Patent Application
20210147929
The official copy of the sequence listing is submitted electronically via EFS-Web as an ASCII-formatted sequence listing with a file named “SLU19007US GENE SEQUENCE LISTING,” created on Oct. 9, 2020, and having a size of 126 kilobytes, and is filed concurrently with the specification. The sequence listing contained in this ASCII-formatted document is part of the specification and is herein incorporated by reference in its entirety.
The invention relates to methods for determining exposure to an immunogenic antigen or a viral infection in a subject. Also provided are methods for evaluating vaccine efficacy.
T cell and B cell responses are responsible for generating the adaptive immune response to vaccines and infections. T cells recognize pathogen-specific peptides in the context of the major histocompatibility complex (MHC) through the T cell receptor (TCR). The TCR is a heterodimeric protein composed of TCRα and TCDβ chains. During T cell development, each TCR chain is generated through quasi-random genetic recombination from the germline loci of the variable (V), diversity (D), and joining (J) gene segments (Manfras et al., 1989, Robins et al., 2010).
In mice, the tcrb locus has approximately 35 different TCRV β segments, 2 TCRD β segments, and 14 TCRJ β segments. During recombination, tcrv, tcrd, and tcrj segments are rearranged together to create and encode complementary determining region 3 (CDR3). CDR3 is the most variable region of the TCR that interacts with foreign peptide. These genetic rearrangement events result in a high degree of diversity in CDR3 of the TCR (Arstila et al., 1999, Cabaniols et al., 2001, Davis and Bjorkman, 1988, Robins et al., 2009).
During an immune response, antigen presentation results in the activation and expansion of T cells with TCR(s) specific to the pathogen (Ishizuka et al., 2009, Venturi et al., 2008b, Venturi et al., 2016). Clonally expanded T cells carry the same unique TCR rearrangement (Manfras et al., 1999). Once the pathogen has been cleared, a subset of T cells with TCRs specific to the pathogen remain as long-lived memory cells. The unique DNA rearrangements have the potential to serve as a stable biomarker, cataloging an individual's functional T cell memory and immunological history (Emerson and DeWitt, 2017, Estorninho et al., 2013).
On average, approximately 107 unique TCRβ chains can be identified from the approximately 101 circulating T cells present in a healthy human adult (Robins et al., 2009). The ability to readily identify identical TCR sequences among multiple individuals (public TCRs) is challenging because an individual has the potential to generate approximately 1018 unique TCR recombinants. Nonetheless, in both humans and murine models, there are examples of public T cell responses to infectious disease (such as cytomegalovirus [CMV] and influenza) and in autoimmunity (Elhanati et al., 2014, Emerson and DeWitt, 2017, Li et al., 2012, Lossius et al., 2014, Marrero et al., 2016, Valkenburg et al., 2016, Venturi et al., 2008b). The presence of virus-specific public TCRs may be due partly to preferential use of specific TCR V and J chains in response to conserved hierarchy of epitope recognition (Chen et al., 2000, Hancock et al., 2015, Kim et al., 2013). Public TCR sequences from antigen-experienced T cells should be readily identifiable within the circulating T cell repertoire because of clonal expansion and the formation of memory T cell populations (Emerson and DeWitt, 2017, Heit et al., 2017).
Identifying antigen-specific T cells and tracking an antigen-specific response over time within individuals is a difficult task, especially against emerging pathogens, in which case precise immunogenic epitopes are not well described. Even when antigens are known, the frequencies of antigen-specific T cell populations are notoriously low and can often be difficult to identify (Douillard et al., 1997, Wolf and DiPaolo, 2016, Lim et al., 2000). This is due partly to a lack of knowledge concerning antigen-specific TCR sequences. Another issue is that antigen-specific TCR identification using many traditional immune assays is restricted to the most high-frequency responders (Wolf and DiPaolo, 2016, van der Velden et al., 2014, van der Velden and van Dongen, 2009). However, advancements in next-generation sequencing are allowing researchers to analyze TCR and B cell receptor (BCR) (Ig) repertoires (immunosequencing) with unprecedented depth and sensitivity, identifying 105-107 individual sequences in humans from a very limited volume of whole blood (DeWitt et al., 2015, Faham et al., 2012, Kirsch et al., 2015, Logan et al., 2014, Robins et al., 2009).
What is needed is a method to evaluate an exposure status of a subject to an immunogenic agent. This method ideally should be able to evaluate whether or not a subject has been exposed to the immunogenic agent and should be sensitive enough to accurately distinguish between closely related immunogenic agents. Further, a method for evaluating the effectiveness of vaccines (particularly their ability to generate a robust immune response) is needed.
Provided herein is a method for determining whether a subject has been exposed to an immunogenic antigen. The method comprises: amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA obtained from T-cells of the subject; identifying unique TCRβ alleles sequences in T-cells of the subject to generate a TCRβ clonotype profile of the subject; comparing the TCRβ clonotype profile of the subject to a database of target associated receptor sequences (TARSs) comprising unique TCRβ alleles identified as associated with exposure to the immunogenic antigen in a cohort of independent test subjects; generating a diagnostic classifier of the subject comprising the number of TARSs identified in the subject relative to the total number of unique TCRβ alleles in the subject; and determining that the subject has been exposed to the immunogenic antigen if the diagnostic classifier exceeds a predetermined threshold, wherein the predetermined threshold is determined by the prevalence of TARSs in the test cohort after exposure to the immunogenic antigen.
Also provided is a method for testing the efficacy of a vaccine. The method comprises: amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA of T-cells obtained from the subject after administration of the vaccine; comparing the TCRβ clonotype profile of the subject to a database of vaccine associated TCRβ sequences (VATSs) statistically associated with vaccination to generate a diagnostic classifier of the subject, wherein the diagnostic classifier comprises the number of VATSs identified in the subject relative to the total number of unique TCRβ alleles in the subject; and determining that the vaccine is effective in generating an immune response if the diagnostic classifier exceeds a threshold determined by the prevalence of VATSs in an independent test cohort after exposure to the vaccine.
Also provided is a method of identifying a viral infection in a subject. The method comprises: amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA of T-cells obtained from the subject; comparing the TCRβ sequences in the subject to one or more databases of virus-associated TCRβ sequences, wherein each database comprises TCRβ sequences statistically associated with one virus and each database is generated according to the methods described herein; and identifying the viral infection of the subject by determining the strength of the association of the TCRβ allele sequences identified in the subject to one or more of the databases.
A further method of identifying an immune response in a subject is provided, the method comprising: identifying in the subject the presence of a significant number of unique TCRβ clonotypes that match a database of TCRβ sequences previously associated with the immune response in an independent cohort.
In all of the methods provided herein, a TCRβ database is generated. Accordingly, a method of generating a TCRβ database is also provided, wherein the TCRβ database comprises TCRβ sequences statistically associated with an immune condition, exposure to a vaccine or immunogenic agent, and/or a pathogen, the method comprising: amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA of T-cells obtained from a cohort of subjects having the immune condition, or having been exposed to the vaccine, immunogenic agent and/or pathogen; and using a machine learning and/or neural network system to analyze the TCRβ allele sequences and statistically associate a subset of the TCRβ sequences to the immune condition, vaccine, immunogenic agent and/or pathogen.
Other objects and features will be in part apparent and in part pointed out hereinafter.
