Patent Application
20210033608
The instant application contains a Sequence Listing which has been filed electronically in ASCII format and is herein incorporated by reference in its entirety. Said ASCII copy, created on Jul. 30, 2020, is named “06026 Seq list_ST25” and is 10,104 bytes in size.
The invention is generally directed toward methods and systems to infer peptide presentation via major histocompatibility complexes (MHC) I and II, and more specifically directed towards methods and systems to analyze peptides of varying length for MHC I and II presentation, and various applications thereof.
The major histocompatibility complex (MHC) is a large locus on vertebrate DNA containing a set of genes that produce cell surface proteins essential for the adaptive immune system. These cell surface proteins are referred to as human leukocyte antigen (HLA) receptors that present small protein fragments (known as peptides) to induce an immune response. There are two major classes of MHCs, MHC I and MHC II, each having a unique set of HLA receptors. MHC I is associated with the HLA-A, HLA-B, and HLA-C receptors, each of which present peptides on the cell surface. MHC II is associated with the HLA-DP, HLA-DQ, and HLA-DR receptors, each of which present peptides on the cell surface. Each HLA receptor has a number alleles that are differentially expressed between individuals of the human population. The differences of expression of the HLA alleles confers unique immunological and allergen responses for each individual, as determined by the peptides the HLA receptor presents.
Various embodiments are directed systems and methods for identification of MHC I or MHC II antigen peptides. In various embodiments, a computational framework incorporates one or more modules that are utilized to determine a MHC I or MHC II presentation score. In various embodiments, MHC presentation scores are utilized to prioritize peptides in downstream applications, including (but not limited to) peptide synthesis, vaccine development, tolerance induction, and T cell therapy.
In an embodiment, the likelihood that a peptide is presented on a human leukocyte antigen (HLA) receptor of a major histocompatibility complex (MHC) is determined. To determine the likelihood that a peptide is presented, one or more peptides sequences for query is obtained. Each queried peptide has a length between 8 and 26 amino acids. A trained peptide presentation module incorporating a recurrent neural network architecture is obtained. The peptide presentation module is capable of determining presentation of peptides having varying length to at least one HLA allele. The one or more peptide sequences is queried one or more peptide sequences. Based on the peptide sequence and the at least one HLA allele assessed, a MHC presentation score for each peptide of the one or more peptide sequences is determined.
In another embodiment, the peptide presentation module is trained utilizing in vivo data derived from human individuals or cell lines that have had their MHC peptide ligand sequences identified by antigen presentation profiling via mass spectrometry.
In yet another embodiment, the peptide presentation module's recurrent neural network has one of the following architectures: fully recurrent, long short-term memory, gated recurrent unit, bidirectional LSTM or hierarchical recurrent network.
In a further embodiment, at least a first peptide sequence and a second peptide sequence are obtained, wherein each of the peptide length of the first peptide is different from the length of the second peptide.
In still yet another embodiment, a trained binding affinity module incorporating a recurrent neural network architecture is obtained. The binding affinity module is capable of determining binding affinity of peptides having varying length to a particular HLA allele. The trained binding affinity module is integrated with the trained peptide presentation module. The one or more peptide sequences is queried utilizing the trained binding affinity module to determine a binding affinity score between each peptide of the one or more peptide sequences and the at least one HLA allele assessed. Based on the peptide sequence, the at least one HLA allele assessed, and the binding affinity score, a MHC presentation score for each peptide of the one or more peptide sequences is determined.
In yet a further embodiment, the binding affinity module is trained utilizing in vitro data derived from the Immune Epitope Database.
In an even further embodiment, the binding affinity module's recurrent neural network has one of the following architectures: fully recurrent, long short-term memory, gated recurrent unit, bidirectional LSTM or hierarchical recurrent network.
In yet an even further embodiment, the flanking amino acid sequences upstream and downstream is determined for each peptide of the one or more peptide sequences. A trained cleavability module incorporating a neural network architecture is obtained. The trained cleavability module is capable of determining the cleavability of peptides based on their flanking amino acids. The trained cleavability module is integrated with the trained peptide presentation module. The one or more peptide sequences is queried utilizing the trained cleavability module to determine a cleavability score for each peptide of the one or more peptide sequences. Based on the peptide sequence, the at least one HLA allele assessed, and the cleavability score, a MHC presentation score for each peptide of the one or more peptide sequences is determined.
In still yet an even further embodiment, the flanking amino acids are determined from a proteome database.
In still yet an even further embodiment, the cleavability module is trained utilizing a ligandome of an antigen presenting cell line.
In still yet an even further embodiment, the gene information for each peptide of the one or more peptide sequences is obtained. A gene expression module incorporating a neural network architecture is obtained. The gene expression module is capable of determining the relative gene expression of peptides based on their gene information. The gene expression module is integrated with the trained peptide presentation module. The one or more peptide sequences is queried utilizing the trained gene expression module to determine the relative expression level for each peptide of the one or more peptide sequences. Based on the peptide sequence, the at least one HLA allele assessed, and the relative gene expression, a MHC presentation score for each peptide of the one or more peptide sequences is determined.
In still yet an even further embodiment, the gene expression module determines relative gene expression empirically from personalized RNA sequencing data.
In still yet an even further embodiment, the gene expression module determines relative gene expression inferentially from external RNA sequencing data.
In still yet an even further embodiment, the gene expression module corrects for low gene expression of extracellular proteins or blood proteins constituents.
In still yet an even further embodiment, the MHC presentation score is for MHC I. The binding affinity module is capable of determining binding affinity of peptides having a length between 8 and 17 amino acids. And the at least one HLA allele is an allele of one of: HLA-A, HLA-B, and HLA-C.
In still yet an even further embodiment, the at least one HLA allele is all alleles of HLA-A, HLA-B, and HLA-C.
In still yet an even further embodiment, the MHC presentation score is for MHC II. The binding affinity module is capable of determining binding affinity of peptides having a length between 8 and 26 amino acids. And the at least one HLA allele is an allele of one of: HLA-DP, HLA-DQ, and HLA-DR.
In still yet an even further embodiment, the at least one HLA allele is all alleles of HLA-DP, HLA-DQ, and HLA-DR.
In still yet an even further embodiment, the MHC presentation score is a basis for utilizing at least one peptide of the one or more peptide sequences in a downstream application.
In still yet an even further embodiment, the downstream application is one of: synthesizing the at least one peptide; developing a vaccine for cancer or an infectious pathogen utilizing the at least one peptide; developing a treatment to induce tolerance to the at least one peptide, wherein the peptide is involved with an autoimmune or allergic response; or developing a T cell therapy to treat cancer based on the at least one peptide.
The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.
The description and claims will be more fully understood with reference to the following figures and data graphs, which are presented as exemplary embodiments of the invention and should not be construed as a complete recitation of the scope of the invention.
Turning now to the drawings and data, embodiments related to utilizing computational frameworks to identify peptides to be presented via major histocompatibility complexes (MHC) I and II as determined by the peptide sequence binding affinity, peptide cleavability, and/or peptide expression level are described herein. Accordingly, various embodiments are directed towards computational frameworks that utilize one or more modules to produce a peptide presentation score for MHC I or MHC II. In some embodiments, a computational framework includes a module to assess a peptide's binding affinity with particular human leukocyte antigen (HLA) allele to determine a likelihood of presentation via MHC I or MHC II. In some embodiments, a computational framework includes a module to assess a peptide's likelihood to be cleaved to determine a likelihood of for presentation via MHC I or MHC II. In some embodiments, a computational framework includes a module to assess a peptide's expression level to determine the likelihood of presentation via MHC I or MHC II.
Various embodiments of computational frameworks include various combinations of modules. Accordingly, in some embodiments, a computational framework utilizes one solitary module to provide a peptide presentation score. In some embodiments, a computational framework integrates two or more modules to provide a peptide presentation score.
Numerous embodiments utilize a peptide presentation score for various analyses and downstream applications. In some embodiments, numerous peptides are assessed for their likelihood to be presented via MHC I or MHC II and then compared on the basis of their peptide presentation score. In some embodiments, peptides with presentation scores over a threshold are utilized in downstream applications. In some embodiments, peptides with higher presentation scores, as compared with other peptides assessed, are utilized in downstream applications. In some embodiments, peptides of a top percentile or a top quantile of presentation scores are utilized in downstream applications.
Numerous downstream applications can be performed on the basis of a peptide having a particular presentation score. As MHC I and II peptide presentation relates to immunological activity, various peptides can be selected on the basis of its presentation score, synthesized, and then utilized in immunological applications. In some embodiments, selected peptides are synthesized and assessed for their ability to be presented via MHC I and II, confirming the computational score. In some embodiments, selected peptides are synthesized and utilized within a vaccine to induce immune responses to the peptide. Accordingly, vaccines can be created for various infectious agents and/or cancers. In some embodiments, selected peptides can be utilized to identify potential targets for T-cell therapies. In some embodiments, selected peptides are synthesized and utilized to induce a tolerance in individuals to the peptide, which can be useful in various autoimmune and allergic disorders. For example, selected wheat gluten peptides can be utilized for tolerance induction in individuals suffering from celiac disease.
Provided in
In some embodiments, a computational framework is designed work for one of: MHC I or MHC II. In some embodiments, a computational framework is designed to consider peptide presentation by alleles of one or more HLA receptors. Particular HLA gene products work in within MHC I and particular HLA gene products work in within MHC II. In humans, the HLA gene products capable of working with MHC I are HLA-A, HLA-B, and HLA-C, each having two number of alleles with unique peptide binding propensities. Likewise, the human HLA gene products capable of working with MHC II are HLA-DP, HLA-DQ, and HLA-DR, each having a number of alleles with unique peptide binding propensities. Accordingly, various embodiments of computational frameworks and/or modules within a framework can be specific for assessment of peptides for particular HLA allele, which may be useful for analysis of human population expressing that particular allele. And in various embodiments, a computational framework and/or modules within a framework can be generalized for a plurality HLA alleles, which may be useful a more generalized population analysis. For instance, in some embodiments, a computational framework determines peptide presentation for the alleles of a single HLA receptor. Accordingly, in various embodiments, a computational framework determines peptide presentation for alleles of HLA-A, HLA-B, HLA-C, HLA-DP, HLA-DQ, or HLA-DR. And for instance, in some embodiments, a computational model determines peptide presentation for all alleles of MHC I or MHC II. Accordingly, in various embodiments, a computational framework determines peptide presentation for the alleles of HLA-A, HLA-B and HLA-C (i.e., the alleles of MHC-I) or for the alleles of HLA-DP, HLA-DQ, and HLA-DR (i.e., the alleles of MHC-II).
Process 100 can begin with obtaining (101) variable length peptide for query. In addition, and dependent on the various modules utilized for analysis, gene data of the peptide and/or HLA allele data for MHC I or MHC II presentation assessment is obtained. The length of peptides to be assessed can vary, and is dependent on MHC I or MHC II presentation. In embodiments for MHC I assessment, peptides between 8 and 17 amino acids can be assessed within the same framework. In embodiments for MHC II assessment, peptides between 8 and 26 amino acids can be assessed within the same framework. The ability to assess multiple lengths of peptides is made possible with the use of a recurrent neural network (RNN), which is discussed in more detail below and within the Exemplary embodiments.
