Determining predicted values of protein-protein interactions in t cell receptor-major histocompatibility complex-peptide complexes

The system generates an ensemble of candidate structures for TCR-pMHC and pMHC complexes, addressing the challenge of predicting protein-protein interactions by extracting structural features, improving accuracy and resource efficiency.

WO2026241707A1PCT designated stage Publication Date: 2026-11-19MEMORIAL SLOAN KETTERING CANCER CENT +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/028481
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-06-20
Filing Date
2026-05-15
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing methods struggle to accurately predict protein-protein interactions in T-cell receptor (TCR)-major histocompatibility complex (MHC)-peptide complexes, particularly for novel sequences or subtle sequence changes, leading to poor generalization and inability to distinguish between binding and non-binding complexes.

Method used

A computing system generates an ensemble of candidate quaternary structures for TCR-pMHC and pMHC complexes, extracting features like confidence metrics and RMSD to determine interaction probabilities, reducing reliance on single structures and sequence similarity.

Benefits of technology

This approach enhances generalization to unseen complexes, improves discrimination of minute structural differences, and conserves computational resources by using a lighter-weight classifier.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026028481_19112026_PF_FP_ABST
    Figure US2026028481_19112026_PF_FP_ABST
Patent Text Reader

Abstract

Presented herein are systems and methods for determining values indicating probability of protein-protein interactions in T-cell receptor (TCR)-peptide-major histocompatibility complex (MHC) (TCR:pMHC) complexes. A computing system may retrieve sequences corresponding to an MHC polypeptide, a peptide epitope, a TCR-a polypeptide, and a TCR-P polypeptide, and generate a plurality of candidate quaternary structures for a TCR-peptide-MHC complex. The computing system may extract, for at least one candidate quaternary structure, a plurality of features including two or more of a confidence metric, a structure metric, or a root mean square deviation (RMSD) metric. The features may be provided to a classifier model to determine at least one value indicating interaction probability for the TCR:pMHC complex.
Need to check novelty before this filing date? Find Prior Art

Description

Atty. Dkt. No.: 115872-3282DETERMINING PREDICTED VALUES OF PROTEIN-PROTEIN INTERACTIONS IN T CELL RECEPTOR-MAJOR HISTOCOMPATIBILITY COMPLEX-PEPTIDE COMPLEXES CROSS REFERENCE TO RELATED APPLICATIONS

[0001] The present application claims the benefit of and priority to U.S. Provisional Patent Application No. 63 / 807,013, filed May 16, 2025, and U.S. Provisional Patent Application No. 63 / 827,804, filed June 20, 2025, each of which is incorporated by reference in their entireties.TECHNICAL FIELD

[0002] The present application relates generally to the field of using machine learning (ML) models to determine values indicating probabilities of protein-protein interactions in T-cell receptor (TCR)-major histocompatibility complex (MHC)-peptide complexes or peptide-MHC complexes.BACKGROUND

[0003] Adaptive immunity depends on T-cell receptors (TCRs) recognizing short peptide antigens displayed by major histocompatibility complex (MHC) molecules on the cell surface. Predicting whether a given peptide will be presented by a given MHC and become a peptide-MHC (pMHC) complex, and whether a given TCR will bind to pMHC to form a TCR-peptide-MHC (TCR:pMHC) may be difficult, as even slight changes in the protein sequences can result in a complex that binds or does not bind.SUMMARY

[0004] Aspects of the present disclosure are directed to systems and methods for determining values indicating probabilities of protein-protein interactions in peptide-major histocompatibility complex (MHC) (pMHC) complexes. One or more processors may retrieve (i) a first sequence corresponding to an MHC polypeptide and (ii) a second sequence corresponding to a peptide epitope that binds to the MHC polypeptide to form a-1- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282pMHC complex. The one or more processors may create, using the first sequence and the second sequence, a plurality of candidate quaternary structures, each of the plurality of candidate quaternary structures comprising (i) a respective MHC structure and (ii) a respective peptide epitope structure for forming the pMHC complex. The one or more processors may generate, for at least one candidate quaternary structure of the plurality of candidate quaternary structures, a plurality of features comprising two or more of (i) a confidence metric, (ii) a structure metric, or (iii) a root mean square deviation (RMSD) metric. The one or more processors may provide the plurality of features as input to a classifier model to determine a value indicating a probability of interaction between the MHC and the peptide epitope in the pMHC complex. The one or more processors may store, using one or more data structures, an association between the respective value and the pMHC complex.(0005] In some embodiments, the one or more processors may determine that the pMHC complex is non-binding based on the value not satisfying a threshold value; and exclude the pMHC complex from provision. In some embodiments, the one or more processors may determine that the pMHC complex is binding based on the value satisfying a threshold value; and select, from a plurality of candidate pMHC complexes, the pMHC complex for provision. In some embodiments, the pMHC complex may be generated using an MHC and a peptide epitope synthesized in accordance with a plurality of sequences defining the pMHC complex selected from the plurality of candidate pMHC complexes.

[0006] In some embodiments, a binding affinity metric may be measured from the pMHC complex to validate as a binding complex, wherein the binding affinity comprises at least one of a monovalent affinity metric or an effective affinity metric. In some embodiments, a subject at risk of cancer or diagnosed with cancer may be administered with neoantigen vaccine comprising the peptide epitope.

[0007] In some embodiments, the cancer may include at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney -2- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof. In some embodiments, the first sequence may correspond to at least one of a wildtype MHC polypeptide or a mutant type MHC polypeptide. The second sequence may correspond to a wildtype type peptide epitope or a mutant type peptide epitope. The first sequence and the second sequence may be derived from at least one mammal comprising one or more of a human or a mouse.

[0008] In some embodiments, the confidence metric may include at least one of (i) a predicted local distance difference test (pLDDT) metric, (ii) a predicted alignment error (PAE) metric, (iii) an interface predicted alignment error (iPAE), (iv) a predicted template modeling (pTM) metric, or (v) an interface predicted template modeling (ipTM) metric. In some embodiments, the structure metric may indicate a probability of contact between at least one amino acid residue of the MHC structure and at least one amino acid residue of the peptide epitope structure. In some embodiments, the structure metric may indicate a probability of contact between at least one amino acid residue of the MHC structure and at least one amino acid residue of the peptide epitope structure. In some embodiments, the RMSD metric may be determined based on a deviation between the candidate quaternary structure and at least one reference quaternary structure for the pMHC complex. In some embodiments, the RMSD metric may be determined based on an aggregation of a plurality of deviations, each of the plurality of deviations between two candidate quaternary structures of a plurality of candidate quaternary structures. In some embodiments, the plurality of features may include a plurality of embeddings generated based on executing a protein language model using a plurality of sequences corresponding to the pMHC complex.

[0009] In some embodiments, the one or more processors may retrieve, from one or more databases or from a sequencing device, (i) a first input file identifying the first sequence corresponding to the MHC polypeptide sequence and (ii) a second input file identifying the second sequence corresponding to the peptide epitope sequence; and-3- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282generate, using the first input file and the second input file, a plurality of output files corresponding to the plurality of candidate quaternary structures.

[0010] In some embodiments, the classifier model may be established using a plurality of training examples, each of the plurality of training examples comprising: (i) a respective example plurality of features comprising two or more of an example confidence metric, an example structure metric, or an example RMSD metric, for an example pMHC complex, and (ii) a respective label indicating the example pMHC complex as one of binding or non-binding. In some embodiments, the one or more processors may execute a folding model using the first sequence and the second sequence to create the plurality of candidate quaternary structures; generate, based on executing the folding model, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective score indicating a respective accuracy of the candidate quaternary structure; and select, from the plurality of candidate quaternary structures, the at least one candidate quaternary structure based on the respective score indicating the respective accuracy. In some embodiments, the folding model comprises at least one of AlphaFold, RoseTTAFold, OpenFold, Boltz-1, OmegaFold, orESMFold.

[0011] In some embodiments, the one or more processors may determine, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective plurality of features comprising two or more: (i) a respective confidence metric, (ii) a respective structure metric, (iii) a respective root mean square deviation (RMSD) metric, or (iv) a plurality of embeddings; generate the plurality of features based on an aggregation of the respective plurality of features for each candidate quaternary structure of the plurality of candidate quaternary structures; and provide the plurality of features to determine the value indicating the indicating the probability of interaction within the pMHC complex.

[0012] In some embodiments, an MHC corresponding to the first sequence comprises at least one of an MHC class I or an MHC class II. The MHC polypeptide may include at least one of HLA-A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP. The MHC class I may include an alpha chain including al domain, a2 domain, and a3 domain, and optionally a P2-microglobulin chain. The MHC class II may include an alpha chain -4- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282having al domain and a2 domain and optionally a beta chain having an pi domain and P2 domain. In some embodiments, the peptide epitope corresponding to the second sequence originates from at least one of a pathogen, a self-antigen, or cancer in a subject. The peptide epitope may include at least one of a class I epitope or a class II epitope, and wherein the class I epitope has a length of 8-11 amino acids, and wherein the class II epitope has a length of 12-25 amino acids.

[0013] Aspects of the present disclosure are directed to systems and methods for determining values indicating probability of protein-protein interactions in T-cell receptor (TCR)-peptide-major histocompatibility complex (MHC) (TCR:pMHC) complexes. One or more processors coupled with memory may retrieve (i) a first sequence corresponding to an MHC polypeptide and (ii) a second sequence corresponding to a peptide epitope that binds to the MHC polypeptide to form an peptide-MHC (pMHC) complex; (iii) a third sequence corresponding to a TCR-a polypeptide and (iv) a fourth sequence corresponding to a TCR-P polypeptide that binds to the TCR-a polypeptide to form a TCR, wherein the TCR binds with the pMHC complex to form a TCR-peptide-MHC complex. The one or more processors may create, using the first sequence, the second sequence, the third sequence, and the fourth sequence, a plurality of candidate quaternary structures for the to form the TCR-peptide-MHC complex, each of the plurality of quaternary structures comprising: (i) a respective MHC structure, (ii) a respective peptide epitope structure, (iii) a respective TCR-a structure, and (iv) a respective TCR-P structure. The one or more processors may generate, for at least one candidate quaternary structure of the plurality of candidate quaternary structures, a plurality of features comprising two or more of a confidence metric, a structure metric, or a root square mean deviation (RMSD) metric. The one or more processors may determine, based on providing the plurality of features as input to a classifier model, at least one value indicating interaction among the MHC structure, the peptide epitope structure, the TCR-a structure, and the TCR-P structure of the at least one candidate quaternary structure for the TCR:pMHC complex. The one or more processors may store, using one or more data structures, an association between the value and the at least one candidate quaternary structure for the TCR:pMHC complex.-5- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0014] In some embodiments, the one or more processors may determine that the TCR:pMHC complex is non-binding based on the value not satisfying a threshold value; and exclude the TCR:pMHC complex from provision. In some embodiments, the one or more processors may determine that the TCR:pMHC complex is binding, based on the value satisfying a threshold; and select, from a plurality of candidate TCR:pMHC complexes, the at least one TCR:pMHC complex for provision. In some embodiments, the TCR:pMHC complex may be generated using a peptide epitope, an MHC molecule, and a TCR synthesized in accordance with a plurality sequences defining the TCR:pMHC complex selected from the plurality of candidate TCR:pMHC complexes. In some embodiments, a binding affinity metric may be measured from the TCR:pMHC complex to validate as a binding complex, wherein the binding affinity metric comprises at least one of a monovalent affinity or effective affinity. In some embodiments, a subject at risk of cancer or diagnosed with cancer may be administered with neoantigen vaccine comprising the peptide epitope corresponding to the peptide epitope structure of the TCR:pMHC complex.

[0015] In some embodiments, the cancer may include at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof.

[0016] In some embodiments, the first sequence may correspond to a wildtype MHC polypeptide sequence. The second sequence may correspond to a wildtype type peptide epitope sequence or a mutant type peptide epitope sequence. The first sequence, the second sequence, the third sequence, or the fourth sequence may be derived from at least one mammal comprising one or more of a human or a mouse. In some embodiments, the -6- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282confidence metric may include at least one of (i) a predicted local distance difference test (pLDDT) metric, (ii) a predicted alignment error (PAE) metric, (iii) an interface predicted alignment error (iPAE), (iv) a predicted template modeling (pTM) metric, or (v) an interface predicted template modeling (ipTM) metric.