The methods provided herein are directed to examining the T-cell receptor (TCR) repertoire of the subject. During T cell development, each TCR chain is generated through quasi-random genetic recombination from the germline loci of the variable (V), diversity (D), and joining (J) gene segments. T-cells express antigen specific TCRs which are expressed from a highly polymorphic TCR gene locus comprising V, D and J gene segments. On average, approximately 107 unique TCRβ chains can be identified from the approximately 1012 circulating T cells present in a healthy human adult. The ability to readily identify identical TCR sequences among multiple individuals (public TCRs) is challenging because an individual has the potential to generate approximately 1018 unique TCR recombinants. Moreover, there is no guarantee that two individuals will express the same TCR to the same antigen. Further, identifying TCR sequences that correlate with an infection can be more difficult the more time passes from the infection as clonally expanded T-cells that were upregulated during the initial immune response are depleted, leaving only a small population of memory T-cells. The present invention addresses each of these issues.
As noted above, a method is provided herein for determining whether a subject has been exposed to an immunogenic antigen. As used herein, the term “immunogenic antigen” comprises any antigen that elicits a robust immune response. In general, the robust immune response comprises humoral and cell-mediated immunity (e.g., upregulation of antigen-specific B- and T-cells in the subject, respectively).
Accordingly, in various embodiments, the methods described herein comprise amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA obtained from T-cells isolated from the subject. TCRβ alleles are well characterized in the art as are methods of amplifying and expanding. For example, the multiplex method of isolating TCRβ genes may be carried out according to previously published methods (e.g., using multiplexed primers targeting all V and J gene segments as described by Carlson et al., 2013, 2013, “Using synthetic templates to design an unbiased multiplex PCR assay. Nat. Comm. 4, 2680 and incorporated herein by reference in its entirety). The genetic diversity of the population (e.g., humans) may require increased sequencing depth. Accordingly, the sequencing may further comprise an ultra-deep sequencing protocol to achieve read depths up of at least about 2 million, at least about 3 million, or at least about 5 million reads. For example, the sequencing can be performed at a depth of from about 2 million to about 100 million reads, from about 2 million to about 10 million reads, from about 2 million to about 5 million reads, from about 4 million to about 100 million reads, from about 4 million to about 10 million reads, from about 4 million to about 6 million reads, or from about 4 million to about 5 million reads.
Once the TCRβ alleles are amplified and sequenced, the method further comprises identifying unique TCRβ alleles in the samples to generate a TCRβ clonotype profile. As used herein, the word ‘unique” means a unique sequence among the total number of TCRβ sequences identified. The word “unique” does not imply that the identified sequences have multiple copies in the original sample.
In various embodiments, the TCRβ clonotype profile (e.g., unique TCRβ allele sequences identified in the sample) is compared to a database of target associated receptor sequences (TARSs) comprising unique TCRβ allele sequences statistically associated with the immunogenic antigen in an independent cohort of test subjects to generate a diagnostic classifier of the sample. The diagnostic classifier comprises the number of TARSs identified in the subject relative to the total number of unique TCRβ alleles in the subject.
In further embodiments, the method comprises determining that the subject has been exposed to the immunogenic antigen if the diagnostic classifier exceeds a predetermined threshold for the diagnostic classifier, wherein the predetermined threshold is determined by the prevalence of TARSs in the test cohort after exposure to the immunogenic antigen.
The method described herein therefore comprises two steps of (a) preparing a database of TCRβ sequences associated with the immunogenic antigen and (b) comparing the TCRβ sequences of the subject to be evaluated with that database. Each of these steps are described in more detail below.
Preparing a Database of TCRβ Sequences Associated with an Immunogenic Antigen
In various embodiments, generating the database of “target associated receptor sequences” TARSs comprises analyzing the shared immune response of an independent cohort of test subjects following an exposure to the antigen. Accordingly, the method can comprise amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA obtained from T cells of the test subjects, wherein the T cells are isolated before and after exposure to the immunogenic antigen; identifying unique TCRβ allele sequences in the cohort of test subjects; performing a Fisher exact test on each unique TCRβ sequence to generate a statistical association (i.e., a p-value) between the TCRβ sequence and the exposure status of the subject at the time the T-cells were obtained (that is, whether the T-cell sample as collected “before” or “after” exposure); generating a database of TARSs comprising unique TCRβ sequences having a p-value that exceeds a p-value threshold.
In various embodiments, the p-value threshold is determined empirically for the cohort of test subjects used. Specifically, the p-value threshold is the p-value that generates a TARSs database having the maximum coverage ratio. As used herein, the term “coverage ratio” is defined as the ratio of “Cp” to “Cn”, wherein “Cp” and “Cn” are, respectively, the proportion of exposed (Cp) or naïve (Cn) samples having at least one TCRβ sequence included in the TARSs database relative to the total number of exposed samples (when calculating “Cp”) or naïve samples (when calculating “Cn”). In other words, a coverage ratio can be calculated using the following equation, where Cv represented Cp as described above, Cn is as described above, and “xi” and yi represent the total number of exposed samples or naïve samples, respectfully, that a single TCRβ is identified in and nv and nn represent the total number of exposed samples or naïve samples, respectfully:
Accordingly, the p-value threshold can be determined by sorting TCRβ sequences into “exposed”-associated or “naïve”-associated groups using a range of p value thresholds (although p values tested should not exceed 0.20). The coverage ratio can be calculated for each p value and the p value that yields a maximum (e.g., highest) coverage ratio can then be selected as the p-value threshold to define the final TCRβ database associated with the immunogenic antigen. Importantly, the final TCRβ database is generally considered to be static and is not altered when an unknown subject must be classified. Accordingly, in some embodiments, the TCRβ database is not regenerated every time an unknown subject is classified.
Preferably, generating the TCRβ database associated with the immunogenic antigen comprises using a machine learning or neural network platform that efficiently sorts TCRβ sequences into “exposed” or “naïve” classes. Although the Fisher exact test is provided as an exemplary statistical test to classify the sequences, other statistical tests and methods may be used. Preferably, generating the TCRβ database comprises using a neural network or machine learning interface that trains on data gathered from naïve or exposed samples and determines relationships between the TCRβ allele sequences and their association with exposure to the antigen.
In various embodiments, the method of generating the TARSs database further comprises validating the database by identifying one or more splenocytes present in the test subjects of the cohort after exposure to the immunogenic antigen that express one or more of the TARSs in the database. In various embodiments, the splenocytes may be identified using an in vitro clonal expansion experiment where splenocytes are exposed to the immunogenic antigen in vitro, clonally expand and are analyzed to determine the sequences of their expressed TCRβ chains. In other embodiments, splenocytes may be analyzed in a flow cytometry procedure where MHC-peptide tetramers are used to bind to and label T-cell receptors on the splenocytes. In this embodiment, the MHC-peptide tetramers are the extracellular binding domain of the major histocompatibility complex (MHC) associated with an antigen peptide. Preferably, the antigen peptide is associated with (or mirrors) the immunogenic antigen used to generate the TARSs database. In various embodiments, the MHC antigen peptide can comprise any one of SEQ ID NO: 675-683. The MHC protein can comprise a human leukocyte antigen peptide (e.g., HLA-A2). Splenocytes that are isolated using this method can be further analyzed to determine their TCRβ sequences to determine whether they match the TCRβ sequences on the database.
As described above, the methods provided comprise classifying a subject of unknown status as either exposed or naïve depending on whether its diagnostic classifier exceeds a predetermined threshold. In various embodiments, the comparison of the diagnostic classifier with the predetermined threshold further comprises applying a probability distribution function that compares the diagnostic classifier of the subject to a distribution of TARSs prevalence in the test subject cohort after exposure to the immunogenic antigen. As used herein, “prevalence” refers to the ratio of unique TARSs identified in each sample relative to the total number of unique TCRβ sequences in each sample.
Accordingly, the methods described herein enable one to evaluate an unknown subject against a predetermined database of TCRβ sequences associated with exposure to the immunogenic antigen. Importantly, since this database is evaluated independently of the test subject, the immune profile of the subject can be re-evaluated through time. Accordingly, the methods described herein can further comprise dynamically tracking an immune response to the subject over time, the method comprising generating a plurality of diagnostic classifier scores using T-cell samples obtained from the subject at different time points and comparing the diagnostic classifiers to a TARSs database associated with the immune response. Further, generating the diagnostic classifiers of the subject does not alter the TARSs database.
