the second specific protein of interest causes the first targeting moiety and second targeting moiety to be in close proximity to each other, wherein when the first targeting moiety and the second targeting moiety are in close proximity to each other, at least one detectable signal is produced; b) imaging the detectable signal or signals in the cell or tissue sample; c) deactivating the detectable signal or signals; d) repeating steps a), b), and c), wherein the first targeting moiety and the second targeting moiety are different in each repetition.
As used herein, the term “protein interaction” refers to a physical contact between at least two proteins. In some aspects, the physical contact can be accomplished by one or more of: hydrophobic bonding, van der Waals forces, salt bridges, ionic interactions, electrostatic forces, hydrogen bonding, electron sharing, disulfide bonds, or other suitable intermolecular or intramolecular forces. In some aspects, a given protein interaction can exist for any length of time depending on the function of the protein interaction and the proteins involved. In some aspects, the function of a protein interaction can be electron transfer, signal transduction, membrane transport, cell metabolism, muscle contraction, cell signaling, or other biological functions.
As used herein, the term “targeting moiety” refers to any molecule or compound which binds only to a specific motif, functional group, or domain in a compound or molecule of interest. In some aspects, each targeting moiety binds to a specific protein of interest. In some aspects, a targeting moiety can leverage a physiologically occurring interaction, or a targeting moiety can be synthetically designed to elicit a specific interaction. In some aspects, a targeting moiety can comprise an antibody. In some aspects, a targeting moiety can further comprise an oligonucleotide. In some aspects, the oligonucleotide can be either covalently or non-covalently bound to the antibody. In some aspects, a part of a first oligonucleotide on the first targeting moiety is complementary to a part of a second oligonucleotide on the second targeting moiety. In some aspects, the first oligonucleotide is not complementary to the second oligonucleotide.
The first targeting moiety and the second targeting moiety being “in close proximity” to each other refers to the first targeting moiety and the second targeting moiety having a distance between them that allows a part of the first targeting moiety and a part of the second targeting moiety to have an interaction between each other that spans at least this distance. In some aspects, this interaction can include hydrophobic bonding, van der Waals forces, salt bridges, ionic interactions, electrostatic forces, hydrogen bonding, electron sharing, disulfide bonds, or other suitable intermolecular or intramolecular forces. In some aspects, the distance can be from about 1 nm to about 50 nm, including exemplary values of about 2 nm, about 3 nm, about 4 nm, about 5 nm, about 6 nm, about 7 nm, about 8 nm, about 9 nm, about 10 nm, about 12 nm, about 14 nm, about 16 nm, about 18 nm, about 20 nm, about 22 nm, about 24 nm, about 26 nm, about 28 nm, about 30 nm, about 32 nm, about 34 nm, about 36 nm, about 38 nm, about 40 nm, about 42 nm, about 44 nm, about 46 nm, and about 48 nm. In some aspects, the distance is equivalent to a space between the first targeting moiety bound to the first specific protein of interest and the second targeting moiety bound to the second specific protein of interest when there is an interaction between the first specific protein of interest and the second specific protein of interest.
As used herein, the term “detectable signal” refers to a chemical compound or complex that can be either directly or indirectly detected by visual or instrumental means. A detectable signal may produce a signal that can be detected, such as a fluorescent, chemiluminescent or radioactive signal. A detectable signal can also be detected visually, for example, colored dyes.
In some aspects, producing a detectable signal comprises: amplifying the oligonucleotides which are in close proximity to each other, wherein the oligonucleotides are amplified only when in close proximity to each other; and applying to the cell or tissue sample a fluorescent-labeled oligonucleotide, wherein the fluorescent-labeled oligonucleotide is complementary to at least a part of the amplified oligonucleotides; wherein the detectable signal comprises the fluorescent-labeled nucleotide coupled to at least a part of the amplified oligonucleotides. In some aspects, the step of amplifying the oligonucleotides further comprises a polymerase enzyme and/or a ligase enzyme. In some aspects, the step of amplifying the oligonucleotides further comprises the addition of an additional oligonucleotide. In some aspects, a first part of the additional oligonucleotide is complementary to a part of a first oligonucleotide on the first targeting moiety and a second part of the additional oligonucleotide is complementary to a part of a second oligonucleotide on the second targeting moiety. In some aspects, the step of amplifying the oligonucleotides comprises rolling circle amplification.
In some aspects, step a) takes from about 30 minutes to about 5 hours, including exemplary values of about 35 minutes, about 40 minutes, about 45 minutes, about 50 minutes, about 55 minutes, about 1 hour, about 1.25 hours, about 1.5 hours, about 1.75 hours, about 2 hours, about 2.25 hours, about 2.5 hours, about 2.75 hours, about 3 hours, about 3.25 hours, about 3.5 hours, about 3.75 hours, about 4 hours, about 4.25 hours, about 4.5 hours, and about 4.75 hours.
In some aspects, step b) is performed using brightfield microscopy, florescent microscopy, confocal microscopy, scanning electron microscopy, transmission electron microscopy, or any other suitable imaging platform.
As used herein, the term “deactivating” refers to a modification of any part of a targeting moiety, any part of a detectable signal, and/or any part of a protein of interest such that a detectable signal can no longer be detected. In some aspects, step c) comprises incubating the cell or tissue sample with a nuclease to remove any oligonucleotides. As used herein, the term “nuclease” refers to any enzyme that cleaves nucleotides. In some aspects, the nuclease is selected from DNase and Nuclease P1. In some aspects, step c) takes from about 30 minutes to about 5 hours, including exemplary values of about 35 minutes, about 40 minutes, about 45 minutes, about 50 minutes, about 55 minutes, about 1 hour, about 1.25 hours, about 1.5 hours, about 1.75 hours, about 2 hours, about 2.25 hours, about 2.5 hours, about 2.75 hours, about 3 hours, about 3.25 hours, about 3.5 hours, about 3.75 hours, about 4 hours, about 4.25 hours, about 4.5 hours, and about 4.75 hours.
In some aspects, the method further comprises a wash step after step a), step b), and/or step c). In some aspects, the wash step uses phosphate-buffered saline (PBS), water, formamide, ethylenediaminetetraacetic acid (EDTA), or other suitable buffers or solvents. In some aspects, the wash step is repeated with the same or different fluid.
In some aspects, steps a), b) and c) are repeated at least from about two times to about thirty times, including exemplary values of three times, four times, five times, six times, seven times, eight times, nine times, ten times, twelve times, fourteen times, sixteen times, eighteen times, twenty times, twenty-two times, twenty-four times, twenty-six times, and twenty-eight times. In some aspects, the first targeting moiety and the second targeting moiety being “different” in each repetition refers to either a different compound used for the first and/or second targeting moieties, or the same compound used for the first and/or second targeting moieties but applied to a portion of the cell or tissue sample that is different spatially or temporally.
In some aspects, step b) further comprises taking images of the cell or tissue sample at multiple spatial and/or temporal locations. As used herein, “multiple spatial locations” refers to multiple regions of the cell or tissue sample that vary in the x-direction, the y-direction, and/or the z-direction. In some aspects, the multiple spatial locations can cover the entire cell or tissue sample. In some aspects, the multiple spatial locations can cover only a portion of the cell or tissue sample. As used herein, “multiple temporal locations” refers to multiple points in time, wherein the multiple points in time can be spread out across minutes, hours, days, weeks, months, or years. In some aspects, the multiple points in time are spread out evenly. In some aspects, the multiple points in time are not spread out evenly.
In some aspects, the method further comprises analysis of the images using a machine learning algorithm. In some aspects, the method further comprises a step of detecting subcellular spatial protein networks in the cell or tissue sample.
In some aspects, the method further comprises: e) applying to the cell or tissue sample at least one imaging moiety which can interact with a specific protein of interest in the cell or tissue sample, wherein when the at least one imaging moiety interacts with a specific protein of interest, it produces a detectable signal; f) imaging the detectable signal or signals in the cell or tissue sample; g) deactivating the detectable signal or signals; h) repeating steps e), f), and g), wherein the at least one imaging moiety is different in each repetition.
As used herein, the term “subcellular spatial protein network” refers to the spatial and/or temporal organization of proteins in the cytoplasm and nucleus of a cell.
As used herein, the term “imaging moiety” refers to any compound, molecule, or complex that can be either directly or indirectly detected by visual or instrumental means. The imaging moiety may produce a signal that can be detected, such as a fluorescent, chemiluminescent or radioactive signal. The imaging moiety can also be detected visually, for example, colored dyes. In some aspects, the at least one imaging moiety comprises an antibody.
In some aspects, the antibody is directly labeled. As such, the antibody which can interact with the specific protein of interest is conjugated to the fluorescent dye before the antibody is applied to the cell or tissue sample. In some aspects, the antibody is indirectly labeled. As such, a first antibody which can interact with the specific protein of interest is not conjugated to the fluorescent dye, and a second antibody which can bind to the first antibody is conjugated to the fluorescent dye before either antibody is applied to the cell or tissue sample, and the second antibody is applied to the cell or tissue sample after the first antibody is applied to the cell or tissue sample.
In some aspects, the step of deactivating comprises fluorescent bleaching. As used herein, the term “fluorescent bleaching” refers to rendering any fluorescent dyes on the cell or tissue sample undetectable. In some aspects, fluorescent bleaching is performed using a solution containing hydrogen peroxide (H2O2), sodium hydroxide (NaOH), or any other suitable bleaches or any combinations thereof.
In some aspects, the method further comprises a wash step after any one of steps e), f), or g).
In some aspects, steps e), f) and g) are repeated at least from about two times to about thirty times, including exemplary values of three times, four times, five times, six times, seven times, eight times, nine times, ten times, twelve times, fourteen times, sixteen times, eighteen times, twenty times, twenty-two times, twenty-four times, twenty-six times, and twenty-eight times. In some aspects, the imaging moiety being “different” in each repetition refers to either a different compound used for the imaging moiety, or the same compound used for the imaging moiety but applied to a portion of the cell or tissue sample that is different spatially or temporally.
In some aspects, subcellular spatial protein networks in the cell or tissue sample are detected.
In some aspects, at least one of the steps is performed using an automated system.
In another aspect, provides is a method of detecting subcellular spatial protein networks in a cell or tissue sample, the method comprising: a) applying to the cell or tissue sample at least one imaging moiety which can interact with a specific protein of interest in the cell or tissue sample, wherein when the at least one imaging moiety interacts with a specific protein of interest, it produces a detectable signal; b) imaging the detectable signal or signals in the cell or tissue sample; c) deactivating the detectable signal or signals; d) repeating steps a), b), and c), wherein the at least one imaging moiety is different in each repetition.
In some aspects, the imaging moiety comprises an antibody. In some aspects, the antibody is preconjugated with a fluorescent dye and the detectable signal comprises the fluorescent dye. In some aspects, the antibody is directly labeled. In some aspects, the antibody is indirectly labeled.
In some aspects, the step of deactivating comprises fluorescent bleaching. In some aspects, fluorescent bleaching is performed using a solution containing hydrogen peroxide (H202), sodium hydroxide (NaOH), or any other suitable bleaches or any combinations thereof.
In some aspects, the method further comprises a wash step after any one of steps a), b), or c).
In some aspects, step a) takes from about 30 minutes to about 5 hours, including exemplary values of about 35 minutes, about 40 minutes, about 45 minutes, about 50 minutes, about 55 minutes, about 1 hour, about 1.25 hours, about 1.5 hours, about 1.75 hours, about 2 hours, about 2.25 hours, about 2.5 hours, about 2.75 hours, about 3 hours, about 3.25 hours, about 3.5 hours, about 3.75 hours, about 4 hours, about 4.25 hours, about 4.5 hours, and about 4.75 hours. In some aspects, step c) takes from about 30 minutes to about 5 hours, including exemplary values of about 35 minutes, about 40 minutes, about 45 minutes, about 50 minutes, about 55 minutes, about 1 hour, about 1.25 hours, about 1.5 hours, about 1.75 hours, about 2 hours, about 2.25 hours, about 2.5 hours, about 2.75 hours, about 3 hours, about 3.25 hours, about 3.5 hours, about 3.75 hours, about 4 hours, about 4.25 hours, about 4.5 hours, and about 4.75 hours.
In some aspects, steps a), b) and c) are repeated at least from about two times to about thirty times, including exemplary values of three times, four times, five times, six times, seven times, eight times, nine times, ten times, twelve times, fourteen times, sixteen times, eighteen times, twenty times, twenty-two times, twenty-four times, twenty-six times, and twenty-eight times. In some aspects, the imaging moiety being “different” in each repetition refers to either a different compound used for the imaging moiety, or the same compound used for the imaging moiety but applied to a portion of the cell or tissue sample that is different spatially or temporally.
In some aspects, step b) further comprises taking images of the cell or tissue sample at multiple spatial and/or temporal locations.
In some aspects, the method further comprises analysis of the images using a machine learning algorithm. In some aspects, at least one of the steps is performed using an automated system.
In another aspect, provided is a method of determining a signaling response in a cell or tissue, the method comprising: a) obtaining data on multiple protein interactions and subcellular spatial protein networks from a plurality of cell and tissue samples; b) analyzing the data using a machine learning algorithm; and c) using a neural network to create a signaling response model to an event in the cell or tissue.
As used herein, the term “signaling response” refers to changes in the spatial and/or temporal organization of proteins and/or protein interactions in a cell or tissue sample to respond to a given event.
In some aspects, the data is obtained via any of the disclosed methods.
In some aspects, the machine learning algorithm is selected from data clustering, t-distributed stochastic neighbor embedding (tSNE), and uniform manifold approximation and projection (UMAP).
In some aspects, the machine learning algorithm is a convolutional neural network (CNN).
In some aspects, the event is administration of at least one therapeutic agent, a method of treatment, or onset of a disease. In some aspects, the method of treatment is exposure to radiation.
In some aspect, the method is used for screening a therapeutic agent, observing response to a therapeutic agent, observing response to a method of treatment, disease diagnosis, or stem cell therapy. In some aspects, the method of treatment is exposure to radiation.
In another aspect, provided is a system for multiplexed imaging comprising: a computer program capable of receiving at least one parameter from a user; a support for a cell or tissue sample; at least one port for delivering fluids to the support for the cell or tissue sample; and at least one port for aspirating fluids from the cell or tissue sample.
In some aspects, each of the at least one port for delivering fluids and the at least one port for aspirating fluids further comprise a valve. In some aspects, each valve is controlled by the computer program.
In some aspects, the system can be coupled to a microscope. In some aspects, the microscope can be selected from a brightfield microscope, florescent microscope, confocal microscope, scanning electron microscope, transmission electron microscope, or any other suitable microscope.
In some aspects, the computer program executes any of the disclosed methods.
In some aspects, the at least one parameters are selected from: spatial position of the support; number of repetitions; time of the step of applying at least one detectable composition; number of images in the step of imaging; time of the step of deactivating; and/or wait time between any of the steps.
In some aspects, each port further comprises a fluidic tube and a needle tip. In some aspects, the system further comprises at least one port for preventing overflow using a vacuum.
Cells: A549 and PC9 cell lines were seeded on coverslips treated with 0.01% poly-l-lysine in a 6-well plate overnight, and cultured for 48 hours respectively at a 37° C. incubator. A549 was purchased from ATCC. The NSCLC-sensitive cell line, PC9, was provided by Dr. Sun Shi-yong (Emory University).