[0017] In some embodiments, the structure metric indicates a probability of contact between two or more of (i) at least one MHC amino acid residue of the MHC structure, (ii) at least one peptide epitope amino acid residue of the peptide epitope structure, (iii) at least one TCR-a amino acid residue of the TCR-a structure, (iv) at least one TCR- P amino acid residue of the TCR-P structure, (v) at least one amino acid residue in the Complementarity Determining Region 3 (CDR3) portion of the TCR-a structure, or (vi) at least one amino acid residue in the CDR3 portion TCR-P structure. In some embodiments, the RMSD metric may be determined based on a deviation between the candidate quaternary structure and at least one reference quaternary structure for the pMHC complex. In some embodiments, the RMSD metric may be determined based on an aggregation of a plurality of deviations, each of the plurality of deviations between two candidate quaternary structures of a plurality of candidate quaternary structures. In some embodiments, the plurality of features may include a plurality of embeddings generated based on executing a protein language model using a plurality of sequences corresponding to the TCR:pMHC complex.

[0018] In some embodiments, the one or more processors may retrieve, from one or more databases or from a sequencing device, (i) a first input file identifying the first sequence corresponding to the MHC polypeptide, (ii) a second input file identifying the second sequence corresponding to the peptide epitope, (iii) a third input file identifying the third sequence corresponding to the TCR-a polypeptide and (iv) a fourth input file identifying the fourth sequence corresponding to the TCR-P polypeptide. The one or more processors may generate, using the first input file, the second input file, the third input file, and the fourth input file, a plurality of output files corresponding to the plurality of candidate quaternary structures.

[0019] In some embodiments, the classifier model may be established using a plurality of training examples, each of the plurality of training examples comprising: (i) a -7- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282respective example plurality of features comprising two or more of a respective example confidence metric, a respective example structure metric, a respective example root mean square deviation (RMSD) metric, for an example TCR:pMHC complex, and (ii) a respective label indicating the example TCR:pMHC complex as one of binding or non-binding.

[0020] In some embodiments, the one or more processors may execute a folding model using the first sequence, the second sequence, the third sequence, and the fourth sequence to create the plurality of candidate quaternary structures. The one or more processors may generate, based on executing the folding model, for each candidate TCR:pMHC complex a plurality of candidate quaternary structures, each with a respective score indicating a respective accuracy of the candidate quaternary structure as determined by the folding model. The one or more processors may select, from the plurality of candidate quaternary structures, the at least one candidate quaternary structure based on the respective score indicating the respective accuracy. In some embodiments, the folding model may include at least one of AlphaFold, RoseTTAFold, OpenFold, Boltz-1, OmegaFold, orESMFold.

[0021] In some embodiments, the one or more processors may determine, by the one or more processors, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective plurality of features comprising two or more: (i) a respective confidence metric, (ii) a respective structure metric, (iii) a respective root mean square deviation (RMSD) metric, or (v) a respective plurality of embeddings. The one or more processors may generate the plurality of features based on an aggregation of the respective plurality of features for each candidate quaternary structure of the plurality of candidate quaternary structures. The one or more processors may provide the input to the classifier model to determine the value indicating the probability of interaction within the TCR:pMHC quaternary structure.

[0022] In some embodiments, the TCR may include a TCRa polypeptide and a TCRP polypeptide that are isolated from at least one of CD8+ T cell or a CD4+ T cell, and the third sequence corresponding to TCR-a polypeptide and the fourth sequence corresponding to the TCR-P polypeptide are determined by next-generation sequencing. In some embodiments, an MHC corresponding to the first sequence comprises at least one of -8- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282an MHC class I or an MHC class II. The MHC polypeptide may include at least one of HLA-A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP. The MHC class I may include an alpha chain including al domain, a2 domain, and a3 domain, and optionally a P2-microglobulin chain. The MHC class II may include an alpha chain having al domain and a2 domain and optionally a beta chain having an pi domain and P2 domain. In some embodiments, the peptide epitope corresponding to the second sequence may originate from at least one of a pathogen, a self-antigen, or cancer in a subject. In some embodiments, the peptide epitope may include at least one of a class I epitope or a class II epitope, and wherein the class I epitope has a length of 8-11 amino acids, and wherein the class II epitope has a length of 12-25 amino acids.BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The foregoing and other objects, aspects, features, and advantages of the disclosure will become more apparent and better understood by referring to the following description taken in conjunction with the accompanying drawings, in which:

[0024] FIGs. 1A-1E: enFoldX architecture for structure-based prediction of TCR-pMHC structures. FIG. 1A, Schematic of enFoldX implementing at scale prediction of TCR-pMHC complex structures and extraction of features predictive of binding. FIGs. 1B-D, Predicted AF3 structures for FIG. IB, an experimentally validated binder and FIGs. 1C and ID, two non-binder complexes, colored by pLDDT, along with the corresponding PAE matrices. The V3D non-binder structure (FIG. 1C) is predicted in the canonical docking pose, while AF3 predicts the complex for non-binder V3E with TCR misdocked on the pMHC complex (FIG. ID). PAE matrices illustrate chain-specific PAE ( TCRa) and chain-pair iPAEs (peptide and TCRP; MHC and TCRa). Means of the PAE submatrices are annotated in corresponding colors and illustrated in the bar plot in FIG. IE, Higher iPAEs suggest lack of confidence in the predicted structure and may potentially reveal unfavorable binding.

[0025] FIGs. 2A-2F: Performance on the Neoantigen Task. FIG. 2A, Crossreactivity dataset with T cell reactivity to mutated peptides estimated by EC50 from activation experiments. One example TCR is shown. (For remaining TCRs, see FIG. 8).-9- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282FIG. 2B, Representative features related to confidence (ipTM, PAE, iPAE), contact probabilities, and peptide RMSD in the predicted structure with the mutant peptide versus the predicted structure with the corresponding wild type peptide (binder) are significantly different for binder and non-binder TCR-pMHC complexes. FIG. 2C, ROC curves for classification on cross-reactivity data using XGBoost or Ridge Regression using the top ranked AF3 structure for each complex (“Best structure”) or using the mean and standard deviation of features across all generated AF3 structures for each complex (“Structure ensemble”). 5-fold cross-validation split was performed, and ROC curves were averaged across folds. Average AUC and confidence intervals across the folds are shown. FIG. 2D, Average AUC for the Ridge models as a function of the number of structures sampled per complex using the best structure approach or the structure ensemble approach. The size of each dot reflects the standard deviation of AUC values across 5-folds. FIG. 2E, ROC curves for Ridge model using subsets of the feature space. Models use confidence features alone, contact probabilities alone (red), RMSD features alone, and the full set of features, with ROC curves and AUC values across the 5-folds as in FIG. 2C. FIG. 2F, Normalized model importances by permutation test for Ridge model. Top: Importance stratified by different types of features. Bottom Left: Importance stratified by substructure. Bottom Right: mean (pink) of features over structures in the ensemble vs. standard deviations (gray). *** p < 0.001 by two-sided t test.

[0026] FIGs. 3A-3G: Performance on predicting TCR-pMHC specificity for data in VDJdb. FIG. 3A, ROC curves for classification of positive versus permutated negative TCR-pMHCs (“permutation task”) or positive versus decoy negative TCR-pMHCs (“decoy task”) in human VDJdb dataset. AUC values are shown as mean and confidence interval over folds using 5-fold cross-validation. FIG. 3B, Normalized feature importances by permutation test on permutation task for human VDJdb dataset. Left: feature importance stratified by feature type. Right: feature importances stratified by substructure.FIG. 3C, Same as a but for mouse dataset. FIG. 3D, Same as FIG. 3B but for mouse dataset. FIG. 3E, Normalized feature importances by permutation test on the decoy task for the mouse dataset, stratified by substructure, highlighting that pMHC substructure-related features, notably pLDDT of the peptide chain, contribute more to distinguishing decoy negatives from positives than distinguishing permuted negatives from positives (FIG. 3D).-10- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282FIG. 3F, Average AUC for the Ridge models on the decoy task on the human dataset as a function of the number of structures generated per complex using the best structure approach or the structure ensemble approach. The size of each dot reflects the standard deviation of AUC values across 5-folds. FIG. 3G, ROC curves for cross-species Ridge models. Models trained on human VDJdb TCR-pMHC complexes can accurately predict binding for mouse VDJdb TCR-pMHC complexes.

[0027] FIGs. 4A-4E: Sequence-based TCR-epitope specificity prediction. FIG.4A, BERTie architecture and input example. FIG. 4B, BERTie performance on held-out data. BERTie models were trained on 5 cross-validation splits of the training portion (80%) of datasets composed of VDJDB, McPAS and MIRA COVID-19 datasets and tested on the remaining 20% of these datasets. The test case consisted of equal proportions of 4 test scenarios: New pair of epitopes seen during training (seen epitope) and TCR seen during training (seen TCR); a pair of seen epitopes and unseen TCR; a pair of seen TCR and unseen epitope; and a pair of unseen (novel) TCR and unseen epitope. FIG. 4C, ROC curve for test examples with unseen epitopes split by similarity of test epitopes to those in the training dataset. FIG. 4D, Accuracy of BERTie predictions as a function of similarity of TCR and epitope sequences to those in training data, measured by Levenshtein distance to the most similar sequences in training data. FIG. 4E, Benchmarking of BERTie and other sequence models on IEDB dataset. Models that included IEDB as part of the training set are highlighted with *.

[0028] FIGs. 5A-5E: Generalizability of models trained on publicly available data to predict TCR specificity for cross-reactivity dataset. FIG. 5A, ROC curve for enFoldX-Ridge model for VDJdb human decoy task for new (unseen) epitopes stratified by epitope similarity to training data shows little dependence of the model’s predictive capacity on sequence similarity. FIG. 5B, Accuracy of the enFoldX-Ridge model on the VDJdb human decoy task by HLA type and number of examples. Size of the dot is proportional to the number of TCR-pMHC binder pairs for each HLA allele. There is no clear dependence of enFoldX’s performance on the number of training examples for alleles in HLA-A or HLA-B, but the lack of data for HLA-C overall may significantly affect performance for HLA-C restricted epitopes. FIG. 5C, enFoldX models trained on VDJdb human data following-11- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282both the decoy and the permuted negatives strategy show higher accuracy in classifying binders vs. non-binders in the cross-reactivity dataset (Neoantigen task) compared to BERTie and other benchmarked sequence models, as illustrated by the ROC curves.Sequence models were trained on various datasets, including VDJdb. FIG. 5D, UMAP embedding and Leiden clustering of enFoldX-Ridge’s feature space, clustered using feature values for the human VDJdb (training) dataset and the cross-reactivity (test) data set. FIG.5E, Normalized feature importance for enFoldX-Ridge model (trained on VDJdb human decoy task) in classifying binders vs. non-binders on the cross-reactivity data set overlaid onto the feature space from FIG. 5D.

[0029] FIG. 6: enFoldX Pipeline Architecture. The AF3 pipeline involves three distinct steps which was run sequentially: structure look-up via template search and multiple sequence alignment (top), structure prediction via diffusion (middle), and feature extraction from predicted structures and confidence metadata, which are used for classification (bottom). These steps was performed in a highly parallelized manner across all unique sequences, thus enabling tractable analysis of large datasets.

[0030] FIGs. 7A-7C: Re-alignment of chains to binder structure reveals a potential steric clash. FIG. 7A, TCR1 was modeled with the wild-type peptide and the V3E mutant. The V3E poses was then extracted for major complex components (epitope, MHC groove, and TCR a and P chains) and aligned each pose with its corresponding prediction for each NLV subunit and the overall NLV complex, FIG. 7B. No meaningful difference was observed between the wild-type and mutant epitope poses (all-atom RMSD = 0.2 A) and MHC grooves (backbone RMSD = 0.5 A) but the aligned TCR poses (backbone RMSD = 0.5 A) indicated greater variability, particularly in the CDR3 loops responsible for epitope recognition. FIG. 7C, Focusing on atoms in the peptide-CDR3p interface, the distances that was found between the epitope M5 methyl carbon atom and neighboring atoms in the CDR3P loop in the reconstructed V3E model are insufficient to accommodate the implicit M5 hydrogen atoms (2.0 A between M5-Gly Ca and 1.4 A between M5-Ala N). The driving force behind the mispositioned TCR in the V3E prediction may be strong steric clashes between residues in the epitope recognition region-12- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282and the CDR3P loop, which could explain experimental data demonstrating the V3E mutation induces loss of T cell activity.

[0031] FIG. 8: Cross-reactivity data for the Neoantigen Task. Full crossreactivity dataset with T cell reactivity to mutated peptides estimated by EC50 from T cell activation experiments. Reactivity is set according to the EC50 threshold as described in the original publication18.

[0032] FIGs. 9A-9E: Detailed Performance on the Neoantigen Task. FIG. 9A, enFoldX model performance on cross-reactivity dataset for “Leave-One-TCR out”. FIG.9B, enFoldX model performance on cross-reactivity dataset for “Leave-One-Epitope out”.FIG. 9C, enFoldX model performance on cross-reactivity dataset for “Leave-One-Position out”. FIG. 9D, Comparison of ROC curves for Logistic regression models with different regularization strategies and XGBoost model for cross-reactivity dataset trained on “best structure” with 5-fold cross-validation. FIG. 9E, Comparison of ROC curves for Logistic regression models with different regularization strategies and XGBoost model for crossreactivity dataset trained on “structure ensemble” with 5-fold cross-validation.