In various embodiments, the methods comprise analyzing a sample of T-cells obtained from the subject up to 9 months after a potential exposure event to the immunogenic antigen. For example, in various embodiments, the sample of T-cells may be obtained around 2 weeks, around 4 weeks, around 6 weeks, around 12 weeks, around 24 weeks, and/or around 36 weeks after the potential exposure event to the immunogenic antigen. In various embodiments, the T cells can comprise CD8+ T cells.
As noted above TCRβ alleles are unique for each T-cell and are generated via thymic recombination of various V, D and J regions of the TCR gene. Accordingly, the TCRβ allele can comprise the associated V region and J region of the TCR gene and the corresponding CDR3 sequence that spans the two. Further, as would be understood by one of skill in the art, once a genomic allele is determined the corresponding amino acid sequence encoded by that allele is easy to obtain. Consequently, as used herein, the word “allele” refers to the gene as provided in DNA or transcribed to mRNA, as well as the gene expressed into protein (amino acid sequence). As used herein, the TCRβ sequences are represented using nomenclature established by the international ImMunoGeneTics (IMGT) system (www.imgt.org). In this system, the variable (v) and joining (j) genes are named and the hypervariable region that spans them (CDR3) is provided as an amino acid sequence. For example, a TCRβ sequence can be represented as: “TCRBV03-01 CASSLGFYEQYF TCRBJ02-07”. In this nomenclature, “TCRBV03-01” and TCRBJ02-07 represent the IMGT classified name for the “v” and J regions, respectfully, and can be identified from public databases (e.g., imgt.org). The sequence CASSLGFYEQYF is the hypervariable CDR3 region and is assigned SEQ ID NO: 121 herein. Accordingly, once provided with an allele name (V-CDR3-J) one can identify the underlying sequence easily using a database such as found on www.imgt.org.
For example, given the V-CDR3-J name, one can obtain the corresponding TRBV and TCRBJ segments (as amino acid sequences) and align the end of the TRBV sequence to the beginning of the CDR3 sequence and the beginning of the TRBJ sequence to the end of the CDR3 sequence to find overlapping amino acid sequences and then combine into a single sequence.
In various embodiments, the TCRβ allele comprises a CDR3 variable region in a recombined TCRβ allele. The CDR3 variable region can comprise an amino acid sequence comprising any one of SEQ ID NOs: 1-674. In further embodiments, the TCRβ allele comprises the V region, the CDR3 variable region and the J region of a recombined TCRβ allele.
The methods described herein may be used to determine whether a subject has been exposed to an immunogenic antigen. In various embodiments, the immunogenic antigen can comprise a pathogen, an allergen, a vaccine, a virus or any immunogenic component or fragment thereof. In some embodiments, the methods comprise identifying an immune response in the subject, provided the immune response is mediated by T-cell upregulation.
In various embodiments, the immunogenic antigen comprises a virus or a vaccine. For example, the immunogenic antigen can comprise an Orthopoxvirus (e.g., smallpox or monkey pox), a coronavirus (e.g., SARS-COV, SARS-COV-2, or MERS), an influenza virus (e.g., Influenza A or Influenza B). As another example, the immunogenic antigen can comprise a vaccine to any of these viruses. So, for example, the immunogenic antigen can comprise an Orthopoxvirus vaccine (e.g., the smallpox vaccine or another Orthopoxvirus vaccine), a coronavirus vaccine (e.g., a SARS COV-2 vaccine) or an influenza vaccine.
In various embodiments, when the immunogenic antigen comprises an Orthopoxvirus (e.g., monkey pox), the TARSs database comprising TCRβ allele sequences associated with the infection (e.g., with monkey pox) can comprise any one of SEQ ID NOs: 1-120. As noted above, the TCRβ allele sequences are annotated to indicate the “V” gene, the “J” gene and the CDR3 amino acid sequence that comprises the final recombined allele. Each of the CDR3 sequences is assigned a SEQ ID NO. For ease of reference, SEQ ID NOs: 1-120 are indicated in Table 1 below. The TARSs associated with monkey pox infection and provided in Table 1 comprise murine TCRβ alleles.
In various embodiments, the immunogenic antigen comprises a vaccine (e.g., a smallpox vaccine). For example, the immunogenic antigen can comprise the ACAM2000 smallpox vaccine. In various embodiments, when a TCRβ allele on the TARSs database that is associated with the smallpox vaccine can comprise any one of SEQ ID NOs: 121-435. For ease of reference SEQ ID NOs: 121-435 are provided in Table 2 below. As above, individual clonotypes are identified using IMGT standard nomenclature (V-CDR3-J). The international ImMunoGenTics database is available (www.imgt.org) and can be used to generate the raw sequences provided below. The TARSs associated with smallpox vaccination and provided in Table 2 comprise murine TCRβ alleles.
In various embodiments, the TARSs database comprising TCRβ sequences associated with smallpox vaccination is generated from a cohort of human subjects. Accordingly, in various embodiments, human TARSs associated with small pox vaccination can comprise any one of SEQ ID NOs: 436-674 (Table 3, below). As above, the TCRβ alleles are provided in IMTG nomenclature and identify the relevant human variable (V) and joining (J) segment that must be combined with the indicated CDR3 sequence to generate the relevant TCRβ allele. Nucleic acid and amino acid sequences for all of the human V and J regions used in this table can be obtained from the International ImMunoGenTics database is available (www.imgt.org).
A method of testing the efficacy of a vaccine is also provided. In various embodiments, a vaccine is considered “effective” if it stimulates a robust immune response. For instance, an effective vaccine would be expected to stimulate T-cell expansion and antibody generation against an immunogenic antigen comprised by the vaccine. The methods provided herein can test the efficacy of a vaccine by identifying TCRβ sequences in the subject that are associated with the vaccination.
In various embodiments, the method of testing the efficacy of a vaccine comprises: (a) amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA of T-cells obtained from a subject after administration of the vaccine; (b) comparing the TCRβ clonotype profile of the subject to a database of vaccine associated TCRβ sequences (VATSs) statistically associated with vaccination to generate a diagnostic classifier of the subject, wherein the diagnostic classifier comprises the number of VATSs identified in the subject relative to the total number of unique TCRβ alleles in the subject; and (c) determining that the vaccine is effective in generating an immune response if the diagnostic classifier exceeds a threshold determined by the prevalence of VATSs in an independent test cohort after exposure to the vaccine.
In various embodiments, the method of testing the efficacy of the vaccine can further comprise administering the vaccine to the subject.
In various embodiments, the vaccine tested can comprise an Orthopoxvirus vaccine (e.g., the smallpox vaccine) a coronavirus vaccine (e.g., a SARS-CoV-2 vaccine) or an influenza vaccine (e.g., Influenza A or Influenza B vaccine) In various embodiments, the vaccine can comprise the smallpox vaccine and a TCRβ allele associated with the vaccination can comprise any one of SEQ ID NOs: 122-435.
Also provided are methods of detecting/identifying a viral infection in a subject. In various embodiments, the method can comprise: (a) amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA of T-cells obtained from the subject; (b) comparing the TCRβ sequences in the subject to one or more databases of virus-associated TCRβ sequences, wherein each database comprises TCRβ sequences statistically associated with one virus and each database is generated according to the methods described above; and (c) identifying the viral infection of the subject by determining the strength of the association of the TCRβ allele sequences identified in the subject to one or more of the databases.