Generally, any peptide sequence within the length requirements can be queried. It may be desirable, however, to assess particular sets of sequences. In some embodiments, endogenously expressed peptides are queried. In some embodiments, cancer neoantigens are queried, which may be useful to identify immunogenic peptides for cancer vaccines or cancer T cell treatments. In some embodiments, autoantigens are queried, which may be useful to develop treatments for autoimmune diseases. In some embodiments, exogenous peptides are queried. In some embodiments, peptides of infectious pathogens (e.g., bacteria, viruses, parasites) are queried, which may be useful to develop vaccines against these pathogens. In some embodiments, exogenous antigens involved with autoimmune disorders (e.g., wheat gluten in celiac disease) and allergies are assessed, which may be useful to develop treatments for autoimmune disorders and allergies.
In addition to peptide sequence, gene data of the peptide and/or HLA allele data may be obtained, as dependent on the modules utilized in a computational framework. As described in greater detail below, a framework includes a peptide presentation module may optionally also include a binding affinity module, a cleavability module, and/or an expression level module. A peptide presentation module utilizes a peptide sequence and HLA allele data to determine the likelihood of a peptide sequence to be presented on the HLA receptor. A binding affinity module utilizes a peptide sequence and HLA allele data to determine the ability of the HLA to bind the peptide sequence. A cleavability module utilizes the sequences flanking upstream and downstream of the peptide to determine the likelihood of a peptide to be cleaved for antigen presentation. Flanking sequences may be derived from any appropriate source that provides such data, such as (for example), gene sequence data which can be utilized to determine the sequence flanking the peptide. An expression module utilizes the peptide gene data to determine the relative expression level of the peptide.
Process 100 utilizes (103) the peptide sequence, gene data, and/or HLA allele data in a peptide presentation module and optionally one or more computational modules. Modules for peptide presentation, binding affinity, cleavability, and expression level are described below.
In numerous embodiments, a peptide presentation module computes the likelihood of a peptide to be presented by MHC I or MHC II based on its sequence, which is dependent on the HLA allele expressed. Accordingly, in some embodiments, a peptide presentation module assesses the peptide's sequence to determine if the sequence has is likely to be presented on a HLA allele (or set of HLA alleles). To determine presentation, in accordance with some embodiments, the module utilizes a recurrent neural network (RNN) that has been trained utilizing in vivo HLA allele/peptide ligand data. Training data may be obtained from patients and/or cell lines. In some embodiments, mass spec data derived from human individuals and/or cell lines to profile MHC-bound peptide ligands bound to an HLA allele. Accordingly, in some embodiments, numerous (e.g., tens of thousands) non-redundant peptide ligand sequences in association with an HLA allele are utilized as positive examples to train the binding affinity module. In some embodiments, randomly selected length matched human peptide sequences are utilized as negative examples to train the binding affinity module. Accordingly, a peptide presentation module learns the patterns of sequences that are presented from the training data to determine the likelihood a queried peptide would be presented. Details on training are provided within Example 1 and 2 within the Exemplary embodiments.
In several embodiments, a peptide presentation module utilizes an RNN, which provides advantages for analyses of variable length sequences. Various architectures of RNN can be utilized. In various embodiments, an RNN architecture is one (or a combination) of: fully recurrent, long short-term memory (LSTM), gated recurrent unit (GRU), bidirectional LSTM and hierarchical recurrent network. By utilizing an RNN, peptides of variable length can be assessed within the same module architecture.
In many embodiments, a peptide (or set of peptides) is queried within the neural network to determine whether the sequence is predicted to be presented by a particular HLA allele, as determined by the data used to train the model. In some embodiments, the output layer contains a single neuron providing a likelihood of presentation between 0 and 1.
In a number of embodiments, a peptide presentation module is integrated with one or more modules: a binding affinity module, a cleavability module, and/or a gene expression module. The results of these modules can be combined with the analysis of sequence data to provide an overall presentation score. In some embodiments, a computational framework is a peptide presentation module integrated with a binding affinity module. In some embodiments, a computational framework is a peptide presentation module integrated with a cleavability module. In some embodiments, a computational framework is a peptide presentation module integrated with a gene expression module. In some embodiments, a computational framework is a peptide presentation module integrated with a binding affinity module and a cleavability module. In some embodiments, a computational framework is a peptide presentation module integrated with a binding affinity module and a gene expression module. In some embodiments, a computational framework is a peptide presentation module integrated with a cleavability module and a gene expression module. In some embodiments, a computational framework is a peptide presentation module integrated with a binding affinity module, a cleavability module, and a gene expression module.
In several embodiments, a binding affinity module computes the likelihood of a peptide to bind to MHC I or MHC II, which is dependent on the HLA allele expressed. Accordingly, in some embodiments, a binding affinity module utilizes the peptide's sequence and is queried to determine if the sequence has affinity to bind to a particular HLA allele (or set of HLA alleles). To determine binding, in accordance with some embodiments, the module utilizes a recurrent neural network (RNN) that has been trained utilizing in vitro HLA allele/peptide binding affinity data. Training data may be obtained from HLA allele/peptide binding affinity experiments. In some embodiments, HLA allele/peptide binding affinity data utilized for training is derived from a database, such as the Immune Epitope Database (IEDB). Accordingly, in some embodiments, numerous HLA/peptide pairs are assessed for their binding affinity as measured by half maximal inhibitory concentration (IC50) to train the binding affinity module. Details on training are provided within Example 1 and 2 within the Exemplary embodiments.
In several embodiments, a binding affinity module utilizes an RNN, which provides advantages for analyses of variable length sequences. Various architectures of RNN can be utilized. In various embodiments, an RNN architecture is one (or a combination) of: fully recurrent, long short-term memory (LSTM), gated recurrent unit (GRU), bidirectional LSTM and hierarchical recurrent network. By utilizing an RNN, peptides of variable length can be assessed within the same module architecture.
In many embodiments, a peptide (or set of peptides) is queried within the neural network to determine whether the sequence is predicted to have affinity for a particular HLA allele, as determined by the data used to train the model. In some embodiments, the output layer contains a single neuron providing a likelihood of affinity between 0 and 1. Results of the model can be utilized in isolation or integrated with a peptide presentation module and optionally other modules to provide a MHC presentation score.
In several embodiments, a cleavability module to determines the likelihood that a cleavage reaction would occur to produce a peptide for MHC presentation. In some embodiments, cleavability is determined by the amino acid sequences that flank the peptide being assessed. Accordingly, in some embodiments, up to 24 flanking amino acids are considered, upstream the peptide and/or downstream the peptide. In various embodiments, 4, 5, 6, 7, 8, 9, 10, 11, or 12 amino acids upstream the peptide and/or 4, 5, 6, 7, 8, 9, 10, 11, or 12 amino acids downstream the peptide are considered.
In many embodiments, a neural network is built to determine cleavability. In some embodiments, a cleavability module determines the flanking amino acids from an appropriate source (e.g., a proteome database with protein sequences), encodes these determined amino acids, a processes them with hidden layers to output a probability score between 0 and 1. To train the model, in accordance with some embodiments, the ligandome of an antigen presenting cell line (e.g., a dendritic or a lymphoma cell line) is utilized to identify the most common flanking amino acids in each flanking position.
In a number of embodiments, a peptide (or set of peptides) is queried within the neural network to determine whether the flanking sequence of the peptide contains enrichment of amino acids are cleaved. In some embodiments, the output layer contains two neurons providing a likelihood of false (F) or true (T) cleavability between 0 and 1. Results of the model can be utilized in isolation or integrated with a peptide presentation module and optionally other modules to provide a MHC presentation score.
In several embodiments, an expression module determines the relative expression level of a peptide. It is now known that expression levels correlate with MHC peptide presentation. When a gene is highly expressed, peptides of the gene product are more likely to be presented than genes expressed at lower rates. Accordingly, in some embodiments, the expression level of a gene from which the peptide is derived is determined, which is utilized to determine the likelihood of presentation.
It was further discovered that some peptides derived from lowly expressed genes are also highly presented. These peptides are often products of extracellular proteins and blood protein constituents. Accordingly, in some embodiments, gene expression of extracellular proteins and blood protein constituents are “corrected” to account for their bias of being highly presented despite their gene being lowly expressed. Any appropriate mechanism to account for this bias can be utilized. In some embodiments, expression of extracellular proteins and blood proteins constituents are artificially set to high expression levels. For example, in various embodiments, gene expression values of genes under one or more of the following GO terms are set to 50 transcripts per million (TPM): extracellular space (0005615), blood microparticle (0072562), secretory granule lumen (0034774), cytoplasmic vesicle lumen (0060205), and/or extracellular matrix (0031012).
Expression of genes can be determined by various methods. In some embodiments, gene expression is determined empirically. Accordingly, the gene expression of the peptide being assessed is determined from personalized RNA sequencing data. In some embodiments, gene expression is inferred. Accordingly, the gene expression of the peptide being assessed is determined from external RNA sequencing data. It was found that gene expression matched tissue, and in some cases, gene expression from unmatched tissue, can provide a robust indication of MHC presentation. When gene expression is inferred, in some embodiments, a gene expression “dictionary” having stored relative expression values of genes is utilized to determine relative expression of a peptide. In some embodiments, a logistic regression model is trained to differentiate highly expressed peptides from decoys.
In many embodiments, a peptide (or set of peptides) is queried within a logistic regression incorporating a gene expression dictionary to estimate the relative peptide expression values. In some embodiments, the output layer contains a single neuron providing an expression value (e.g., TPM). Results of the model can be utilized in isolation or integrated with a peptide presentation module and optionally other modules to provide a MHC presentation score.
In several embodiments, a peptide presentation module and one or more modules are integrated within a computational framework to produce a MHC I or MHC II presentation score. As explained within the Exemplary embodiments, integrating a binding infinity module and/or a cleavability module and/or an expression module with a peptide presentation module can result in better MHC presentation prediction. To integrate modules, in some embodiments, input data, including peptide sequence, HLA allele, and/or peptide gene data, is shared amongst the modules. In addition, in some embodiments, an integrated framework includes an RNN layer to encode each ligand peptide sequence. In some embodiments, the output of each module is concatenated to merge their information to yield an ultimate output layer indicating likelihood of presentation.
Process 100 determines (105) a MHC I or MHC II presentation score for each peptide sequence queried within the computational framework. Accordingly, in various embodiments, presentation scores are determined from one or more computational modules utilized. In some embodiments, the output presentation score is between 0 and 1 indicating how likely a query peptide is to be presented by a specific HLA allele. To increase human interpretability and enable comparison across different peptide lengths, in some embodiments, the framework's output can be represented as a percentile score. In some embodiments, a percentile score is generated by comparing the raw output score to a score distribution generated from length-matched random human peptides. The higher the percentile, the more likely the peptide will be presented by a cell HLA-DR complex. Based on the presentation score, in accordance with some embodiments, peptides are compared and/or ranked.
Process 100 optionally performs downstream applications on one more peptides on the basis of their presentation score. In some embodiments, based on its presentation score, a peptide is synthesized for further analysis. In some embodiments, based on its presentation score, a peptide sequence is utilized to develop cancer vaccines. In some embodiments, based on its presentation score, a peptide sequence is utilized to develop vaccines against infectious pathogens. In some embodiments, based on its presentation score, a peptide sequence is used to develop T cell therapeutics. In some embodiments, based on its presentation score, a peptide sequence is utilized to develop a treatment to induce tolerance to autoantigen. In some embodiments, based on its presentation score, a peptide sequence is utilized to develop a treatment to induce tolerance to exogenous antigen that is involved in autoimmunity and/or allergies.