Drugs: The Osimertinib was provided by Dr. Sun Shi-yong (Emory University) at 10 mM concentration. The cells were treated with 0, 20, 40, and 60 nM Osimertinib for 48 hours in a 37° C. incubator. Following that, the cells were then fixed and permeabilized with 1.6% formaldehyde in 1×PBS for 10 minutes at RT and 0.5% Triton X-100 for 10 minutes at RT, respectively (
Tissues: FFPE lung cancer microarray tissue blocks were purchased from Biomax with 61 cores and 21 cases (BS04081a). The slide was baked in a 60° C. oven for 2 hours followed by deparaffinization and heat-induced epitope retrieval (HIER). Sectioned slices were ready for multiple rounds of staining, imaging, and bleaching.
Antibodies: A total of 26 antibodies were used. Cell segmentation included Phalloidin (A34055, Thermofisher Scientific), wheat germ agglutinin (WGA, W32466, Thermofisher Scientific), Concanavalin A (C11252, Thermofisher Scientific), β-tubulin (sc-5274, Santa Cruz Biotechnology), and β-actin (4970BF, Cell Signaling Technology). For tissue tumor region identification: pan-cytokeratin (914204, BioLegend) was used. 11 Markers targeting the WNT/β-catenin pathway included non-β-catenin (Ser45) (70034S, Cell Signaling Technology), APC (ab239828, Abcam), DKK1 (sc-374574, Santa Cruz Biotechnology), Cyclin E (sc-247, Santa Cruz Biotechnology), EMMPRIN (sc-21746, Santa Cruz Biotechnology), WNT1 (sc-514531, Santa Cruz Biotechnology), p-β-catenin (Ser45) (44-208G, Thermofisher Scientific), RNF43 (ab84125, Abcam), DKK2 (PA5-11616, Thermofisher Scientific), AXIN1 (2087BF, Cell Signaling Technology), Cyclin D1 (ab156448, Abcam). 6 Markers involved AKT/mTOR pathway included p-AKT (Ser473) (4060BF, Cell Signaling Technology), AKT (4691BF, Cell Signaling Technology), p-mTOR (Ser2448) (5536BF, Cell Signaling Technology), mTOR (2983BF, Cell Signaling Technology), p-EGFR (Tyr1086) (36-9700, Thermofisher Scientific), EGFR (sc-373746, Santa Cruz Biotechnology). 3 epigenetic markers included H3K27me3 (9733BF, Cell Signaling Technology), H3K9ac (ab203951, Abcam), H3k4me3 (ab227060, Abcam). For rapid staining of the sample, the BSA-free primary antibodies were pre-conjugated with Alexa Fluor 488, 555, or 647 using the Lightning-Link™ Rapid Conjugation kit (ab236553, ab269820, ab269823). Before conjugation, the concentrations of the antibodies were measured using Nanodrop to ensure the required antibody amount was in the range. The overall time for the conjugation process was around 20 minutes. 15 antibodies involved in signaling pathways were first tested in IF of A549 cells. For conjugated antibodies, IF was tested under four conditions, indirect/direct labeling, overnight 4° C./RT 1 hour, respectively. For antibodies conjugated to Alexa Fluor dyes, the direct labeling was tested at overnight 4° C./RT for 1 hour. To improve the staining quality, titrations were performed on antibodies, and IF was evaluated at two different dilution rates. The improved staining condition of each antibody was determined based on the SNR enhancement. The SNR of each condition was measured based on calculations of the max and min pixel values of five random cell regions in ImageJ (TABLE 1). For IF staining on cell culture, the antibodies and DAPI (62248, Thermofisher Scientific) were diluted in CSM containing 0.5% BSA, and 0.02% sodium azide in PBS. For IF staining on tissue, the antibodies and Hoechst 33342 (H3570, Thermofisher Scientific) were diluted in a protein block buffer (DAKO, X0909).
RapMIF: The RapMIF setup consisted of an autosampler that transferred antibodies and performed washes of the sample automatically and controlled by a custom-written python code (
Cell multiplexing: Each coverslip prepared with cells was stained with 25 markers using a total of 11-12 cycles (
To examine the efficiency of bleaching, the signaling levels of p-AKT-488, Cyclin E-555, and EGFR-647 before and after bleaching were quantified (
Tissue multiplexing: RapMIF is compatible with tissues mounted on coverslips and the tissue holder would be processed on a cooled stage holder (
Imaging: A wide-field microscope, Keyence BZ-X700, was used for fluorescence and brightfield imaging. For each cycle of imaging, four channels were captured with 40% excitation light: Channel 1 detects DAPI/Hoechst at 360 nm, Channel 2-4 detects fluorophores at Alexa Fluor 488 nm (FITC), 555 nm (TRITC), and 647 nm (Cy5), respectively. The exposure time was varied, but the exposure time for each marker across control and drug-treated PC9 cells was consistent. The exposure time for Channel 1, DAPI/Hoechst was around 1/400s, and Channel 2-4 varied from is to 1/20s. The sample for cells grown on the coverslip was imaged at a 40× oil lens for 48-62 region-of-interest (ROI), and each ROI was imaged across 25-30 z-stacks with 0.4 μm/stack. To determine the optimal step size for imaging, A549 cells stained with WNT1 marker were imaged at four different step sizes, 0.1-μm, 0.4-μm, 0.8-μm, and 1.2-μm. The step size of 0.4-μm was chosen for the imaging system to reveal more continuity of the dim signaling molecules. Even though the acquisition time was around 1 hour for each cycle, and slower compared to images at a larger step size, the signaling molecules could be better resolved. The 40× oil lens achieved a high-resolution acquisition at 0.18872 μm/pixel. The tissue microarray was imaged at a 20× dry lens, each ROI was imaged using autofocus. The 20× dry lens provided a resolution of 0.37742 μm/pixel. The best focus z level was determined by calculating the blur score of each DAPI channel image. The blur score of each image was calculated by taking the fast Fourier transform (FFT) transform of the image and filtering out low frequencies, the resulting inverse Fourier transform image's mean magnitude was calculated. All z-level images of one location were then ranked based on the mean magnitude as a blurriness score. The best focus images around +/−2 z-stacks were used for signaling analysis. 3D subcellular spatial signaling levels were also assigned into 26-pixel clusters in 3D across 21 z-stacks, demonstrating the 3D distributions of signaling proteins as a proof of concept.
Background correction: To correct the uneven illumination caused by the flat and dark field, a BASIC algorithm in the ImageJ plugin was used to remove the line artifacts in each tissue core (Peng et al., 2017). Since the patterns of the uneven illumination across different channels were different, and the shape and size of tissue cores varied, the correction was applied per tissue core for each channel separately across 12 IF cycles. The software estimated the flat and dark fields and subtracted the raw images from the two fields to generate corrected images. Then, images were post-processed for quantification of signaling networks and intensity levels.
Segmentation: For cell culture, single-cell nuclei regions were segmented using Cellpose (Stringer et al., 2021) by cell painting markers (Bray et al., 2016). The best focus images around +/−2 z-stacks were used for segmentation of the cells. DAPI or Hoechst marker was used to segment cell nuclei boundaries. DAPI and segmentation markers were used to register the cells across the multiple cycles. In the A549 experiment, the Phalloidin channel showed strong background noise due to the effect of the bleaching solution, therefore p-EGFR, Concanavalin A, B-actin, WGA, APC, and WNT1 were combined to outline the cell boundary for segmentation. For PC9 data, Phalloidin marker was used for cell boundary segmentation. For FFPE lung cancer microarray tissue, DAPI was used to segment the nuclei region. The cytosol region was calculated by expanding the nuclei segmented region by 10 pixels. The 10 pixels were determined empirically by looking at nuclei segmentation and cytosol segmentation markers.
Image intensity: From nuclei and cytosol cell masks, the mean intensity of markers in each cell area was calculated. Each cell nuclei and cytosol region were separated and mean intensity for each marker was calculated and cell masks with a colormap showing the intensity were generated to visualize the variation across signaling markers within and outside the nuclei area.
Correlation analysis: Correlation between pairs of cells was performed based on the mean intensity level for a pair of markers across all the cells (
Marker intensity prediction: Per cell intensity of individual markers was predicted from other available markers using a Random Forest regressor with a 5-fold cross-validation training. The training set was composed of 2885 single-cell data points with a testing set of size 572. The regression was evaluated by the r-score between true intensity per cell compared to predicted intensity per cell. The r-value found was 0.76 with a p-value<0.01. Feature importance was plotted using the impurity index.
Pixel clustering: Pixel phenotypes were clustered in a two-step clustering pipeline (
Spatial signaling network: The network representation of spatial signaling was represented using the spatial information of pixel clustering. Each cluster was represented by a node with node size in the network proportional to the number of pixels in the corresponding cluster. For each cluster, the distribution of the neighbor cluster was calculated by looking at the 2-hop neighbor's cluster of all pixels in the cluster. The edge color represented low (blue) to high (red) neighboring probability between two clusters.
Subpopulation clustering: A549 and PC9 dataset subpopulations were defined based on signaling phenotype using unsupervised clustering with the Leiden algorithm from multiplex protein mean expression level per cell (Traag et al., 2019). Both A549 and PC9 datasets were clustered into 12 clusters.
Receptor discontinuity: The centroid of nuclei of each cell was calculated from each cell mask. Then each cell was divided into regions defined by the absolute distance from the centroid. In each region, the mean intensity of the marker was quantified as a function of distance from the nuclei centroid. Distances from nuclei centroid in each cell were shown as a radius map and overlaid with marker intensity with a custom colormap. The colormap for marker intensity was transformed from matplotlib package ‘greens’ with alpha set using the NumPy function linspace going from 0 to 1 with N equal to the number of RGB quantization levels in the colormap.
Cell survival: The effect of Osimertinib on PC9 cell proliferation and survival was measured using ImageJ. The cell survival was measured based on the cell number on the coverslip. The cells were seeded on coverslips at the same starting concentration, 0.833×105, followed by 48-hour treatment of Osimertinib at 0, 20, 40, and 60 nM. The cells were stained with p-EGFR, and β-tubulin and three ROIs were imaged for each concentration at 40×. The cell number on each ROI was counted in ImageJ following the steps: import the Channel image, make binary, convert it to masks, analyze particles, and then choose “masks” under the “show” option.
Statistical testing: The details of statistical tests employed in each case were provided in the figure captions. All P values were corrected for multiple testing and the statistical testing method was indicated in the figure captions. The following convention was used to indicate significance with asterisks: not significant (ns) (P>0.1), * (0.1>P>0.01), ** (0.01>P>0.001), *** (0.001>P>0.0001), and **** (P<0.001).
Cells: A549 cells were used for antibody optimization. The cells were seeded on coverslips treated with 0.01% poly-l-lysine in a 6-well plate overnight at a 37° C. incubator. A549 was purchased from ATCC. NSCLC-sensitive cell line, HCC827 was provided by Dr. Sun Shi-yong (Emory University). HCC827 cells were seeded on coverslips in a 6-well plate overnight, followed by Osimertinib treatment.
Drugs: The Osimertinib was provided by Dr. Sun Shi-yong (Emory University) at 10 mM concentration. HCC827 cells were treated with 100 nM Osimertinib at different time points in a 37° C. incubator. Following that, the cells were then fixed and permeabilized with 1.6% formaldehyde in 1×PBS for 10 minutes at RT and 0.5% Triton X-100 for 10 minutes at RT, respectively. The cells were then blocked using a cell staining medium (CSM) containing 0.5% BSA, and 0.02% Sodium Azide in PBS. The effect of Osimertinib on p-ERK(T202/Y204) was assessed. HCC827 cells were stained with p-ERK overnight followed by 1-hour incubation of secondary antibodies at RT, and 10 mins DAPI staining. The cells were then imaged at 40× magnification, and the intensity levels were analyzed. For the multiplexing experiment, HCC827 cells were treated with 100 nM Osimertinib for 12 hours in a 37° C. incubator, followed by RapMPI. The coverslip was mounted to an acrylic holder for multiple cycles of staining.
Tissues: HCC827-derived mouse xenografts were provided by Dr. Sun Shi-yong (Emory University). The mice were treated with Osimertinib for 1 week and 2 months respectively. Osimertinib was given to mice daily. Mice were then euthanized, and the tumor was embedded in OCT for sectioning. The OCT tissues were fixed in acetone, rehydrated, permeabilized in 0.4% Triton X-100, blocked, and ready for multiplexed experiments.
Antibodies: A total of 27 antibodies were used. 9 PPIs targeting the AKT/mTOR, MEK/ERK, and YAP/TEAD1 pathways included TEAD1 (12292BF, Cell Signaling Technology), YAP (ab172373, Abcam), Cyclin E (sc-247, Santa Cruz Biotechnology), CDK2 (sc-6248, Santa Cruz), p-ERK (T202+Y204) (ab242418, Abcam), c-Myc (5605BF, Cell Signaling Technology), p-AKT (Ser473) (4060BF, Cell Signaling Technology), mTOR (2983BF, Cell Signaling Technology), Mcl-1 (66157BF, Cell Signaling Technology), Bak (12105BF, Cell Signaling Technology), Cyclin D1 (66467, Cell Signaling Technology), CDK4 (23972, Cell Signaling Technology), NF-κB p65 (69994SF, Cell Signaling Technology), p-p90RSK (Ser380) (11989BF, Cell Signaling Technology), Bim (26184SF, Cell Signaling Technology), Tom20 (sc-17764, Santa Cruz), Oct4 (ab240358, Abcam), Sox2 (ab243909, Abcam). 10 of protein markers were used for RapMIF, including p-EGFR (Y1068) (ab205827, Abcam), Tom20 (sc-17764, Santa Cruz), Ki67 (ab283699, Abcam), Pan-cytokeratin (53-9003-82, Invitrogen), Golph4 (ab197595, Abcam), NBD-C6 (N22651, Thermofisher), Cox IV (ab197491, Abcam), Phalloidin (A34055, Thermofisher Scientific), wheat germ agglutinin (WGA, W32466, Thermofisher Scientific), and Concanavalin A (C11252, Thermofisher Scientific). For RapMIF, the antibodies were either purchased in preconjugated versions or conjugated with Alexa Fluor 488, 555, or 647 using the Lightning-Link™ Rapid Conjugation kit (ab236553, ab269820, ab269823). For RapMPI, the carrier-free antibodies were pre-conjugated with Duolink Probemaker (single color: DU092009 Sigma; multicolor: DU096020 Sigma). Before conjugation, the concentrations of the antibodies were measured using Nanodrop to ensure the required antibody amount was within the range. All antibodies involved in signaling pathways were first tested by IF in A549 cells. For unconjugated antibodies, IF was tested at overnight 4° C./RT 1-hour conditions. To improve the staining quality, titrations were performed on antibodies, and IF was evaluated at two different dilution rates. To optimize the antibody conditions, A549 cells were seeded on coverslips for antibody staining, followed by DAPI (62248, Thermofisher Scientific). For IF staining on tissue, the antibodies and Hoechst 33342 (H3570, Thermofisher Scientific) were diluted in a protein block buffer (DAKO, X0909).
RapMPI on cell cultures: For single color detection, each coverslip prepared with cells was stained with multiple PPIs with one PPI per cycle. The cells were fixed, permeabilized, and then ready for the following staining steps: 1) block the sample with Duolink blocking solution; 2) incubate it with one pair of proteins, conjugated to one PLUS and one MINUS oligonucleotide at 4° C. overnight. The antibodies were diluted in Duolink antibody diluent at the preferred dilution rate; 3) incubate the sample with ligase for 30 mins at 37° C.; 4) amplify the signals for 100 mins at 37° C.; 5) stain the sample with DAPI, and the sample was ready for imaging.