In various embodiments, the strength of the association can comprise performing a probability distribution function to determine whether the TCRβ clonotype profile of the subject is statistically similar to the TCRβ clonotype distribution in naïve or virus infected samples.
Advantageously, the method described herein can be used to distinguish between viruses that present with similar symptoms and etiology but stimulate clonal expansion of different T cell populations.
In various embodiments, the viral infection can comprise a smallpox infection. In various embodiments, the method can distinguish between a smallpox virus and a Zika virus.
In various embodiments, the viral infection can comprise a coronavirus infection. Exemplary coronaviruses include Severe Acute Respiratory Syndrome coronaviruses (e.g., SARS, including the new SARS-CoV-2 strain) and Middle Eastern Respiratory Syndrome (MERS) coronavirus. One useful application for this method is to identify individuals infected with SARS-CoV-2 (i.e., COVID-19). In additional embodiments, the viral infection can comprise influenza (e.g., Influenza A or Influenza B). For example, the influenza virus can comprise an H1N1 Influenza A strain.
In various embodiments, when the methods provided herein comprise analyzing a sample obtained from a subject, the subject can be a mammal. In various embodiments, the subject is a mouse. In other embodiments, the subject is a human.
The immune repertoire of a human is orders of magnitude larger than that of a mouse (particularly a model organism kept in immune privileged conditions and genetically identical to other subjects). This presents unique challenges in generating the TARSs database of TCRβ sequences associated with an immune response in humans.
Accordingly, a method is also provided for generating a TCRβ database comprising TCRβ sequences statistically associated with an immune condition, exposure to a vaccine or immunogenic agent and/or pathogen. The method comprises: (a) amplifying and sequencing TCRβ alleles in mRNA and/or genomic DNA of T-cells obtained from a cohort of subjects having the immune condition, or having been exposed to the vaccine, immunogenic agent and/or pathogen; and (b) using a machine learning and/or deep neural network system to analyze the TCRβ allele sequences and statistically associate a subset of the TCRβ sequences to the immune condition, vaccine, immunogenic agent and/or pathogen. The machine learning and/or deep neural network can perform the Fisher exact tests described above, or may perform different statistical tests.
As one of the most powerful machine learning methods, the deep learning neural network has been substantially employed to explore the high-level features hidden in biomedical data. As provided herein, the deep learning framework is used to train the deep learning models for diagnostic discrimination. A multi-layer neural network is used to extract hidden patterns from the input features through differing numbers of hidden layers (i.e., using more than three layers, more than four layers, or more than five layers, etc.). The extracted hidden features are finally fed into the last layer of logistic regression to classify the sample into binary classes. See
In this work, the predictive ability of the DNN method for diagnostic discrimination of viral infection are evaluated to understand how immune system features can diagnose viral infection status. The frequency counts of all CDR3 amino acid sequences (a.k.a. peptides) can be calculated from quantified TCR beta chain sequence data and used as input features for machine learning methods to build discriminative classifiers. Each negative sample (pre-inoculation) or positive sample (post-inoculation) can be described as a vector of frequency counts, each representing the number of CDR3 amino acids found in the sequence data of the sample.
The analysis starts with the data partition. A stratified sampling method can be used to randomly divide the data into subsets according to the status of infection (pre- or post-introduction). The datasets can then be further partitioned into a training set and a testing set. In various embodiments, the training set can comprise at least about 50%, at least 60%, at least 70%, or at least 75%, of the data. For example, the training set can comprise about 50% to 90%, about 50% to 80%, about 50% to 75%, about 60% to 90%, about 60% to 80%, or about 70% to 80% of the data. Data not incorporated into the training set can be reserved to test the model (e.g., see
Once the data is divided, a repeated multi-fold cross-validation can be used to estimate the optimal parameters of each machine learning algorithm on the training dataset. The best training parameters selected by cross-validation are used to retrain the whole training dataset to derive the final model for evaluation. The independent testing subset is only seen when the final model of each algorithm is determined. After the training data is collected, several data normalization schemes are attempted before applying machine learning algorithms for model learning. Due to the different experimental conditions (i.e., sequence depth) and sample variations, the number of frequency counts for amino acids might vary in magnitude. Normalization might be necessary to remove inherent bias for different machine learning methods. For example, the data may be transformed using (1) peptide-based normalization that normalizes counts across all training samples within each amino acid sequence; (2) sample-based normalization that normalizes counts of amino acids within individual samples; and/or (3) the benchmark data that uses original counts without any normalization. The Minimum-Maximum transformation is adopted to convert counts into the range between zero and one when the normalization is needed. The normalized/original features are then used to train different machine learning models for infection diagnosis.
As a nonlimiting example, the model may comprise 5 hidden layers, 90 nodes (neurons), and 1000 max iterations. The model can show high prediction accuracy (e.g., greater than 80%, greater than 85%, greater than 90%, greater than 95%, greater than 96% or about 97%) when tested on previously unseen (independent test) samples while retaining 100% accuracy when identifying previously seen (training) samples. In addition, the parameters can be properly configured by randomized hyper-parameter search strategy since the DNN algorithm may affect proposed model's effectiveness. If inappropriate parameters are selected, the weights or coefficients in the deep neural network do not converge, rendering the trained models unusable.
The accuracy of this protocol in humans, which have significantly more genetic diversity than lab mice, can be improved by increasing accuracy of the sequencing of TCRβ alleles in the T-cells of the population. One way to increase sequencing accuracy, provided herein, is to use an ultra-deep sequencing protocol. In ultra-deep sequencing, the number of independent reads of a given sequence often exceed 1 million or even 2, 3, 4 or 5 million. Accordingly, in various embodiments, any steps provided herein that require amplifying and sequencing TCRβ alleles may be performed using an ultra-deep sequencing protocol. In various embodiments, the sequencing of the TCRβ alleles is performed at a depth of at least about 2 million, at least about 3 million, or at least about 5 million reads. For example, the sequencing of the TCRβ alleles can be performed at a depth of from about 2 million to about 100 million reads, from about 2 million to about 10 million reads, from about 2 million to about 5 million reads, from about 4 million to about 100 million reads, from about 4 million to about 10 million reads, from about 4 million to about 6 million reads, or from about 4 million to about 5 million reads.
Having described the invention in detail, it will be apparent that modifications and variations are possible without departing from the scope of the invention defined in the appended claims.
The following non-limiting examples are provided to further illustrate the present invention.
Materials and reagents used in the following examples are provided in Table 4 below. Common methods also used are provided herein below.
HLA-A2.1 AAD C57BL/6 male and female mice were purchased from Jackson Laboratories and maintained under specific pathogen conditions (Gilchuck et al., 2013). The HLA-A2.1 AAD mice express a transgenic HLA-A2.1 chimeric molecule containing the human β-2 microglobulin and HLA-A2.1 α1 and α2 domains with a mouse 3 and transmembrane domain (AAD HLA-A2). Mice entered the study at approximately 6-8 weeks of age. Male mice weighed between 18 and 23 g, female mice weighed between 16-21 g. Mice were housed in groups of 3 to 5 mice per cage. All animal work has been conducted in accordance with the Guide for Care and Use of Laboratory Animals of the National Institute of Health with approval from the Saint Louis University Institutional Animal Care and Use Committee.