While specific examples of processes for determining a MHC presentation score are described above, one of ordinary skill in the art can appreciate that various steps of the process can be performed in different orders and that certain steps may be optional according to some embodiments of the invention. As such, it should be clear that the various steps of the process could be used as appropriate to the requirements of specific applications. Furthermore, any of a variety of processes for determining a MHC presentation score appropriate to the requirements of a given application can be utilized in accordance with various embodiments of the invention.
Various embodiments are directed towards utilizing MHC I or MHC II peptide presentation scores in various applications. In some embodiments, MHC presentation scores can identify immunogenic peptides, which can be useful in vaccine development and/or T cell therapies. In some embodiments, MHC presentation scores can identify peptides that could cause autoimmunity and/or allergies and thus treatments can be designed to induce tolerance to those peptides.
Peptides can be synthesized chemically by a number of methods. One common method is to use solid-phase peptide synthesis (SPPS). In many embodiments, a peptide is synthesized first as a linear peptide utilizing solid-phase peptide synthesis (SPPS). Any appropriate SPPS protocol can be utilized. The solid support can be any appropriate solid support, such as (for example) the Merrifield resin, the PAM resin, the Wang resin, or 2-chlorotrityl resin. Any appropriate protecting groups can be utilized, such as (for example) Fmoc or Boc.
Peptides can also be synthesized utilizing molecular tools and a host cell. Nucleic acid sequences corresponding with antigenic peptides can be synthesized. In some embodiments, nucleic acids synthesized in in vitro synthesizers (e.g., phosphoramidite synthesizer), bacterial recombination system, or other suitable methods. Furthermore, synthesized nucleic acids can be purified and lyophilized, or kept stored in a biological system (e.g., bacteria, yeast). For use in a biological system, synthetic nucleic acid molecules can be inserted into a plasmid vector, or similar. A plasmid vector can also be an expression vector, wherein a suitable promoter and a suitable 3′-polyA tail is combined with the transcript sequence.
Embodiments are also directed to expression vectors and expression systems that produce antigenic peptides or proteins. These expression systems can incorporate an expression vector to express transcripts and peptides in a suitable expression system. Typical expression systems include bacterial (e.g., E. coli), insect (e.g., SF9), yeast (e.g., S. cerevisiae), animal (e.g., CHO), or human (e.g., HEK 293) cell lines. RNA and/or peptides can be purified from these systems using standard biotechnology production procedures.
A number of embodiments utilize methods to develop and administer vaccines against antigenic peptides. In various embodiments, infectious pathogen antigen and/or cancer neoantigen peptides are screened by their MHC presentation score then utilized to develop a vaccine against the peptide. Accordingly, various embodiments contemplate administering immunogenic compositions to individuals, proposed to be suitable for use as a vaccine, prepared using one or more antigenic peptides that comprise peptide sequences selected on the basis of their MHC presentation score. In some embodiments, antigenic peptides would further include flanking sequences that confer high cleavability. In some embodiments, antigenic peptides can be used in combination with other secreted virulence proteins, surface proteins or immunogenic fragments thereof. In certain aspects, antigenic material is extensively dialyzed to remove undesired small molecular weight molecules and/or lyophilized for more ready formulation into a desired vehicle.
The preparation of vaccines that contain polypeptide or peptide sequence(s) as active ingredients is generally well understood in the art, as exemplified by U.S. Pat. Nos. 4,608,251; 4,601,903; 4,599,231; 4,599,230; and 4,596,792; each of which is incorporated herein by reference. Typically, such vaccines are prepared as injectables either as liquid solutions or suspensions: solid forms suitable for solution in or suspension in liquid prior to injection may also be prepared. The preparation may also be emulsified. The active immunogenic ingredient is often mixed with excipients that are pharmaceutically acceptable and compatible with the active ingredient. Suitable excipients are, for example, water, saline, dextrose, glycerol, ethanol, or the like and combinations thereof. In addition, if desired, the vaccine may contain amounts of auxiliary substances such as wetting or emulsifying agents, pH buffering agents, or adjuvants that enhance the effectiveness of the vaccines. In specific embodiments, vaccines are formulated with a combination of substances, as described in U.S. Pat. Nos. 6,793,923 and 6,733,754, each of which is incorporated herein by reference.
Vaccines may be conventionally administered parenterally, by injection, for example, either subcutaneously or intramuscularly. Additional formulations which are suitable for other modes of administration include suppositories and, in some cases, oral formulations. For suppositories, traditional binders and carriers may include, for example, polyalkalene glycols or triglycerides: such suppositories may be formed from mixtures containing the active ingredient in the range of about 0.5% to about 10%, preferably about 1% to about 2%. Oral formulations include such normally employed excipients as, for example, pharmaceutical grades of mannitol, lactose, starch, magnesium stearate, sodium saccharine, cellulose, magnesium carbonate and the like. These compositions take the form of solutions, suspensions, tablets, pills, capsules, sustained release formulations or powders and contain about 10% to about 95% of active ingredient, preferably about 25% to about 70%.
Vaccine compositions would normally be administered as pharmaceutically acceptable compositions that include physiologically acceptable carriers, buffers or other excipients. As used herein, the term “pharmaceutically acceptable” refers to those compounds, materials, compositions, and/or dosage forms which are, within the scope of sound medical judgment, suitable for contact with the tissues of human beings and animals without excessive toxicity, irritation, allergic response, or other problem complications commensurate with a reasonable benefit/risk ratio. The term “pharmaceutically acceptable carrier,” means a pharmaceutically acceptable material, composition or vehicle, such as a liquid or solid filler, diluent, excipient, solvent or encapsulating material, involved in carrying or transporting a chemical agent.
An effective amount of therapeutic or prophylactic composition is determined based on the intended goal. The term “unit dose” or “dosage” refers to physically discrete units suitable for use in a subject, each unit containing a predetermined quantity of the composition calculated to produce the desired responses discussed above in association with its administration, i.e., the appropriate route and regimen. The quantity to be administered, both according to number of treatments and unit dose, depends on the protection desired.
Precise amounts of the composition also depend on the judgment of the practitioner and are peculiar to each individual. Factors affecting dose include physical and clinical state of the subject, route of administration, intended goal of treatment (alleviation of symptoms versus cure), and potency, stability, and toxicity of the particular composition.
Typically, vaccines are administered in a manner compatible with the dosage formulation, and in such amount as will be therapeutically effective and immunogenic. The quantity to be administered depends on the subject to be treated, including the capacity of the individual's immune system to induce a T cell response and the degree of protection desired. Precise amounts of active ingredient required to be administered depend on the judgment of the practitioner. However, suitable dosage ranges are of the order of several hundred micrograms of active ingredient per vaccination. Suitable regimes for initial administration and booster shots are also variable, but are typified by an initial administration followed by subsequent inoculations or other administrations.
Upon formulation, solutions will be administered in a manner compatible with the dosage formulation and in such amount as is therapeutically or prophylactically effective. The formulations are easily administered in a variety of dosage forms, such as the type of injectable solutions described within.
The manner of application may be varied widely. Any of the conventional methods for administration of a vaccine are applicable. These are believed to include oral application within a solid physiologically acceptable base or in a physiologically acceptable dispersion, parenterally, by injection and the like. The dosage of the vaccine will depend on the route of administration and will vary according to the size and health of the subject.
In certain instances, it will be desirable to have multiple administrations of the vaccine, e.g., 2, 3, 4, 5, 6 or more administrations. The vaccinations can be at 1, 2, 3, 4, 5, 6, 7, 8, to 5, 6, 7, 8, 9, 10, 11, 12 twelve week intervals, including all ranges there between. Periodic boosters at intervals of 1-5 years will be desirable to maintain protective levels of the antibodies. The course of the immunization may be followed by assays for antibodies against the antigens, as described in U.S. Pat. Nos. 3,791,932; 4,174,384 and 3,949,064, each of which is incorporated herein by reference.
A given composition may vary in its immunogenicity. It is often necessary therefore to boost the host immune system, as may be achieved by coupling a peptide or polypeptide to a carrier. Exemplary and preferred carriers are keyhole limpet hemocyanin (KLH) and bovine serum albumin (BSA). Other albumins such as ovalbumin, mouse serum albumin, or rabbit serum albumin can also be used as carriers. Means for conjugating a polypeptide to a carrier protein are well known in the art and include glutaraldehyde, m-maleimidobenzoyl-N-hydroxysuccinimide ester, carbodiimide, and bis-diazotized benzidine.
The immunogenicity of polypeptide or peptide compositions can be enhanced by the use of non-specific stimulators of the immune response, known as adjuvants. Suitable adjuvants include all acceptable immunostimulatory compounds, such as cytokines, toxins, or synthetic compositions. A number of adjuvants can be used to enhance an antibody response against antigenic peptides. Adjuvants can (1) trap the antigen in the body to cause a slow release; (2) attract cells involved in the immune response to the site of administration; (3) induce proliferation or activation of immune system cells; or (4) improve the spread of the antigen throughout the subject's body.
Examples of adjuvants include, but are not limited to, complete Freund's adjuvant (a non-specific stimulator of the immune response containing killed Mycobacterium tuberculosis), incomplete Freund's adjuvants, aluminum hydroxide, oil-in-water emulsions, water-in-oil emulsions, mineral salts, polynucleotides, and natural substances. Others adjuvants or methods are exemplified in U.S. Pat. Nos. 6,814,971, 5,084,269, 6,656,462, each of which is incorporated herein by reference.
In some embodiments, vaccines are directed towards dendritic cell therapy, which provokes anti-tumor responses by causing dendritic cells to present tumor antigens to lymphocytes, which activates them, priming them to kill other cells that present the antigen. Dendritic cells are antigen presenting cells (APCs) in the mammalian immune system. In cancer treatment they aid cancer antigen targeting.
One method of inducing dendritic cells to present tumor antigens is by vaccination with MHC II presenting peptides. Peptides can be administered as described herein, especially in combination with adjuvants (highly immunogenic substances) to increase the immune and anti-tumor responses. Other adjuvants include proteins or other chemicals that attract and/or activate dendritic cells, such as granulocyte macrophage colony-stimulating factor (GM-CSF).
A number of embodiments are directed towards identify MHC presenting peptides involved with autoimmune and/or allergic responses and then using those peptides to induce tolerance (i.e., desensitize response) to the peptides. It is known that a number of HLA alleles (especially MHC II HLA alleles) are involved in autoimmune and allergic responses. Autoantigens in type I diabetes, multiple sclerosis, rheumatoid arthritis, and systemic lupus erythematosus have been identified to correlate with various MHC II HLA alleles. Likewise, celiac disease, gluten sensitivity, dust mite allergies, pet allergies, and peanut allergies have been identified to correlate with various MHC II HLA alleles. Accordingly, peptides associated with these autoantigens and allergens can be screened to identify peptides responsible for these responses.
In some embodiments, individual are treated for an autoimmune disease or allergy by administering to an individual MHC II presented peptide antigens in way to sensitive the individual to the antigens. Intradermic, subdermal, and/or intramuscular administrations of peptides in ascending doses can be utilized to induce tolerance to certain autoimmune diseases and allergies.