[0033] FIGs. 10A-10J: Additional Analysis of sequence-based specificity prediction model, BERTie. FIG. 10A, Performance for different BERTie encoding schemes on held-out test data. ROC curves for BERTie models encoding sequences using single amino acids, 2-mers, and 3-mers show slightly different performance, and the ensemble of models outperforms each individual sequence encoding approach. FIG. 10B, BERTie model performance on held-out test data by epitope scenario, as measured by precision-recall curve and average precision (AP). FIG. 10C, Per-epitope AUROC scores for BERTie and other benchmarked sequence models on the IEDB test set (epitopes with at least 2 cognate TCRs) shows lack of consistency in predictions among sequence models, including models trained on the IEDB data. FIG. 10D and FIG. 10E, BERTie per-epitope AUROC score is consistent among the 5 cross-validation folds for seen (FIG. 10D) and new (FIG. 10E) epitopes. Each dot is the per-epitope AUROC score of a BERTie model trained on one of the 5 cross-validation splits; the mean of the AUROC scores for the 5 CV folds is shown in orange. Only epitopes with at least 2 cognate TCRs are shown. FIGs.10F-H, Per-epitope AUROC score for seen epitopes is correlated with the number of -13- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282positive examples (“binder” TCRs) for that epitope in the training data (FIG. 10F), TCR motif score that detects a consistent motif in binder TCRs in training data for this epitope (FIG. 10G), and KL divergence between distribution of amino acids at each position in binder TCRs compared to a background, measuring how different the binder TCRs are from the rest of TCRs in training data (FIG. 10H). FIGs. 101 and 10J, Per-epitope AUROC score for new epitopes is negatively correlated with Levenshtein distance to the most similar epitope in the training data (FIG. 101) and positively correlated with the number of binder TCRs for each epitope included in the training data (although paired with other epitopes in the training data) (FIG. 10J).

[0034] FIGs. 11A-11P: enFoldX-Ridge Feature Space. FIG. 11 A, UMAP embedding and Leiden clustering of enFoldX-Ridge’s feature space, clustered using feature values for the human VDJdb decoy (training) dataset and the cross-reactivity (test) data set.FIG. 11B, Normalized feature importance for enFoldX-Ridge model (trained on VDJdb human decoy task) in classifying binders vs. non-binders on the cross-reactivity data set overlaid onto the feature space from FIG. 11 A. C, enFoldX’s features stratified by substructure of the TCR-pMHC complex overlaid onto the feature space from FIG. 11 A. FIG. 11D, enFoldX’s features stratified by method of aggregation across multisample draw from AlphaFold3’s generative model overlaid onto the feature space from FIG. 11 A. FIG. HE, Normalized feature importance for enFoldX-Ridge model (trained on VDJdb human decoy task) in classifying binders vs. non-binders on the cross-reactivity stratified by Leiden cluster from FIG. 11 A. FIG. HF, enFoldX’s features stratified by feature type overlaid onto the feature space from FIG. 11 A. FIG. 11G, Composition of feature types that make up the Leiden clusters from FIG. 11 A. FIG. 11H, Normalized feature importance by feature type. FIGs. 11I-11P, same as FIGs. 11A-H but for the enFoldX trained with human VDJdb permutation data.

[0035] FIG. 12: Comparison of enFoldX classifier architectures for generalization task. Model performance for enFoldX-Ridge regression vs. enFoldX-XGBoost trained on publicly available data in VDJdb human + decoy negatives on binder classification in the cross-reactivity data.-14- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0036] FIG. 13: a receiver operating characteristic (ROC) curve using Boltz-2 steering potentials. The graph includes curves for classifier performance using (i) all confidence-derived features, (ii) PAE-only features, and (iii) a reduced PAE feature set including minimum and standard deviation statistics aggregated across predicted structural samples. This graph demonstrates the applicability of the EnfoldX to models other than AlphaFold3.

[0037] FIG. 14 depicts a block diagram of a system for determining values predicting probabilities of protein-protein interactions in T-cell receptor (TCR)-major histocompatibility complex (MHC)-peptide complexes or peptide-MHC complexes, in accordance with an illustrative embodiment.

[0038] FIGs. 15A and 15B depict block diagrams of a process for training classifier models in the system for determining values predicting probabilities of protein-protein interactions in accordance with an illustrative embodiment.

[0039] FIG. 16 depicts a block diagram of a process for creating candidate complex structures in the system for determining values predicting probabilities of protein-protein interactions in accordance with an illustrative embodiment.

[0040] FIG. 17 depicts a block diagram of a process for executing classifier models in the system for determining values predicting probabilities of protein-protein interactions in accordance with an illustrative embodiment.

[0041] FIG. 18 depicts a block diagram of a process for providing outputs in the system for determining values predicting probabilities of protein-protein interactions in accordance with an illustrative embodiment.

[0042] FIG. 19 depicts a flow diagram of a method of determining values predicting probabilities of protein-protein interactions in T-cell receptor (TCR)-major histocompatibility complex (MHC)-peptide complexes in accordance with an illustrative embodiment.-15- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0043] FIG. 20 depicts a block diagram of a computing environment according to an example implementation of the present disclosure.DETAILED DESCRIPTION

[0044] It is to be appreciated that certain aspects, modes, embodiments, variations and features of the present methods are described below in various levels of detail in order to provide a substantial understanding of the present technology. It is to be understood that the present disclosure is not limited to particular uses, methods, reagents, compounds, compositions or biological systems, which can, of course, vary. It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only, and is not intended to be limiting.

[0045] Adaptive immunity may rely on the ability of T-cell receptors (TCRs) to recognize peptide antigens presented by major histocompatibility complex (MHC) molecules on the cell surface. Computational prediction of such interactions may have applications in immunology, including identifying pathogen-specific immune responses, self- or non-self-discrimination, autoimmunity, selecting neoantigens for cancer vaccines, and developing T-cell-based therapies, among others. Accurately predicting whether a peptide will bind an MHC molecule and whether a TCR will bind with the peptide-MHC complex may be technically challenging, particularly in cases involving previously unseen sequences or subtle sequence changes. For one, predicting TCR specificity for novel peptides may be inherently difficult, because TCR binding is highly diverse and structurally variable. For another, the binding may depend strongly on local interfaces, such as the TCR-peptide interface, TCR-MHC interface, and chain-pair relationships, among others.

[0046] Current approaches for predicting binding probability may suffer from poor generalization to novel TCRs or peptides and strong dependence on sequence similarity, and thus exhibit inability to distinguish between binding and non-binding complexes even with slight changes (e.g., when peptides differ by one amino acid). Under sequence-based approaches, a language model may be used to infer binding probabilities from amino acid sequence information. These methods, however, may depend strongly on sequence similarity and can show substantially reduced performance when applied to novel epitopes -16- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282or TCRs that were not present during training. These models may also fail to capture whether the substitution meaningfully alters recognition.

[0047] In addition, under structure-based approaches, a three-dimensional arrangement of the relevant biomolecules, for example by co-folding or docking TCRs and peptide-MHC complexes may be modeled to predict binding accuracies. These methods can provide information that is more directly related to molecular recognition than sequence similarity alone. Reliance on single predicted structure, however, can be misleading because a structure predictor may generate a plausible bound-like conformation even for a non-binding complex. These methods also may struggle to generalize to out-of-distribution data, such as unseen epitopes, unseen TCRs, or data from a different species.

[0048] To address these and other technical challenges, a prediction binding model may be trained to predict binding probabilities of TCR-MHC-peptide or peptide-MHC complexes using information about the local interfaces and structures of the complexes. A computing system may collect sequences for components of the complex, including the TCR a chain, TCR P chain, the peptide epitope, and the MHC molecule. The sequences may be collected from a database or read from a sample from a subject. The computing system may perform template search and multiple sequence alignment (MSA) preprocessing for the unique protein chains, particularly the TCR and MHC chains, in a parallelized manner so that repeated chains across many candidate complexes need not be processed redundantly. Using the collected sequences, the computing system may apply a folding model (e.g., AlphaFold3) to generate multiple structural predictions for each candidate complex.

[0049] For each predicted structure, the computing system may extract a set of features, including confidence metrics, structural metrics, and root mean square deviation (RMSD) metrics. The confidence metrics may be for the entire complex and for individual chains, and may include predicted local distance difference test (pLDDT) metric, a predicted alignment error (PAE) metric, an interface predicted alignment error (iPAE) metric, a predicted template modeling (pTM) metric, or an interface precited template modeling score (ipTM) metric, among others. The structure metric may indicate a probability that residues across chains lie within a contact threshold. The RMSD metric -17- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282may be between residues of the predicted structure and a reference structure (e.g., by comparing a predicted mutant complex to a corresponding predicted wild-type complex) or a pair-wise comparison among the predicted structures.

[0050] With the extraction, the computing system may aggregate the set of features across the predicted structure for the complex. For example, the computing system may combine corresponding features across all predicted structures for the complex, by calculating the mean and standard deviation of each feature across the ensemble. Based on the combination, the computing system may generate a single aggregated set of features that captures both the central tendency and the variability of the predicted structural behavior. The computing system may apply the prediction binding model to the aggregated set of features to output a value indicating a binding probability in the complex. When the complex is predicted to be non-binding, the computing system may discard the candidate structures from the complex, and repeat the process with another candidate complex. When the complex is predicted to be binding, the complex may be used in any number of applications. For example, the peptide in the complex determined to binding may be included in a neoantigen vaccine, and the neoantigen vaccine may be administered to a subject with cancer, from which samples were taken for the MHC and TCR cells.

[0051] In this manner, this framework may shift TCR specificity prediction away from pure sequence similarity and away from reliance on a single folded complex. Rather than relying on a single predicted structure, the computing system may produce an ensemble of candidate structures for the same complex. The ensemble may capture structural uncertainty and interface variability that are not reflected in sequence similarity alone. Furthermore, by extracting interface and structural metrics, the computing system may capture information that are the physical determinants of whether the complex will bind. Overall, the framework presented herein may support better generalization to unseen complexes and finer discrimination of structures with minute distinctions that otherwise would not be apparent in other approaches. This framework may thus conserve computing resources (e.g., processor and memory) that otherwise would have been wasted in less accurate and discriminative approaches. The binding prediction model may also be lighter--18- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282weight supervised classifier, compared to language models or structural-based models, further saving computational resources.Definitions

[0052] Unless defined otherwise, all technical and scientific terms used herein generally have the same meaning as commonly understood by one of ordinary skill in the art to which this technology belongs. As used in this specification and the appended claims, the singular forms “a”, “an” and “the” include plural referents unless the content clearly dictates otherwise. For example, reference to “a cell” includes a combination of two or more cells, and the like. Generally, the nomenclature used herein and the laboratory procedures in cell culture, molecular genetics, organic chemistry, analytical chemistry and nucleic acid chemistry and hybridization described below are those well-known and commonly employed in the art.

[0053] Unless defined otherwise, all technical and scientific terms used herein generally have the same meaning as commonly understood by one of ordinary skill in the art to which this technology belongs. As used in this specification and the appended claims, the singular forms “a”, “an” and “the” include plural referents unless the content clearly dictates otherwise. For example, reference to “a cell” includes a combination of two or more cells, and the like. Generally, the nomenclature used herein and the laboratory procedures in cell culture, molecular genetics, organic chemistry, analytical chemistry and nucleic acid chemistry and hybridization described below are those well-known and commonly employed in the art.

[0054] As used herein, the term “about” in reference to a number is generally taken to include numbers that fall within a range of 1%, 5%, or 10% in either direction (greater than or less than) of the number unless otherwise stated or otherwise evident from the context (except where such number would be less than 0% or exceed 100% of a possible value).

[0055] As used herein, the “administration” of an agent or drug to a subject includes any route of introducing or delivering to a subject a compound to perform its intended function. Administration can be carried out by any suitable route, including but -19- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282not limited to, orally, intranasally, parenterally (intravenously, intramuscularly, intraperitoneally, or subcutaneously), rectally, intrathecally, intratumorally or topically. Administration includes self-administration and the administration by another.

[0056] The terms “cancer” or “tumor” are used interchangeably and refer to the presence of cells possessing characteristics typical of cancer-causing cells, such as uncontrolled proliferation, immortality, metastatic potential, rapid growth and proliferation rate, and certain characteristic morphological features. Cancer cells are often in the form of a tumor, but such cells can exist alone within an animal, or can be a non-tumorigenic cancer cell. As used herein, the term “cancer” includes premalignant, as well as malignant cancers. In some embodiments, the cancer is bladder cancer, breast cancer, colorectal cancer, esophagogastric cancer, gynecological cancer (e.g., uterine cancer, cervical cancer, ovarian cancer), head and neck cancer, hepatobiliary cancer, high-grade glioma, low-grade glioma, lung cancer, melanoma, pancreatic cancer, prostate cancer, renal cancer, or soft tissue sarcoma. In some embodiments, the cancer is lung cancer (e.g., non-small cell lung cancer (NSCLC) such as adenocarcinoma, squamous cell carcinoma, or large cell carcinoma).