The ACAM2000 (Acambis, Inc.) smallpox vaccine is a live virus derived from the original Dryvax (Wyeth Laboratories, Inc.). MPXV is a member of the Orthopoxvirus family and is 95% genetically identical to the smallpox vaccine. Vaccination with the smallpox vaccine confers protection against MPXV infection (Handley et al., 2009). MPXV Zaire 79 was obtained from the Saint Louis University School of Medicine, department of molecular microbiology and immunology Biosafety Level 3/Select Agent program. ACAM2000 smallpox vaccine was a gift from the center for disease control (Atlanta, Ga.). Both MPXV and ACAM2000 smallpox vaccine titers and infectivity were estimated plaque forming assay (Handley et al., 2009; Parker et al., 2014). 1×105 BSC-1 cells were plated in DMEM+ 10% FCS in 24-well plates at 0.5 mL final volume and cultured for 24 hours at 37° C. and 5% CO2. Stock viral suspensions were serially diluted 1:10 in DMEM+1% FCS, and 300 μL of supernatant from BSC-1 cells was removed. 100 μL of diluted virus solutions were added to BSC-1 cells in triplicate, gently swirled, and incubated for 1 hour at 37° C. and 5% CO2. Overlay media (1% carboxyl methyl cellulose in DMEM with 5% FCS) was warmed to 37° C. and 1 mL added to each well and cultures were returned to the incubator. Cultures were maintained for 3-4 days at 37° C. until plaques were visible. Plaques were visualized and virus inactivated with 200 μL 0.3% crystal violet/10% formalin solution for 1 hour. All liquid was aspirated and plates were washed with water, inverted, and dried. Plaques were counted for each viral dilution, the average plaque count is divided by product of the dilution and volume of virus overlay.
Spleens from male and female mice previously vaccinated with the ACAM2000 smallpox vaccine were mechanically disrupted to form single cell suspensions. Cells were filtered through a 40 um nylon mesh cell strainer and washed with complete RPMI supplemented with 10% fetal bovine serum (cRPMI10F, 1% Penicillin/Streptomycin (Sigma-Aldrich, P0871, 10,000 U/10 mg per mL), 1% L-glutamine (Sigma-Aldrich, G7513, 200 mM), 1% non-essential amino acids (Sigma-Aldrich, M7145, 100×), 1% HEPES (Sigma-Aldrich, H3537), 1% sodium pyruvate (Sigma-Aldrich, S8636, 100 mM)). Red blood cells were lysed by incubation with 5 mL 1×ACK (ammonium-chloride-potassium) lysing buffer for 5 minutes at 37° C. and 5% CO2. Cells were washed with cRPMI10F and cell count determined. Splenocytes were cultured in cRPMI10F at a concentration of 1×106/mL at 37° C. and 5% CO2.
At 6-8 weeks of age mice were anesthetized by intraperitoneal injection with 0.01 mL/g body weight Ketamine (6 mg/mL)/Xylazine (0.5 mg/mL) cocktail and intranasally administered sub-lethal doses of either the ACAM2000 smallpox vaccine (approximately 5×104 PFU) or MPXV Zaire-79 strain (0.5×104 PFU) in 25 μL total volume (12.5 μL per naris) (Moutaftsi et al., 2009, Parker and Buller, 2013, Parker et al., 2014, Stabenow et al., 2010). Male and female mice were entered into each treatment group with approximate equality by cage. All mice recovered from the viral challenge. Blood samples were collected via the submandibular cheek bleed 1 week prior to vaccination or infection and at 2 weeks, 8 weeks, 16 weeks, and 9 months post-vaccination or post-infection.
Serum from HLA-A2 mice was collected by aspiration from whole blood prior to vaccination or infection and at 2-weeks and 8-weeks after viral exposure by centrifugation at 11,000 g for 5 min. Samples were tested for vaccinia-specific serum antibody by neutralizing antibody ELISA (Frey et al., 2002). 96-well maxSorp plates (ThermoFisher 44-2404-21) were coated with crude extract from lysed BSC-1 cells infected with live ACAM2000 smallpox vaccine. Plates were washed 3× with PBS and wells were coated with approximately 5×104 PFU live ACAM2000 diluted in carbonate coating buffer (0.1M Na2CO3 0.1M NaHCO3 pH 9.3) in 100 ul overnight at 4° C. Wells were washed and blocked with blocking buffer containing 5% BSA in PBS for 30 min at room temperature (RT). Plates were washed 3× with PBS and serum samples diluted 1:20 in blocking buffer were added (in duplicate) to wells and incubated for 2 hours at room temperature (RT). Plate was washed 3× with PBS, then 100 μL anti-mouse HRP-conjugated antibody (Sigma A8924) was diluted 1:2500 in blocking buffer was added and allowed to incubate for 1 h. Plates were washed 3× with PBS and 75 μl of True Blue peroxidase substrate (KPL, 71-00-65) was added to each well and incubated for 15 min in the dark at room temperature (RT). 75 μl of 1N HCl was added to each well and OD measured at 450 nm.
Genomic DNA was extracted and purified using the QIAGEN Blood and Tissue Kit (item #69506). Genomic DNA was amplified using multiplexed primers targeting all V and J gene segments as described previously (Carlson et al., 2013). tcrb CDR3 regions were amplified and sequenced using ImmunoSEQ (Adaptive Biotechnologies). Synthetic templates mimicking natural V(D)J rearrangements were used to measure and correct potential amplification bias (Carlson et al., 2013, Wolf et al., 2016). CDR3 segments were annotated according to the International ImMunoGeneTics (IMGT) collaboration (Yousfi Monod et al., 2004), identifying V, D, and J genes contributing to each rearrangement.
Cells were maintained in cRPMI10F media alone or supplemented with 0.2MOI live ACAM2000 smallpox vaccine and allowed to incubate at 37° C. for 5 days. T cell blasting and proliferation were observed prior to DNA extraction for Immunosequencing. Cultured cells were pelleted by centrifugation at 1500 rpm for 10 minutes and supernatant aspirated. Pelleted cells were re-suspended in 200 μl PBS prior to DNA extraction.
PE-conjugated HLA-A2.1 chimeric tetramers (HLA-A2 tetramers) loaded with vaccinia-derived peptides (Gilchuk et al., 2013) were a kind gift from Dr. Sebastian Joyce (Vanderbilt). Pooled splenocytes from previously vaccinated mice were stained with T cell and B cell lineage markers CD4 (ebioscience, PerCP-efluor710, clone GK1.5), CD8 (BD Horizon, BV450, clone 53-6.7), and CD19 (BD Horizon, BV605, clone D3) for 20 min at 4° C. Cells were washed 2× in PBS+2% FBS, then cells were incubated with vaccinia-peptide loaded HLA-A2 tetramers (25 μg/mL, 2 μL per 1×106 splenocytes) for 1 h at RT. TCR-epitope-tetramer binding CD4-CD19-CD8+ T cell populations were purified by FACS into tetramer− and TCR-epitope-tetramer binding tetramer+ populations. Tetramer sorted T cells were centrifuged at 1500 rpm for 10 minutes. Supernatant was aspirated and cells re-suspended in 200 μl PBS prior to DNA extraction.
Alignment of shared and non-shared TCRO sequences was completed using ImmunoSEQ software provided by Adaptive Biotechnologies. Alignments of all 2-week and 8-week post-ACAM2000 smallpox vaccination TCR repertoires were used to identify public TCR sequences. The list of public TCR sequences was compared to alignments of all TCRO repertoires from naive samples in order to perform an association analysis to identify a set of TCRβ sequences that had significantly increased incidence among vaccinated but not naive TCRβ repertoires. For the association analysis, we performed a one-tailed Fisher's Exact test on all sequences, comparing the number of naive and vaccinated samples each TCRβ sequence was present in.
To determine an optimal p value threshold for identifying VATS, we applied a heuristic test that selected the optimal p value threshold based on the “coverage” provided by the library for both vaccinated (Cv) and naive samples (Cn). “Coverage” is defined as the summation of the number of samples containing each VATS divided by the number of samples. In the equations below, xi denotes the number of vaccinated samples a single TCRβ is identified in (yi denotes naive samples) and nv represents the number of samples in the training data (nn represents naive samples).