For multicolor detection, after permeabilization, the cells underwent the following steps: 1) block the sample with Duolink blocking solution; 2) incubate it with two pair of proteins, conjugated to one pair of oligonucleotides at 4° C. overnight. The antibodies were diluted in Duolink antibody diluent at the preferred dilution rate; 3) incubate the sample with ligase for 30 mins at 37° C.; 4) amplify the signals for 100 mins at 37° C.; 5) incubate the sample with the detection buffer for 30 mins at 37° C.; 6) stain the sample with DAPI, and the sample was ready for imaging. After imaging, the samples were incubated with DNase at a 1:50 dilution rate for 4 hours at RT, followed by 3 times 30% formamide washes and 3 times 1×PBS washes.
RapMPI on tissues: RapMPI is compatible with tissues mounted on slides. The staining settings were the same as the RapMPI in cell cultures. The blocking step between cycles for tissue multiplexing was skipped to avoid hiding the signals. The tissue samples were mounted with 10% glycerol made in 1×PBS and covered the slide using a 24×50 mm No. 1 coverslip (3322, Thermo Scientific) to prevent evaporation during imaging. To de-coverslip the sample after imaging, the slide was placed in a vertical jar containing 1×PBS for around 10 minutes, and the coverslip was released due to gravity.
RapMIF: Following RapMPI, RapMIF was performed to profile Pan-cytokeratin, ki67, Tom20, p-EGFR, Golph4, Bim, Concanavalin A, Phalloidin, and WGA. The settings were the same as described previously20. Between cycles, the fluorophores were deactivated using and 3% H2O2 and 20 mM NaOH mixture made up in 1×PBS for 1 hr at RT in the presence of white light, followed by 3 times 1×PBS wash (for cell cultures: 3% H2O2 and 20 mM NaOH; for tissues: 4.5% H2O2 and 24 mM NaOH). For tissue multiplexing, after IF staining, H&E staining was performed at the last cycle.
Nuclease P1 stripping: Nuclease P1 has been examined as an alternative to DNase I. Following the imaging of PPI, the samples were incubated with nuclease P1 for 30 mins at 37° C. The samples were then washed with 20 mM EDTA three times, followed by 3 times 1×PBS wash.
H2 buffer: Add 30 ml of 5 M NaCl solution, 10 ml of 1 M Tris (pH 7.5), 0.943 ml of Triton X-100, 2.03 g of MgCl2·6H2O and 0.02% (wt/vol) NaN3 to 960 ml of ddH2O335.
DMSO stripping solution: DMSO has also been tested to remove oligos and PLA signals. Following the imaging of PPI, the samples were incubated with Hybridization buffer (100 ml of DMSO with 400 ml of H2 buffer) for 1 min, followed by stripping buffer (62.5 ml of H2 buffer to 187.5 ml of DMSO) for 10 mins at RT. The samples were then washed three times with 1×PBS 35.
Imaging: A wide-field microscope, Keyence BZ-X800, was used for fluorescence and brightfield imaging. The fluorescent signals were detected by five filters with an excitation spectrum of Alexa Fluor 360 nm, 488 nm (FITC), 555 nm (TRITC), 590 nm (Texas Red), and 647 nm (Cy5). The exposure time was varied, but the exposure time for each marker across control and drug-treated HCC827 cells was consistent. The sample for cells grown on the coverslip was imaged at a 40× oil lens for 36 region-of-interest (ROI), and each ROI was imaged across 25-30 z-stacks with 0.4 m/stack. The whole slide tissue was imaged at a 20× dry lens, each ROI was imaged using autofocus and imaged at a 40× oil lens across 25-30 z-stacks with 0.4 m/stack. The resolution for the 20× dry lens and 40× oil lens was 0.37742 m/pixel and 0.18872 m/pixel respectively.
Image Processing: For 2D maximum projection images, stitched images provided by BZ-X800 Analyzer were used. The Hoechst channel from each cycle was used to register the images using a phase cross-correlation algorithm. On the other hand, for 3D per z-stack image processing, ROI images of 1024 by pixels were stitched by using a 30% overlap ratio using a code based on ASHLAR39. After stitching all cycle images were registered per z-stack as the microscope captured each ROI z-stack image at once.
Cell segmentation: Two distinct methods were used for single-cell segmentation for cell culture and tissue images. In cell culture images, Cellpose40 deep learning algorithm was used whereas in tissue images Mesmer41 algorithm from Deepcell42 package was used for single-cell segmentation. Single cells in cell culture are more homogenous with more defined cell boundaries whereas cells in mouse tissue exhibited more variation in cell shapes and sizes. Cellpose was chosen for cell culture data and Mesmer for tissue data because the two algorithms were pre-trained on corresponding data modalities. Hoechst was used for nuclei segmentation and p-EGFR as a cytosolic marker.
PPI detection: PPI signals were detected using a custom algorithm leveraging a traditional image processing pipeline. More specifically, each PPI image was preprocessed as followed: images were first transformed using a top-hat filter of size 3 pixels to reduce the noise around PPI signals, then Laplacian of Gaussian images was used to detect bright local maxima as PPI signals. In 3D images, double detection of PPI was filtered out at the same position in consecutive Z stacks. After detecting all the PPI locations in 3D, PPI was observed within a 2.5-pixel radius (0.045 um) by creating a PPI neighborhood graph of radius threshold 2.5 pixels and getting all connected components in the graph. Finally, the mean position of the unique connected components in the graph was extracted as identified unique PPI signals in 3D.
PPI cell assignment in tissue samples: In the mouse tissue data, the cell membrane segmentations were not capturing the whole cell area. To avoid loss of PPI information due to cell segmentation error, a nearest-distance-based assignment method was used to assign the PPI signal detected and not in any cell segmentation mask. For each PPI signal not in any segmented cell regions, the nearest cell mask pixel was observed and if the distance was lower than a user-defined value, it was assigned to the cell to which the nearest cell mask pixel belonged.
Spatial Graph Construction: For each detected PPI in a cell, the corresponding 2D or 3D localization was extracted, and a node was assigned in the created graph. The node labels are assigned by creating a one-hot encoding of the corresponding PPI detected for the node. Delaunay triangulation was used to create edges connecting the nodes in the graph and therefore create a PPI spatial graph for each cell. The models are trained in a multi-instance learning framework, that is a cell label for each instance was assigned based on the cell treatment condition of a group.
SpPPI-GCN model: For the graph neural network, a 2-layer network was used consisting of graph convolutional layers of embedding size 16 for 5 PPI dataset 32 for 9 PPI dataset. The input of the model is the generated PPI graph for each cell with node feature represented by the PPI one-hot encoding. Each layer transforms the input as the following function: Hl+1=f(WlHlA*) with l the corresponding layer, f the activation function, Hl the node embedding matrix, Wl the weight matrix of the layer l, and A* the spectral normalized adjacency matrix. The spectral normalized adjacency matrix can be obtained with the following formula: A*=D−1/2AD1/2 with A the corresponding adjacency matrix and D the degree matrix of A. Finally, the node embeddings are then aggregated by a pooling layer and two dense layers were then used to obtain prediction at the cell level (i.e., graph-level prediction).
Multi-instance learning baseline: A multi-layer perception baseline was used to compare the spPPIGCN network. In the same multi-instance learning framework, a class label from the cell treatment condition was assigned. The input of the model is the generated PPI graph for each cell with node feature represented by the PPI one-hot encoding. Here, stacked dense layers of embedding size 16 for 5 PPI dataset and 32 for 9 PPI dataset were used to obtain a node embedding. Each layer transforms the input as the following function: Hl+1=f (WlHl) with l the corresponding layer, f the activation function, Hl the node embedding matrix and Hl the weight matrix of the layer l. Finally, the node embeddings are then aggregated by a pooling layer and two dense layers were then used to obtain prediction at the cell level (i.e., graph-level prediction).
Multi-layer perception baseline: On the other hand, another multi-layer perception model was trained on the total PPI count per event for each cell using the same embedding dimension and number of layers. The input is the sum of each PPI class in (1) whole cell (5 and 9 input size for 5 PPI and 9 PPI dataset respectively) or divided by (2) cytosol and nuclei regions (10 and 18 input size for 5 PPI and 9PPI dataset respectively). The model output is the predicted treatment condition.
Machine learning baseline: Several machine learning models were used for a baseline on a total number of PPI events which includes Naïve Bayes, Random Forest, AdaBoost, Decision Tree, Support Vector Machine (SVM) and Gradient Boosting. The scikit-learn python library with default setting was used when training and testing these machine learning models. The total number of PPI events per whole cell was tested (1) and divided by subcellular regions (2) (Nuclei/Cytosol). Therefore, the input is the sum of each PPI class in (1) whole cell or divided by (2) cytosol and nuclei regions. The model output is the predicted treatment condition.
Prediction metrics: To compare the models' prediction abilities, a 5-fold cross-validation setting was used by separating the dataset into an 80% training set and a 20% validation set. Accuracy, ROC AUC score, and F1 score were used as metrics to evaluate the data prediction in the validation sets.
Statistical testing: The details of statistical tests employed in each case were provided in the figure captions. All P values were corrected for multiple testing and the statistical testing method was indicated in the figure captions. The following convention was used to indicate significance with asterisks: not significant (ns) (P>0.1), * (0.1>P>0.01), ** (0.01>P>0.001), *** (0.001>P>0.0001), and **** (P<0.0001).
Spatial signaling networks in single cells can validate previously established signaling alterations and provide new signaling coordination in EGFRm NSCLC cells (e.g., PC-9) under control, Osimertinib treatment only, MEK/ERK, or YAP inhibition only, and Osimertinib combined with MEK/ERK or YAP inhibition conditions. Signaling pathways mediate cell communication and cell functions such as cell proliferation, differentiation, and migration. Although the correlation between signaling markers' expression levels was extensively studied, signaling factors' spatial relation is still unclear. Also, the cell-to-cell heterogeneity in the signaling transduction could not be systematically profiled due to technical limitations.55 Therefore, RapMIF can be utilized to resolve the subcellular spatial protein analysis in the MEK/ERK and YAP/Hippo signaling pathways (i.e., chosen because MEK/ERK and YAP use independent and dependent activation that can lead to secondary resistance) in established EGFRm PC-9 (responsive), PC-9/AR (first-line treatment resistance) and PC-9/GR/AR (second-line treatment resistance) human cell lines as key models to test drug treatments for EGFRm lung cancers. EGFR-TKI drug Osimertinib has been clinically tested. A portion of the patients suffers from resistance to the Osimertinib after first-line treatments. This resistance poses a challenge in clinical treatments of NSCLC patients with EGFR mutations. YAP in the Hippo pathway remains active under EGFR TKI and EGFR/MEK inhibitions, promoting the cells to enter a senescence-like dormant state in the absence of EGFR downstream signaling.10 Thus, recent combinatorial interventions using MEK/ERK8 or YAP10 inhibitors have demonstrated promising results to overcome the acquired resistance to Osimertinib. Thus, the establishment of spatial signaling states per cell in PC-9 and H1975 can translate the findings to combinatorial signaling therapies potentially that can later be used in translational settings.
Design and approach: RapMIF targets signaling protein markers (50-plex) including ERK-pERK, ERK1/2, pERK1/2, Mcl-1, Bim, p90RSK, and p-p90RSK to quantify the signaling crosstalk of ERK, AKT, mTOR, EGFR, and Wnt pathways. The previously established targets7-9,11,56 for assessing the efficacy of combination drugs (e.g., ACK1, c-Myc, ERK1/2, MEK1/2, YAP1, CD74, and AXL), Hippo signaling related upstream factors57,58 (MST1/2, LATS1/2, YAP/TAZ, 14-3-3, and MOB1), and YAP-related apoptosis targets10 (e.g., BMF, BCL2L11, SAT1, ERBB3, TNFSF10, and PDCD40) is used to reconstruct the spatial signaling network. These targets are benchmarked by mass spectrometry and RNA sequencing results in NSCLC culture and in vivo models, allowing to define on a “targeted panel” from the transcriptome or proteome-wide datasets in Osimertinib therapy. The number of cells needed to accurately reconstruct the spatial signaling network can be determined. The accuracy of the network depends on the estimated pairwise distances of clusters, which is defined as the percentage of different clusters within the neighborhood of a given cluster. It is estimated that with 1000 cells (TABLE 4), the standard deviation of the distance estimates will be within 5% of the true value. Thus, 1000 cells, can accurately construct the network.
Signaling networks depend on the quality of detection antibodies and inhibitor efficacy in the image-based IF staining experiments. Thus, genetic perturbations7 can be used to validate one antibody and one inhibition target at a time before multiplexing experiments. Specifically, a signaling protein of interest (e.g., ACK1, c-Myc, ERK1/2, MEK1/2, YAP1, CD74, BMF, and AXL) can be targeted using small hairpin RNA (shRNA) or small interfering RNA (siRNA) to perform gene knockdown, providing robust suppression of the gene expression and reduction in the protein production levels. A sequence-specific shRNA (Millipore Sigma) for the gene of interest can be performed in pLKO1. lentiviral vector to knock down the gene target, providing the genetically perturbed PC-9 validation for the RapMIF. Similarly, a siRNA (Santa Cruz Biotechnology) for a gene of interest can be transfected into the PC-9 cell lines in 6-well plates using the HiPerFect transfection reagent (QIAGEN; Germantown, MD). After 48 h, the cells were collected and re-plated in 24-well plates for RapMIF experimental validations. Non-silencing scramble control siRNA (siCtrl) can be performed as before59.
Established human cell lines EGFRm PC-9 (responsive), PC-9/AR (first-line treatment resistance), and PC-9/GR/AR (second-line treatment resistance) serve as useful models to test drug treatments for EGFR mutant lung cancers. EGFR-TKI drug Osimertinib has been clinically tested. A portion of the patients suffers from resistance to the Osimertinib after first-line treatments. This resistance poses a challenge in clinical treatments of NSCLC patients with EGFR mutations. Thus, the establishment of signaling states per cell line translates the findings to clinical applications. It has been shown that PC-9 cell cultures can be used to study synergistic combinations of Osimertinib with ERK (VRT752271), MEK (Trametinib), or YAP/TEAD (MYF-01-37) inhibitors7,8,10. Cell apoptosis is enhanced by Osimertinib combined with VRT, Tram, and YAP/TEAD inhibitor drugs. Cell viability counts are performed by sulforhodamine B (SRB). Control experiments can be performed by cells growing in DMSO. Apoptosis can be measured by caspase/PARP cleavage using western blot (WB), and Annexin-V can be quantified by flow cytometry. Drug treatments can be performed in the 0-10 nM for Osim, 0-100 nM for Tram, 0-1000 nM for VRT, and 0-1000 nM for MYF-01-37 in PC-9 cell lines. Thus, RapMIF experiments are designed for 50-plex markers in this range to profile signaling networks in PC-9, PC-9/AR, and PC-9/GR/AR cells for a combination of these four drugs, yielding strategies to overcome the acquired resistance to Osimertinib. Given the marker expressions for each pixel from all cells, a regression-based analysis is run for all markers using drug concentrations and drug combinations as covariates (TABLE 4). This analysis provides markers that are associated with drug treatment.
RapMIF can perform 50-plex signaling measurements in PC-9, PC-9/AR, and PC-9/GR/AR cells using 24-combinatorial drug treatments with four different concentrations, each in a glass-bottom 96-well format during 0.5-2 days cell cultures. Quantification of cell survival is performed by SRB assay. Immunostaining of Annexin-V and Caspase 3 provides cells committing to apoptosis. In this analysis, the data from all the conditions is pooled together and clustered at once. This process reveals signaling networks on a reference map using heatmaps, t-distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP) representations. Spatial cross-correlation can determine the spatial similarity of multiple cell groups on the colored cluster images. RapMIF is then provide subcellular and cell-to-cell heterogeneity of these synergistic combinations to shed light on treatment for NSCLCs with EGFRm.