[0057] As used herein, the term “effective amount” refers to a quantity sufficient to achieve a desired therapeutic and / or prophylactic effect, e.g., an amount which results in the prevention of, or a decrease in a disease or condition described herein or one or more signs or symptoms associated with a disease or condition described herein. In the context of therapeutic or prophylactic applications, the amount of a composition administered to the subject will vary depending on the composition, the degree, type, and severity of the disease and on the characteristics of the individual, such as general health, age, sex, body weight and tolerance to drugs. The skilled artisan will be able to determine appropriate dosages depending on these and other factors. The compositions can also be administered in combination with one or more additional therapeutic compounds. In the methods described herein, the therapeutic compositions may be administered to a subject having one or more signs or symptoms of a disease or condition described herein. As used herein, a “therapeutically effective amount” of a composition refers to composition levels in which the physiological effects of a disease or condition are ameliorated or eliminated. A therapeutically effective amount can be given in one or more administrations.-20- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0058] Gene” as used herein refers to a DNA sequence that comprises regulatory and coding sequences necessary for the production of an RNA, which may have a noncoding function (e.g., a ribosomal or transfer RNA) or which may include a polypeptide or a polypeptide precursor. The RNA or polypeptide may be encoded by a full length coding sequence or by any portion of the coding sequence so long as the desired activity or function is retained. Although a sequence of the nucleic acids may be shown in the form of DNA, a person of ordinary skill in the art recognizes that the corresponding RNA sequence will have a similar sequence with the thymine being replaced by uracil, i.e., “T” is replaced with “U ”

[0059] “Next-generation sequencing or NGS” as used herein, refers to any sequencing method that determines the nucleotide sequence of either individual nucleic acid molecules (e.g., in single molecule sequencing) or clonally expanded proxies for individual nucleic acid molecules in a high throughput parallel fashion (e.g., greater than 103, 104, 105or more molecules are sequenced simultaneously). In one embodiment, the relative abundance of the nucleic acid species in the library can be estimated by counting the relative number of occurrences of their cognate sequences in the data generated by the sequencing experiment. Next generation sequencing methods are known in the art, and are described, e.g., in Metzker, M. Nature Biotechnology Reviews 11 :31-46 (2010).

[0060] As used herein, a “sample” refers to a substance that is being assayed for the presence of a nucleic acid of interest. Processing methods to release or otherwise make available a nucleic acid for detection are well known in the art and may include steps of nucleic acid manipulation. A biological sample may be a body fluid or a tissue sample. In some cases, a biological sample may consist of or comprise of blood, plasma, sera, urine, feces, epidermal sample, vaginal sample, skin sample, cheek swab, sperm, amniotic fluid, cultured cells, bone marrow sample, tumor biopsies, aspirate and / or chorionic villi, cultured cells, and the like. Fresh, fixed or frozen tissues may also be used.

[0061] As used herein, the terms “subject”, “patient”, or “individual” can be an individual organism, a vertebrate, a mammal, or a human. In some embodiments, the subject, patient or individual is a human.-21- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0062] As used herein, the term “therapeutic agent” is intended to mean a compound that, when present in an effective amount, produces a desired therapeutic effect on a subject in need thereof.

[0063] “Treating” or “treatment” as used herein covers the treatment of a disease or disorder described herein, in a subject, such as a human, and includes: (i) inhibiting a disease or disorder, i.e., arresting its development; (ii) relieving a disease or disorder, i.e., causing regression of the disorder; (iii) slowing progression of the disorder; and / or (iv) inhibiting, relieving, or slowing progression of one or more symptoms of the disease or disorder. In some embodiments, treatment means that the symptoms associated with the disease are, e.g., alleviated, reduced, cured, or placed in a state of remission.

[0064] As used herein, an “antigen” refers to a molecule to which an immunoglobulin-related composition (e.g., antibody or antigen binding fragment thereof or T Cell Receptor) can selectively bind. The target antigen may be a protein, carbohydrate, nucleic acid, lipid, hapten, or other naturally occurring or synthetic compound. In some embodiments, the target antigen may be a peptide / MHC complex. An antigen may also be administered to an animal to generate an immune response in the animal.

[0065] As used herein, the term “amino acid” refers to naturally occurring and non-naturally occurring amino acids, as well as amino acid analogs and amino acid mimetics that function in a manner similar to the naturally occurring amino acids. Naturally encoded amino acids are the 20 common amino acids (alanine, arginine, asparagine, aspartic acid, cysteine, glutamine, glutamic acid, glycine, histidine, isoleucine, leucine, lysine, methionine, phenylalanine, proline, serine, threonine, tryptophan, tyrosine, and valine) and pyrolysine and selenocysteine. Amino acid analogs refer to agents that have the same basic chemical structure as a naturally occurring amino acid, i.e., an a carbon that is bound to a hydrogen, a carboxyl group, an amino group, and an R group, such as, homoserine, norleucine, methionine sulfoxide, methionine methyl sulfonium. Such analogs have modified R groups (such as, norleucine) or modified peptide backbones, but retain the same basic chemical structure as a naturally occurring amino acid. In some embodiments, amino acids forming a polypeptide are in the D form. In some embodiments, the amino acids forming a polypeptide are in the L form. In some embodiments, a first plurality of amino -22- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282acids forming a polypeptide are in the D form, and a second plurality of amino acids are in the L form.

[0066] Amino acids are referred to herein by either their commonly known three letter symbols or by the one-letter symbols recommended by the IUPAC-IUB Biochemical Nomenclature Commission. Nucleotides, likewise, are referred to by their commonly accepted single-letter code.

[0067] As used herein, a “major histocompatibility complex” or “MHC” refers to specific cell-surface proteins that help the immune system recognize peptide fragments derived from pathogens or mutated proteins. MHC molecules are found in all higher vertebrates, and are also known as the human leukocyte antigen (HLA) system in humans. An MHC may comprise a class I MHC, a class II MHC, a non-classical MHC, a naturally occurring MHC, a recombinant MHC, an engineered MHC, a mutant MHC, a variant thereof, a fragment thereof, or any combination thereof. In some embodiments, an MHC may comprise one or more alpha chains, one or more beta chains, p2-microglobulin, or any domain, subunit, or associated component thereof. In some cases, an MHC may be an HLA allele, including HLA-A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP. An MHC may be represented by sequence information, structural information, predicted coordinates, confidence information, or any combination thereof. As used herein, “HLA” is the human form of “MHC”.

[0068] As used herein, the terms “MHC light chain” and “MHC heavy chain” refer to portions of the MHC molecule. Structurally, class I molecules are heterodimers comprised of two noncovalently bound polypeptide chains, a larger “heavy” chain (a) and a smaller “light” chain (P2-microglobulin or P2m). The polymorphic, polygenic heavy chain (45 kDa), encoded within the MHC on chromosome six, is subdivided into three extracellular domains (designated 1, 2, and 3), one intracellular domain, and one transmembrane domain. The two outermost extracellular domains, 1 and 2, together form the groove that binds antigenic peptide. Thus, interaction with the TCR occurs at this region of the protein. The 3 domain of the molecule contains the recognition site for the CD8 protein on the killer CD8 T-cell; this interaction serves to stabilize the contact between the T-cell and the antigen presenting cell (APC). The invariant light chain (12 kDa),-23- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282encoded outside the MHC on chromosome 15, consists of a single, extracellular polypeptide. The terms “MHC light chain”, “P2-microglobulin”, and “32m” may be used interchangeably herein.

[0069] As used herein, “peptide epitope,” “epitope,” or “peptide” refer to a portion of an antigenic peptide that is recognized by the immune system in the appropriate context, specifically by antibodies, B cells, or T cells. Epitopes may include B cell epitopes (e.g., predicted B cell reactive epitopes) and T cell epitopes (e.g., predicted T cell reactive epitopes). T-cell epitopes (e.g., predicted T cell reactive epitopes) are peptide sequences which are required for recognition by specific T-cell receptors expressed by T-cells. T cell epitopes (e.g., predicted T cell reactive epitopes) are processed intracellularly and presented on the surface of antigen presenting cells (APCs), where they are bound to MHC molecules including MHC class II and MHC class I molecules. In some embodiments, an epitope is comprised of a plurality of chemical atoms or groups on an antigen. In some embodiments, such chemical atoms or groups are surface-exposed when the antigen adopts a relevant three-dimensional conformation. In some embodiments, such chemical atoms or groups are physically near to each other in space when the antigen adopts such a conformation. In some embodiments, at least some such chemical atoms are groups are physically separated from one another when the antigen adopts an alternative conformation (e.g., is linearized). In some embodiments, the peptide may include, for example, class I epitope or a class II epitope. A class I epitope has a length of 8-11 amino acids. A class II epitope may have a length of 12-25 amino acids.

[0070] As used herein the term "neoepitope" or "neoantigen" is understood in the art to refer to an epitope that emerges or develops in a subject after exposure to or occurrence of a particular event (e.g., development or progression of a particular disease, disorder or condition, e.g., infection, cancer, stage of cancer, etc). As used herein, a neoantigen is one whose presence and / or level is correlated with exposure to or occurrence of the event. In some embodiments, a neoantigen is one that triggers an immune response against cells that express it (e.g., at a relevant level). In some embodiments, a neoantigen is one that triggers an immune response that kills or otherwise destroys cells that express it (e.g., at a relevant level). In some embodiments, a relevant event that triggers a neoantigen is or comprises-24- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282somatic mutation in a cell. In some embodiments, a neoantigen is not expressed in noncancer cells to a level and / or in a manner that triggers and / or supports an immune response (e.g., an immune response sufficient to target cancer cells expressing the neoantigen).

[0071] As used herein, the terms “self-antigen” or “self-epitope” may be used interchangeably and refer to an epitope that is derived from a non-mutated self-protein (e.g., isolated from a subject’s own tissues). In some embodiments, the self-antigen or selfepitope may exhibit ubiquitous expression, tissue-specific expression, or inducible expression under specific triggers or conditions

[0072] As used herein, the terms “T-cell” or “T cells” include naive T cells, CD4+ T cells, CD8+ T cells, memory T cells (including central memory T cells, stem-cell-like memory T cells (or stem-like memory T cells), and two types of effector memory T cells: e.g., TEM cells and TEMRA cells), activated T cells, anergic T cells, tolerant T cells, chimeric B cells, Regulatory T cells (also known as suppressor T cells), Natural killer T cells, Mucosal associated invariant T cells, and y6 T cells, and antigen-specific T cells.

[0073] As used herein, “T-cell receptor” or “TCR”, is a protein complex found on the surface of T cells, that is responsible for recognizing fragments of antigen as peptides bound to major histocompatibility complex (MHC) molecules. TCR is composed of two disulfide-linked protein chains. Cells expressing a TCR containing the highly variable alpha (a) and beta (P) chains are referred to as aP T cells. Cells expressing an alternate TCR, formed by variable gamma (y) and delta (8) chains, are referred to as y6 T cells. When the TCR engages with antigenic peptide and MHC (peptide / MHC), the T lymphocyte is activated through signal transduction, that is, a series of biochemical events mediated by associated enzymes, co-receptors, specialized adaptor molecules, and activated or released transcription factors. In some embodiments, the TCR is a native T cell receptor that is endogenous to the immune cells. In some embodiments, the TCR is an artificial receptor that mimics native TCR function, i.e., recognizing peptide antigens of key intracellular proteins in the context of MHC on the cell surface. In some embodiments, a TCR cell corresponding to the TCR-a polypeptide and the TCR-P polypeptide may be isolated from at least one of a killer CD8 T-cell or a helper CD4 T-cell, among others. In some embodiments, the TCR-a polypeptide may include at least a segment corresponding to -25- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282CDR3a domain. In some embodiments, the TCR-P polypeptide may include at least a segment corresponding to CDR3P domain. When the TCR binds its target, it triggers signaling through associated CD3 proteins to initiate T-cell activation.

[0074] As used herein, an “peptide-MHC complex” or “pMHC complex” refers to any complex, assembly, arrangement, or association comprising at least an MHC molecule and a peptide epitope. An MHC-peptide complex may be a structure formed when a peptide epitope is bound in the peptide-binding groove of an MHC molecule. The MHC may act as a platform, positioning parts of the peptide so they can be recognized by a TCR. Once formed, the complex may be transported to the cell surface where the peptide-MHC complex can be interrogated by T cells. The density and stability of these complexes can influence how strongly T cells respond.