The ratio of Cv to Cn is determined for each p value. Additionally, the Cv and Cn of each p value (rounded to the nearest whole integer), are applied to a one-tailed Fisher's Exact test against the total number of sequences in the prospective library to determine if there is sufficient coverage to distinguish vaccinated from naive samples (p<0.05). The p value with the largest Cv:Cn ratio and offers significant coverage to distinguish vaccinated from naive samples was chosen.
To distinguish between vaccinated and naive samples, the proportion of VATS present in a sample was compared against the normal distribution of the naive and vaccinated training data. Normal distribution for our purposes is used to measure the distance a sample is from the mean. The normal distributions for the naive and vaccinated populations in our training data were calculated based on a function of the difference between a single sample value (x) and the mean of a set of data (μ) over the standard deviation of that set of data (σ). The greater the value, the greater association that sample has with the training group. By comparing a sample against the normal distribution of vaccinated and naive training groups, we can determine which group a sample is more statistically associated with.
The Leave-one-out (LOO) analysis was completed as previously described (Emerson and DeWitt, 2017). Briefly, all samples associated with a single mouse were removed from the training data and the VATS library was re-derived using the remaining training cohort. The % VATS was calculated for all samples and used to train the diagnostic classifier.
Alignment of shared and non-shared TCRs was completed using ImmunoSeq software provided by Adaptive Biotechnologies. Graphical analyses were created using GraphPad Prism 5.0. 1-way ANOVA and Bonferroni's multiple comparison test was accomplished using GraphPad Prism 5.0. Pearson correlation was calculated using GraphPad Prism 5.0.
An extensive TCR sequence database was generated from a large cohort of HLA-A2 transgenic (AAD) mice before and up to 9 months after administration of the ACAM2000 smallpox vaccine or infection with MPXV (
Consolidated data referencing the total number of mice, unique TCRβ sequences (clonotypes), total number of rearranged TCRβ genes sequenced for each time point for naive, ACAM2000 vaccinated, and MPXV infected samples
TCR repertoires from whole blood of mice pre- and post-vaccination were analyzed to computationally identify TCR clonotypes present post-vaccination but absent pre-vaccination versus sequences present pre- and post-vaccination (
Next it was determined whether the computationally identified VATS contained TCR clonotypes that functionally expanded in response to smallpox vaccine. Splenocytes of mice from the original cohort 12 weeks after vaccination were cultured with or without the smallpox vaccine for 5 days to induce expansion of vaccine-specific T cells in vitro. Intra-mouse analysis of the TCR repertoires pre- and 2 and 8 weeks post-vaccination were compared with the libraries from splenocytes cultured with or without ACAM2000. It has been previously shown that the smallpox vaccine does not induce bystander activation of CD8+ T cells, which leads us to conclude that TCR sequences from proliferating T cells in this experimental design are virus specific (Miller et al., 2008). The relative abundances of vaccine-associated and non-vaccine-associated TCR clonotypes were compared between vaccine-stimulated and unstimulated cultures. Post-vaccine-associated sequences were significantly expanded (8.9-fold, p<0.0001) in the vaccine-stimulated versus unstimulated controls; this is significantly greater (p<0.0001) than the expansion (0.94-fold) measured in non-VATS (
To distinguish between TCR repertoires from naive and exposed samples, a vaccine-associated public TCRβ library was generated. Pre- and post-vaccination TCRβ sequence libraries were analyzed, computationally detecting the virus-specific T cell response by identifying sequences that were statistically associated with post-vaccination samples. TCRβ sequences from all naive (n=32) and 2 and 8 week post-vaccination TCRβ repertoires (n=58) were used for this analysis. Each TCRβ clonotype identified from the 58 post-vaccinated samples (approximately 576,000) was analyzed using a one-tailed Fisher's exact test for association with vaccinated libraries compared with naive libraries (
A diagnostic classifier was developed by calculating the number of VATS present relative to the total number of unique TCRβ clonotypes present for each sample. It was observed that the number of VATS present in a sample was significantly correlated with the total number of unique TCRβ clonotypes in both vaccinated and naive samples, indicating that the number of TCRβ clonotypes present directly affects the number of VATS identified (
A binary classification system was constructed to differentiate naive and vaccinated samples on the basis of the normal distribution of % VATS from the TCR repertoires of the naive or vaccinated groups. In this way, the TCR repertoires from the vaccinated and naive samples act as “training data,” teaching the diagnostic classifier to predict the vaccination status of samples (See the Materials and Methods provided before Example 1 herein).
A comparison of naive and smallpox-vaccinated samples showed a 43-fold increase in the % VATS in vaccinated repertoires (average 0.248±0.047%) compared with naive repertoires (average 0.0057±0.0039%) (
To test whether the VATS diagnostic classifier was capable of identifying the vaccine-specific T cell response after the primary infection and generation of long-term memory, blood collected 16 weeks and 9 months after vaccination was analyzed. Sixteen week and 9-month post-vaccination samples were collected from the same mice 2 and 8 week samples used for the training data. TCRβ sequences from 16 week and 9 month post-vaccination samples were not used to generate the VATS library or as part of the training data. Enrichment of VATS was observed in the 16 week and 9 month post-vaccinated samples. The VATS library occupied on average 0.091±0.019% and 0.105±0.043% of TCRβ sequences from 16 week and 9 month post-vaccination samples, respectively, compared with naive repertoires (0.0057±0.0039%). Compared with the determination threshold calculated by the diagnostic assay, 100% of 16 week (18 of 18) and 96% of 9 month (22 of 23) post-vaccinated samples were correctly differentiated from naive samples (
The accuracy of the VATS library and diagnostic classifier was tested using an unrelated cohort of mice infected with a highly related Orthopoxvirus, MPXV (
To determine whether the platform used to generate the VATS could be replicated independent of the ACAM2000 analysis, the same protocol was used with the TCRβ sequences identified in the MPXV-infected mice to generate a separate library of MPXV-associated TCR sequences (MATS). A total of 120 MATS were identified (Table 7).
Using the same diagnostic approach implemented with the VATS library, a diagnostic classifier using the MATS library was generated. The proportion of a sample's unique TCRβ clonotypes occupied by MATS (% MATS) was calculated and used to distinguish between naïve and infected or vaccinated samples. Compared with naïve samples (0.0009+/−0.0018%), there were significant increases in the % MATS of MPXV-infected samples (0.114+/−0.037%, 126.7-fold increase) and ACAM2000-vaccinated samples (0.036%+/−0.02%, 40-fold increase,
The diagnostic assay has shown the ability to monitor and track the presence of sequences from the VATS library over time in vaccinated or infected mice. Using sequence analyses, the relative frequencies of TCR sequences from the VATS library within the circulating T cell repertoire was determined. The frequencies of VATS sequences significantly decrease in mice over time from 2 weeks (0.35±0.13%) through 9 months (0.11±0.05%) after exposure (
TCR sequences identified in the VATS library were examined for sequences specific for known HLA-A2 epitopes previously identified in VACV-immune humans and HLA-A2 transgenic mice (Gilchuk et al., 2013). HLA-A2 tetramers loaded with nine different vaccinia peptides were used to identify and isolate HLA-A2-restricted vaccinia-specific T cells (Table 8). The vaccinia peptides loaded onto HLA-A2 tetramers had been previously shown by Gilchuk et al. (2013) to elicit strong CD8+ T cells responses in HLA-A2 transgenic mice. Mice approximately 6 months post-vaccination were boosted with ACAM2000; after 4 days, splenocytes were isolated, and tetramer-binding CD8+ T cells were isolated using fluorescence-activated cell sorting (FACS) with pooled tetramers (
In the experiments shown in Examples 1 to 5, high-throughput TCR repertoire analyses from a large cohort of mice (n=58) were used to identify and track TCR sequences responding to either the ACAM2000 smallpox vaccine or infection with MPXV. In total, >2.8×106 unique TCRβ clonotypes were analyzed from 245 individual blood samples collected before and after exposure. Data from mice administered the ACAM2000 smallpox vaccine were used to identify a library of 315 VATS. The VATS library acted as a diagnostic classifier, differentiating between naive and vaccinated or infected mice on the basis of the presence or absence of the public TCRβ sequences. The VATS library correctly identified samples from mice vaccinated with the smallpox vaccine and infected with MPXV from 2 weeks up to 9 months post-vaccination or infection. Overall, the diagnostic classifier was capable of distinguishing between vaccinated or infected samples and naive repertoires with >95% accuracy, which was replicated with MPXV-infected mice and the generation of the MATS library.