Several embodiments are directed towards identifying MHC presenting peptides that can be utilized in a T cell therapy, which may be utilized in the treatments of various cancers. T cell receptors (TCR) recognize MHC I and MHC II complexes with HLA alleles and presented peptide. When an antigenic peptide is presented, the TCR induces an immune response against the antigenic peptide. Accordingly, presented peptides, as identified in embodiments described herein, can be utilized in T cell therapies.
In some embodiments, T cell can be screened to identify T cells with a T cell receptor (TCR) that recognizes a MHC with a particular presented peptide, such as peptide that would be presented as neoantigen on a cancer cell. Once T cells capable of recognizing peptides are identified, they can be isolated and propagated for a T cell therapy. In some embodiments, a patient's own T cells are removed from the patient and then screened and then propagated.
In some embodiments, a T cell can be engineered to express a TCR-like receptor to recognize a MHC with a particular presented peptide, such as peptide that would be presented as neoantigen on a cancer cell. Engineered T cells that recognize particular presented peptides in the MHC are typically referred to as TCR-like chimeric antigen receptor (CAR) T-cells, as they include a chimeric receptor similar to a TCR that is specifically designed to identify a particular presented peptide. Engineered T cells can be propagated for T cell therapies. In some embodiments, a patient's own T cells are removed from the patient and then engineered and then propagated.
In some embodiments, propagated T cells capable of recognizing MHC complex with a particular presented neoantigen are utilized to treat a patient having cancer. Accordingly, T cells are administered to the patient intravenously and/or intratumorally. The T cells would recognize the MHC complex with a particular presented neoantigen on the cancer cells, inducing the patient's immune system to respond to and attack the cancer.
The embodiments of the invention will be better understood with the several examples provided within. Many exemplary results of processes that identify HLA antigen peptides are described. Validation results are also provided.
Major histocompatibility complex class II (MHC-II) is a glycoprotein complex on the surface of professional antigen-presenting cells that displays short antigen peptides to CD4+ helper T cells. Human antigen-presenting cells, such as dendritic cells and B cells, rely in large part on HLA class II (HLA-II) for the presentation of antigens to CD4+ T cells. This human form of MHC-II can also be conditionally expressed by many other human cell types, including tumor cells. Antigen presentation by these HLA-II molecules on human cells involves three loci on chromosome 6 (DR, DQ and DP) which encode the corresponding heterodimeric proteins through combinations of alpha and beta chains.
Such HLA-II presentation of endogenous and exogenous antigenic peptides is essential for robust immune responses against diverse pathogens, and is also of major significance for autoimmunity and antitumor immunity. For example, recent mass spectrometry (MS)-based studies have shown that lymphoma and melanoma cells present somatically mutated cancer peptides (neoantigens) in the context of HLA-II. CD4+ T cell recognition of neoantigens is commonly observed across diverse human tumor types and in animal models, which underscores the potential clinical relevance of HLA-II-restricted neoantigens for cancer immunotherapy. Furthermore, neoantigens presented by HLA-II elicit potent antitumor responses in T cells from immunized patients. Reliably identifying presentation by HLA-II would allow prioritization of vaccine candidates and potentially identify likely responders to immune therapies.
Owing to the high cost and technical challenge of experimentally testing all possible peptide candidates, researchers have attempted to computationally identify HLA-II peptides with machine-learning algorithms. However, nearly all current HLA-II prediction methods rely on in vitro binding affinities of recombinant HLA-II molecules as surrogates, and therefore ignore other contributing factors including gene expression and protease cleavage preferences. When combined with the remarkably variable length of HLA-II peptides and their binding promiscuity, this deficiency makes HLA-II antigen presentation prediction task especially challenging. For example, the latest benchmarks report average receiver operating characteristic area under the curve (ROC-AUC or AUC) of ˜0.83 for current prevailing HLA-II prediction models, even when validated on in vitro binding data.
In this study, a deep neural network trained to accurately predict the likelihood of a peptide being presented by HLA-II complexes is described, which is referred to as MARIA throughout. Rather than relying on in vitro binding affinities alone, MARIA is trained on naturally presented HLA-II peptides (ligands) identified from human samples profiled by liquid chromatography-tandem mass spectrometry (LC-MS/MS). Despite some inherent limitations of MS methods, peptide ligand sequences identified by antigen presentation profiling currently provide the closest sample population to the true presented ligands. Such training data could enable new prediction models to consider multiple relevant features including expression and binding affinities. Here it is shown that MARIA allows robust and more accurate HLA-II prediction, and that its performance gains are achieved by combining these improved training data with a new supervised machine learning model using a multimodal recurrent neural network (RNN).
Immunoprecipitation of MHC molecules followed by peptide elution and LC-MS/MS analysis enables direct recovery of peptides presented by HLA-II in primary cells. In comparison to traditional in vitro binding affinity assays, MS-based profiling methods can rapidly yield large datasets of peptides actively presented by cells or tissues. The Immune Epitope Database (IEDB), the largest public depository of results of HLA binding assays, contains quantitative HLA-DR binding affinities for ˜12,000 non-redundant peptide sequences (as of December 2018) (see R. Vita Nucleic Acids Res. 43, D405-D412 (2015), the disclosure of which is incorporated herein by reference). By comparison, two studies employing HLA-DR immunoprecipitation and MS analysis identified >23,000 and >16,000 non-redundant peptide sequences, respectively (
The performance of the HLA binding affinity prediction tool was tested. The tool was trained on in vitro binding data to identify HLA-DR ligands presented by human antigen-presenting cells. NetMHCIIpan (R. Marty, et al., Cell 175, 416-428 (2018), the disclosure of which is incorporated herein by reference), a widely used HLA-II binding prediction method, was applied to predict the binding affinity of HLA-DR ligands experimentally identified from 18 mantle cell lymphomas (MCLs) representing 16 HLA-DR alleles (Table 1 in
The performance of NetMHCIIpan was further tested on >6,000 HLA-DR ligands discovered through deep profiling of the MCL-derived JeKo-1 cell line. When using the recommended threshold of binding affinities (top 10% ranks), NetMHCIIpan labeled only ˜22% of the ligands as positive (
To improve upon previous HLA-II prediction methods MARIA was developed to predicting utilizing active HLA-II presentation in vivo, rather than from in vitro binding affinities. It was therefore examined whether HLA-II prediction could be refined by learning directly from MS-based antigen presentation profiling datasets, in addition to traditional HLA binding affinity data. Additionally, it was tested whether gene expression and protease cleavage signatures also have utility in predicting HLA-II peptide presentation. MARIA was trained using the HLA-II ligands identified by MS-based antigen presentation profiling, along with empirical in vitro HLA binding measurements, and gene expression levels (
It was observed that gene expression levels of recovered HLA-DR ligands were significantly higher than both non-presented genes and random genes (
RNA sequencing (RNA-seq) gene expression profiles for six patients with MCL was analyzed. MARIA AUCs did not differ significantly between using patient-specific RNA-seq and an external RNA-seq profile (
It was next assessed whether information from the flanking residues of a peptide could further improve predictive performance. Both significant enrichment and depletion of certain amino acids at specific residues immediately upstream of the ligand N terminus or downstream of its C terminus was observed (
On the basis of the findings above, MARIA was developed with an integrative strategy to better predict HLA-II presentation. The model takes in three input values: the query peptide sequence, the patient or cell HLA-DR allele(s) and the corresponding gene name (
To systematically compare the performance of MARIA with alternative methods, antigen presentation profiling was applied to directly identify HLA-DR ligands from a human cell line (K562) expressing single HLA-DR alleles (DRB1*01:01, DRB1*04:04). Given the myeloid hematopoietic origin of this cell line (in contrast to the lymphoid tumors using for MARIA training), this allowed an assessment of both the effects of HLA-II allelic variation and the cell-of-origin on performance (
DRB1*01:01 was selected for initial testing in this system, as it has the most abundant training data for existing binding prediction methods, including NetMHCIIpan3.1 (M. Andreatta, et al., Immunogenetics 67, 641-650 (2015)), SMM Align (M. Nielsen, C. Lundegaard, and O. Lund, BMC Bioinformatics 8, 238 (2007)), NN Align (M. Nielsen and O. Lund BMC Bioinformatics 10, 296 (2009)), Sturniolo (T. Sturniolo, et al., Nat. Biotechnol. 17, 555-561 (1999)), CombLib (J. Sidney, et al., Immunome Res. 4, 2 (2008)), and IEDB Consensus3 (P. Wang, et al., BMC Bioinformatics 11, 568 (2010)). The performance of MARIA was benchmarked against these six methods when predicting the presentation of 1,331 DRB1*01:01 ligands empirically and distinguishing them from length-matched decoys (
The influence of neural network structure on prediction performance was also explored. Using the same MCL dataset of HLA-DR ligands used for training MARIA, a shallow neural network (SNN) similar to NetMHCIIpan with a single hidden layer was trained, as well a deep RNN model. These two models only considered peptide sequences, yet both outperformed NetMHCIIpan on external validation data from K562 (
HLA-DQ2.2 (DQA1*02:01 and DQB1*02:02) and HLA-DQ2.5 (DQA1*0501 and DQB1*0201) are known to present wheat gluten peptides and to predispose patients to celiac disease. To test the MARIA prediction framework in the context of the HLA-DQ locus, MARIA was trained on 11,482 HLA-DQ2.2 human peptide ligands identified from previously profiled cell lines (
In addition to the presentation of human peptides described above, the presentation of diverse gluten peptides by DQ2.2 has also previously been profiled by MS (see S. Dorum, et al., J. Immunol. 193, 4497-4506 (2014), the disclosure of which is incorporated herein by reference. Sixty-nine of the wheat peptides presented and 69 decoys were scored with both NetMHCIIpan and MARIA, which was exclusively trained on human peptides. NetMHCIIpan identified 6% of positive gluten peptides with 88% specificity at the recommended cut-off (90th percentile). By comparison, MARIA identified 49% of positive gluten peptides with 92% specificity (
A small number of reported natural HLA-DP ligands were also identified (n=20; Table 12 in
The ability of MARIA to identify immunogenic neoantigens in cancer was assessed. Ideal antitumor neoantigen candidates should be both presentable by HLA complexes and capable of inducing proinflammatory responses by interacting with T cell receptors. Most current cancer vaccine platforms prioritize candidate neoantigens for vaccine production by selecting only highly expressed candidates with high predicted binding affinity for self-HLA alleles. Yet, many vaccine peptides do not elicit T cell responses upon vaccination, despite rigorous candidate selection. It was therefore tested whether MARIA could better select neoantigens that were most likely to induce a T cell response upon vaccination.