[0075] As used herein, a “TCR-peptide-MHC complex” or “TCR:pMHC complex” refers to any complex, assembly, arrangement, or association comprising a T-cell receptor and a peptide-MHC complex. The TCR-peptide-MHC complex may include multicomponent assembly created when a TCR binds to an MHC-peptide complex (e.g., on a cell surface of an APC). This interaction may be central to antigen recognition because the TCR simultaneously contacts exposed residues of the peptide epitope and conserved regions of the MHC molecule. The binding event helps determine whether the T cell will become activated, remain tolerant, or ignore the interaction, depending on affinity, duration, and context. In some embodiments, co-receptors (CD4 or CD8) may bind the MHC at the same time, strengthening the interaction and helping initiate signaling. Formation of the TCR-peptide-MHC complex may trigger downstream immune responses such as cytokine production and cytotoxic activity.

[0076] As used herein, a “quaternary structure” refers to any three-dimensional arrangement, assembly, association, organization, or relative positioning of two or more molecular components, including proteins, polypeptides, peptides, domains, chains, subunits, or portions thereof. The quaternary structure may define assembly of two or more polypeptide chains (subunits) into a single functional protein complex. In some embodiments, the quaternary structure may include a three-dimensional arrangement of the-26- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282MHC, TCR, peptide, or any combination thereof (e.g., TCR-MHC -peptide complex or MHC -peptide complex).

[0077] As used herein, a “wildtype sequence” refers to a standard or reference form of a gene sequence, nucleic acid sequence, or protein sequence that is most commonly found in a natural population or is used as the baseline in experiments. As used herein, a “mutant type sequence,” “mutant sequence,” or similar term refers to any sequence that differs from a wildtype sequence, a reference sequence, or another comparison sequence by one or more substitutions, insertions, deletions, rearrangements, fusions, truncations, duplications, modifications, edits, or other variations. In some embodiments, the MHC, TCR, or peptide epitope may comprise a wildtype or mutant sequence.

[0078] The term “vaccine” as used herein is a preparation used to enhance protective immunity against cancer, or infectious agents such as viruses, fungi, bacteria and other pathogens. A vaccine may be useful as a prophylactic agent or a therapeutic agent. Vaccines contain cells or antigens which, when administered to the body, induce an immune response with the production of antibodies and immune lymphocytes (T-cells and B-cells).

[0079] It is also to be appreciated that the various modes of treatment of disorders as described herein are intended to mean “substantial,” which includes total but also less than total treatment, and wherein some biologically or medically relevant result is achieved. The treatment may be a continuous prolonged treatment for a chronic disease or a single, or few time administrations for the treatment of an acute condition.NGS Platforms

[0080] In some embodiments, high throughput, massively parallel sequencing employs sequencing-by-synthesis with reversible dye terminators. In other embodiments, sequencing is performed via sequencing-by-ligation. In yet other embodiments, sequencing is single molecule sequencing. Examples of Next Generation Sequencing techniques include, but are not limited to pyrosequencing, Reversible dye-terminator sequencing, SOLiD sequencing, Ion semiconductor sequencing, Helioscope single molecule sequencing etc.-27- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0081] The Ion Torrent™ (Life Technologies, Carlsbad, CA) amplicon sequencing system employs a flow-based approach that detects pH changes caused by the release of hydrogen ions during incorporation of unmodified nucleotides in DNA replication. For use with this system, a sequencing library is initially produced by generating DNA fragments flanked by sequencing adapters. In some embodiments, these fragments can be clonally amplified on particles by emulsion PCR. The particles with the amplified template are then placed in a silicon semiconductor sequencing chip. During replication, the chip is flooded with one nucleotide after another, and if a nucleotide complements the DNA molecule in a particular microwell of the chip, then it will be incorporated. A proton is naturally released when a nucleotide is incorporated by the polymerase in the DNA molecule, resulting in a detectable local change of pH. The pH of the solution then changes in that well and is detected by the ion sensor. If homopolymer repeats are present in the template sequence, multiple nucleotides will be incorporated in a single cycle. This leads to a corresponding number of released hydrogens and a proportionally higher electronic signal.

[0082] The 454TM GS FLX ™ sequencing system (Roche, Germany), employs a light-based detection methodology in a large-scale parallel pyrosequencing system.Pyrosequencing uses DNA polymerization, adding one nucleotide species at a time and detecting and quantifying the number of nucleotides added to a given location through the light emitted by the release of attached pyrophosphates. For use with the 454™ system, adapter-ligated DNA fragments are fixed to small DNA-capture beads in a water-in-oil emulsion and amplified by PCR (emulsion PCR). Each DNA-bound bead is placed into a well on a picotiter plate and sequencing reagents are delivered across the wells of the plate. The four DNA nucleotides are added sequentially in a fixed order across the picotiter plate device during a sequencing run. During the nucleotide flow, millions of copies of DNA bound to each of the beads are sequenced in parallel. When a nucleotide complementary to the template strand is added to a well, the nucleotide is incorporated onto the existing DNA strand, generating a light signal that is recorded by a CCD camera in the instrument.

[0083] Sequencing technology based on reversible dye-terminators: DNA molecules are first attached to primers on a slide and amplified so that local clonal colonies are formed. Four types of reversible terminator bases (RT-bases) are added, and non--28- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282incorporated nucleotides are washed away. Unlike pyrosequencing, the DNA can only be extended one nucleotide at a time. A camera takes images of the fluorescently labeled nucleotides, then the dye along with the terminal 3' blocker is chemically removed from the DNA, allowing the next cycle.

[0084] Sequencing by synthesis (SBS), like the "old style" dye-termination electrophoretic sequencing, relies on incorporation of nucleotides by a DNA polymerase to determine the base sequence. A DNA library with affixed adapters is denatured into single strands and grafted to a flow cell, followed by bridge amplification to form a high-density array of spots onto a glass chip. Reversible terminator methods use reversible versions of dye-terminators, adding one nucleotide at a time, detecting fluorescence at each position by repeated removal of the blocking group to allow polymerization of another nucleotide. The signal of nucleotide incorporation can vary with fluorescently labeled nucleotides, phosphate-driven light reactions and hydrogen ion sensing having all been used. Examples of SBS platforms include Illumina GA and HiSeq 2000. The MiSeq® personal sequencing system (Illumina, Inc.) also employs sequencing by synthesis with reversible terminator chemistry.

[0085] In contrast to the sequencing by synthesis method, the sequencing by ligation method uses a DNA ligase to determine the target sequence. This sequencing method relies on enzymatic ligation of oligonucleotides that are adjacent through local complementarity on a template DNA strand. This technology employs a partition of all possible oligonucleotides of a fixed length, labeled according to the sequenced position.Oligonucleotides are annealed and ligated and the preferential ligation by DNA ligase for matching sequences results in a dinucleotide encoded color space signal at that position (through the release of a fluorescently labeled probe that corresponds to a known nucleotide at a known position along the oligo). This method is primarily used by Life Technologies’ SOLiD™ sequencers. Before sequencing, the DNA is amplified by emulsion PCR. The resulting beads, each containing only copies of the same DNA molecule, are deposited on a solid planar substrate.

[0086] SMRT™ sequencing is based on the sequencing by synthesis approach. The DNA is synthesized in zero-mode wave-guides (ZMWs)-small well-like containers with the -29- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282capturing tools located at the bottom of the well. The sequencing is performed with use of unmodified polymerase (attached to the ZMW bottom) and fluorescently labeled nucleotides flowing freely in the solution. The wells are constructed in a way that only the fluorescence occurring at the bottom of the well is detected. The fluorescent label is detached from the nucleotide at its incorporation into the DNA strand, leaving an unmodified DNA strand.Artificial Intelligence (Al) Platforms for Predicting Structures of Proteins

[0087] AlphaFold 3 is an extension of AlphaFold and is a deep learning model for predicting biomolecular structures and interactions, extending beyond proteins to complexes involving nucleic acids, ligands, ions, and post-translationally modified components.Architecturally, it builds on the transformer-based and pairwise geometric reasoning ideas from AlphaFold 2, and introduces a diffusion-based generative framework that refines candidate molecular arrangements into physically plausible 3D structures. Its main function is to predict accurate atomic coordinates for multi-component biological assemblies, often including binding interfaces and interaction geometries. Outputs include, for example, fullatom 3D structures, confidence estimates for local and global regions, and interface-quality information for complexes. The model works by integrating sequence information and molecular context, constructing internal representations of residue-residue and atom-level relationships, and then iteratively denoising structural states to produce a final conformation.

[0088] RoseTTAFold is a deep learning model for protein structure prediction that uses a three-track architecture linking ID sequence information, 2D residue-pair relationships, and 3D coordinates. These tracks exchange information continuously, allowing the model to reason about sequence patterns, contacts, and spatial arrangement at the same time. Its main function is to predict protein 3D structures, and it can also be used for complexes and design-related tasks. Outputs include predicted atomic coordinates, residue-residue geometries, and confidence estimates. RoseTTAFold works by using evolutionary information from multiple sequence alignments and, when available, structural templates. It then jointly refines pairwise constraints and 3D geometry to generate a folded structure.-30- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0089] OpenFold is an implementation of the AlphaFold 2 architecture for high-accuracy protein structure prediction. Its structure includes an Evoformer module, which processes multiple sequence alignment and pairwise residue features, and a structure module, which converts these learned relationships into 3D coordinates. The model’s main function is to predict protein tertiary structure from sequence, with outputs including atomic coordinates and confidence metrics such as per-residue reliability and pairwise error estimates. OpenFold works by collecting evolutionary information from homologous sequences and optional templates, then using attention-based updates to infer residueresidue relationships. The structure module iteratively refines these constraints into a final 3D model.

[0090] Boltz- 1 is an open biomolecular structure prediction model designed for proteins and molecular complexes, including interactions with ligands and nucleic acids. Its architecture combines sequence-based representations with pairwise and atom-level geometric reasoning to capture both folding and binding interactions. The model’s main function is to predict 3D atomic structures for single proteins and complexes. Outputs, for example, include full structural coordinates and confidence estimates, especially for interfaces between molecular partners. Boltz- 1 works by encoding all molecular components together, learning spatial constraints among residues and atoms, and iteratively refining coordinates into a consistent 3D arrangement.

[0091] OmegaFold is a protein structure prediction model that can generate 3D structures directly from a single amino acid sequence. Its architecture uses protein language model-style sequence representations together with geometry-aware modules that infer residue relationships and construct 3D coordinates. The model’s main function is to predict protein tertiary structure, with outputs including backbone or full-atom coordinates and confidence-related scores. OmegaFold works by embedding the sequence into contextual features learned from large protein datasets, then translating those features into spatial constraints and iteratively assembling the protein in 3D.

[0092] ESMFold is a protein structure prediction model that combines a large transformer-based protein language model with a structure prediction head. Its architecture uses sequence embeddings from the ESM language model, which capture long-range and -31- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282evolutionary -like patterns directly from raw protein sequences, and then converts them into 3D coordinates. Its main function is rapid prediction of protein tertiary structure without requiring multiple sequence alignments or template searches. Outputs include a 3D structural model and confidence scores for predicted regions. ESMFold works by first encoding the amino acid sequence using the language model, then using a folding module to infer residue contacts and spatial geometry.Protein Language Models for Generating Protein Embeddings

[0093] ESM-2 is a transformer-based protein language model trained on large protein sequence databases using masked language modeling. Its architecture uses selfattention layers to learn contextual relationships between amino acid residues across a protein sequence. The model’s main function is to generate protein embeddings that capture structural, biochemical, and evolutionary information. Outputs include residue-level and protein-level embeddings for downstream tasks such as protein function prediction, mutation effect prediction, residue contact prediction, and structure prediction. ESM-2 works by masking residues during training and learning to infer missing sequence information from surrounding context.

[0094] ESM-C is a transformer-based protein language model family designed for efficient protein representation learning and transfer learning applications. Its architecture is derived from the ESM family and uses self-attention mechanisms to encode contextual residue relationships while reducing computational complexity relative to larger models. The model’s main function is generation of protein embeddings for downstream biological prediction tasks. Outputs include contextual sequence representations for protein classification, annotation, and structure-related prediction. ESM-C works by learning residue dependencies from large protein databases using masked language modeling.

[0095] ProtT5 is a transformer-based protein language model derived from the T5 architecture and trained on large-scale protein sequence datasets. Its architecture uses encoder-decoder transformer blocks with attention mechanisms that learn contextual residue relationships across protein sequences. The model’s main function is generation of transferable protein embeddings for downstream prediction tasks. Outputs include residue- -32- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282level and sequence-level embeddings for protein function prediction, localization prediction, secondary structure prediction, and mutation effect modeling. ProtT5 works by learning contextual sequence representations from large protein datasets.