It was confirmed that the VATS library represented the public vaccine-specific T cell population by comparing the VATS library with TCRs expanded after in vitro culture with ACAM2000 and from vaccinia-specific HLA-A2.1 tetramer sorted T cells. The overlap between the TCR repertoires identified by in vitro expansion or tetramer sorting and the VATS library was limited. This was expected given the large number of immune-recognized Orthopoxvirus epitopes, various MHC molecules in mice, and that the majority of antigen-specific TCR clonotypes are private (specific to the individual mouse). This was previously shown in the human TCR repertoire, analyzing HLA-A2-restricted CD8+ T cells recognizing Epstein-Barr virus- and CMV-specific epitopes (Venturi et al., 2008a). Although only a small number of VATS sequences were found in the tetramer+ TCRβ repertoire, those sequences could be readily tracked and identified in mice up to 9 months after vaccination. Thus, this approach allowed the development of a tool to follow the virus-specific response over sequential time points in an aging population of mice. The frequencies at which the tetramer+ VATS were found in the circulation by TCR sequencing were as low as 1:50,000, meaning that using the VATS library to probe the TCRβ repertoire can be more sensitive than other technologies such as tetramer staining, intracellular cytokine staining, or allele-specific oligonucleotide PCR (Campana, 2010, Faham et al., 2012, Wolf and DiPaolo, 2016, van der Velden et al., 2014, van der Velden and van Dongen, 2009).
The low-frequency virus-specific memory response previously could only be readily measured through immune assays screening for serum antibodies against vaccinia and other pox viruses, which have historically been shown to be a very powerful and accurate tool for determining an individual's prior exposure to Orthopoxvirus (Frey et al., 2003, Newman et al., 2003, Yin et al., 2013). However, we have shown that immunosequencing of the TCRβ chain can differentiate between naive and vaccinated or infected individuals with approximately the same level of accuracy (Hammarlund et al., 2005). Considering the nature of the two methods, there are some key differences between immunosequencing of the TCR repertoire and serum antibody profiling. A major difference is that immunosequencing relies on genomic DNA or cDNA. In areas of the world where resources are limited, the molecular stability and shelf life of DNA compared with serum and antibodies offers significant benefit when collecting and transporting samples. Additionally, once the appropriate pathogen-associated TCR sequence libraries are developed, an individual's TCR repertoire could easily be used to determine prior exposure to a host of different pathogens with virtually no increase in effort, whereas testing serum for antibodies would require multiple tests for subsequent infectious agents (Emerson and DeWitt, 2017). Although the initial effort of collecting and analyzing large training cohorts for each target pathogen may be substantial, as sequencing data from subsequent studies become available, the resources required to produce large datasets becomes less (Emerson and DeWitt, 2017). Immunosequencing data, when published, are archived in public databases. This allows researchers to use previously published TCR repertoire data to increase diagnostic power while lowering the resources required to achieve large sample sizes (DeWitt et al., 2016).
Using this technology, it may be possible to develop panels of TCR sequence libraries capable of determining individuals' prior exposure to multiple pathogens simultaneously. The ability to discern an individual's immunological history has significant clinical benefits. Additionally, it is known that multiple viruses are able to undergo rapid mutation to evade immune detection. Analysis of the TCR repertoire could potentially be used in an attempt to track viral variants. However, we recognize that there are significant challenges involved in recapitulating this approach in human populations. Compared with genetically identical mice, humans display significant diversity in their HLA haplotypes, including rare HLAs, and TCR repertoires will likely limit the detection of public TCRs. Additionally, human populations are under constant exposure to different commensals, pathogens, and environmental stimuli, which can make identifying TCR sequences recognizing specific pathogens significantly more difficult. However, by acquiring larger blood volumes (5 mL in human versus 100 μL in mouse) and performing sequencing at greater depths (e.g., using ultra-deep sequencing), generating TCR repertoires of hundreds of thousands of clonotypes per sample, it is possible to perform similar studies in humans. This has been shown to be possible using populations of CMV+ and CMV− human populations (Emerson and DeWitt, 2017). Future studies are focused on adapting the current methodology to determine prior pathogen-specific exposure in human populations.
In summary, we have demonstrated that immunosequencing is a powerful and highly versatile tool for analyzing the TCR repertoire. In this study, an extensive database of TCRβ sequences was generated from the circulating TCR repertoires from a large cohort of mice (n=58) and used to identify and track the vaccine-specific T cell response over time. This allowed a comprehensive analysis of Orthopoxvirus-associated TCR sequences and shows that analyses of TCRβ repertoires can be used to determine individuals' prior exposure to ACAM2000 or MPXV with a high degree of accuracy and is capable of tracking the virus-specific populations present at ultra-low frequencies long after primary exposure resolved.
iCAT is a user-friendly, graphical-interface software that takes exposed and non-exposed samples of T-cell receptor (TCR) clonotypes as input and identifies pathogen-specific TCR sequences. Using these sequences, iCAT can also classify independent samples of TCR clonotypes. iCAT's backend methodology is based on performing Fisher's exact tests to find informative TCR sequences. When tested on mice samples from a recent publication, iCAT was able to identify vaccine-associated TCR sequences with 95% accuracy. With iCAT, we capitalize on the power of TCR sequencing to simplify infection diagnosis and further investigation of immunological history.
Software (iCAT) Framework
iCAT provides a graphical user interface in the form of a web-app by the power of R-Shiny. The user can upload multiple TCR sequence repertoires from negative (control) and positive (experimental) cohorts. iCAT accepts tab-delimited files with the size limit of 10 gigabytes per file. The user can select amongst three unique options to define TCR clonotypes within samples as well as parameters of the analyses: nucleotide sequences (“nucleotide”), CDR3 amino acid sequences (“aminoAcid”), or a combination of TCRV name, CDR3 sequence, and TCRJ name (“vGeneName”, “amino-Acid”, and “jGeneName”) where column name is represented in parentheses. Users can change an upper p-value threshold for performing Fisher Exact tests, which are used to identify TCR sequences of interest.
Clicking “Train” will start the pipeline to statistically identify a subset of target-associated sequences (TARSs) that give signal about the identity of the samples, negative or positive. As this is not typically an instantaneous process and can often be the bottleneck of analyses with large data, a graphical progress bar is implemented to provide status updates about the iCAT pipeline. Upon training, iCAT's main tab provides a table summary of the data, a figure shows the distribution of TARSs between the positive and negative samples, and a classification matrix shows the expected accuracy if those sequences were used for characterization, which is estimated from the training data.
A separate tab, “Library”, is unlocked upon training and shows a table where each row describes a TARS and its presence in the positive and negative samples. All tables and figures are supplemented with a custom button for easy download. The third tab of iCAT, “Prediction”, also unlocks after training and allows the user to upload one or more independent TCR-sequencing samples for classification. iCAT will provide a downloadable table with the predictions if more than one sample was uploaded.