Using antigen presentation profiling, hotspots within specific immunoglobulin (Ig) regions are presented by HLA-DR and associated with antitumor CD4+ T cell responses to lymphoma neoantigens. It was therefore tested whether MARIA could accurately identify potential Ig antigens as potential lymphoma-specific targets for immunotherapy. For this test, all Ig-derived peptides were intentionally excluded for the training. This Ig-naive version of MARIA was applied to predict presentation of Ig sequences in the tumors. The resulting MARIA-predicted presentation scores were significantly correlated with MS-identified HLA-DR ligand frequencies across the full-length heavy and light chains (
Immunoglobulin heavy chain variable regions represent challenging test examples as most HLA ligand prediction algorithms including MARIA were trained on wild-type peptides. To further address the utility of MARIA for predicting presentation of mutated peptides, MARIA predictions of HLA-II intrinsic ligand (CLIP) were assessed with and without specific point mutations. For diverse CLIP variants, MARIA scores consistently correlated with stabilizing versus destabilizing structural changes, while NetMHCIIpan did not (Supplementary Note 3 and
Personalized protein-coding somatic mutations are attractive cancer vaccine candidates in melanoma owing to the high mutation burden of patients with melanoma. MARIA was assessed whether it could prioritize vaccine candidates for melanoma. 10,513 melanoma self-antigens identified by MS were analyzed, generated from two bulk melanoma tumors with distinct HLA-DR alleles (Mel15, DRB1*03:01 and DRB1*07:01; Mel16, DRB1*13:01 and DRB1*08:03). Each melanoma-presented ligand or decoy was scored using both NetMHCIIpan and MARIA (trained on lymphoma data). Even without patient-specific gene expression data, MARIA outperformed NetMHCIIpan when differentiating melanoma HLA-II ligands from decoys (
After confirming the performance of MARIA in non-hematopoietic tissue, such as melanoma, MARIA was used to analyze two sets of personalized melanoma vaccine neoantigens with corresponding immune response data (ex vivo CD4+ T ell enzyme-linked immunospot (ELISPOT) test; n=121 and 96). Gene expression levels of neoantigens alone for T cell reactive and non-reactive neoantigen candidates were largely indistinguishable (
Conversely, peptides with the highest MARIA scores (>99.5% MARIA percentiles) were more likely to elicit a T cell response upon vaccination (
To further validate MARIA performance for predicting immune responses, seven additional cancer-related CD4+ T cell response studies were assessed. Each of these studies individually identified a small number of cancer-associated CD4+ T cell epitopes across diverse cancer types. Using a 95% MARIA percentile cut-off (
HLA-I and HLA-II both play central roles in antigen recognition and adaptive immune responses. HLA-II gene expression analysis in the Cancer Genomic Atlas (TCGA) cohorts suggests abundance of antigen-presenting cells or tumor HLA-II presentation in various cancer types (
The SysteMHC Atlas was recently constructed to consolidate HLA-I and HLA-II ligand sequences from a diverse set of studies. MARIA was designed to be capable of integrating additional training from emerging sources such as the SysteMHC Atlas and expect its performance to improve as such ligand datasets continue to grow. For example, recurrent patterns in HLA-II cleavage signatures was observed, including the enrichment of ligand flanking sequences for tyrosines and their depletion for prolines. However, cleavage signatures from different cell types showed subtle variation in motif sequences (
The results also demonstrate how MARIA might allow researchers to better identify immunogens relevant to autoimmunity and to antitumor immunity. Given the inherent challenges limiting the accuracy of previous methods for characterizing tumor-derived HLA-II ligands, MARIA should allow researchers to explore less well-studied HLA-II neoantigens. Specifically, MARIA is useful for directly identifying and prioritizing cancer vaccine candidates from patient sequencing data.
Candidate peptides were synthesized with N-terminal 2,4-dinitrophenyl (DNP) tags joined by a 6-am inohexanoic acid linker (Sigma). Biotinylated HLA-DR recombinant proteins (HLA-DRB1*04:03 and HLA-DRB1*07:01) molecules were provided by the NIH tetramer core. Intrinsic CLIP peptide was cleaved from the HLA-DR molecules with human rhinovirus 3C protease. DNP-tagged peptides were supplied in molar excess to encourage efficient exchange of binders and incubated overnight at 32° C. or 37° C. (pH 4.5). Exchange reactions were then neutralized with 1 M Tris, pH 8.0 and biotinylated HLA-DR molecules were bound to streptavidin microspheres (Polyscience). Microspheres were washed and stained with allophycocyanin (APC)-labeled anti-HLA-DR (clone L243; BD Biosciences, 340549) and anti-DNP (clone 2-9(4); Abcam, ab6306) followed by rat anti-mouse IgE FITC secondary antibody (clone R35-72; BD Biosciences, 553415). Microspheres that were positive for HLA-DR and DNP-tagged peptide were detected by flow cytometry. Peptides were considered to be binders if both HLA-DR and DNP signals were detectable above an HLA-DR unexchanged control.
Cell lines expressing single HLA-DR alleles were prepared from K562 cells, which do not express surface class I or II HLA, by lentiviral transduction. Sequences for the DR α-chain and the relevant β-chain alleles (DRB1*01:01 and DRB1*04:04) separated by a 2A peptide sequence were encoded in the N103 lentiviral vector backbone (kindly provided by J. Crabtree, Stanford University) and used to produce lentiviruses in HEK293 cells. To enhance expression levels of HLA-DR in the K562 cell lines, the top 1% of cell populations were selected and expanded on the basis of surface HLA-DR signal with fluorescence-activated cell sorting (clone L243; BD Biosciences 347367). Expression of HLA-DR was confirmed by flow cytometry before and after sorting (
HLA-DR immunopeptidomes were extracted from the K562 HLA-DRB1*01:01 and K562 HLA-DRB1*04:04 cell lines. HLA-DR molecules were isolated and the associated peptides were extracted. See Supplementary Note 4 for detailed HLA-DR immunopeptidome purification and MS analysis.
Detailed ligand sequence data sources are listed in
HLA alleles of patients with MCL were identified with PHLAT from patient tumor exome sequencing data. HLA alleles of melanoma patients were identified with HLAVBSeq from patient exome sequencing data. When patient alleles were not available, HLA-DRB1*07:01 and HLA-DRB1*01:01 were used as they are the most common alleles in general populations (www.allelefrequencies.net/).
All specimens were obtained with informed consent in accordance with the Declaration of Helsinki and this study was approved by Stanford University's Administrative Panels on Human Subjects in Medical Research. Samples were collected from patients as part of a clinical trial of autologous tumor vaccination (NCT00490529). Patient leukocytes were collected by leukapheresis approximately 2 weeks after a series of autologous tumor vaccinations. Cells were cultured in a 1:1 mix of AIM-V medium and RPMI1640 (Thermo Fisher) with 10% pooled human AB sera (Gemini Bio) and 50 μM β-mercaptoethanol. Neoantigen peptides were synthesized (ElimBio) and added to a final concentration of 10 μg m1−1. In one patient (MCL052), cells were treated concurrently with two predicted neoantigen peptides. As a positive control, cells were stimulated with a mixture of pathogen-associated peptides, CEFT pool (JPT Peptide Technologies). Cells were incubated for 30 h before flow cytometry analysis. CD137 (clone 4B4-1; BD Biosciences, 561702) and CD69 (clone L78; BD Biosciences 341652) expression was assessed on live CD4+ (clone RPA-T4; BD Biosciences, 562659) T cells using a FACS Aria sorter (BD Biosciences). See
Only minor differences in gene expression profiles were observed when using personalized versus inferred gene expression levels with modest impacts on MARIA prediction results (
Gene expression profiles of melanoma in the cancer vaccine studies were provided when analyzing how original neoantigen gene expressions correlated with responses. A fixed TPM of 50 was used as input for MARIA to evaluate how MARIA scores with post-vaccine responses as cancer vaccines made each neoantigen peptide readily available to antigen-presenting cells. A fixed TPM of 50 was also used when evaluating how HLA-DRB1*01:01 binds to CLIP. Gene expression values were not needed when MARIA analyzed HLA-DQ2.2 or HLA-DP peptides, as only peptide sequences were used during training.
The gene expression value (TPM) of each HLA-DR-peptide-producing gene was estimated with RNA-seq of samples from patients with MCL or MCL cell lines. A gene was allowed to be counted multiple times if multiple peptides were identified from the same gene. Each gene expression value was converted into logarithmic space with log10(X+10−6) and represented as a violin plot (
Using a similar approach, the gene expression profiles of all genes were analyzed in patients with MCL and MCL cell lines as a background distribution. Each gene with a known RNA-seq gene expression value was converted into logarithmic space, and each gene was counted once. The median was calculated and a Mann-Whitney U test was conducted on gene expressions between MCL HLA-DR-presented genes and the background distribution.
To investigate sources of MCL HLA-DR peptides with extremely low gene expression values, Gene Ontology (GO) term enrichment analysis was conducted on HLA-DR peptide genes with TPM <0.1 (
To test the ability of gene expression values to differentiate HLA-II ligands from decoys, a logistic regression model was built with gene expression values as a single feature to differentiate HLA-DR peptides from a random human peptide decoy. Decoy gene lists were generated from a Uniport reviewed human protein list and were enriched for expressed genes. Specifically, human peptide decoys contain roughly 40% genes with TPM >10, 10% genes with unknown expression levels and 50% genes with TPM <10. Predictive performance of gene expression values under different assumptions is shown in
See Supplementary Note 5 for an overview of neural networks considered and their implementation. Neural networks in this study were implemented with Keras 2.0.3 (keras.io/) using the Tensorflow backend and Python 2.7. For training, an NVIDIA Tesla K80 GPU with CUDA 7.5.18 was utilized on the Stanford XStream GUP cluster to speed up the gradient descent.
Each amino acid in a peptide sequence is represented by a 21-number one-hot vector (20 common amino acids+X) A one-hot vector is a vector of zeros with a single one that indicates the amino acid species in an arbitrary but consistent mapping (for example, position one=alanine;
The majority of existing HLA-II peptide prediction algorithms use data on the binding of recombinant HLA-II protein to peptides as a surrogate to rank potential HLA-II peptides presentation or HLA-II neoantigens. MARIA aims to directly predict the probability of a given peptide being presented by a HLA-II complex in a cell or patient context. Rather than in vitro measurements of the binding of HLA protein to peptides, HLA-II ligands identified by MS/MS peptide sequencing data from antigen-presenting cells were used as the ground truth.
The MARIA model takes peptide sequences (8-26 amino acids long), patient or cell HLA-DR alleles and corresponding gene names to predict peptide presentation by the HLA-DR complex (
Besides estimating binding affinities, cleavage scores and gene expression values, MARIA encodes each ligand peptide sequence using an RNN layer known as LSTM. All of these values are integrated with two fully connected dense layers to estimate the probability of a peptide-gene pair being presented in a allele- and gene-expression-specific manner (
LSTM networks with 32, 64 and 128 neurons were assessed and from one to four layers deep with a 9:1 training: validation scheme (
MARIA was trained with the Tensorflow GPU backend to enable parallel calculation of gradient. For general user applications, the CPU backend is sufficient. Ten thousand peptides take ˜80 s with a 2.8 GHz Intel Core Xeon CPU or 11.3 s with one NVIDIA Tesla K80 GPU.
MARIA assumes the main influence on HLA-DR allele is from the change in HLA-DR-peptide binding affinity. A pan-allele regression model was trained with the publicly available IEDB data to estimate binding affinity given a peptide-allele pair. Training data was curated with 33,909 peptide-allele pairs for HLA-DR. Each HLA-DR allele is converted into a 19-amino-acid pseudopeptide sequence reflecting 19 amino acid residues in HLA-DRB1 interacting with ligand peptides in the binding groove. Each peptide-allele pair has a corresponding binding affinity measured in half maximal inhibitory concentration (IC50, in nM). For computational efficiency, IC50 was converted into log space with the formula (1−log50,000(IC50)). The model includes an LSTM layer followed by two dense layers and a single output neuron (
To evaluate the performance of the LSTM-based binding affinity predictor with NetMHCIIpan3.1, MARIA was evaluated on the same held-out set of in vitro binding. Similar ROC-AUC scores and Pearson's correlation coefficients (predicted versus measured) were revealed.