[0096] ProtBERT is a BERT-style transformer protein language model trained on large protein sequence databases using masked token prediction. Its architecture includes bidirectional self-attention layers that learn contextual relationships among amino acid residues. The model’s main function is protein sequence representation learning for downstream biological applications. Outputs include contextual protein embeddings useful for protein classification, function prediction, and interaction modeling. ProtBERT works by masking residues during training and learning to infer missing sequence elements from surrounding context.

[0097] ProteinBERT is a protein language model architecture designed specifically for biological sequences. Its architecture combines local residue-level representations with global protein-level representations to capture both sequence motifs and whole-protein functional properties. The model’s main function is protein representation learning for downstream prediction and annotation tasks. Outputs include contextual embeddings representing residue-level and protein-level biological information. ProteinBERT works by jointly modeling local sequence context and global protein attributes through attentionbased updates.

[0098] TCR-BERT is a transformer-based language model designed for T cell receptor sequence representation learning. Its architecture is based on BERT-style bidirectional self-attention layers trained on large TCR repertoire datasets to learn contextual relationships among amino acid residues within TCR sequences. The model’s main function is generation of TCR embeddings for downstream immunological prediction tasks. Outputs include contextual sequence representations useful for TCR specificity prediction, repertoire analysis, clustering, and immune classification tasks. TCR-BERT works by masking residues during training and learning to infer missing sequence information from surrounding TCR sequence context.Examples-33- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282The present technology is further illustrated by the following Examples, which should not be construed as limiting in any way.Example 1: Ensembles of Predicted Structures Enable T Cell Peptide-MHC 1 Binding Prediction

[0099] Adaptive immunity relies on T-cell receptor (TCR) recognition of non-self epitopes, short peptides presented by the Major Histocompatibility Complex (MHC) on the cell surface. Accurate computational prediction of TCR-epitope binding would simultaneously deepen the fundamental understanding of self / nonself discrimination, pathogen recognition, and autoimmunity; while unlocking development of targeted immunotherapies, such as cancer vaccines and T cell therapies. An ensemble approach (enFoldX) was created that leverages structure prediction models to build sensitive binding predictors, from which T cell reactivity was distinguished between peptides that differ by a single amino acid substitution, as needed for cancer neoantigens. enFoldX utilizes a customized highly parallelized workflow which allows us to produce ensembles of predicted protein structures at scale and train classifiers to infer reactivity based on distributions of engineered structure features and alignment confidence metrics. In contrast, the predictive power of sequence-based models was evaluated, including a Large Language Model (LLM) that was created. While sequence-based approaches could predict well for observed TCRs and epitopes close in sequence to training data, their applicability to novel sequences was limited. Conversely, the ensemble approach is the only model that showed true generalizability. By leveraging the entire protein universe at scale, structure ensembles therefore enable classifiers that reflect physical free energies, providing a tractable path towards self / nonself discrimination at the sensitivity required for cancer neoantigen prediction and imparting lessons for a wide array of complex binding problems.INTRODUCTION

[0100] T cells play an essential role in adaptive immunity due to their ability to recognize and kill virus infected cells and are fundamental for immune surveillance of tumors and their clearance by immunotherapies1. Their specificity is encoded by their T-cell receptors (TCRs) which can recognize epitopes presented by Major Histocompatibility-34- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282Complex (MHC) molecules. While the typical human immune system will generate 109-1O10possible TCR sequences through V(D)J recombination2,3, somewhere between ten to thousands of TCRs have been estimated to recognize and respond to a specific epitope, a segment of a protein processed by the cellular machinery and presented on the cell surface by an MHC molecule. While the adaptive immune system likely evolved to defend against pathogens4, T cells can also recognize cancer neoantigens, including mutated self-peptides which may differ from self by only one amino acid substitution5 7. The ability to predict which TCR binds a specific peptide-MHC (pMHC) has broad applications, ranging from understanding which features drive viral clearance, to predictive biomarkers for autoimmune diseases, a greater understanding of self-tolerance, and the design of personalized cancer vaccines and T cell therapies.

[0101] The advent of high throughput TCR sequencing has enabled nascent attempts to predict whether a given TCR can recognize a specific epitope8,9. Such binding prediction models have recently gained great attention, including but not limited to graph-based models, large language models (LLMs), protein language models, and Al-based structural prediction and docking methods10 14While some methods have shown promise, generalization to novel antigens remains a major challenge and a clear path forward is lacking15,16. It has been established the principle that biophysical neoantigen qualities can discriminate immunogenic neoantigens in cancer that differ from their wild-type counterpart by one mutation, in part by leveraging viral antigen data - suggesting structure based biophysical binding models are a promising path17 l9.

[0102] The generic case of predicting TCR specificity for a new TCR-epitope pair remains challenging due to the diversity of binding modes and the lack of available training data20. Promising results have come from using protein structure simulators to jointly fold the TCR:pMHC structure and leverage the results to estimate if the TCR and pMHC can form a complex21 25. However, such studies tended to focus on specific TCRs or specific epitopes but often performed poorly with out-of-distribution data. Here, a high throughput was designed, highly parallelized computational approach that was called Ensemble of Folded complexes (enFoldX). enFoldX enables the evaluation of large datasets of predicted structures simultaneously as it extracts distributions of high dimensional confidence-35- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282features, including customized biophysical features that was designed specifically for this task. These features was found vary significantly between TCR:pMHC pairs of differing T cell reactivity, implying that binders and non-binders are separable in feature space. The resultant ensemble of features was then used to create a binding model, analogous to an energy function in statistical physics. The enFoldX approach generalizes in a dataset agnostic way - for example, models trained on human data can classify binder TCR:pMHC pairs in mouse data. Moreover, it was found, for the first time, that the structure-guided models trained on publicly available data can generalize to discriminate epitopes that differ by only one mutation away from self, the resolution needed for predicting mutation-derived cancer neoantigens.RESULTSenFoldX: Ensemble Approach to Classify Binders and Non-Binders[0103| enFoldX involves four distinct steps, run sequentially in a highly parallelized manner: structure look-up by template search and multiple sequence alignment (MSA), structure prediction via diffusion, feature extraction from predicted structures and confidence metadata, and classification of predicted complexes as bona fide stable versus unstable binders (FIG. 1A, FIG. 6). enFoldX is implemented leveraging AlphaFold326(AF3) and applied to predict TCR:pMHC pairs, but the general framework can be adapted to support other structure prediction methods (ESMFold27, RoseTTAFold28) or protein language models like OmegaFold29. While TCR:pMHC was the focus on pairing here, the general strategy can be adapted to other protein complexes. The pipeline enables analysis of large datasets of TCR:pMHC pairs in a tractable manner by leveraging parallel computing environments to speed up dataset-level predictions by a factor proportional to the amount of compute available.

[0104] A TCR:pMHC complex is made up of four chains: the a and P chains of the TCR, the MHC molecule, and the peptide. The predicted 3D structures of the complexes of non-binders occasionally show obvious signs of incorrect docking. However, it is not always clear from the predicted structures alone whether a given complex will bind (FIGs.1B-D) So, to assess binding, other outputs of AF3 was considered. Like other protein -36- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282folding methods, AF3 outputs confidence measures that can be used to assess the reliability of its predicted structure. These measures include the PAEs (predicted alignment errors), pairwise residue scores that estimate error (in Angstroms) in their relative positions; pLDDT (predicted local distance difference test), a per-atom confidence score; pTM (predicted template modeling), an estimate of the overall accuracy of the protein structure (in RMSD) and ipTM (interface predicted template modeling) matrices, measuring the accuracy of predicted protein-protein interfaces and complexes; as well as other global summary metrics that indicate model confidence. These confidence measures was extracted for the complex overall, each chain individually, and where applicable, for each pair of chains. Bespoke local features was further constructed corresponding to known sites of biophysical interactions, such as the TCR-MHC interface or the TCR-peptide interface (FIG. IB) As will be shown below, these bespoke interchain features have more predictive information than the global summary metrics alone. For example, averages over pairs of inter-protein residues in the PAE matrix can predict binding30. This “iPAE” (also sometimes referred to as interface PAE or interaction PAE) has been used as a binding heuristic in the design of TCR mimics binding to pMHC complexes31 33and de novo pMHC complexes binding to biological TCRs34. Given the TCR:pMHC complex has many crucial interfaces, the iPAEs was calculated between all pairwise groups of chains (TCRa, TCRb, MHC, and peptide) in the model.

[0105] First, the paired ipTM scores was collected that are output directly by AF3. Next, iPAE scores are built based on the off-diagonal blocks of the PAE matrix (FIG. 1B-D). A variety of ways of collapsing these submatrices was included, including taking the average, maximum, minimum, and standard deviation (Methods). Finally, interchain contact probabilities (<8A) was calculated from the matrix that is returned from AF3’s distogram head. For discrimination between two peptides that are one substitution away, additional structural features was included, by computing the root mean square deviation (RMSD) between the predicted TCR:PMTMHC structure and the predicted TCR:pwrMHC structure. These RMSD features are omitted from any analysis where there is not a reference pwrMHC against which to compare.-37- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0106] Thus, for a single AF3 predicted output, a range of features was extracted that incorporate global structure- wide confidence, local per chain confidence, and interchain alignment errors and contact probabilities (Extended Data Table 1). Two different methods of integrating this multisample information was compared: taking the average and standard deviation of the feature’s value across outputs (“structural ensemble”) versus using features from only the top ranked structure based on AF3’s ranking score (“best structure models”). Because of the highly parallel nature of the pipeline (FIG. 6), multiple structures was predicted per protein complex using multiple AF3 seeds and generate such feature matrices across large databases of sequences. With this large amount of training examples, classifiers like Ridge regression35or XGBoost36can be trained to distinguish binders from non-binders using this rich feature space. The dataset can either be split out a portion into a training set or use one database to train the classifier to generalize and make predictions on an independent dataset. These classification models can then serve as a proxy energy function to predict binding versus non-binding.The Neoantigen Task

[0107] To assess whether confidence metrics from ensembles of predicted structures could predict TCR:pMHC binding, it was first assessed whether it could distinguish binding complexes from non-binders that differ by a single amino acid change in the epitope. This “Neoantigen Task” was implemented using a T cell cross-reactivity dataset that was previously created that shows differing levels of T cell activity for peptides that differ by one amino acid18. Starting from a set of peptides and their known cognate TCRs, a library of peptides was generated which were one amino acid substitution away, resulting in various degrees of TCR activation (“cross-reactivity”). For example, the TCR labelled TCR1 is specific for the wild-type epitope NLVPMVATV (“NLV”) presented on HLA-A*02:01. When every possible substitution was introduced to its residues, certain mutant epitopes conferred a complete loss in activity while others retained recognition (FIG. 2A).Importantly, the dataset contains positive and negative pairs, which avoids any biases that come from synthetic approaches to creating negative data37. The full dataset includes 7 human TCRs experimentally tested against the same total mutational scan of all possible substitutions at each position in the seed wild-type epitope (FIG. 8). Using this T cell-38- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282activation as the approximation of TCR:pMHC binding, many of the enFoldX-extracted features was found differed significantly between binding and non-binding complexes (FIG. 2B), indicating that simple classifiers trained on this feature space could potentially predict binding.

[0108] The machine learning classifiers was therefore assessed, including Ridge regression and XGBoost, trained on the full feature features on the task of classifying TCR:pMHC pairs as binders or non-binders. These classifiers were able to discriminate binders from non-binders accurately (FIG. 2C, FIGs. 9A-E). The cross-reactivity dataset was split into 5 folds, each further divided into 80% training set and 20% test set so that all TCR:pMHC pairs are included as test cases in at least one of the folds and achieved an average Area Under the ROC Curve (AUROC) over 0.8 across the five folds. As described in a later section, this task was particularly challenging for sequence-based models, which performed only slightly better than chance. The structural ensemble model outperformed the best structure model (FIG. 2C), although the exact performance depended on the number of structures sampled from AF3, where the ensemble strategy benefits significantly from more structures being included (FIG. 2D). Thus for challenging structural predictions like the TCR:pMHC complex, multiple stochastically generated models may be required to enable robust enough sampling24’38, although there is a tradeoff between the performance boost and the computational cost of including more seeds. In terms of model architecture, XGBoost offered slightly better performance on this dataset as well as on other datasets described below, but since it generalized poorly across datasets (FIG. 12), further analysis was focused on the Ridge regression models.