The statistical methodology of iCAT is based on identifying a subset of TARSs that informs classification. TCR sequences significantly associated with positive samples as opposed to negative samples are identified by performing Fisher's exact test. iCAT determines the optimal p-value cutoff based on the idea of coverage ratio describes above in Materials and Methods. Coverage was defined as the sum of samples containing each TARS divided by the total number of samples. iCAT calculates the coverage for negative samples (Cn) and positive samples (Cp). The coverage ratio, Cn:Cp, is calculated for each p-value. The optimal p-value is thus defined as the p-value with the maximum coverage ratio.
To classify an independent sample, iCAT first determines the percentage of TARSs in the sample. This percentage is compared to the normal distributions of negative and positive samples previously observed. The probability density function of the independent TARS percentage is calculated for the positive and negative normal distributions, with an internal normalization factor to reduce potential overfitting of the classifier. Classification is determined by the strength of samples' association with the positive or negative training data. Independent sample classification is used as a method to cross-validate the diagnostic accuracy of the classifier and for performing Leave-one-out analyses by removing all samples from a single source (mouse, individual, etc.) from the training data and re-training the classifier, using the removed samples in the prediction tab.
iCAT was used to identify vaccine-associated receptor sequences in mice injected with the smallpox vaccine (Wolf, et al., 2018). 32 pre-exposure (naïve) samples were analyzed, which included 714,522 clonotypes and 2,049,383 unique CDR3 amino acid sequences. 58 samples taken 2- and 8-weeks post-vaccination were analyzed, which included 573,612 and 1,581,619 unique CDR3 amino acid sequences (
From the training data, iCAT correctly classified 32 of 32 naïve samples as “unexposed” and 58 of 58 vaccinated samples as “exposed” (100% accuracy). TCR repertoires from 10 mice pre- and 2-weeks postvaccination were used as an independent cross-validation cohort and iCAT classified them with 95% accuracy. Overall, this data displays that the iCAT platform computationally identifies TCR sequences associated with exposure to a pathogen, training a diagnostic classifier to distinguish between exposed and unexposed samples with a high degree of accuracy. See
A human cohort was vaccinated with the ACAM2000 vaccine. T-cells were isolated and analyzed before and after the vaccine administration. As above, the genomic DNA of these T-cells were amplified and sequences to generate TCRβ clonotype profiles of the test subjects. The data generated was fed to a neural network program which trained on the data to identify unique TCRβ alleles statistically associated with small pox vaccination, using methods similar to those described above with changes as described below. The alleles identified are presented in Table 3 above. A detailed explanation is provided below.
During an infection or vaccination, T-cells that carry receptors specific to a certain pathogen become activated and each receptor is encoded by a uniquely rearranged DNA sequence. Even after the infection is eliminated, these activated immune cells remain and serve to prevent secondary infection of that pathogen. As a result of this persistence in the body, analyzing the large and diverse TCR repertoire may help us better understand immune system features and disease progression. In addition, the unique DNA rearrangements could be stable biomarkers for reliable diagnosis of infectious diseases. Recently, high-throughput NGS techniques were employed to analyze the diverse immune cell repertoire. A few research labs including us have attempted to develop statistical methods to identify and classify TCR sequences corresponding to a specific viral infection; however, it is challenging and critical to increase the accuracy of identifying viral infection from the diverse TCR repertoire over time and within the same individual. Especially, predicting human viral infection requires fast diagnosis and high accuracy. Satisfying both speed and accuracy requires much effort and skill.
After the analysis of the mouse dataset, we tried to analyze much larger and complex samples from smallpox vaccinated human cohorts (
In detail, we first preprocessed 129 samples were preprocessed to generate 2,525,775 unique TCR sequences and their frequency in each sample. This data (both the sequences and their frequency) was used as input features in a classifier to train it to identify pre- and post-vaccination of smallpox vaccine from 96 of the 129 samples. The remaining 33 samples were saved for a later independent test. Preliminary optimization results showed that using 5 hidden layers, 90 nodes (neurons), and 1000 max iterations (
As one of the most powerful machine learning methods, the deep learning neural network has been substantially employed to explore the high-level features hidden in biomedical data. In this example, the deep learning framework was used to train the deep learning models for diagnostic discrimination. A multi-layer neural network (i.e., more than three layers) was used to extract hidden patterns from the input features through differing numbers of hidden layers. The extracted hidden features were finally fed into the last layer of logistic regression to classify the sample into binary classes. The deep learning model can be optimized through minimizing the binary cross-entropy objective function in the process of standard error backward propagation. Similar to the other three methods, the training dataset was equally split into five sets, and five-fold cross-validation was used to train and validate the model's robustness. Several parameters of neural networks were also adjusted using the repeated cross-validation, including the number of hidden layers, the number of hidden nodes in each hidden layer, and the types of activation functions for the hidden nodes. Several hyper-parameters were also tuned, including the dropout rate for regularization, learning rate and momentum used in different types of optimization algorithms. The classification accuracy is calculated for each round of five-fold cross-validation, and the accuracy scores are averaged over a total of 50 rounds to select the best parameter set for final testing. The deep neural network was implemented using the Tensorflow library (www.tensorflow.org), along with the cross-validation and parameter tuning available in the Scikit-learn library.
In this work, the predictive ability of the DNN method for diagnostic discrimination of viral infection was evaluated to understand how immune system features can diagnose viral infection status. The frequency counts of all CDR3 amino acid sequences (a.k.a. peptides) were calculated from quantified TCR beta chain sequence data and used as input features for machine learning methods to build discriminative classifiers. Each negative sample (pre-inoculation) or positive sample (post-inoculation) was described as a vector of frequency counts, each representing the number of CDR3 amino acids found in the sequence data of the sample.
The analysis started with the data partition. A stratified sampling method was applied to randomly divide the data into subsets according to the status of infection (pre- or post-introduction). Each of the three datasets were partitioned into a training set and testing set with a ratio of 75%/25%. The repeated 5-fold cross-validation was used to estimate the optimal parameters of each machine learning algorithm on the training dataset. The best training parameters selected by cross-validation were used to retrain the whole training dataset to derive the final model for evaluation. The independent testing subset is only seen when the final model of each algorithm is determined. After the training data was collected, several data normalization schemes were attempted before applying machine learning algorithms for model learning. Due to the different experimental conditions (i.e., sequence depth) and sample variations, the number of frequency counts for amino acids might vary in magnitude. Normalization might be necessary to remove inherent bias for different machine learning methods. To this end, the training data was transformed in three ways: (1) peptide-based normalization that normalizes counts across all training samples within each amino acid sequence; (2) sample-based normalization that normalizes counts of amino acids within individual samples; and (3) the benchmark data that uses original counts without any normalization. The Minimum-Maximum transformation was adopted to convert counts into the range between zero and one when the normalization is needed. The normalized/original features was then used to train different machine learning models for infection diagnosis.
When introducing elements of the present invention or the preferred embodiments(s) thereof, the articles “a”, “an”, “the” and “said” are intended to mean that there are one or more of the elements. The terms “comprising”, “including” and “having” are intended to be inclusive and mean that there may be additional elements other than the listed elements.
In view of the above, it will be seen that the several objects of the invention are achieved and other advantageous results attained.
As various changes could be made in the above methods without departing from the scope of the invention, it is intended that all matter contained in the above description and shown in the accompanying drawings shall be interpreted as illustrative and not in a limiting sense.
This application claims the benefit of U.S. Provisional Application Ser. No. 62/914,169, filed Oct. 10, 2019, the contents of which are incorporated by reference herein.
This invention was made with Government support under DJF-15-1200-P-0001007 awarded by Federal Bureau of Investigations. The Government has certain rights in the invention.
| Number | Date | Country | |
|---|---|---|---|
| 62914169 | Oct 2019 | US |