To understand amino acid preference for HLA-II peptide cleavage, amino acid frequency around HLA-II peptide cleavage sites was compared with a background distribution. Cleavage sites included six amino acids upstream of the HLA-II peptide N terminus (−6 to −1) and six amino acids downstream of the HLA-II peptide C terminus (+1 to +6) assuming a N terminus to C terminus direction. An equal number of gene-matched and length-matched peptides was randomly generated on the basis of the HLA-II peptides, and the amino acid frequency from the same cleavage sites was used as the background distribution. Enrichment and depletion level were calculated as HLA-II peptide amino acid i frequency at j position divided by background distribution amino acid i frequency at j position (
A neural network was built to quantitatively estimate cleavage scores given the flanking regions of a query gene-peptide pair (
Raw output of MARIA is a score between 0 and 1 indicating how likely a query peptide is to be presented by a specific HLA-II. To increase human interpretability and enable comparison across different peptide lengths, MARIA's output can be represented as a percentile score. A percentile score is generated by comparing the raw output score to a score distribution generated from length-matched random human peptides. The higher the percentile, the more likely the peptide will be presented by a cell HLA-DR complex.
Specifically, for each peptide length of 8 to 26, 20,000 random natural peptides were sampled from the human proteome. MARIA was run on each set of random peptides and used the output as empirical distributions for normalizing query peptides with a certain length. For example, MARIA output of a 15-amino-acid query peptide will be compared against the scores of 20,000 random 15-amino acid oligomers to obtain a percentile score. NetMHCIIpan used a similar approach to generate ranks, and NetMHCIIpan percentiles=100%−NetMHCIIpan ranks.
To fully evaluate the performance of MARIA and related methods, a set of independent HLA-DR ligandomes was obtained from various cell types and MS instruments (
To evaluate how NetMHCIIpan performed on MS-identified MCL HLA-DR ligands, NetMHCIIpan3.1 was run on a set of 18 MCL samples with a minimum of 200 peptides identified (Table 1 in
MARIA was benchmarked with six commonly used HLA-II prediction algorithms on non-MCL data: NetMHCIIpan3.1, SMM Align, NN Align, Sturniolo/TEPITOPE, Comblib and IEDB Consensus. All six algorithms were trained on in vitro recombinant protein binding data but differ in allele coverage and machine learning algorithms. Held-out data included K562 myeloid cells and primary melanoma patient samples. Because IEDB Concensus (Concensus3) is only compatible with 50 HLA-DR alleles and peptide sequences longer than 14, MARIA was compared with NetMHCIIpan in most cases.
NetMHCIIpan and NN align both use a dense neural network to scan through a given query peptide with a 9-amino-acid oligomer sliding window. The best 9-amino-acid oligomer score is reported. NetMHCIIpan incorporates important amino acid sequences on HLA-II alleles as a part of input features to train a universal algorithm for all HLA-II alleles. NN Align learns binding patterns of each HLA-II allele separately. SMM Align and Comblib both use a stabilized matrix method. Sturniolo/TEPITOPE used a combination of expert rules and assembled matrices. IEDB Consensus (Concensus3) is an ensemble method that is based on scores from NetMHCIIpan, NN Align, SMM Align and Sturniolo. Depending on the HLA-II allele, IEDB Consensus combines scores from one to three methods to report a ranking score.
A separate model was trained for HLA-DQ2.2-presented peptide that was based on two available HLA-DQ ligandomes for HLA-DQ2.2 and HLA-DQ2.5. The HLA-DQ model neural network architecture is identical to the sequence-only LSTM model for HLA-DR (
The ability of the DQ model was assessed to differentiate DQ2.2-binding wheat gluten peptide (n=69) from length-matched random human peptide (n=138). NetMHCIIpan3.1 was run on the same test set with the allele input of HLA-DQA10201 and HLA-DQB10202. The performance of these two methods was compared by measuring their sensitivity when the specificity was set to 90% (cut-off at the 90th percentile for both MARIA DQ model and NetMHCIIpan3.1; percentile=100%−rank).
The HLA presentation of Ig from patients with MCL was estimated on the basis of how each amino acid position was covered by experimentally identified or predicted peptide ligands. HLA-presented Ig peptides from 18 MCL samples were obtained from a previous MS-based study. All MS-identified peptides were aligned against a standard Ig template sequence with the Biopython pairwise alignment function. The MS-identified ligand number at each amino acid position is equal to the number of peptide ligands covering the position.
To generate predicted HLA-DR ligands, Ig sequences were broken down from 18 MCL samples into consecutive 15-amino-acid sliding windows. Each sliding window peptide was scored by NetMHCIIpan or MARIA (a version not trained on Ig sequences) with matched patient alleles. High-scoring peptides were aligned against the same Ig template sequence as described above. Scores (S) for peptide sequences were based on their predicted presentation percentiles (p) and the numbers of ligands identified from the corresponding patient samples (q):
where w is a constant to scale the predicted heat map and MS-identified heat map into a similar scale for plotting purposes. The predicted ligand number at each amino acid position is the sum of scores across 18 MCL samples.
Analyzing Cancer Neoantigen Candidates with MARIA
To score each somatic mutation in two cancer vaccine trials with MARIA, each somatic mutation was represented by a sequence that includes 14 amino acids upstream and downstream of the mutation site, such that the typical sequence length is 29 amino acids (14+1+14). 15-amino-acid sliding windows along these sequences were combined each window with the HLA alleles and gene expression values of each patient to produce a MARIA score. In therapeutic cancer vaccine studies, 50 TPM were used to reflect the high concentration of mutated peptides introduced by vaccines, otherwise median TCGA tissue-matched RNA-seq gene expression was used. The highest score of all sliding windows was used as the score for a given somatic mutation. Neoantigen examples with known unsupported alleles (for example, HLA-DQB1*06:01) were excluded from the analysis. The same analysis was performed with the IEDB CD4 immunogenicity tool for comparison.
N annotates the number of a group of peptide ligands. Sensitivity (also known as recall) was calculated as:
Specificity was calculated as:
Positive predictive value (also known as precision) was calculated as:
Negative predictive value was calculated as:
ROC-AUC scores were calculated on the basis of the area under sensitivity and 1−specificity curves and implemented with Python scikit-learn. To generate precision and recall curves, pairs of recalls and specificities were first calculated across a range of MARIA cut-offs (70th to 99.9th percentile). Precision for each recall was then calculated using an assumption of positive peptide prevalence (1%; Supplementary Note 2).
Sequence logo plots and amino acid frequency enrichment were generated with IceLogo. Plotting in this study was done in matplotlib and seaborn. Two-tailed paired t tests were used in
Supplementary Note 1: Importance of Gene Expression for HLA-II Presentation and Robustness of MARIA with Gene Expression References
In this study, the importance of gene expression or protein abundance for HLA-II presentations was demonstrated. Strikingly, gene expression alone obtained AUC of 0.81 using gene expression as a predictor in the validation set (
Genes with high expression or unknown expression were enriched for the decoy genes/peptides in this validation set. When the human genome was sampled with a uniform distribution, AUC of gene expression went up to 0.84 since more non-expressed genes were selected. Conversely, the predictive performance of gene expression values dropped when more highly expressed genes were selected as a negative set (
To understanding whether MARIA requires accurate or even personalized gene expression profiles to predict patients antigen presentations, the following two experiments were performed. For six MCL patients with personalized RNA-Seq data available, the predictive performance of MARIA was evaluated with external RNA-Seq values in comparison to patient matched values. The results yielded a little drop in MARIA performance when using external values (
For cancer vaccine applications, personalized RNA-Seq profiling for each patient is theoretically desired. However, correlations between CD4 T cells responses and gene expression values of targeted patient neoantigens were not observed (
A 1% prevalence was assumed for presented HLA-II ligands based on the following two CD4 T-cell epitope studies. In a high throughput screen for melanoma neoantigens the researchers identified 4 immunogenic mutated peptides from 458 candidates. In a melanoma vaccine trial, the researchers identified 18 immunogenic mutated peptides from 97 candidates. Both studies were systematic in their approach to screening, and neither of these studies filtered candidates through use of predicted HLA-II binding scores.
HLA-II ligand presentation is essential for CD4 T-cell responses. However, peripheral tolerance, regulatory T-cells, tumor immune editing and low sensitivity of assays can all lead to negative T-cell responses for presented antigens. Further, presented HLA-II neoantigens sometimes stimulate regulatory T-cells rather than conventional CD4 T-cells, which led to negative read-out for common interferon gamma markers. Thus these two studies (0.8% and 19%) suggest a lower bound of the presented ligands for HLA-II. To be on the conservative side, 1% was chosen. This number is higher than directly observed ligand numbers from some MS HLA-ligand studies likely due to limited sensitivities of the current MS technology. A more comprehensive understanding of the fraction of presented HLA-II antigens that are able to elicit strong CD4 T-cell responses will likely require prospective screening of a very large number of antigens in many subjects in future clinical trials.
Docking method: Docking of CLIP variants was accomplished using the FlexPepDock, which estimated how peptides interact with a protein complex given structure files (e.g. PDB files) of protein (e.g. HLA-DR1) and peptides (e.g. CLIP). Using PyMOL mutagenesis function to introduce single amino acid mutation in the wild type structure of HLA-DR1 with bound CLIP14 (PDB ID: 3PDO), the conformation with the least strain was selected as the FlexPepDock input file. Structure changes of each mutant complexes were determined based on top three output poses from FlexPepDock.
Docking results: It was tested whether MARIA scores could predict expected changes in binding affinity for single amino acid mutations in peptide ligands as relevant to neoantigens and minor antigens. This is important to test because such mutant peptides were rare in most of validation studies. NetMHCIIpan and MARIA were evaluated when predicting binding/presentation of CLIP (fragment of CD74, PVSKMRMATPLLMQALP) of HLA-DRB1*01:01 complex and its mutated counterparts (
Using in silico docking experiments and previous structural studies, 7 mutations that enhance, impede, or have little effect on binding were considered. To allow comparison between methods, both raw MARIA and NetMHCIIpan scores were normalized into percentiles by comparing raw scores to random human peptides, where higher percentiles reflect better predicted binding. NetMHCIIpan yielded binding percentiles exceeding 99.85% for all CLIP mutants with the exception of a single CLIP mutant showing a modestly decreased NetMHCIIpan score (R108D, 97.50%). In comparison, MARIA scores for these same peptide mutants were more consistent with the expected structural changes (75.51-97.96%,
Supplementary Note 4: Identifying HLA-DR with Mass Spectrometry
HLA-DR immunopeptidome purification: HLA class-DR molecules were isolated and the associated peptides extracted. In brief, cells were lysed for 20 min on ice in 20 mM Tris-HCl (pH 8), 150 mM NaCl, 1% CHAPS, 0.2 mM PMSF, 1× Halt Protease and Phosphatase Inhibitor Cocktail (Thermo Fisher Scientific) supplemented with complete protease inhibitor cocktail (Roche). The lysate was centrifuged (2×30 min, 13,200 rpm at 4° C.) and the resulting supernatant was precleared for 30 min using recombinant Protein A Sepharose fast-flow beads (GE Healthcare). Precleared lysate was incubated with the HLA-DR specific antibody L24317 coupled to rProtein A Sepharose fast-flow beads for 5 h at 4° C. Following the immune-capture of HLA-II molecules, beads were washed with TBS (pH 7.4) and peptides were eluted from the purified HLA molecules using 10% acetic acid. The eluate was then passed through a 10 kDa MWCO size filter and stored at −80° C. until LC-MS/MS analysis.