[0109] To understand which components of the output space are most predictive of TCR specificity, a family of models that only incorporated subsets of the feature space (FIG. 2E) was considered. Models trained using only features extracted from the model confidence matrices performed the best, outperforming models using only contact probabilities or structural RMSD features. Nevertheless, the model using the maximal feature set still outperformed any of the subset models, indicating that the different types of features offer some level of independent predictive value. The approach was further evaluated using leave-one-out-style cross-validation strategy where it was left out: one of-39- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282the 7 TCRs, one of the 3 WT peptides, one of 9 mutated positions. While the classifier performance varied between specific scenarios, in all cases it was significantly better than random (FIGs. 9A-E). Feature importance analysis was performed via feature permutation and noted that many of the most important features were iPAEs, and that a substantial proportion of that signal was specifically at the TCR:pMHC interface (FIG. 2F). Both means and standard deviations of features across predicted structures for the same complexes were important for the structural ensemble models. In conclusion, both the breadth of the feature space and the aggregation strategy across ensembles of predicted structures enabled accurate classification of binder versus non-binder structures in the Neoantigen Task.Classification of TCR specificity in publicly available human and mouse datasets[OHO] EnFoldX’s scalability was then leveraged to tackle large human and mouse datasets of TCR-epitope pairs in the VDJdb, a curated database of TCRs with known antigen specificity, mostly from viruses39. Since VDJdb contains only binder pairs, a common method37of creating negatives was followed by randomly permuting TCRs and epitopes between complexes within the same HLA. A second negative dataset was further created by replacing reported binder peptides with “decoy” peptides predicted by NetMHC4.040 (for mouse data) or NetMHCPan41(for human data) to be presentable on the same MHC allele, to attempt to control for the difference between a negative pair not binding due to lack of MHC presentation versus not binding due to lack of recognition of a bound pMHC by a TCR. In both the human and mouse VDJdb datasets, enFoldX classifiers achieved high classification performance in distinguishing positives for both permuted negatives and for decoy negatives (FIGs.3A-D). As in the Neoantigen Task, a variety of feature types were important for the classification, including iPAEs, particularly at the TCR:pMHC interface.

[0111] Across both the mouse and human VDJdb datasets, the decoy task that was found was a lot easier for the models than the permutation task. The extremely high accuracy in classifying positives from NetMHC decoys, particularly for the VDJdb mouse dataset, could imply several of the curated peptides were false positives that were incorrectly predicted to bind by NetMHC. Indeed, the feature importance analysis for this -40- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282task showed that many of the significant features, particularly for the VDJdb mouse classifier, were related to the peptide-MHC interface rather than the TCR interface (FIG.3E). This suggests both that structural modeling approaches can be helpful for the task of predicting MHC presentation independent of TCR recognition42, and that performance on this decoy task10’22may be overly reliant on MHC presentation-related features and may not represent realistic performance on predicting bona fide TCR specificity. As before, the structure ensemble strategy outperformed the best structure strategy, with the performance gain depending on the number of structures sampled (FIG. 3F). Significantly, enFoldX models trained on human data from VDJdb were able to classify binder TCR:pMHC pairs in mouse data from VDJdb (FIG. 3G). In other words, the classifiers built from ensembles of features displayed generalizability.Limits of sequence-based classification

[0112] To contrast the structure ensemble approach of enFoldX with the sequencebased models, the ability that was set out to evaluate of large language models (LLMs) to predict TCR-epitope specificity from sequence and the sufficiency of publicly available TCR specificity datasets for this purpose. BERTie (BERt for TCRs and Immunogenic Epitopes) was developed specifically for this task, an adaptation of the widely adopted transformer architecture BERT (Bidirectional Encoder Representations from Transformers)43and compared its performance to the parallelized structure prediction approach and to other published sequence based models44 47(Methods, FIG. 4A). A limitation of these approaches is the requirement of a large training corpus of functionally validated paired sequences of TCRs and peptides, while structural models such as AlphaFold3 are pre-trained on the universe of protein structures available in the Protein DataBank48.

[0113] BERTie was trained on single chain TCRP and peptide datasets and used only the CDR3 portion of the TCRP chains. 80% of the combined deduplicated dataset composed of MIRA49, VDJdb39, and McPAS50datasets was used for training and 20% was reserved for testing. The test dataset was carefully created so that it included equal proportions of 4 progressively more challenging test case scenarios: a new pairing of a TCR and an epitope both seen during training; a new (unseen) epitope paired with a TCR -41- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282seen during training; an unseen TCR paired with a seen epitope; and a new pairing of an unseen TCR and an unseen epitope. Several embedding regimes were investigated, including single amino acid embedding (char), 2-gram embedding where the input sequences are split into overlapping subsequences of length 2; 3-grams; and 4-grams. An ensemble of 5 3-gram models was then created by splitting the training dataset into 5 training and validation cross-validation (CV) folds, training one model on each fold, and then averaging predictions for test data across the five models. The same was repeated for the single amino acid embedding models. The combined ensemble of 3-gram, char, and 2-gram achieved the best performance of 0.87 AUROC (FIG. 10A). As expected, BERTie produces the most accurate predictions on the easier test case involving a new pair of a seen epitope and a seen TCR; however, it can generalize to the more challenging test cases, performing significantly better than chance even on the most challenging case of unseen epitope paired with an unseen TCR (AUROC=0.64) (FIG. 4B).

[0114] BERTie was benchmarked against several best-in-class methods on Immune Epitope Database (IEDB)51and cross-reactivity datasets. Notably, BERTie trained only on pairs of peptides and CDR3P sequences - and not trained on IEDB - has comparable performance to other methods trained on paired u / p chains (Tulip, NetTCR2.2) and methods trained on IEDB dataset (Tulip, ATM-TCR, NetTCR2.2) (FIG. 4E). Secondly, there is a clear lack of a concordance between different methods, including those trained on IEDB, as demonstrated by wide ranges of per-epitope AUROC values for the same epitope by different methods in most cases (FIG. 10C). Factors were to be investigated that explain BERTie’ s predictive ability on seen and unseen epitopes. Firstly, AUROC was computed for each epitope separately and noted that the mean AUROC for epitopes seen during training (but paired with new TCRs, either seen or unseen during training) was 0.85 and that for 64% of these epitopes AUROC was 1.0 (FIG. 10D, 10E). For almost 10% of unseen epitopes, BERTie also performed very strongly, with AUROC over 0.9, while the mean AUROC for unseen epitopes was 0.61 overall. For seen epitopes, BERTie’s performance is significantly correlated with the number of that epitope’s binder TCRs in the training data and how consistent TCR sequences of those binder TCRs were (FIG. 10F). Additionally, TCR motif score was calculated for each epitope based on detectable motifs in the cognate TCR sequences and calculated Kullback-Leibler (KL) divergence of binder TCRs versus -42- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282background (all other TCRs in the data), measuring how self-similar binder TCRs are and how different they are from background. Both the motif score and the KL divergence were correlated with BERTie’s per-epitope AUROC (FIGs. 10G and 10H).

[0115] For unseen epitopes, BERTie’s ability to correctly predict TCR specificity was largely dependent on the similarity of these epitopes to those included in the training data (FIG. 4C, FIG. 101). For new epitopes within Levenshtein distance of up to 2 amino acids from those included in training data, the AUROC was very high (0.94), but the ability to generalize to novel epitopes significantly deteriorated as epitope sequences diverged further from those seen during training. Similarly, accuracy of predictions also depends on sequence similarity to training data, both for the epitope and the TCR (FIG. 4D). Notably, accuracy is higher for epitopes at Levenshtein distance of 2 than for distance of 1, perhaps suggesting that in cases when a single amino acid substitution breaks TCR reactivity, the over-reliance on sequence similarity causes sequence-based models to continue predicting this non-binder TCR-epitope pair as cognate.Structure-guided approach overcomes reliance on sequence similarity

[0116] By contrast, the structure-guided enFoldX model does not suffer from this overdependence on sequence similarity: prediction accuracy does not depend heavily on how similar epitopes in the test set are compared to those included in the training set. AUROC curves for the enFoldX Ridge model trained on the human VDJdb decoy task show only a slight drop in performance for epitopes up to 2 amino acid substitutions and no further deterioration in performance for more dissimilar epitopes (FIG. 5A). It was further investigated whether HLA alleles and the number of training examples for each allele had a strong impact on enFoldX’ s performance. FIG. 5B shows that within HLA-A and HLA-B groups, there is no correlation between the data available for a specific HLA allele and enFoldX’ s accuracy. However, enFoldX had very low accuracy in predicting epitopes presented on alleles in HLA-C, likely due to insufficient training data for any HLA-C allele (including for AF3) and the difficulty generalizing across HLA groups.

[0117] An added challenge in predicting TCR specificity for cancer neoantigens is that they tend to be very close to self-epitopes, for which cognate TCRs are eliminated -43- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282during thymic selection. Therefore, to have meaningful predictive ability for neoantigen epitope specificity, models need to be sensitive enough to recognize cross-reactivity or lack thereof produced by single substitutions in epitope sequences. Performance of the enFoldX Ridge models trained on human VDJdb data to BERTie and other sequence models on the Neoantigen task defined above (on cross-reactivity dataset) was compared. BERTie and other evaluated sequence-based methods performed slightly better than chance (FIG. 5C).For this task, BERTie was re-trained on all available human TCR:pMHC pairs from VDJdb, McPAS and MIRA dataset.

[0118] This analysis highlights the difficulty current sequence-based modeling approaches encounter in predicting TCR cross-reactivity. Predictions was hypothesized due to being inherently based on epitope sequence similarity to those seen during training. In contrast, enFoldX retained its ability to generalize, transferring learning from one dataset to another, demonstrating meaningful predictive power for this challenging task (AUC=0.71 and AUC=0.67 for decoy and permutation training datasets respectively). Notably, while enFoldX with XGBoost classification had a slightly better performance on predicting within the same datasets, it generalized worse to the cross-reactivity dataset (FIG. 12), indicating potential over-fitting. By learning from more general protein structure data, enFoldX-Ridge can leverage a much more diverse training set to learn an approximate biophysical energy function52.

[0119] To compute this approximate energy function, enFoldX utilizes the high dimensional feature space described above. This space has a far lower effective dimensionality, as many features are correlated and cluster together (FIG. 5D, FIGs. 11A-P). When the features was embedded for the VDJdb and the cross-reactivity datasets in a lower dimensional space, the most important features tended to cluster together (FIG. 5E), and these high-importance clusters were mostly related to the TCR:pMHC interface (FIGs.11A-P) Taken together, these results indicate that enFoldX can effectively compress this high dimensional feature space to approximate the key properties of binding, thus learning a more biophysically accurate representation of the inputs which allows for generalizability to novel, unseen data regimes.Applicability of enFoldX Pipeline to Other Predictive Models-44- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0120] Referring now to FIG. 13, depicted is a receiver operating characteristic (ROC) curve using Boltz-2 steering potentials. The graph includes curves for classifier performance using (i) all confidence-derived features, (ii) PAE-only features, and (iii) a reduced PAE feature set including minimum and standard deviation statistics aggregated across predicted structural samples. This graph demonstrates the applicability of the EnfoldX to models other than AlphaFold3.DISCUSSION

[0121] enFoldX introduces a new paradigm in predictive modeling of protein binding. Instead of using the sequences themselves, like LLM-based TCR-epitope predictors, or analyzing a single predicted co-complex structure, enFoldX makes its predictions based almost exclusively on distributions of interface confidence metrics across predicted structural ensembles. Because it considers averages of quality metrics of the ensemble, enFoldX can generalize to very distinct datasets, or even to different organisms, whereas the generalization performance of LLM-based models depends heavily on sequence similarity. In contrast to prior work53enFoldX’ s success illustrates that models like AlphaFold3 can be leveraged to make accurate protein binding predictions, however, these models must be used with care. First, it is important to focus on the binding interface, and not necessarily on the structure far from the binding sites, though allosteric effects can be important. Furthermore, the use of ensembles is critical; AlphaFold3 (AF3) can produce plausible structures for non-binders, however, it does not do so as consistently as for binders, and AF3 can occasionally produce implausible “hallucinatory” structures for nonbinders. enFoldX, by using multiple draws from AF3, and prioritizing evaluation of points of biophysical contact, can thus robustly distinguish binders from non-binders. There are clear generalizations of this paradigmatic shift for other binding prediction problems.

[0122] A key feature of enFoldX, in contrast to sequence-based methods, is that it can resolve reactivity differences between peptides which differ by one mutation. This unique capability is important for detecting mutation-derived neoantigens in cancer and the many close-to-self viral antigens54. enFoldX generalizes to TCRs and epitopes that are very dissimilar in amino acid sequence to those in its training set; this capability is critical for capturing the combinatorial diversity of TCRs and epitopes. enFoldX also generalizes well -45- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282to unseen MHC sequences, but only within the same HLA group. The results suggest that this weakness may be addressed by training protein structure prediction methods on more HLA-C structures which are currently uncommon in the protein structure databases.