Experimental results: A panel of 8 cycles consisting of 20-markers that includes p-EGFR, Phalloidin, WGA, p-AKT, p-β-catenin, RNF 43, non-p-β-catenin, APC, DKK1, Cyclin E, EMMPRIN, Wnt1, AKT, p-mTOR, mTOR, DKK2, AXIN1, EGFR, Concanavalin A, and Cyclin D1 was designed for RapMIF experiments in EGFRm PC-9 cells. To reduce sample incubation time in indirect IF, the antibodies were conjugated with one of three Alexa dyes using Lightning-Link Rapid Alexa Fluor Antibody Labeling Kit (Novus). The order of markers in the experiment was decided based on the marker's brightness and expression level. After mounting, permeabilizing, and blocking the cells on a coverslip, the coverslip was ready for cyclic experiments. In the eight-cycle, the sample with antibodies was incubated for 30 minutes or 1 hour. Following imaging, the dyes were bleached in the NaOH—H2O2 NaBH4, and LiBH4 cocktails, and other fluorophore inactivation methods and antibody stripping methods were tested to make rapid switching per cycle. Even though the base-hydrogen peroxide mixture bleaches Alex 488, 555, and 647 at different rates, 30-45 min of treatment is sufficient. In the experiment, cell loss was not observed in any of the cycles. Following fluorophore inactivation, the coverslip was re-blocked with cell staining media, and the next process was performed.
An eight-cycle RapMIF experiment yielded 18-channel images of the samples at the single-cell level. The multiplex single-cell measurement reveals the interactions of signaling proteins in the subcellular networks. Pixel-level correlations of signaling protein expressions in the concatenated RapMIF data (see TABLE 4) provided co-regulation on a heatmap using z-scored spatially resolved data analysis. Note that pixel-level analysis is used to compute local protein density in each cell.
The RapMIF experiment was then performed in EGFRm NSCLC sensitive cell line, PC-9 cells. Osimertinib has shown a prominent effect in inhibiting PC-9 cells, and the cell survival rate has decreased as the concentration has increased from 0 to 10 nM under a 3-day treatment7,8. PC-9 cells responding to Osimertinib treatment have been shown to develop acquired resistance. A preliminary study investigated whether p-EGFR (Tyr1086) would be inhibited by Osimertinib in PC-9 cells. Osimertinib would then affect the proliferation and growth of the PC-9 cells. Under the incubation of 48-hour drug treatment, the number of cells adhering to the coverslips has decreased as the drug concentration has increased. Following drug treatments, the RapMIF data provided the imaging maps of up to 25-plex signaling proteins in PC-9 with (n=451 cells) and without (n=531 cells) the treatment of 40 nM Osimertinib. Validation experiments decreased the p-EGFR expression as the drug concentration increased from 0 to 20 nM to 40 nM, whereas p-EGFR exhibited relatively higher expression under 60 nM than 40 nM. Thus, 40 nM Osimertinib was chosen for RapMIF multiplexing. The Osimertinib treatment increased the WNT1 signaling, while the p-EGFR (Tyr1086) alterations remained minimally enhanced in the heatmap of combined clustering from the signaling factor maps of untreated and 40 nM Osimertinib PC-9 samples. The heatmap indicated the single-cell alterations of signaling in response to Osimertinib treatment, providing sub-groups of control (blue) and drug-treated (red) in the normalized heatmap (dark blue to light orange). The results validated that signaling intensities concurrently alter in Osimertinib drug treatments. After showing similarities of intensities in the multiplexed data at the single-cell level, the pixel-level co-regulation of signaling subsets was determined using clustering analysis in the next analysis.
Spatial signaling factor expression levels in the pixel-level RapMIF datasets were concatenated together with untreated (control) and drug-treated cells, allowing the comparisons of pixel clusters (identified by UMAP) containing a subset of signaling networks modules across different conditions of PC-9 experiments. An interesting finding was that the spatial signaling network maps of the PC-9 cells after 40 nM Osimertinib drug treatment exhibited more dispersed (i.e., less aggregated) distribution in the cytosol, noted as clusters 0, 1, and 2, compared to the control sample. Subcellular spatial networks (identified by pixel neighbors of clusters) were represented as a physical neighborhood map for the assessment of protein-protein proximity at the pixel level for RapMIF data from the control (untreated) and 40 nM Osimertinib treated PC-9 cells. Each node corresponds to one of the cluster phenotypes of pixels (see TABLE 4, Cluster number #column) in the form of cluster number. The color of the node matched with the cluster color assigned to the pixel's identity. The length of the connected line represented the pixel's spatial neighborhood in the form of cluster groups e.g, 1, 5, 3 (see TABLE 4, Network column). The size of the node showed the percentage of the cluster out of all clusters. The results indicated how signaling subsets co-localize or co-regulate in the same or neighbor pixels.
A study was conducted with the following goals:
1) Compare subcellular image-based protein interaction assays using multiplexed proximity ligation assay in sensitive and resistant EGFRm NSCLC lines (HCC827 vs HCC827/AR and PC9 vs PC-9/AR and PC9/GR/AR) using control, Osimertinib only, a MEK/ERK or YAP inhibitor only, and their combinations.
2) Model alterations in cellular protein interaction maps built from distinct subcellular locations (endoplasmic reticulum, mitochondria, nucleus, and cytosol) of sensitive and resistant EGFRm NSCLC lines.
Spatial signaling interactome in single cells can validate previously established signaling alterations and provide new signaling coordination, and combination drug-specific ratiometric protein interactors in sensitive and resistant EGFRm NSCLC lines (HCC827 vs HCC827/AR and PC9 vs PC-9/AR and PC9/GR/AR) under control, Osimertinib treatment only, MEK/ERK or YAP inhibition only, and Osimertinib combined with MEK/ERK or YAP inhibition conditions. PPI assays have shown the efficacy of Osimertinib treatments on blocking the specific EGFR downstream signaling (e.g., Grb2) using a single-molecule-based pulldown assay60. This method generated in vitro fluorescence images when PPI was achieved within two pairs of proteins through antibody-based detection and isolation procedures of cell extracts, causing the loss of spatial features in cell shapes, morphology, and signaling. While single-molecule PPI provides alterations of signaling in cell extracts, protein interaction pairs necessitate in-situ labeling and imaging technology to build a direct PPI in intact cells and tissues.
Besides, recent pieces of evidences61 have suggested the role of EGFR's tyrosine kinase protein interaction partners on the subcellular localization of PPIs across 13 distinct regions of signaling interactions identified by a guilt-by-association pipeline in NSCLCs61. The correct predictions (green) of subcellular protein positions were limited to less than half of the experimental and prediction datasets, demonstrating a challenge to accurately reconstruct PPIs' subcellular distributions. Thus, the ultimate validation is lacking to co-localize subcellular compartments and PPIs in the multiplexed single-cell data.
Design and approach: Identification and validation of multiplexed signaling interactor targets. Single-cell spatial interaction of PPIs can be targeted for EGFR-based network activity (n+) that contains EGFR, CDH1, P IK3R3, DOK5, TNC, PRKD1, VAV3, PARK7, ANKS1B, ALB, EGF, RB14, SLFN13, AP2B1; EGFR-related resistance markers including FAM157A, SCYL2, ZNF503, AP2B1, TMSBX4, KRT7, FOXA1, CDKN2AIPNL, ERBB3, AP2A2, RBM10, PEBP1, NME2, PLEC, KRT18, and TMSB10, all identified in the EGFR network oncogenesis results61 Additional targets from the previous Osimertinib drug-induced protein alterations7,8 can also be included in this interaction panel.
RapMPI can then profile 25-protein pairs in sensitive and resistant EGFRm NSCLC lines (HCC827 vs HCC827/AR and PC-9 vs PC-9/AR and PC-9/GR/AR) using control, Osimertinib only, a MEK/ERK or YAP inhibitor only, and their combinations. This study can include 5 cell types, 5 drug combinations, and 4 concentrations that can be performed in a 96-well and cultured for 12-24 hours, followed by fixation and RapMPI analysis of protein pairs. In a small pilot experiment, the RapMPI can validate the known and previously uncovered PPIs among proteins of the network1 (WNT1, DKK1, DKK2, RNF43, APC, AXI N1, EGFR, AKT, p-AKT, mTOR, p-mTOR, β-Catenin, Cyclin D1, Cyclin E, and EMMPRIN) and network 2 (YAP/TAZ, pYAP/TAZ, ERK, MEK, RAF, EGFR, RAS, RASSF, MERLIN, NF2, MOB1, MST1/2, and LATS1/2) targets in the same single cells of each cell culture and condition within this 96-well plate. This pilot PPI module and similar experiments can decipher pairwise PPI ratios (e.g., PPI1/PPI2 in the same/different pixel pairs) in many other PPI modules as a function of microwell descriptions. Ratios are calculated to model stochiometric protein balance62,63 in the signaling interactome. A multivariate binomial generalized linear model can be run to study the association of ratiometric interactions and covariates including different cell types, drug combinations, and concentrations (TABLE 5). The result can identify ratios that are associated with the covariates.
The reconstructed RapMPI data can then associate the alterations in subcellular protein interaction maps to the distinct organelles (e.g., Golgi: Sortilin & GOLPH4, Endoplasmic reticulum: ATF6, Mitochondria: TOM20, nucleus: DAPI, and cytosol: Actin, β-Tubulin) of sensitive and resistant EGFRm NSCLC lines. The RapMPI experiment can be followed by an organelle identifier RapMIF experiment using these established markers, yielding and validating the subcellular localizations of PPIs (see TABLE 5 last column) in individual cells rather than prediction pipelines previously reported and widely used in cancer signaling for evaluating the drug response and resistance.
Spatial maps of subcellular PPIs and subcellular localization patterns can then be modeled as spatial signaling velocity maps, akin to the RNA velocity concept45. RapMPI results provide multiplexed PPIs in receptor, cytosol, and nuclear locations. Computational modeling can include position (x-y-z) and signaling interaction (e.g., A-B, C-D, and E-F PPIs). Using the priori information on signaling cascades obtained from the bulk and imaging assays, an arrow strategy can be used to trace signaling information flow in each lung cancer cell. For instance, a simple scenario depicts the probabilistic signaling interactions denoted by arrow fields within a cell using the signaling cascade across A-B→C-D→E-F, wherein each PPI was colored by magenta, green, and blue, respectively (
Experimental results: The proof-of-principle of RapMPI was performed by using the two sets of protein targets using one set of distinct antibodies (Cyclin E & CDK2, PPI1) and another total and phosphorylated site of Akt (PPI2), providing unique PPI within the same cell during phosphorylation and Cyclin E and CDK2 interactions within proximity. For instance, the interaction of Cyclin E and CDK2 controls cell cycle regulation, and, at the same time, the interaction between p-AKT(Ser473) and AKT indicates the phosphorylation and activation of p-AKT by PI3K. Therefore, the interaction between p-AKT (Ser473) and AKT indicates the full activation of AKT. To probe the PPI1 of Cyclin E and CDK2, two sets of PLA assemblies were performed in the fresh PDX lung tumors, PC9 cultures, and EGFRm FFPE human tissues, yielding ubiquitous PPI in every single cell denoted by red dots (i.e., the same color for two distinct PPI validated multiplexing). After DNAse I treatment, the PLA was removed. A second cycle was then performed by a different protein pair of Akt and p-Akt (PPI2) to detect previously unidentified PPIs. The PLA in 2 cycles provided PPIs in different subcellular locations (
The PLA removal was also effective and the subsequent cycle was then used to detect immunofluorescence of pan-cytokeratin and B-actin markers, indicating single-cell and tumors (CK+). This demonstration sets the practicability of multiplexing PLAs in tissues. Thus, RapMPI is feasible both in cultures and tissues.
Discussion: Spatial interaction ratios of PPIs from single-cells (RapMPI) and their subcellular localization on distinct organelles (RapMIF) is projected on high-dimensional visualization maps (tSNE & UMAP) to define necessary and sufficient “resistance” or “sensitivity” to drug combinations for rationally designed NSCLC therapies. The signaling velocity concept can be an emerging approach to quantify signaling responses by tracing PPIs across the subcellular volumes quantitatively. The proposed signaling framework is widely applicable to many other TKI proteins and their spatial interactome analysis.
To overcome any issues of antibody already bound, which would lead to the second staining suffering from the previous antibody blocking the e.g., A antigens, two strategies can be used to overcome this issue. The first one is to strip the antibody instead of PLA removal, making the antigens from the previous cycles accessible. The second one is to access distinct phosphorylation sites of the same antibody to make the relevance of tri-protein interactions. These strategies open up directions in scaling up the barcoding space of RapMPI assay from F×N (linear) to FN (exponential), where F is the number of colors and N is the number of cycles. This further advances the capacity of spatial PPI mapping in single-cells to the proteome level as next-generation spatial proteomics technology.
A study was conducted with the following goals:
1) Construct single cell and subcellular networks and protein interactions in tissues from Osimertinib-treated PC-9 xenografts or EGFRm PDX models that are sensitive to EGFR-TKIs including pretreatment or baseline condition (n=5), 1-week post-treatment (n=5), and after 1-month treatment (n=5) or until the resistance occurs using 10 tissue slices per condition.
2) Develop a predictive signaling response tool using in vivo signaling networks and interaction ratios from human tissues (n=20) obtained from EGFRm patients collected as the baseline pretreatment condition.
Single cell and subcellular networks and protein interactions in tissues extracted from Osimertinib-treated PC-9 xenografts or an EGFRm PDX model and EGFRm human biopsies can confirm or reform signaling co-regulation and interactors from in-vitro cell cultures in physiologically relevant in-vivo models and translational settings. While cell cultures shed light on signaling mechanisms to evaluate the acquired resistance to Osimertinib, molecular heterogeneity of in vivo tumors further contributes to the differences in the resistance mechanisms. Osimertinib-treated PC-9 xenografts or EGFRm PDX models provide signaling response studies in both pre and post-treatment conditions, while the human studies are only limited to pre-treatment biospecimens. Thus, this study can employ RapMIF and RapMPI to profile tissues from PC-9 xenografts or an EGFRm PDX model that is sensitive to EGFR-TKIs including pretreatment or baseline condition, 1-week post-treatment, and after 1-month treatment or until the resistance occurs.
EGFR-TKI inhibitors' role in improving NSCLC patients' treatment outcomes has been identified in phase 3 clinical trial. Osimertinib (80 mg once daily) showed longer progression-free survival and overall survival than other EGFR-TKI drugs, while the tolerability was improved compared to that of others3. However, a fraction of patients show resistance to first-line therapy, and some others also exhibit secondary resistance in other signaling and genetic mechanisms.4 Previous work on immunohistochemistry (IHC) and plasma analysis of these patients with EGFR mutations yielded significant variability between the patients and lung cancer subtypes64,65. The selection of patients for Osimertinib can benefit from personalized molecular profiling of patient tissues for clinical applications. Thus, RapMIF and RapMPI profile spatial signaling networks and interactions in baseline pre-treatment patient biopsies with EGFR mutations as a molecular screening method as a potential companion diagnostic tool using multivariate classification.