Mass spectrometry analysis of HLA-DR-associated peptides: Isolated HLA peptides were reconstituted in 12 μl of 0.1% formic acid and analyzed on a Fusion Lumos mass spectrometer (Thermo Fisher Scientific). Peptides were separated by capillary reverse phase chromatography on a 24 cm reversed phase column (100 μm inner diameter, packed in-house with ReproSil-Pur C18-AQ 3.0 m resin). The Fusion Lumos was equipped with a Dionex Ultimate 3000 LC-system and used a two-step linear gradient with 4-25% buffer B (0.1% (v/v) formic acid in acetonitrile) for 80 min followed by 25-45% buffer B for 10 min. Data were acquired in top speed data dependent mode with a duty cycle time of 3 s. Full MS scans were acquired in the Orbitrap mass analyzer with a resolution of 120 000 (FWHM) and m/z scan range of 340-1540. Precursor ions with mass range of 700-2760 and charge state 2-6 and intensity threshold above 50,000 were selected for fragmentation using higher-energy collisional dissociation (HCD) with quadrupole isolation, isolation window of 1.6 m/z and normalized collision energy of 30%. HCD fragments were analyzed in the Orbitrap mass analyzer with a resolution of 15,000 (FWHM). Fragmented ions were dynamically excluded from further selection for a period of 30 seconds. Each sample was measured twice, once with above described HCD method and a second analysis using a method which toggled HCD and electron transfer dissociation (ETD) fragmentation modes for each isolated precursor using the following parameters for ETD: charge state 2 was excluded, calibrated charge dependent ETD parameters were enabled and 25% of supplemental collision energy was used. The AGC target was set to 400000 and 50000 for full FTMS scans and FTMS2 scans. The maximum injection time was set to 50 ms and 200 ms for full FTMS scans and FTMS2 scans.
Computational identification of immunopeptidomes from mass spectra: All tandem mass (MS/MS) spectra were processed using Proteome Discoverer (v 2.2.0.388) and queried against a “target-decoy” sequence database19 consisting of the human UniProt proteome (June 2016) with added common contaminants (e.g. keratins and staphylococcus protein A) using the SEQUEST search engine. SEQUEST search parameters were set as follows: spectrum matching was set to one for b and y ions for HCD and c and z for EThcD. Parent mass error tolerance was set to 10 ppm and fragment mass error tolerance to 0.02 Da. Enzyme specificity was set to none, peptide length was set to 7-25 amino acids, and oxidation of methionines and deamidation (N,Q), cysteinylation, and phosphorylation (S, T, Y) were considered as variable modification. High-confidence peptide identifications were selected at a 1% false discovery rate (FDR) using the Percolator algorithm with a validation based on the q-value. During machine learning, modified amino acid residues were encoded as the original amino acids.
Neural networks or artificial neural networks mimic natural neural networks by constructing multiple layers of gating neuron units to enable a signal processing model with high complexity and flexibility. Neural networks have achieved many successes across a diverse set of classification problems and are fundamental building blocks of deep learning. For a binary classification model, the output layer contains two neurons representing the probability of the input being positive (T) or negative (F) (
Σi=1np(xi)log(q(xi))
where n is the total sample size, p(xi) is the true label of the input xi (0 or 1) and q(xi) is the predicted presentation score for input xi (0-1). Presentation score of an input sequence was defined by the value of positive output neuron (T). The model was trained to minimize the loss function. Softmax functions enable the positive and negative output neuron values to sum up to one.
For a regression model (e.g. predicting binding affinities), the output layer contains a single neuron without additional activation function (
A recurrent neural network is a special type of neural network layer designed to process variable length-sequence data (
Supplementary Note 6: Predicting HLA-II Ligands with Sequence Information Only
With deep RNN/LSTM: A deep neural network model was built solely based on peptide sequences recovered from HLA-DR MS. The model takes in variable length of peptide amino acid sequences (8-26 AA long) and encodes them with one-hot encoding as described (
With shallow neural network (hidden layer=1): The prediction performance of LSTM was compared with a shallow network consisting only of fully connected hidden layers such as NetMHCIIpan. A version of NetMHCII was trained on the same training and validation data as the sequence only LSTM model. Briefly, NetMHCII feeds the following information into a conventional dense neural network: peptide length, peptide N-terminal amino acid, peptide C-terminal amino acid, 9 amino acid sequence (9mer) of predicted binding core, sequence length left to the binding core (left peptide flanking region), sequence length right to the binding core (left peptide flanking region). These information are fed into one hidden layer of 40 neurons 27, and the output is two neurons with binary cross categorical entropy as the cost function. To determine the 9mer binding core, all possible 9mer sliding windows for a query peptide (minimum 9AA long) are generated and fed into the neural network independently. The 9mer window with the highest predicted binding score was chosen as the binding core for the query peptide. No HLA-DR allele information was included in training to be consistent with the LSTM sequence-only model.
The goal is to build a model that takes into account the biological mechanisms that result in antigen processing and presentation. At the beginning of the antigen presentation pathway, genes are variably expressed and their mRNA products are translated into proteins. To be presented by MHC-I, these proteins must be cleaved by the cytosolic proteasome, carried by the Transporter associated with Antigen Processing (TAP) into the endoplasmic reticulum, and loaded onto the MHC molecule that is transported to the cell membrane for display. In recognition of the selective narrowing of the number of candidate antigens, features were created to reflect the biological patterns within the class I presentation pathway (
Gene expression influences the availability of protein available for presentation and is significantly correlated with MHC-I antigen presentation. MHC-I presented peptides had lower gene expression compared to MHC-II presented peptides and peptides presented by both MHC-I and MHC-II had higher expression levels compared to non-presented peptides (
MARIA-I (MHC Analysis with Recurrent Integrated Architecture for Class I) Design
MARIA-I (MHC Analysis with Recurrent Integrated Architecture for class I) combines the features above into a single is a model to predict MHC-I antigen presentation (
To utilize MARIA-I, the user inputs a query peptide sequence, its corresponding gene name, and the relevant MHC-A, B, and C alleles (6 total). MARIA-I obtains a gene expression value from either a TCGA tissue-matched expression dictionary or user-provided expression values as transcripts per million (TPM). A cleavage sub-model was trained on flanking residues of naturally presented peptides, which uses the gene and peptide to create a cleavage score. The binding affinity sub-model, trained on in-vitro binding data, uses the six alleles and peptide sequence to calculate six binding scores. The query peptide is also encoded in a separate MS-sequence sub-model that is trained to differentiate between naturally presented and decoy peptides based only on amino acid sequences. MARIA-I then merges the gene expression, cleavage, and binding affinity scores with the encoded peptide sequence to generate a presentation score. This presentation score is transformed into a percentile based on a background distribution of 10,000 random human peptides for each MHC allele.
The relative contribution of each feature helps explain its relative importance in predicting antigen presentation. Gene expression had a modest contribution to presentation prediction (
MARIA-I was assessed and compared with other methods to predict antigen presentation as identified by MS. Other methods for comparison included netMHCpan4.0, MixMHCpred, and MHCflurry, each of which are trained on similar MS antigen data. Three diverse external validation data were utilized for the assessment: 1 ovarian cancer cell line, 12 meningioma samples, and 17 tumors including three chronic lymphocytic leukemia, one ovarian, six glioblastoma, and seven melanoma samples. These three datasets include 209,876 presented MHC-I peptides from a diverse set of MHC-I alleles and 534,431 decoys generated for testing. To truly reflect the rarity of MHC-I neoantigens, a prevalence of 1% was used to calculate positive predictive values (PPV).
For the ovarian cancer sample, MARIA-I and three alternative methods were applied utilizing 1,377 positive and 4,131 decoy ligands from the MS-identified SK-OV-3 cell line ligands. MARIA-I had a significantly higher PPV compared to netMHCpan4.0, the next best method among other existing methods (PPV 0.594 vs. 0.512, p-value=7.25×10{circumflex over ( )}-9,
Existing murine tumor models have proved vital insights of tumorigenesis and cancer immunology. Recent studies have indicated needs for MHC antigen prediction tools that can prioritize neoantigens in mouse models. Using MS-identified, publicly available peptides presented on MHC-I haplotype b and MHC-II haplotype I-Ab, Mouse MARIA was developed, an adaptation of the human MARIA models to predict murine antigen presentation. Mouse MARIA models are trained on MS-identified and decoy peptide sequences (11,919 MHC-I ligands and 3,709 MHC-II ligands).
Performance of the MARIA MHC-I haplotype b (genotype) on an external validation set of 512 MS-identified ligands and 5,769 decoys (IEDB) is comparable to netMHCpan4.0 (
MARIA-I Scores are Associated with CD8+ T Cell Responses
Given the promising results on antigen presentation, MARIA-I was applied to prioritize cancer neoantigens. These peptides are generated by cancer mutations and most likely absent from the normal human genome. They represent a class of promising antigens that may drive anti-tumor T cell responses in immunotherapies. First, a pipeline was created to process publicly available whole exome sequencing (WES) into candidate peptides that can be scored by MARIA-I and other methods (
MARIA-I was evaluated on a study that tested antigen immunogenicity using in vitro T cell assays on TILs from gastrointestinal cancers (colorectal, bile duct, pancreas, stomach) (n=7,422,). Each candidate antigen derived from exome sequencing was binned into one of three percentile categories (Low: <95, Medium: 95-99.5, High: >99.5). In the absence of immune stimulatory agents, only a small fraction of synthesized neoantigen peptides induced immune responses in their patient TILs (0.74%), but MARIA-I scores helped enrich robust CD8+ T-cell epitopes. Peptides that MARIA-I ranked as high, with scores >99.5 percentile, had a 3% response rate, as compared to 0.5% and 1.1% in the low and medium category, respectively (
Cancer vaccines can leverage T-cell mediated cellular immunity through recognition of cancer neoantigens. Clinical trials have shown their promise in anti-tumor immune responses and improved sequencing methods and lowered the barriers to identifying cancer-specific mutations. Identifying strong neoantigens from a pool hundreds of candidates, by their MHC presentation or T cell recognition profiles, is vital for therapeutic success. Previous vaccines designs have prioritized candidate peptides using binding affinity prediction tools like NetMHCpan, which have yielded limited success in generating T cell responses.
Using the same three percentile cut-offs as above, MARIA-I's ability to select immunogenic neoantigens was evaluated on personalized melanoma vaccine antigens with matched CD8+ T cell response data (ex vivo IFN-gamma release, n=91). 40% of neoantigen peptides in the high category were able to induce a post-vaccination CD8+ T-cell responses measured by ex vivo IFN-γ release assays. In comparison, MARIA-I-ranked antigens in the low and medium bins were less likely to elicit T cell responses (10% and 12%, respectively;
To explore whether more promising neoantigens existed within their patient mutation profile, all nonsynonymous SNVs were analyzed as potential neoantigens with MARIA-I from the six patients vaccinated with immunizing peptides. (Fraction of mutations fall into MARIA-I high) On average, 83.8% of mutations labelled as “highly presentable” (>99.5 percentile) by MARIA-I were unused in their vaccine design. Each patient has X-Y promising neoantigen candidates yet utilized by the original study. 16.2% of mutations are labelled as “highly presentable” (>99.5 percentile) by MARIA-I, on average, were used to vaccinate patients with melanoma, suggesting the possibility of missed mutations that may contribute to a strongly presented peptide (
In summary, MARIA-I scores are significantly correlated with robust CD8+ T-cell responses despite the model being trained on antigen presentation data.