Design and approach: RapMIF and RapMPI can be performed to reconstruct single cell and subcellular networks and protein interactions in tissues from Osimertinib-treated PC-9 xenografts or an EGFRm PDX model that is sensitive to EGFR-TKIs including pretreatment or baseline condition (n=5), 1-week post-treatment (n=5), and after 1-month treatment (n=5) or until the resistance occurs using 10 tissue slices per condition (n=5 tumors). Previously validated8 drug concentrations can include Osim (15 mg/kg, once per day), Tram (1 mg/kg, once/2 days), or the combination of Osim and Tram. The YAP inhibitor (MYF-01-37) can be optimized in the range of 1-15 mg/kg once/1-3 days for establishing the efficient pharmacologic inhibition of Osim, Tram, and MYF. Tumor size (0.5 length×width2) can be measured with caliper quantities. Fresh tissues (10 slices per condition) can be extracted and stored at −80° C., followed by 50-plex RapMIF and RapMPI experiments. Spatial image-based data can be processed by joint bio statistical modeling for comparing distinct conditions. It is estimated that with 1000 cells, the standard deviation of the distance estimates will be within 5% of the true value. Thus, with 1000 cells, the network can be accurately constructed.
Signaling networks in tissues of EGFR mutant patients in response to EGFR TKIs treatments would help predict the drug response and isolate signaling mechanisms responsible for drug resistance. RapMIF & RapMPI can image and visualize 50-plex signaling biomarkers to predict development, progression, and response to drug treatment of lung cancers in retrospective studies involving pre-treatment biopsies and Osimertinib treatment outcomes. The study focuses on baseline because practically obtaining post-treatment biopsies are limited to small biopsies used for clinical purposes and not sufficient for research investigations. Initially, RapMIF & RapMPI can quantify a 50-plex antibody panel in 27 tissue cores from lung cancer patients with EGFR mutations and another wild-type group (TriStar Tech Group, SKU: 69572826-2826). The feasibility of tissue imaging can be tested, and the results can be compared to cell colonies obtained in cancer cultures. This comparison can be performed by clustering all the data into one pool and comparative analysis of signaling expression levels and subcellular patterns. Next, clinically archived tissues (n=20 patients, 5 tissue sections each) from EGFRm NSCLC patients can be used for signaling network validations. These retrospective tissues can be from two patient groups comprising Osimertinib resistant (n=10) and responder (n=10) cohort, providing opportunities to correlate signaling networks and clinical outcome/survival. RapMIF & RapMPI can then profile FFPE & fresh frozen tissue sections after fixing and performing to validate tissue staining and analysis protocols (
Experimental results: Continuous and intermittent pharmacologic inhibition of MEK/ERK signaling was demonstrated to delay the emergence of Osimertinib resistance in PC-9 xenografts in nude mice. Complementary study10 showed Osimertinib and trametinib treatment of mice bearing xenograft tumors from YAP1 knock-out (KO) PC-9 or, from the corresponding control cells, led to a durable response for the entire 4-week treatment period.
RapMIF screened up to 25-plex signaling proteins in a tissue lung adenocarcinoma microarray (BS04081a, Biomax) that contained 63 cores, 21 patient cases, varying from normal, stage I, II, III tumor. The tissue cores were then classified into pan-cytokeratin positive and negative regions. Patients with higher tumor stages demonstrate upregulated WNT and AKT pathways with larger variation in the pan-cytokeratin-positive region. The p-β-catenin was upregulated in tumor regions while the constituent activation of WNT and AKT pathway could promote the stabilization of non-p-β-catenin and the degradation of p-β-catenin.66 Multiplexed signaling maps across 55 cores were classified by stages and pan-cytokeratin expression at the single-cell level. The signaling markers in pan-cytokeratin-positive regions exhibited higher expression levels than normal tissue regions. The majority of signaling markers, RNF43, EGFR, H3K27me3, WNT1, p-EGFR, AKT, mTOR, and cyclin D1, demonstrated upregulation in the tumor regions.
Discussion: Non-small cell lung cancer (NSCLC) comprises 80% to 85% of lung cancer, and it is the second leading cause of death after cardiovascular disease.1 While there is an increasing number of treatments for NSCLC, the survival rates are slightly increasing from 15% to 20%.2 NSCLC patients with epidermal growth factor receptor (EGFR) gene mutations, harboring 19 deletions (19del) and exon 21 point mutation (L858R), are treated with EGFR tyrosine kinase inhibitors (EGFR-TKIs). One of these drugs has been Osimertinib that extended patients' progression-free survival and overall survival in phase 3 clinical trial settings.3 However, a subset of patients have primary resistance to these therapies and the initial responders will ultimately develop secondary or acquired resistance to Osimertinib.4 Tumors' molecular and cellular alterations such as the appearance of new EGFR resistant mutations and signaling rewiring cause emergence of acquired resistance and eventual cancer progression. Bulk genomics and proteomics assays have yielded significant signaling mechanisms in cells and tissues. However, these signaling assays fail to capture single-cell and subcellular features in the spatial context in cancer specimens. They also lack cell-to-cell variations due to the averaging of samples during the measurements. Thus, there is an essential need for single-cell imaging and analysis methods to visualize individual cells' signaling dynamics and protein interactions of intracellular labels under drug perturbations.
This proposed system directly visualizes signaling processes in single cells as signaling factors interact with each other in fixed samples. Image-based data spatially resolve cell membrane, cytosol, organelles, and nucleus to reconstruct direct signaling maps in EGFR mutant (EGFRm) cell models. EGFRm cell lines such as PC-9 are sensitive to Osimertinib and the expected signaling distributions were obtained from western blot data, providing opportunities to validate the proposed signaling visualization technology. In-vivo validations of signaling responses can be obtained from PC-9 engrafted mouse models and human tissues from EGFRm NSCLC patients. The proposed platform can inform signaling therapy design or Osimertinib resistance at ultimate resolution and sensitivity.
Multiplexed subcellular protein imaging methods are desirable for spatial signaling networks as a screening platform for clinical use. Provided is an advanced multiplexed protein imaging method to provide rapid experiments, automated microfluidics-based re-labeling and removal of signal, and imaging on the same platform. Here, a rapid multiplexed immunofluorescence (RapMIF) method can be developed to visualize many molecules (30 to potentially 100 parameters) at high optical resolution in cells and tissues. The multiplexing three-color experiments label three antibodies for unique targets at a time. In this experiment, “n cycles” can provide “n×3” markers in the reconstructed image. For instance, ten cycles can provide 30 biomarkers per cell. The experiment is performed on a wide-field microscope using 40× to 60× optical magnification to capture subcellular details. Fixed (1.6% formaldehyde) and permeabilized (0.5% Triton-PBS1×) cells are then labeled, washed, and re-labeled (after signal removal by rapid chemistry40) using an autosampler, valve, and fluidic components. The sample is positioned on a custom microscope stage that allows both sample treatments and imaging on the same RapMIF platform.
Thirty signaling protein factors are defined based on prior literature on signaling cascades. Both total and phosphorylated proteins are targeted in the RapMIF experiments for mapping spatial distributions of signaling patterns across the cytosol and nucleus of a cell. RapMIF maps thousands of cells to target thirty to hundreds of markers in 24 hours over three dimensions after integrating automated labeling, imaging, and analysis pipelines. Spatial bioinformatics models were also developed to create a digital spatial signaling network model of cells as a result of pixel-level clustering of signaling factor distributions measured at a high-optical resolution up to super-resolution41 maps.
Deciphering the PPIs within Osimertinib-related signaling pathways reveals the potential targets for overcoming the acquired resistance. Proximity ligation assay (PLA), also referred to as Duolink PLA technology,42,43 provides an approach to detect direct PPIs in situ at endogenous protein levels. For the detection of one PPI, the cells need to be stained with two primary antibodies targeting two different proteins or proteins phosphorylated at different sites. These primary antibody pairs are conjugated to PLA probes with one PLU, and another one to MINUS. The PLA probe, attached to the heavy chain of the primary antibody, contains a unique oligonucleotide. When the protein of interest interacts with each other, the DNA probes from two antibodies hybridize and ligate to form circular DNA. The amplified circular DNA can be visualized using fluorescence microscopy. The PLA detection kit allows visualization and quantification of the individual PPI. Multiplexing PLA is limited to 3 protein pairs. Here, a highly multiplexed PLA assay, termed rapid multiplexed protein interactions (RapMPI), can be developed based on sequential PLA labeling of 3 antibody pairs, removal of PLA assembly using DNase I, and re-labeling of the same cell with a different set of 3 antibody pairs. This process is then repeated “n” times, to create a multiplexed protein interaction map in the same single cell. For instance, 10 cycles would provide 30 protein pairs identified by 60 antibodies specific to protein pairs of interest. RapMPI is akin to sequential FISH (our previous invention to multiplex RNAs),44 but it has a unique DNA assembly to detect protein-protein proximity within 20-nm physical range. In the RapMPI assay, the proteins are targeted by the same antibodies used for RapMIF, however, antibodies are conjugated to plus and minus oligos for building a rolling circle amplification (RCA) to make individual protein interactions detectable as a diffraction-limited spot in the microscopic image of a cell.
Compared to other PPI assays using overexpression of vector engineered cells,39 the native state of protein interactions are detected by the RapMPI assay, an important feature of PPIs in physiologically relevant conditions. The PLA assay was assembled using the commercial DuoLink platform, but RapMPI utilizes Co-detection by Indexing (CODEX) protocols to develop custom PLA panels for highly multiplexed subcellular protein interactions maps. Automated microfluidics and spatial bioinformatics analysis are shared between the RapMPI and RapMIF platforms using the same automated valve, fluidic, and autosampler setup mounted on the sample holder located on the microscope. One key difference is to spatially trace subcellular signaling cascades using prior information on downstream signaling molecules. This allows the creation of “signaling velocity” maps akin to RNA velocity45 implemented by the differential equations of biomolecular conversions.
The proposed multiplexed in-situ signaling analysis and data mining can validate or restructure baseline signaling states in pre-treatment samples of retrospective biorepositories and how they would respond to Osimertinib to further develop predictive pipelines for incoming patient's signaling therapy design to overcome acquired resistance. A multivariate model can provide regression statistics to link aggregated data (X signals) from all RapMIF and RapMPI experiments in all conditions to the outcome (Y response).
The predictive signaling response studies can benefit from tissue biopsies during and post-therapies. Once a large library (n>40) is collected, only RapMIF & RapMPI images in tissues can be analyzed using machine learning (e.g., Data clustering, tSNE & UMAP) and convolutional neural networks (CNN) on the results of clusters per tissue type. Using 70/30 data split, CNN can train and test the Osimertinib response in available prospective patient specimens, providing opportunities to implement deep learning for the prediction of combinatorial regimens from the signaling maps in tissues of NSCLC patients with EGFRm status and clinical outcome data.
Methods and scientific rigor are provided in individual aims to validate results using multiple complementary assays. The mean percent of individual cell types and single cells is compared between groups using t-tests or Mann-Whitney U tests or ANOVA with posthoc test. Spatial protein expression, network, or interactors are associated are tested with clinical, demographic, and pathologic variables of interest using Fisher's exact/chi-squared tests or Wilcoxon rank-sum/ANOVA tests for categorical and continuous variables, respectively. Survival is estimated using the Kaplan-Meier method and evaluate the impact of race, spatial signaling network, and interactome on survival using Cox proportional hazards models. Longevity, proliferation, and signaling expression levels on drug-sensitive and resistant cell lines or control samples in triplicate are analyzed with a linear mixed-effects model with fixed effects of 5 drug combinations, 4 concentrations, spatial neighboring cells, and their interactions with a random cell line effect to control for variability between cell lines. With 5 mice per combination of treatment and cell line, an effect size difference of 1.3 between any two treatment arms can be detected with 80% power and 0.05 significance level. All experiments are done in triplicate. >50% cell viability in untreated cells. >3-days cultures are used for cell viability. >4-replicates are used for each condition. SD and Mean (CV %) values can be used for comparisons. <0.05 p-value can be significant in cell counts. >80% cell confluency can be used for drug treatments. N, the number of cells can be normalized per drug condition. >1,000 cells per well can be imaged and analyzed. % fold change as designed in the drug concentration. >80% imaging reproducibility in experiments, >0.05 p-value in two-sided T-tests, >0.9 accuracies for pathology validation using Hematoxylin and Eosin (H&E) stain, >0.5 correlation between 30-markers in similar tissue types. Statistical >95% confidence using ANOVA with Bonferroni or Tukey's test can be used to compare spatial signaling network profiles of human and mouse tissues. Spatial signaling network feature analysis of tumors aims to achieve a power level of 0.95 based on a single false positive, desired 2-fold difference, and standard deviation of 1. The scientific rigor can be established by ensuring that variables such as age and sex are considered.
To resolve the subcellular spatial protein networks in the WNT/β-catenin and AKT/mTOR pathways, the presented RapMIF method was performed on A549 cells, EGFR-mutant PC9 cells, and lung adenocarcinoma tissues. RapMIF technology has been built upon the validated protocols of the CycIF method that typically utilizes robust fluorophore inactivation and repeated labeling of dyes to overcome the limitations of the microscopes constrained with 4-6 color channels in conventional IF (Lin et al., 2016). The RapMIF approach incorporates an autostainer setup and pre-conjugated antibodies to expedite the sample processing for highly multiplexed imaging of cells or tissues located on coverslips (
For the multiplexing experiment, RapMIF utilized a panel of 11 cycles, resulting in a total of up to 25 markers, including 5 structural, 17 signaling, and 3 epigenetic markers. For robust segmentation, the panel included Concanavalin A (Endoplasmic reticulum), Phalloidin (Actin), wheat germ agglutinin/WGA (Golgi and plasma membrane) as cell painting markers (Bray et al., 2016), and 3-actin (Actin filaments) and β-tubulin (Microtubules) to accurately delineate nucleus and cell body using an automated algorithm. To improve the staining dilution rate, titration was performed on antibodies. IF experiments evaluated antibody staining using two different dilution rates. The signal-to-noise ratios (SNRs) of these IF images was compared, yielding the condition with higher SNR used for RapMIF experiments (TABLE 1). The staining dilutions were consistent with the manufacturers' suggested dilution range. To reduce the time of sample incubation in indirect IF, most antibodies were conjugated with one of the 488, 555, and 647 Alexa dyes using Lightning-Link Rapid Alexa Fluor Antibody Labeling Kit (Novus/Abcam). These markers were evaluated in four conditions based on conjugated antibodies, unconjugated antibodies, overnight incubation at 4° C., and 1-hour incubation at room temperature (RT). The comparisons of antibody staining experiments at overnight and RT temperature incubation yielded similar signaling patterns, denoted in the cross-sections of plots across pixel histograms. For each round of staining, the sample was incubated with antibodies conjugated to distinct dyes for 1 hour at RT after blocking with cell staining media (CSM). Following imaging of labeled cells, the dyes were bleached in the NaOH—H2O2 mixture under white light for 1 hour (
To reduce the need for conjugated antibodies for each marker, both indirect labeling and direct labeling were combined in the panel. To minimize the effect of unspecific binding of secondary antibodies to conjugated primary antibodies, the indirect labeling of p-EGFR was designed in the first cycle. Indirect labeling of 0-tubulin and 3-actin allowed cell segmentation and pan-cytokeratin distinguished the cancerous tissue regions in the last cycle. The images were taken at 40× with a step size of 0.4-μm to reveal the slowly varying signaling molecules across the z-axis of a cell. The best focus images were generated from +/−2 z-stacks and they were used for post-processing of the single cells. To avoid the effect of the dye inactivation solution on Phalloidin labeled structures, p-EGFR targeting antibodies were directly incubated between the first and second cycles for the cell and tissue experiments without an intermediate bleaching step.