The prior example in predicting MHC-II antigen presentation showed that combining features derived from the MHC peptide processing pathway, including gene expression, cleavage signature, and binding affinity, with MHC-bound peptides sequenced through MS yielded significant improvements in performance over existing class II predictors (see Example 1). MARIA-I replicated this result in MHC-I antigen presentation. With respect to gene expression, as was the case in MHC-II prediction, tissue-specific levels provide a small improvement over unmatched gene expression levels but allow the model to maintain a high prediction performance despite variations in measurement of this feature.
It is also noted that the majority of published MHC ligand data are from heterozygous samples (not mono-allelic cell lines), in which the 1:1 relationship of peptide to MHC is unknown. This is one of the bottlenecks of training MHC presentation prediction algorithms. NetMHCpan4.0 trains on unambiguous MHC allele data, restricting training data to those peptides that have an identified restricting MHC. MixMHCpred2.0 deconvoluted allele specificity by assigning allele-specific motifs using shared ligands between samples with shared MHC alleles. MARIA-I allows learning from heterozygous samples (e.g. primary tumor samples) to better understand the landscape of physiological MHC presentation beyond mono-allelic cell lines. The results indicate that MARIA-I outperforms existing methods, including netMHCpan4.0, MixMHCpred, and MHCflurry, in diverse tissue types. As MS techniques continue to advance, continued dissemination of peptide presentation data will serve as a cornucopia to improve MHC-I antigen prediction.
Applying MARIA-1 to T-cell epitope prediction resulted in improved presentation performance associated with antigenicity. MARIA-I scores of neoantigens were significantly correlated with their likelihood of eliciting CD8+ T cell responses in ex-vivo assays. However, a key hurdle to building a machine learning algorithm for T-cell epitope prediction is the lack of appropriate negative sets. Appropriate negative training examples are peptides with good presentation properties that do not elicit a T-cell response, which avoids building a model heavily relying on presentation as the main feature.
Training data: MCL MHC-A, B, and C ligandomes were retrieved from tumors of 17 patients and two cell lines (JEKO, L128). Publicly available peptide data (n=18,288, 104,966, 47,023), from a diverse set of samples including cell lines derived from B-cell leukemia, basal like breast cancer, and colon carcinoma as well as melanoma tumor tissue, were also included in training.
Independent validation data: SK-OV-3 ovarian cancer cell line ligands (n=1,377). 12 peptidomes identified from meningioma samples were obtained from a previous study (n=45,371,). A recent study profiled 11 tumors with known MHC alleles and exposed 6 of those to additional IFN-gamma treatment, for a total of 17 tumor peptidomes evaluated in this work (n=163,128).
In both training and validation data, peptide:HLA-I complexes were purified using anti-b2M W6/32 antibodies.
Patients' MHC alleles were identified by analyzing tumor exome sequencing of MCL samples through the PHLAT bioinformatics pipeline. Given the wide range of HLA-I alleles and their selective presentation patterns, analyses were restricted to peptide data with known HLA alleles.
Amino acids were coded using “one-hot encoding.” Each residue was represented by a vector of length 21, allowing 20 amino acids and 1 separator. Each vector is a list of zeros and a one in a specific position indicating the identity of the amino acid. This method provides internally consistent map of amino acids and allows peptide sequences to be represented by a matrix of size X,21, with X indicating the number of amino acids in the sequence.
Gene expression profiles for MCL patients and JeKo-1 cell line were obtained from RNA-sequencing results. The gene expression dictionary for MCL was created by using the median gene expression for each gene across the patients with MCL. Expression values were normalized to TPM for ease of comparison and simplifying user input for MARIA-I.
To ensure the model is robust against gene expression noise and does not rely on patient-specific gene expression profiles, tissue specific was use rather than patient specific RNASeq values (e.g. median TPM across a cohort of MCL patients). For estimated gene expression levels of tumor tissues, the median of TCGA RNA-sequencing results from the closest tissue type were used. For example, melanoma peptides were assigned median TCGA gene expression levels for skin cutaneous melanoma (SKCM). However, for cell lines, available expression data was downloaded from the ENCODE database.
The analyses show minor differences in presentation prediction performance when using tissue-matched gene expression profiles compared to unmatched gene expression levels.
Finally, for the analysis of candidate peptides in vaccine studies, a fixed TPM of 25 was used to ensure consistently high presentation, reflecting the readily available peptides for cells' antigen presentation machinery in a therapeutic context.
MARIA-I takes as input a peptide sequence (8-17mer), six patient MHC-A,B,C alleles, and Hugo Symbol gene names. Peptide sequence and MHC alleles are inputs to the binding affinity neural network, and peptide sequence and gene name are inputs to the cleavage score network. Gene names are used to obtain gene expression estimates based on an external TPM dictionary, either specified by the user or defaulting to the MCL dictionary. Peptide sequence is separately one-hot encoded into a 17×21 matrix, with padding and masking to ensure a consistent number of rows despite varying peptide length. The peptide sequence matrix is fed into a special RNN layer called LSTM, which can manage longer associations in sequential data. Normalized outputs from the binding affinity model, cleavage model, and gene expression dictionary are concatenated into a fixed 8×1 vector (6 binding affinities, 1 cleavage score, and 1 gene expression value). The LSTM output is concatenated with the fixed vector of scores of binding, cleavage, and gene expression and inputted into two fully connected dense layers separated by dropout layers. The output of these layers is connected to a softmax activation layer that outputs the likelihood of peptide presentation in the context of predicted binding affinity, cleavage scores, and gene expression level.
The same neural numbers (64 for LSTM, 32 for dense) and drop-out (0.35) were used as in Example 1. Very minor improvements in model performance were observed when doubling the neural number. Categorical cross-entropy is used as the loss function.
The binding affinity LSTM was trained on publicly available data of 185 k peptides (tools.immuneepitope.org/main/datasets/) curated by the NetMHCpan3.0 research team. IC50 scores were normalized to log scale using the following formula: 1-log 50,000*IC50. Each MHC allele was condensed into a 34-amino-acid sequence, termed a pseudosequence, representing key residues that mediate the MHC-peptide interaction. This interaction is one-hot encoded into a 52×21 matrix, with 52 as the maximum length of the combined pseudosequence (34 amino acids) and peptide (maximum of 17 amino acids). This matrix is fed into an LSTM layer followed by 2 dense layers (neuron number=32) and an output neuron. Mean squared error is used as the loss function and the output is the predicted binding affinity for a given allele-peptide pair.
This feature provides information about cleavage patterns around the peptide and its relationship to presentation. This pattern was defined as 6 amino acids upstream and downstream of the peptide. To create negative labeled data, length- and gene-matched random peptides were derived from the positive peptide. The cleavage data was one-hot encoded into a 12,21 matrix, with 12 corresponding to the combined number of upstream and downstream amino acids. This matrix was used as input into the cleavage score neural network, which processed the input and used a sigmoid activation function to output the probability that a peptide with the given flanking regions would be presented.
The MARIA-I softmax activation output provides an expected likelihood of presentation on a 0 to 1 scale. These values are not human-interpretable as they provide little context for the overall distribution of peptide presentation scores. Thus, in addition to the raw score, a percentile score was created to rank the presentation likelihood of a given sample amongst a background distribution of peptides. 10,000 random human peptides were generated for each kmer from 8 to 17, spanning the breadth of variation in peptide length in the MS training data, and calculated MARIA-I scores for each set of 10,000 with a random selection of 6 MHC alleles.
To create a random peptide, a gene is selected from a list of 26,013 genes with a probability distribution corresponding to the expression level across genes. Next, a random amino acid position is picked as the first amino acid to generate a peptide with the appropriate length. To prevent the generating random peptides that appear like identified positive peptides, the new peptide is not identical to any known MHC-I presented peptides.
To judge the generalizability of MARIA-I's performance, datasets of independent ligands were obtained, processed under a variety of MS methods in different research groups, and tested against existing methods for predicting class I antigen presentation. Peptides below 8 amino acids and above 17 amino acids were excluded from all analyses due to the model's constraints, eliminating a small minority of peptides. Random length-matched human peptides were generated for negative data.
When evaluating independent peptides using MARIA-I and NetMHCpan4.0, MixMHCpred2.0, and MHCflurry, all peptides below 8 amino acids and above 14 amino acids were excluded due to model constraints.
For NetMHCpan4.0, six predictions were generated for each peptide, one for each MHC allele, and the allele with the lowest rank (highest likelihood of presentation) was chosen as the predictor's output for the likelihood of peptide presentation.
MixMHCpred2.0 automatically runs all allele predictions for each peptide. For each sample's peptides, all peptides and all six MHC alleles were input. In the output file, for each peptide, the “%Rank_bestAllele” was selected as MixMHCpred2.0's likelihood of presentation.
Similarly, MHCflurry runs all allele predictions for each peptide. All peptides and MHC alleles were input and, for each peptide, the “mhcflurry_prediction_percentile” column value was selected as the MHCflurry's likelihood of presentation. MHCflurry was excluded from the analysis of the held-out 17 tumor peptidomes due to lack of support for several MHC alleles.
Analysis of T Cell Responses with MARIA-I Scores
Each somatic mutation tested in the two T cell response studies was represented as a 17mer sequence, sampling 8 amino acids upstream and downstream of the mutated residue. Within that sequence, all 9-amino-acid sliding windows were generated from each 17mer, and produced corresponding MARIA-I scores. 25 TPM was used to reflect the high concentration present in a pulsed peptide assay. Each mutation was assigned the highest MARIA-I score calculated from its sub-sampled peptides (a 9mer window). Each peptide was then assigned a presentation category, low, medium, or high based on its maximum MARIA-I percentile score: less than 95th percentile, 95th-99.5th percentile, or above 99.5th percentile.
Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) of a sensitivity vs. 1-specificity graph were implemented with Python scikit-learn.
Positive predictive value (PPV), or precision, was calculated as:
(Prevalence*Recall)/[(Prevalence*Recall+(1−Prevalence)*(1−Specificity)]
Most plots were generated using python packages, matplotlib and seaborn. Motif analysis plots, represented as amino acid frequencies, were generated using IceLogo.
Unless otherwise noted, p-values for differences in statistical distributions were calculated using Student's t-test. For
While the above description contains many specific embodiments, these should not be construed as limitations on the scope of the invention, but rather as an example of one embodiment thereof. Accordingly, the scope of the invention should be determined not by the embodiments illustrated, but by the appended claims and their equivalents.
This application claims priority to U.S. Provisional Application Ser. No. 62/880,566, entitled “Neural Network for Accurate Prediction of HLA Class II Antigen Presentation,” filed Jul. 30, 2019, which is incorporated herein by reference in its entirety.
This invention was made with Government support under contract CA 194389 awarded by the National Institutes of Health. The Government has certain rights in the invention.
| Number | Date | Country | |
|---|---|---|---|
| 62880566 | Jul 2019 | US |