To study the spatial distributions of proteins, a single-pixel analysis was performed to demonstrate the nuclear and cytosolic expression levels and translocations of signaling proteins in WNT/β-catenin and AKT/mTOR pathways. The measurement of the expression level of each marker in two regions, cytosol and nucleus, was quantified separately (
Discussion: Signaling pathways mediate cell communication, proliferation, differentiation, and migration. The pathways are usually regulated through protein complexes, controlled via PPIs. To avoid the need of modifying samples genetically, which is not feasible with patients' samples, proximity ligation assay has shown the capability of detecting the PPI in situ (distance <40 nm) at endogenous protein levels (Alam, 2018). Förster resonance energy transfer (FRET) has also been used for observing dynamics and reversible PPIs in vivo by transferring the fluorophores from a donor to an acceptor in its close proximity. Time-resolved FRET advances the current FRET methods with a faster, more sensitive, and less complex detection of PPIs (Maurel et al., 2008; Rajapakse et al., 2010). Due to spectral bleed and limited energy transfer efficiency, it is difficult to implement FRET with super-resolution imaging methods. Biomolecular fluorescence complementation (BiFC) has shown the feasibility of detecting PPIs based on fluorescent complementation reporters. Implementing BiFC on a super-resolution imaging approach, photoactivated localization microscopy (PALM), achieves imaging of the subcellular dynamics of PPIs at high spatial-temporal resolution (Liu et al., 2014).
Phosphorylation can be used to modulate PPIs, thereby regulating signaling transduction through either modifying proteins post-translationally or producing secondary messengers. Protein kinases play an important role in activating cellular processes, and the target proteins can be phosphorylated at different sites resulting in heterogeneous functions (Xue et al., 2012). Conventional methods for identifying kinase substrates suffer from several limitations. Kinase assay is one of the methods to identify kinase-substrate pairs by using analog-sensitive alleles. The entire in vitro substrate can be assayed for a particular kinase. Various synthesized analogs provided a panel for substrate identification and protein kinase function (Elphick et al., 2007). However, this method is not applicable when the introduced analog does not affect the kinase's function (Mok et al., 2009). Protein microarray analysis can also reveal the potential phosphorylation reactions as well as novel functions. To identify potential substrates, the protein kinases are added to arrays of immobilized proteins and generate signals upon detection (Meng et al., 2008). However, the relationships are identified in vitro and may not represent the relationships in vivo (Mok et al., 2009). Kinase substrate tracking and elucidation (KESTREL) has also been used to identify protein kinase substrates. The high background phosphorylation in the cell extracts can obscure the detection of putative phosphorylation, however, KESTREL achieves the identification of physiologically relevant substrates of proteins in crude cell extracts with an improved signal-to-noise ratio by limiting the phosphorylation reaction time and increasing concentrations of protein kinase of interest. Performing chromatography steps before the analysis could also greatly improve the SNR ratio. Also, to establish high physiological relevance, two or three closely related protein kinases with similar binding specificities were used for one substrate detection. However, the temporal and spatial information is largely ignored in the cell lysate (Knebel et al., 2001). However, the assays have experienced challenges with the background in cell extracts and medium-to-low specific activity to kinases (Knebel et al., 2001; Troiani et al., 2005).
High-throughput techniques such as mass cytometry have been widely used to identify the phosphorylation sites of the target protein. Although the bulk level correlation between the protein expression levels of markers was previously mapped out, the spatial correlations and the precise association between the kinases and their direct substrates are still unclear (Xue et al., 2012). Also, the cell-to-cell heterogeneity in the signaling transduction resulting from factors crucial for cancer development could not be systematically profiled due to technical limitations (Lun and Bodenmiller, 2020). The study could advance the current single-cell mass cytometry (CyTOF) in detecting kinases involved in signaling pathways using additional antibodies in the multiplexed imaging data (Han et al., 2015; Leelatian et al., 2015; Mingueneau et al., 2014). RapMIF provides spatial information on the protein distributions compared to traditional protein analysis assays. Since kinase assay requires the synthesis and introduction of ATP-analogs and is only applicable when the introduced analogs do not interfere with kinase’ function (Mok et al., 2009), RapMIF circumvents the complex process of synthesizing ATP-analogs by directly detecting protein activity from multiple cycles of IF. Also, unlike KESTREL, RapMIF conserves the subcellular structure by avoiding lysate in the cells.
The multiplexed single-pixel analyses have quantified the subcellular protein correlations, indicating potential direct physical neighborhood among signaling protein subsets. Distinct multiplexed intensity profiles of signaling markers revealed the crosstalk between heterogeneous signaling pathways. The presented RapMIF method simplified multiplexed protein analysis by pre-conjugating the antibodies with commercially available dyes. Also, it was observed that the bleaching solution interfered with Phalloidin signals in A549 cells. Therefore, in the A549 data analysis, a combination of p-EGFR, Concanavalin A, B-actin, WGA, APC, and WNT1 markers was used to outline the cell boundary for segmentation. Also, to avoid the bleaching effect, cycle 2 (Phalloidin and WGA) staining was done directly after cycle 1 (p-EGFR) without an intermediate bleaching step in PC9 multiplexing samples. A549 cells showed a larger cytosolic area than that of PC9 cells, providing a more detailed spatial network model for studying the protein distribution in the cytosol. The spatial maps demonstrated the translocation patterns of signaling markers corresponding to their functions in both cell types. The combination of ERK or MEK inhibition with Osimertinib could delay or overcome the acquired Osimertinib resistance (Gu et al., 2020a). Thus, the spatial signaling patterns in the MEK/ERK pathway, another downstream pathway of EGFR, regulating cell proliferation and survival could further be determined using additional antibody targets.
Even in the same population of cells, the cells may display heterogeneous signaling profiles due to the difference in signaling-acting pathways. Autocrine signaling involves the production and secretion of an extracellular mediator, acting on the same cell, while paracrine signaling occurs between different types of cells. Both signaling pathways mediate tumor growth, invasion, and metastasis. During cancer development, the growth factors are usually produced in an autocrine manner, allowing the cell to simulate itself in positive feedback (Caicedo, 2013; Ungefroren, 2021). Current methods have been developed to decipher the cell-cell communication in single cells, however, they failed to demonstrate the heterogeneity in autocrine and paracrine signaling interactions. CellChat as a predictive model for cell-cell communication has been achieved to study the autocrine-acting vs. paracrine-acting pathways in the same cell type or same subpopulation of cells. However, this approach highly relies on the accuracy of the ligand-receptor databases and it produces fewer interactions than other methods when detecting communications from distant cells (Jin et al., 2021). Therefore, combining predictive models and proteomics could optimize cell-cell communication methods. Also, the complex crosstalk among signaling pathways could contribute to the heterogeneity in signaling profiles. Fluorescent reporters can be introduced and utilized to track protein activity in the subpopulation of cells following drug treatment (Kurppa et al., 2020). However, the current study focuses on the design of signaling pathway analysis using molecular snapshots, instead of the subpopulation validation in real-time.
RapMIF multiplexed 17 signaling markers and converted their distributions into 33-pixel spatial clusters. Physical proximity analysis between two distinct proteins yielded a total of 272 (17 choose 2 combinations) possible pairwise neighborhoods, yielding their correlations in cytosolic and nuclear expression levels at the single-pixel level. Pairwise analysis of protein neighborhoods could reveal potential critical signaling proteins or networks (Ramana, 2019). Besides, coordination of the multi-protein neighborhood can be extracted from the multiplexed signaling data. For instance, the tri-protein interactions in cytosol and nucleus are assessed with 1360 (17 choose 3 combinations) possible three-way neighborhoods. Combinatorial signaling analysis would predict the dynamics of signaling networks when multiple proteins physically interact, a task that would be challenging to achieve using bulk signaling assays.
The system design achieved a fully automated staining system including blocking, staining, imaging, and bleaching. The fluidic automation takes the antibody diluted in CSM media in the multi-well in a 2.5-fold dilution. Therefore, to save the antibody volume, the antibody loading step by step could be done by manually pipetting the antibodies to the sample. If needed, the automated system could be further optimized to reduce the dilution rate of the antibody during the staining step. Aspiration settings and washing speed could be adjusted to minimize cell loss across cycles. The p-EGFR signaling receptors in PC9 cells were dependent on the sample quality due to the step of drying the coverslip before mounting it to the holder. The freshness of the sample may affect the morphology of p-EGFR. Thus, the multiplexing experiments in PC9 cells were initiated on the same day of cell fixation to make sure the quality of the samples. Thus, the RapMIF could provide a precision oncology framework for defining the signaling dynamics in both intensities and spatial distribution of human cancers.
For multiplexing experiments, two factors are concerned to ensure cell quality, the efficiency of the bleaching approach and the total spent time for one set experiment. RapMIF utilizes fluorophore-conjugated antibodies for direct labeling. Fluorophore-conjugated antibodies generally produce relatively weaker signals than indirect labeling. However, indirect labeling raised some problems in the context of multiplexed experiments. It is constrained by the availability of the reactivity of secondary antibodies to limited species (e.g., mouse, human, donkey). Also, multiplexing indirect IF requires antibody stripping rather than a regular bleaching approach. The harsh process may cause cell loss and tissue detachment. Compared to the antibody stripping approach removing the whole secondary antibody, the RapMIF bleaching approach is effective in removing dyes while preserving the cells. One concern of the RapMIF is whether the direct IF staining at RT for 1-h for each cycle can label all the antigens in the cell. This trade-off can be addressed by improving the antibody concentration to generate similar signals in 1-h RT staining compared to overnight incubation. The approach aims to achieve rapid profiling of signaling pathways while maximizing the efficiency of protein detection. The staining settings were kept consistent across different samples, making it reliable for studying drug efficiency. Also, staining the sample for 1 hour instead of overnight reduces the 12-cycle multiplexing experiments from 24 days to 6 days, and prevents cell degradation. Another drawback of the RapMIF on FFPE tissues is that FFPE tissue generates a high autofluorescence background and is affected by shading or vignetting due to imaging artifacts. Signaling proteins are expected to express nearly everywhere, making it harder to distinguish between background and real signals. Therefore, background correction steps were performed using the BASIC algorithm to reduce the contribution of background on real signals (Peng et al., 2017). Also, post-bleach imaging data serve as a control of native autofluorescence signals and confirm the background corrected images using the stained raw tissue image that is subtracted from the post-bleach image.
In terms of antibody specificity validation, a positive control (e.g., a cell line expressing the target) and negative control (e.g., a cell line not expressing the target) can be utilized to confirm the selective binding of the antibody. Sometimes, it is difficult to find cell lines that do not express the target. Therefore, knock-out (KO) models such as CRISPR-Cas9 and siRNA can serve as negative controls, and be applied to various approaches, including mass spectrometry, western blot, IHC, and IF (Kurppa et al., 2020; Uhlen et al., 2016). In previous studies, independent antibody strategies are also suggested to confirm antibody specificity. When two antibodies bind to different epitopes of the same target, they can yield correlated signals, suggesting both antibodies recognize the same target (Uhlen et al., 2016). However, this is beyond the scope of current work and independent validations of antibody specificity are lacking in the current study. In the study, the antibody staining was evaluated using conventional IF protocol before performing the multiplexing to improve the staining quality of the antibodies.
Another important consideration is the physical dimensions of signaling protein visualizations. The diffraction limit of light microscopy is around 200 nm, while the antibody size is around 10 nm (Erickson, 2009; Tan et al., 2008; Xing et al., 2016). Due to the low spatial resolution, standard light microscopy fails to resolve the protein complex interactions based on the fluorescent signals. To determine the colocalization of two proteins, their spatial distributions are needed. When two proteins are colocalized with one another, the overlay of the two spatial distributions should show a correlation of intensity at the pixel level. In this case, the pixel size is around 188 nm using an oil immersion 40× objective lens with a 1.4 numerical aperture and the protein molecule is around 2-9 nm (Erickson, 2009). For a single cell, the number of protein molecules is a median of 170,000 (Li et al., 2014) to 2×109, considering the number of ribosomes, tRNA, synthesis rate, and translation rate, among others (Princiotta et al., 2003). Each pixel would contain around a few to 15,700 proteins. Thus, due to the limitation on the resolution of the microscope, it is unable to resolve the single protein interaction. Additional color deconvolution methods might be needed to reach single protein detection. The pixel-level correlations can distinguish the random color overlap due to protein compartmentalization from protein colocalization and indicate that true colocalization only occupied 3% of total colocalization (Costes et al., 2004). To remove nonspecific interactions in data analysis, a probabilistic protein interaction network has previously been developed to identify protein interactions based on the normalized spectral abundance factors using Bayes' approach. The probabilistic model predicts each pair of proteins by providing quantitative information on the preference of each interaction. It takes advantage of quantitative proteomics by measuring spectral counts to generate a probabilistic network of protein interactions, providing an alternative framework to quantify probabilistic pixel-level protein interactions in the RapMIF data (Sardiu et al., 2008). While it is not implemented, it can provide an alternative data analysis approach to quantify probabilistic pixel-level protein distributions in the same pixel area in the RapMIF data. Current studies focus on the development of the RapMIF pipelines for studying the crosstalk of multiple signaling pathways and the protein neighborhood likelihood using spatial networks by studying the neighboring protein distribution in the same pixel area. However, super-resolution can be implemented in this technology to further resolve protein complexes. Protein interaction detection assays, such as proximity ligation assays, can be multiplexed to achieve multiparameter detection of PPIs.
To optimize the staining protocols and data analysis pipeline, RapMIF was performed to quantify the expression and spatial distribution of signaling markers involved in WNT/β-catenin and AKT/mTOR pathways in the A549 NSCLC cell line (
To capture the spatial information among the two pathways, signaling protein maps of the multiplexed panel were converted into 33-pixel clusters and uniquely colored back onto the original images of the single cancer cells (
The colocalization of WNT/β-catenin and p-AKT pathway components was also observed in the spatial signaling network maps, demonstrating the physical neighborhood between two pathways in the same pixel area (
Next, a machine learning model was designed to predict the β-catenin phosphorylation activity based on the 20 biomarkers using a Random forest (RF) algorithm (
Cell cycle analysis was also conducted based on the quantification of nuclear cyclin D1 expression level, classifying the cells into G1, S, and G2 phases (
To reconstruct the signaling networks in EGFR mutant cells under drug perturbations, the RapMIF was used to profile multiple signaling molecules in EGFR mutant NSCLC-sensitive PC9 cell lines. Osimertinib demonstrated a prominent effect in inhibiting PC9 cells, and the cell survival rate decreased as the concentration was increased from 0 to 10-nM under a 3-day treatment (Gu et al., 2020b, 2020a). PC9 cells responding to Osimertinib treatment can develop the acquired resistance. However, in previous work, the western blot experiments of signaling proteins in PC9 cells treated with the combination of Osimertinib and either ACK1 inhibitor or MEK inhibitor exhibited effectiveness to overcome the acquired resistance to Osimertinib (Gu et al., 2020b). Besides, it showed that Osimertinib can effectively inhibit the p-AKT (S473) level in PC9 cells, but not AKT itself. Similar results were obtained from the signaling studies in H1975 cells, a drug-resistant cell line (Gu et al., 2020c). Osimertinib inhibited the phosphorylation of EGFR in a phosphorylation site-dependent manner and decreased the p-EGFR (Y1173) level but failed to alter the p-EGFR (Y1068) in PC9 cells (Gu et al., 2020b, 2020a). However, Osimertinib could affect the downstream effector of EGFR, Grb2, by suppressing its binding to p-EGFR (Tyr1086) in H1975 cells (Lee et al., 2018).
In a study, it was first asked whether the EGFR pathway would be inhibited by Osimertinib in PC9 cells, eventually affecting the proliferation and growth of the PC9 cells. A 48-hour Osimertinib treatment reduced the number of adherent PC9 cells compared to untreated ones. Higher drug concentrations further decreased the cell viability compared to lower drug treatments. Following drug treatments, the RapMIF was used to generate the multiplexed imaging maps of up to 25-plex signaling proteins in PC9 with and without the treatment of 40 nM Osimertinib (
To evaluate the effect of Osimertinib on EGFR's downstream proteins, it was investigated whether p-EGFR would correlate with p-AKT expression level. The p-EGFR signaling proteins demonstrated a high correlation to p-AKT after drug treatment based on the mean-intensity correlation of single-cells (
Next, it was asked whether Osimertinib would alter the WNT pathway in PC9 cells. Prior studies on breast tumors and HC11 mammary epithelial cells yielded constituent activation of WNT1 to transactivate the EGFR pathway by increasing the ligand availability (Civenni et al., 2003; Faivre and Lange, 2007; Schlange et al., 2007). However, in the RapMIF results, WNT1 showed a low correlation with p-EGFR in both conditions, whereas a high correlation with p-AKT in the presence of the drug was obtained (
The effect of Osimertinib in spatial distributions and networks in PC9 cells was then analyzed. The spatial signaling network maps of the PC9 cells after drug treatment exhibited more dispersed distribution in the cytosol, noted as clusters 0, 1, and 2, compared to the control sample (
To visualize the heterogeneity of signaling expression levels, subpopulation analysis was performed on both EGFR-wild type (A549) and EGFR-mutant (PC9) NSCLC cell lines. First, A549 was clustered into 12 subpopulations based on their single-cell mean intensity levels. Each subpopulation represented a distinct signaling profile across 17 signaling markers. A549 cells in subpopulation 3 displayed relatively lower expression levels of all signaling proteins. In contrast, cells in subpopulation 0 and subpopulation 6 showed higher signaling expression levels and highly expressed WNT1, p-EGFR, cyclin E, mTOR, p-AKT, and non-p-β-catenin, indicating these cells may have higher activation of WNT/β-catenin and AKT/mTOR signaling. To decipher the cell heterogeneity, a similar analysis was performed on PC9 cells in the presence and absence of drug treatment (
The signaling network concept and subpopulation analysis can be extended to other cell types. Thus, RapMIF examined a 7-plex protein panel in Umbilical cord mesenchymal stem cells (UC-MSC) (
To further verify the feasibility of signaling measurements in situ, RapMIF on FFPE tissues was investigated. The RapMIF on cell culture resolved the subcellular distribution among target proteins. However, pixel-level signaling analysis in tissues was challenging due to the densely packed cells with a limited spatial barcoding space in the cytosol of each cell. Thus, single-cell signaling maps were generated in tissues to decipher the architecture of normal and diseased cellular distributions. Haematoxylin and eosin (H&E) staining and Immunohistochemistry (IHC) have been widely used to diagnose disease (Lin et al., 2018). However, the information from H&E is limited, and IHC is a single-channel imaging method. Tissue imaging typically requires intricate spatial analysis for feature extraction and cell classification algorithms due to the native tissue architecture. The RapMIF made it possible to study potential signaling imaging biomarkers to predict drug response using the baseline tissues obtained from the patients before targeted therapies.
RapMIF screened up to 25-plex signaling proteins in a tissue lung adenocarcinoma microarray (BS04081a, Biomax) that contained 63 cores, 21 patients' cases, varying from normal, stage I, II, and III tumors. The images were first processed to correct the line artifacts caused by the uneven illumination from the flat field and dark field using the BASIC function in ImageJ (Peng et al., 2017). Background corrections improved the downstream image analysis. Single cells were then segmented in the tissue cores and classified into pan-cytokeratin positive and negative regions (
Next, the signaling proteins were assessed across various cancer stages. Multiplexed signaling maps across 55 cores were classified by cancer stages and pan-cytokeratin expression at the single-cell level. The signaling markers in pan-cytokeratin-positive regions exhibited higher expression levels than normal tissue regions (
To investigate the spatial distribution of signaling proteins, a single-cell clustering analysis was performed in the pan-cytokeratin-positive areas. The malignant tissue exhibited more variance in signaling expression from the same patient (
The next step of RapMIF in tissues would be to study how AKT/mTOR and WNT/0-catenin pathways coordinate in wild-type, EGFR mutant FFPE tissues, and fresh frozen tissues. Fresh frozen tissues may provide varying staining results, dependent on the conditions of fixation and antigen retrieval (Shi et al., 2008). Signaling networks in tissues of EGFR mutant patients in response to EGFR TKI treatments would help predict the drug response and isolate signaling mechanisms responsible for drug resistance. Considering the complexity of tumor samples, combinations of additional signaling targets would enhance the prediction accuracy of patient classification using deep learning methods.
Discussion: Signaling pathways mediate cell communication, proliferation, differentiation, and migration. The pathways are usually regulated through protein complexes, controlled via PPIs. To avoid the need of modifying samples genetically, which is not feasible with patients' samples, proximity ligation assay has shown the capability of detecting the PPI in situ (distance <40 nm) at endogenous protein levels (Alam, 2018). Förster resonance energy transfer (FRET) has also been used for observing dynamics and reversible PPIs in vivo by transferring the fluorophores from a donor to an acceptor in its close proximity. Time-resolved FRET advances the current FRET methods with a faster, more sensitive, and less complex detection of PPIs (Maurel et al., 2008; Rajapakse et al., 2010). Due to spectral bleed and limited energy transfer efficiency, it is difficult to implement FRET with super-resolution imaging methods. Biomolecular fluorescence complementation (BiFC) has shown the feasibility of detecting PPIs based on fluorescent complementation reporters. Implementing BiFC on a super-resolution imaging approach, photoactivated localization microscopy (PALM), achieves imaging of the subcellular dynamics of PPIs at high spatial-temporal resolution (Liu et al., 2014).
Phosphorylation can be used to modulate PPIs, thereby regulating signaling transduction through either modifying proteins post-translationally or producing secondary messengers. Protein kinases play an important role in activating cellular processes, and the target proteins can be phosphorylated at different sites resulting in heterogeneous functions (Xue et al., 2012). Conventional methods for identifying kinase substrates suffer from several limitations. Kinase assay is one of the methods to identify kinase-substrate pairs by using analog-sensitive alleles. The entire in vitro substrate can be assayed for a particular kinase. Various synthesized analogs provided a panel for substrate identification and protein kinase function (Elphick et al., 2007). However, this method is not applicable when the introduced analog does not affect the kinase's function (Mok et al., 2009). Protein microarray analysis can also reveal the potential phosphorylation reactions as well as novel functions. To identify potential substrates, the protein kinases are added to arrays of immobilized proteins and generate signals upon detection (Meng et al., 2008). However, the relationships are identified in vitro and may not represent the relationships in vivo (Mok et al., 2009). Kinase substrate tracking and elucidation (KESTREL) has also been used to identify protein kinase substrates. The high background phosphorylation in the cell extracts can obscure the detection of putative phosphorylation, however, KESTREL achieves the identification of physiologically relevant substrates of proteins in crude cell extracts with an improved signal-to-noise ratio by limiting the phosphorylation reaction time and increasing concentrations of protein kinase of interest. Performing chromatography steps before the analysis could also greatly improve the SNR ratio. Also, to establish high physiological relevance, two or three closely related protein kinases with similar binding specificities were used for one substrate detection. However, the temporal and spatial information is largely ignored in the cell lysate (Knebel et al., 2001). However, the assays have experienced challenges with the background in cell extracts and medium-to-low specific activity to kinases (Knebel et al., 2001; Troiani et al., 2005).
High-throughput techniques such as mass cytometry have been widely used to identify the phosphorylation sites of the target protein. Although the bulk level correlation between the protein expression levels of markers was previously mapped out, the spatial correlations and the precise association between the kinases and their direct substrates are still unclear (Xue et al., 2012). Also, the cell-to-cell heterogeneity in the signaling transduction resulting from factors crucial for cancer development could not be systematically profiled due to technical limitations (Lun and Bodenmiller, 2020). This study could advance the current single-cell mass cytometry (CyTOF) in detecting kinases involved in signaling pathways using additional antibodies in the multiplexed imaging data (Han et al., 2015; Leelatian et al., 2015; Mingueneau et al., 2014). RapMIF provides spatial information on the protein distributions compared to traditional protein analysis assays. Since kinase assay requires the synthesis and introduction of ATP-analogs and is only applicable when the introduced analogs do not interfere with kinase’ function (Mok et al., 2009), RapMIF circumvents the complex process of synthesizing ATP-analogs by directly detecting protein activity from multiple cycles of IF. Also, unlike KESTREL, RapMIF conserves the subcellular structure by avoiding lysate in the cells.
The multiplexed single-pixel analyses have quantified the subcellular protein correlations, indicating potential direct physical neighborhood among signaling protein subsets. Distinct multiplexed intensity profiles of signaling markers revealed the crosstalk between heterogeneous signaling pathways. The presented RapMIF method simplified multiplexed protein analysis by pre-conjugating the antibodies with commercially available dyes. Also, it was observed that the bleaching solution interfered with Phalloidin signals in A549 cells. Therefore, in the A549 data analysis, a combination of p-EGFR, Concanavalin A, B-actin, WGA, APC, and WNT1 markers was used to outline the cell boundary for segmentation. Also, to avoid the bleaching effect, cycle 2 (Phalloidin and WGA) staining was done directly after cycle 1 (p-EGFR) without an intermediate bleaching step in PC9 multiplexing samples. A549 cells showed a larger cytosolic area than that of PC9 cells, providing a more detailed spatial network model for studying the protein distribution in the cytosol. The spatial maps demonstrated the translocation patterns of signaling markers corresponding to their functions in both cell types. The combination of ERK or MEK inhibition with Osimertinib could delay or overcome the acquired Osimertinib resistance (Gu et al., 2020a). Thus, the spatial signaling patterns in the MEK/ERK pathway, another downstream pathway of EGFR, regulating cell proliferation and survival could further be determined using additional antibody targets.
Even in the same population of cells, the cells may display heterogeneous signaling profiles due to the difference in signaling-acting pathways. Autocrine signaling involves the production and secretion of an extracellular mediator, acting on the same cell, while paracrine signaling occurs between different types of cells. Both signaling pathways mediate tumor growth, invasion, and metastasis. During cancer development, the growth factors are usually produced in an autocrine manner, allowing the cell to simulate itself in positive feedback (Caicedo, 2013; Ungefroren, 2021). Current methods have been developed to decipher the cell-cell communication in single cells, however, they failed to demonstrate the heterogeneity in autocrine and paracrine signaling interactions. CellChat as a predictive model for cell-cell communication has been achieved to study the autocrine-acting vs. paracrine-acting pathways in the same cell type or same subpopulation of cells. However, this approach highly relies on the accuracy of the ligand-receptor databases and it produces fewer interactions than other methods when detecting communications from distant cells (Jin et al., 2021). Therefore, combining predictive models and proteomics could optimize cell-cell communication methods. Also, the complex crosstalk among signaling pathways could contribute to the heterogeneity in signaling profiles. Fluorescent reporters can be introduced and utilized to track protein activity in the subpopulation of cells following drug treatment (Kurppa et al., 2020). However, the current study focuses on the design of signaling pathway analysis using molecular snapshots, instead of the subpopulation validation in real-time.
RapMIF multiplexed 17 signaling markers and converted their distributions into 33-pixel spatial clusters. Physical proximity analysis between two distinct proteins yielded a total of 272 (17 choose 2 combinations) possible pairwise neighborhoods, yielding their correlations in cytosolic and nuclear expression levels at the single-pixel level. Pairwise analysis of protein neighborhoods could reveal potential critical signaling proteins or networks (Ramana, 2019). Besides, coordination of the multi-protein neighborhood can be extracted from the multiplexed signaling data. For instance, the tri-protein interactions in cytosol and nucleus are assessed with 1360 (17 choose 3 combinations) possible three-way neighborhoods. Combinatorial signaling analysis would predict the dynamics of signaling networks when multiple proteins physically interact, a task that would be challenging to achieve using bulk signaling assays.
Our system design achieved a fully automated staining system including blocking, staining, imaging, and bleaching. The fluidic automation takes the antibody diluted in CSM media in the multi-well in a 2.5-fold dilution. Therefore, to save the antibody volume, the antibody loading step by step could be done by manually pipetting the antibodies to the sample. If needed, the automated system could be further optimized to reduce the dilution rate of the antibody during the staining step. Aspiration settings and washing speed could be adjusted to minimize cell loss across cycles. The p-EGFR signaling receptors in PC9 cells were dependent on the sample quality due to the step of drying the coverslip before mounting it to the holder. The freshness of the sample may affect the morphology of p-EGFR. Thus, the multiplexing experiments in PC9 cells were initiated on the same day of cell fixation to make sure the quality of the samples. Thus, the RapMIF could provide a precision oncology framework for defining the signaling dynamics in both intensities and spatial distribution of human cancers.
For multiplexing experiments, two factors are concerned to ensure cell quality, the efficiency of the bleaching approach and the total spent time for one set experiment. RapMIF utilizes fluorophore-conjugated antibodies for direct labeling. Fluorophore-conjugated antibodies generally produce relatively weaker signals than indirect labeling. However, indirect labeling raised some problems in the context of multiplexed experiments. It is constrained by the availability of the reactivity of secondary antibodies to limited species (e.g., mouse, human, donkey). Also, multiplexing indirect IF requires antibody stripping rather than a regular bleaching approach. The harsh process may cause cell loss and tissue detachment. Compared to the antibody stripping approach removing the whole secondary antibody, the RapMIF bleaching approach is effective in removing dyes while preserving the cells. One concern of the RapMIF is whether the direct IF staining at RT for 1-h for each cycle can label all the antigens in the cell. This trade-off can be addressed by improving the antibody concentration to generate similar signals in 1-h RT staining compared to overnight incubation. This approach aims to achieve rapid profiling of signaling pathways while maximizing the efficiency of protein detection. The staining settings were kept consistent across different samples, making it reliable for studying drug efficiency. Also, staining the sample for 1 hour instead of overnight reduces the 12-cycle multiplexing experiments from 24 days to 6 days, and prevents cell degradation. Another drawback of the RapMIF on FFPE tissues is that FFPE tissue generates a high autofluorescence background and is affected by shading or vignetting due to imaging artifacts. Signaling proteins are expected to express nearly everywhere, making it harder to distinguish between background and real signals. Therefore, background correction steps were performed using the BASIC algorithm to reduce the contribution of background on real signals (Peng et al., 2017). Also, post-bleach imaging data serve as a control of native autofluorescence signals and confirm the background corrected images using the stained raw tissue image that is subtracted from the post-bleach image.
In terms of antibody specificity validation, a positive control (e.g., a cell line expressing the target) and negative control (e.g., a cell line not expressing the target) can be utilized to confirm the selective binding of the antibody. Sometimes, it is difficult to find cell lines that do not express the target. Therefore, knock-out (KO) models such as CRISPR-Cas9 and siRNA can serve as negative controls, and be applied to various approaches, including mass spectrometry, western blot, IHC, and IF (Kurppa et al., 2020; Uhlen et al., 2016). In previous studies, independent antibody strategies are also suggested to confirm antibody specificity. When two antibodies bind to different epitopes of the same target, they can yield correlated signals, suggesting both antibodies recognize the same target (Uhlen et al., 2016). However, this is beyond the scope of current work and independent validations of antibody specificity are lacking in the current study. In this study, the antibody staining has been evaluated using conventional IF protocol before performing the multiplexing to improve the staining quality of the antibodies.
Another important consideration is the physical dimensions of signaling protein visualizations. The diffraction limit of light microscopy is around 200 nm, while the antibody size is around 10 nm (Erickson, 2009; Tan et al., 2008; Xing et al., 2016). Due to the low spatial resolution, standard light microscopy fails to resolve the protein complex interactions based on the fluorescent signals. To determine the colocalization of two proteins, their spatial distributions are needed. When two proteins are colocalized with one another, the overlay of the two spatial distributions should show a correlation of intensity at the pixel level. In this case, the pixel size is around 188 nm using an oil immersion 40× objective lens with a 1.4 numerical aperture and the protein molecule is around 2-9 nm (Erickson, 2009). For a single cell, the number of protein molecules is a median of 170,000 (Li et al., 2014) to 2×109, considering the number of ribosomes, tRNA, synthesis rate, and translation rate, among others (Princiotta et al., 2003). Each pixel would contain around a few to 15,700 proteins. Thus, due to the limitation on the resolution of the microscope, it is unable to resolve the single protein interaction. Additional color deconvolution methods might be needed to reach single protein detection. The pixel-level correlations can distinguish the random color overlap due to protein compartmentalization from protein colocalization and indicate that true colocalization only occupied 3% of total colocalization (Costes et al., 2004). To remove nonspecific interactions in data analysis, a probabilistic protein interaction network has previously been developed to identify protein interactions based on the normalized spectral abundance factors using Bayes' approach. The probabilistic model predicts each pair of proteins by providing quantitative information on the preference of each interaction. It takes advantage of quantitative proteomics by measuring spectral counts to generate a probabilistic network of protein interactions, providing an alternative framework to quantify probabilistic pixel-level protein interactions in the RapMIF data (Sardiu et al., 2008). While it is not implemented, it can provide an alternative data analysis approach to quantify probabilistic pixel-level protein distributions in the same pixel area in the RapMIF data. Current studies focus on the development of the RapMIF pipelines for studying the crosstalk of multiple signaling pathways and the protein neighborhood likelihood using spatial networks by studying the neighboring protein distribution in the same pixel area. However, super-resolution can be implemented in this technology to further resolve protein complexes. Protein interaction detection assays, such as proximity ligation assays, can be multiplexed to achieve multiparameter detection of PPIs.
To detect complex PPI networks, a multiplexed PLA assay, RapMPI was developed that provides higher sensitivity and is compatible with both cells and tissues (frozen or FFPE). The detection can be either indirect or direct. For the detection of one PPI using direct labeling, the cells need to be stained with two primary antibodies targeting two different proteins or proteins phosphorylated at different sites17,18. These primary antibody pairs are conjugated to one of the PLA probes, one to PLUS, and another one to MINUS (
RapMPI achieves profiling of multiple PPIs in the multiplexed single-cell and reconstructed their subcellular distributions. By incorporating the localization information of organelles and proliferation proteins using Rapid Multiplexed Immunofluorescence (RapMIF) (
To reconstruct the signaling networks and PPI in EGFRm cells under drug perturbations, RapMPI was used to profile 5 PPIs involved in the AKT/mTOR, MEK/ERK, and YAP/TEAD pathways in NSCLC EGFRm Osimertinib-sensitive cell line, HCC827 cells. In a study, it was first investigated whether the EGFR pathways would be affected by Osimertinib in HCC827 cells, eventually affecting cell proliferation and growth. HCC827 cells were treated with and without 100 nM O for 6, 12, and 24 h. p-ERK was used to indicate the efficacity of Osimertinib treatment. P-ERK was suppressed initially after 6-hour treatment, however, Osimertinib even increased p-ERK after 12-hour treatment, suggesting Osimertinib may exert transient inhibitory effects on MEK/ERK pathway. To reveal the PPIs dynamics under Osimertinib treatment, RapMPI was performed on HCC827 cells treated without and with 100 nM Osimertinib for 12 h. To determine the best staining conditions for both RapMIF and RapMPI, titrations were performed on each antibody in the multiplexed panel. Immunofluorescence (IF) was performed to evaluate antibody staining using two different dilution rates. The SNRs of these IF images were compared, and the conditions with higher SNRs were used for either RapMPI or RapMIF experiments (TABLE 6). The staining dilutions were consistent with the manufacturer's suggested dilution range.
1:400
1:200
1:400
1:200
1:200
1:100
1:200
1:200
1:500
1:400
RapMPI demonstrated and compared the PPIs dynamics among proteins including YAP & TEAD1, Cyclin E & CDK2, p-ERK & c-MYC, Mcl-1 & Bak, and p-AKT & mTOR. The PPI distributions have been visualized using their spatial localization. Each node represents one detected PPI event, and for each single cell, a spatial graph of PPI events is constructed using Delaunay triangulation which captures the underlying spatial neighboring information of PPI events. The cell boundary was obtained from segmentation on the p-EGFR signaling protein (
The effect of Osimertinib on cell apoptosis and proliferation was first examined by comparing the Cyclin E/CDK2 and Mcl-1/Bak interactions. MCL-1 is known as an anti-apoptotic factor, and it can be phosphorylated by ERK, resulting in enhanced proteasome-dependent degradation. Osimertinib downregulates Mcl-1, therefore, enhancing cell apoptosis21. Mcl-1 can sequester Bak activity via direct interaction, therefore preventing cell apoptosis22. The interaction between Mcl-1 and Bak is related to the inhibition of cell apoptosis. After Osimertinib treatment, the cells did not exhibit an obvious change in the Mcl1/Bak interaction in either the cytoplasm or the whole cell (
The effects of Osimertinib on the EGFR-related signaling pathways were next determined. Osimertinib as an inhibitor to EGFR downregulated the p-AKT/mTOR pathway reducing the interactions in the cytoplasm. Osimertinib can induce the degradation of c-MYC, which regulates cell growth and proliferation in EGFRm-sensitive cells, and ERK can phosphorylate c-Myc at S62 linked to c-MYC's stabilization24. However, the inhibitory effect is limited in Osimertinib-resistant cell lines, and the upregulation of c-Myc is related to the acquired resistance to Osimertinib24. In terms of the YAP/TEAD1 pathway, YAP can remain active under EGFR TKIs and EGFR/MEK inhibitions contributing to the tumor dormancy in EGFRm NSCLC4. It was observed that after treatment the cells exhibited significantly increased interactions of YAP/TEAD1 and P-ERK/c-Myc (
To resolve the spatial proteomics and PPIs, RapMIF was also into the panel20. RapMIF was utilized to detect the localization of distinct organelles (e.g., Golgi: WGA, Endoplasmic reticulum: concanavalin A, Mitochondria: TOM20, nucleus: DAPI). The organelle information can be associated with the distribution of PPIs from RapMPI to identify the subcellular localization of PPIs (receptors, cytosol, or nucleus) in individual cells (
PPI events distribution as predictive features to respond to drug treatment was investigated by training models to predict the treatment output based on the PPI event counts. First, six machine learning models and one MLP model were trained to evaluate the accuracy, Area Under the Curve (AUC), and F1 scores to compare subcellular feature importance. The MLP model consists of stacked fully connected layers with input the same as the machine learning models. The separation of PPI events into cytoplasmic and nucleic compartments showed improvement in all metrics across models (
To confirm the feasibility of scaling up the PPI detection, 9 PPIs were profiled in HCC827 cells treated with and without Osimertinib for 12 hours. The cells were performed with two cycles of multicolor detection followed by five cycles of single detection of PPIs (YAP/TEAD1, Cyclin E/CDK2, p-ERK/c-Myc, p-AKT/mTOR, Mcl-1/Bak). The multicolor detection allows detection of 2 to 3 pairs of PPIs in a single cycle. Followed by RapMPI, RapMIF was performed to detect organelle locations, cell proliferation, and tumor cells. The PPI distribution in every single cell can be visualized using the spatial network (
In addition to cyclin E/CDK2, cyclin D1/CDK4 was used to detect the GUS transition in the cell cycle27. Osimertinib exhibited an inhibitory effect on cyclin D1/CDK4 (
Cyclin E/CDK2 colocalizes in the Golgi with high co-expression of NBD-C6 (
It was observed that the later cycles of RapMPI such as TEAD1/YAP and Cyclin E/CDK4 exhibited much fewer signals. That may be because of Dnase digesting DNA-connected proteins in the nucleus, especially transcription factors. To minimize the Dnase effect on nuclear proteins, the order of PPI detections was changed, and TEAD1/YAP and Cyclin E/CDK4 PPIs were not affected in the first two cycles. It was observed that multicolor detection was less sensitive to DNase and bleaching, and the residuals of PPI could be re-detected after 6 times DNase and bleaching.
To preserve the quality of single-color PPI detection, nuclease P1 as an alternative reagent to DNase to digest single-strand DNA was tested on five pairs of PPIs using single-color detection in HCC827 untreated cells (
Similarly, machine learning (
The small groups of interconnections among PPIs were also visualized using network motifs. These network motifs are statistically significant patterns within large spatial signaling networks. Increased autoregulation of p-ERK/c-Myc PPI activity was observed after Osimertinib treatment in motif 9 (p-ERK/c-Myc & p-ERK/c-Myc), and the PPI of p-ERK/c-Myc occupies a large proportion in the network (
To verify the feasibility of detecting PPIs in situ, RapMPI was investigated on HCC827-derived mouse xenografts (CDX). The RapMPI on cell culture resolves the PPI at the subcellular level, however, it fails to consider the architecture of cellular distributions. A study utilized RapMPI on tissue achieving to generate single-cell signaling maps within the context of the tumor microenvironment.
HCC827-derived xenografts in mice were treated with and without Osimertinib for a sustained period, and OCT sections were obtained to compare the emergence of resistance. Osimertinib was given to mice daily. HCC827 xenografts in mice receiving Osimertinib treatment were effectively inhibited for the first week. The tumor almost disappeared after a 5-day Osimertinib treatment. As treatment continued, the tumor grew back and larger, indicating the development of acquired resistance to Osimertinib. Therefore, in the study, RapMPI was performed on mouse HCC827 cells-derived xenografts tissues in the presence of Osimertinib treatment for 1 week and 2 months, representing responders and non-responders. The dynamics of 5 PPIs related to organelle localization, proliferation markers, and gene expression were detected (
A spatial network has been used to visualize the PPI distributions at the single-cell level in the tissues (
To demonstrate the feasibility of detecting PPIs using a super-resolved microscope, the RapMPI detected by a widefield microscope and Zeiss 900A with Airy scan was compared. The negative control with one MINUS probe alone confirmed the validity of the RapMPI protocol.
Discussion: RapMPI is an image-based multiplexing approach to detect PPI at the subcellular level. The iterative processes of ligation, amplification, imaging, and DNasing allows detection of 9 PPIs in cell cultures and 5 PPIs in tissues. The feasibility of integrating RapMPI with RapMIF to profile both PPIs and signaling, proliferation, and organelles' markers is demonstrated herein. The upregulation of Tead1/YAP and p-ERK/c-Myc PPIs after Osimertinib treatment in HCC827 cells may indicate the activation of YAP and the p-ERK pathway as compensatory pathways to EGFR inhibition. The co-expression analysis evaluated the co-localization of signaling markers with organelles. Also, it was observed that Dnase interfered with Phalloidin signaling in HCC827 cells. Therefore, a combination of p-EGFR, concanavalin, and WGA was used for cell segmentation. DNase digests both single-strand and double-strand DNA in the nucleus, suggesting that it may change the localization of proteins connected with DNA and reduce PPI detection in later cycles. To reduce the negative effect of Dnase on PPI detection, further experiments were conducted on examining the effects of nuclease P1 and DMSO stripping solution on deactivating oligos and fluorophores. Nuclease P1 and DMSO exhibited comparable effects on digesting DNA as Dnase. Nuclease P1 targeting only single-strand DNA serves as a good alternative to nuclease P1. Also, it preserves the decent staining of Phalloidin, indicating that it has minimal effect on changing the localization and structure of proteins.
Using machine learning models, it was shown that subcellular information allows better predictivity of single-cell drug perturbation states. To better incorporate spatial information into the predictive model the single-cell PPI events were transformed into graphs, and scPPI-GCN, a predictive pipeline, was developed for determining the drug treatment outcome from single-cell PPI data. The scPPIGCN predictive pipeline was benchmarked using spatial PPI graphs with ML, MLP, and MIL models and showed that spatial information plays an important role in improving the prediction of single-cell states. It was also shown that 3D graphs of PPI resulted in better cell state predictive abilities compared to 2D graphs of PPI.
To select the PPI of interest, a bioluminescence resonance energy transfer (BRET) based differential PPI discovery platform can be performed to reveal the differential interactions between wildtype NSCLC and EGFRm NSCLC cells. This quantitative high-throughput PPI screening platform can detect direct PPI with a proximity of <10 nm37. By comparing the PPI between wild-type and mutant cells, mutation-directed novel PPIs can be discovered which function in potential alternative signaling pathways. The selected PPIs are involved in the EGFRm-related signaling pathways. The selection includes the upstream, downstream, and effectors in the AKT/mTOR, ERK/MEK, and YAP/TEAD1 pathways. They not only indicate the activity of the signaling pathway, but they also demonstrate the cell response including cell proliferation and apoptosis in response to drug perturbations. The selection of PPIs in the panel is based on the literature review, the OncoPPi, and Bioplex interactome networks.
Another method to determine potential PPIs is to prescreen interesting pairs using affinity purification-mass spectrometry. The underlying bait-prey pairs can be selectively purified and quantified by mass spectrometry analysis. This can help determine the interactome of target proteins between mutant and wide-type cells to understand the EGFRm-resistant mechanisms. The potential target proteins include p-AKT, ERK, and YAP. Due to the unknown biological questions involved in the signaling cascades, it is hard to determine one pair of proteins that do not interact with each other at all. To examine the specificity of RapMPI, protein knockout could be used as a negative control.
The results discussed herein exhibited an image-based approach to detect multiple PPI events and signaling markers in both cell culture and tissues. The conventional PPI detection approaches using MS cause the loss of spatial information due to the sample preparation step and peptide extraction38. Unlike coimmunoprecipitation (Co-IP), PLA preserves the spatial information of proteins without cell lysis and can be performed on both cell cultures and tissues. The efficiency of detection can be improved using UnFold probes, which prevent cross-reactive detection of irrelevant proteins using a hairpin loop structure12.
In conclusion, RapMPI illustrates the feasibility of multiplexing PLA and detecting multiple PPI distributions at the subcellular level. It also demonstrates the value of modeling drug treatment outcomes with graph-based PPI inputs, which integrates both the quantification and spatial information of protein interactions. This predictive model would predict patients' treatment outcomes with the signaling network inputs, which overcomes the limitation of lack of spatial details of signaling pathways using bulk signaling assays.
The following patents, applications and publications as listed below and throughout this document are hereby incorporated by reference in their entirety herein.
This invention was made with government support under Grant No. P50CA217691, awarded by National Institutes of Health. The government has certain rights in the invention.
| Number | Date | Country | |
|---|---|---|---|
| 63399427 | Aug 2022 | US |