[0123] Here, input that was illustrated features critical to the success of the binding predictor are distributions of interface confidence metrics derived from a predicted structural ensemble. Indeed, it was found that incorporating more structures in the ensemble led to better performance than using only the AlphaF old-ranked best structure. This may highlight the limitation of using noisy heuristics to approximate optimality in search problems beyond biomolecular complexes55,56, but it could also point to other reasons that the ensemble is important. First, using an ensemble may help to identify confident hallucinations made by neural network models, including AF3, when applied to out-of-distribution data: examining the ensemble reveals the disagreement among these random guesses. There may also be a biophysical explanation for the value of the ensemble. Under the folding funnel hypothesis57, a rugged energy landscape demonstrates barrier ridges where misfolded intermediates fall in kinetic traps and fail to achieve their final conformation at the global minimum. This is avoided by thermal noise, i.e. entropy, which kicks the partially folded protein over the ridge until it reaches a true minimum. When sampling from the AF3 generative diffusion model, the occasional misfolded intermediate (FIG. ID) was also found and often find that it models non-binding complexes in the bound state which are presumably less energetically favorable than the unbound state (FIG. 1C). However, increasing the number of draws from the model allows for a more robust sampling of this rugged space. In this view, the model is not always “wrong” but is often accurately drawing from a configuration and trapped in a true local minimum. This points to a deep relationship between rugged energy landscapes and sampling of neural networks, suggesting ensembling may prove useful in other applications of generative modeling in which true samples and hallucinations are hard to distinguish58.

[0124] There may also be a biophysical explanation for the importance of the interface residue confidence features: the predicted confidence metrics could reflect either uncertainty in the reconstruction, due to model errors, or actual uncertainty due to residual entropy in the positions, or relative positions, of interface residues. Under this-46- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282interpretation, that the classifier was suggested, whose most important features were largely the customized points of biophysical contact, could capture a true free energy landscape, and ensembling approximates the entropic favorability of the true minima over the many false minima along the way to protein folding.

[0125] The fact that enFoldX when trained on VDJdb was predictive for the Neoantigen Task leads us to conclude that the current state-of-the-art structure prediction models, when used to generate ensembles, can transfer knowledge from the large universe of resolved protein structures to highly specific complex binding tasks. This also supports the notion that viral T cell epitopes are not fundamentally different from neoantigens, and that models that are trained using the more widely available data from viruses will continue to provide value in understanding T cell responses in cancer17,18’54. While the work leveraged AlphaFold3, the same principles could apply for any protein structure predictor, as the structure prediction software was taken as an input, along with the TCR and pMHC. This was an intentional design decision, as the protein modeling community consistently trains new models and focuses on developing open source models59, and the hope was for the tool to be widely compatible and usable. As a result, the path was hypothesized to improved computational self / nonself discrimination runs alongside the path to improved structure prediction. Following was envisioned that this joint path can lead to fundamental immunological discoveries and novel precision immunotherapies.METHODSDATASETSCross-Reactivity Data

[0126] The T-cell “cross-reactivity” dataset used for the Neoantigen task is publicly available as part of the previous paper18. The half maximal effective concentration (EC50) fitted to the measured avidity curve for each of the 1204 TCR:pMHC combinations was binarized so that TCR:pMHC complexes with EC50 greater than 0.1 pg / mL were considered as non-reactive (non-binders), using the same threshold as in the original publication. This resulted in 342 binder TCR:pMHCs and 862 non-binders (FIG. 2A, FIG.8). The measured output in the original experiment was TCR activation, therefore some -47- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282pMHCs potentially did not elicit TCR activation because peptide failed to be presented on the MHC; however, for the purposes of this analysis TCR:pMHC was considered a complex “binder” if the peptide is both presented on the MHC and recognized by the TCR (via activation or other functional or binding assay). For the purposes of AF3 modeling, full length TCR sequences was generated using Stitchr60for a and P TCR chains.Mouse VDJdb Data

[0127] VDJdb39version 2025 / 02 / 21 was downloaded from vdjdb.cdr3.net. For the mouse data set, the data was filtered to ‘species’==‘MusMusculus’ and only kept rows with non-null TCRs for both chains, non-null MHC, and a non-null peptide. It was further filtered and kept only data for the MHC -I alleles H-2-Kb and H-2-Db. Stitchr60was used to stitch the TCR sequences from the V and J gene and CDR3 information, and filtered out any rows for which Stitchr failed or inserted a stop codon. This yielded 1896 positive TCR:pMHC combinations, 718 of which were for H-2-Kb (38%) and 1178 of which were for H-2-Db (62%), corresponding to 24 unique peptides. The amino acid sequences for the MHC molecules were downloaded from Uniprot61.

[0128] To create negative pairs, two alternative approaches was tried. (1) For the decoy task, each of the 1896 rows was matched in the data with a decoy row, where the peptide was swapped out with a random decoy peptide from the mouse proteome that was predicted by netMHC-4.O40to be a strong binder to the MHC allele for that row. Strong binders was defined as having Rank < 0.5 as is standard. (2) For the permutation task, each of the 1896 rows was matched in the data with a permuted negative where the peptide was swapped out with a different peptide from the dataset, where the other row had a different TCRa and TCRP but the same MHC. In principle, this means that the peptide was effectively presented on that MHC to a different TCR, but has no prior relationship to the TCR it was matched with, and thus can be assumed to be non-binding37.Human VDJdb Data

[0129] As described above, the VDJdb39version 2025 / 02 / 21 was downloaded from vdjdb.cdr3.net and filtered to include only the rows with ‘species’=‘ HomoSapiens’ and entries that contain TCR a and P chains, a valid human MHC class I allele, and a non-null -48- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282peptide. As with the mouse dataset, Stitchr60was used to generate full-length TCR sequences from the V gene, J gene, and CDR3 information for a and P chains, and filtered out any rows for which Stitchr failed or inserted a stop codon. This yielded 8452 positive TCR:pMHC combinations. The same strategy was then followed as described above to create the negative (non-binder) examples, either via the decoy approach or the permutation approach. For the decoy task, peptides were downloaded from IEDB51 (marl0_2025.csv version) and filtered to include peptide sequences between 8 and 12 amino acids long with “Object type”=“linear peptide”; peptides that were already included in the VDJdb human dataset were filtered out. MHC presentation for each of the MHC alleles in the VDJdb Human dataset was predicted using NetMHCPan4.141, and peptides predicted as “Strong Binders” were substituted for each positive TCR:pMHC example to create the decoy negative TCR:pMHC examples. For the Permutation task, the negative examples were created as described above. HLA sequences were obtained from HLA sequences were obtained from the IPD-IMGT / HLA62.enFoldX METHODStructure Prediction Pipeline

[0130] The pipeline involves three distinct steps which was run sequentially: structure look-up via template search and multiple sequence alignment (MSA), structure prediction via diffusion, and feature extraction from predicted structures and confidence metadata (FIG. 1A, FIG. 6). All unique TCR a and P chains was first collected and MHC sequences in the dataset and run the first stage (“MSA”), where AF3 queries the input sequence against multiple sequence databases to find similar sequences and collects co-evolutionary information for structure prediction63. This step was performed in a parallelized manner across all unique sequences (FIG. 6). The MSA output files was then processed and generate input files in JSON format for the second step: structure prediction. These JSON input files include TCRa, TCRP, MHC, and peptides as four separate sequence chains, and MSA and template search results from the previous step for all chains except the peptide, which is generally too short to yield any MSA or template matches. The protein folding step was then run in parallel across all input files. For each input file, AF3 then produces multiple structures corresponding to the number of random “seeds” and the -49- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282number of “samples”, which are set as hyperparameters in the AlphaFold input. The total number of predicted structures output by the model will be N seeds x 5 samples per input.Feature Extraction

[0131] For each of the predicted structures, both confidence features was extracted and structural features. To collect confidence features, the “confidence” JSON files output was leveraged by the AF3 container. These JSON files contain a slew of metrics that represent model confidence on the per structure level, per chain level, per residue level, per pairs of chains, or per pairs of residues.Confidence Metrics

[0132] For the metrics that are on the per structure level, per chain level, or per pairs of chains was included directly as components in the feature set. For residue or atom level confidence metrics, like pLDDT, the average was extracted over the pLDDT matrix of the entire structure as well as the average over specific chains (e.g. avg_plddt_MHC).

[0133] The predicted alignment error (PAE) matrix gives a per residue pair confidence score. Consider a PAE matrix of size NxN (for N input residues). The submatrix subPAE(X,Y) for chains X and Y was defined such that it was only considered the indexes from the matrix that correspond to the residues in chain X and chain Y respectively. When X=Y was called the pae_X, and it represents the alignment confidence of residues of a chain to other residues in the same chain. When X != Y was called the “interaction” PAE matrix, which represents the alignment confidence of residues of a chain to residues in another chain.

[0134] For a simple example, consider a protein complex with N=8 residues, where Chain A runs from indices 1-2, chain B from 3-4, chain C from 5-6, and chain D from 7-8:

[0135] Residue chain Ids: [A, A, B, B, C, C, D, D]

[0136] The PAE matrix will be a 8x8 matrix:-50- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282PAE1 1PAE12PAE1 3PAE14PAE15PAE16PAE17PAE18~ PAE2,I PAE22PAE23PAE24PAE28PAE26PAE27PAE2 8PAE34PAE32PAE33PAE34PAE38PAE36PAE37PAE38PAE41PAE42PAE43PAE44PAE4SPAE46PAE47PAE48PAESI1PAESI2PAE83PAE84PAE88PAE86PAE57PAE88PAESI PAE62PAE63PAE64PAE68PAE66PAE67PAE68PAE71PAE72PAE73PAE74PAE75PAE76PAE77PAE78_PAES 1PAE82PAE83PAE84PAE88PAE86PAE87PAE88_

[0137] The matrix subPAE(B,B) is thus defined as:P E33PAE34PAE43PAE44

[0138] Several features were defined for the classifier by collapsing this submatrix:

[0139] avg_pae_B = mean(subPAE(B,B))

[0140] std_pae_B = std(subPAE(B,B))

[0141] max_pae_B = max(subPAE(B,B))

[0142] min_pae_B = min(subPAE(B,B))

[0143] Similarly, the matrix subPAE(B, D) is defined as:~PAE3 7PAE38PAE47PAE48PAE73PAE83PAE74PAE84_

[0144] Again, several features was defined for the classifier by collapsing this submatrix:

[0145] avg_pae_interaction_B_D = avg(subPAE(B,D))

[0146] std_pae_interaction_B_D = std(subPAE(B,D))

[0147] max_pae_interaction_B_D = max(subPAE(B,D))

[0148] min_pae_interaction_B_D = min(subPAE(B,D))-51- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0149] As has been shown previously30, taking the average of the elements in these submatrices can be informative. Given that the TCR:pMHC complex has many crucial interfaces, the average iPAEs was included between all pairwise groups of chains (a, P, MHC, and peptide), as well as some subchains (CDR3a, CDR3P) as features in the model. This iPAE is also sometimes referred to as interface PAE or interaction PAE, and has mostly been leveraged in the context of predicting de novo proteins binding to protein targets64,65. It has also been used as a general predictor of protein complex formation66in place of traditional docking methods. As shown in the above example, other ways of collapsing these submatrices was included, including taking the minimum, maximum, or standard deviation of the entries. Importantly, while average iPAE has been shown to be a good predictor of classifying complexes as “binding” or “non-binding”, it seems to lack power to predict the precise affinity of binding complexes67.Structure Metrics

[0150] From the predicted 3D structures, AlphaFold provides contact probabilities between pairs of residues in the complex. The contact probability between two residues indicates the predicted probability that their representative atoms are within 8A of each other. Like in the PAE example above, the contact probability matrix was built, and aggregate informative features for the task. For example, one such feature in the model, max_contact_probs_CDR3 Alpha Peptide, is the maximum over the submatrix of contact probabilities between any residue on the CDR3a and any residue on the peptide.RMSD Metrics

[0151] For the cross-reactivity data set, an additional set of structural metrics was also computed. The predicted structure of the wild type TCR:pMHC was denoted complex as WT and the predicted structure of the complex with a single amino acid substitution in the peptide as MT. The Root Mean Square Deviation was computed between the two predicted structures, RMSD(MT,WT) by aligning the two proteins and then computing the distance between the representative atoms of each protein. Several variations of this approach was included, where chains that was aligned on can vary, and which chains was used to calculate the RMSD. All of the alignments and computations are done using -52- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282Biopython’s68PDB module69. For the other datasets, there is no WT comparator for a structure, so RMSD features are not utilized.Multiple Structures

[0152] For each input, AF3 produces multiple structures from multiple draws from a generative model. Thus, each feature that is extracted from an input TCR:pMHC complex has N copies corresponding to the N draws from AF3. Various ways was tried to integrate information across these N predicted structures, including (1) taking the average and standard deviation of the feature’s value across N structures and (2) using features from only the top ranked structure based on AF3’s ranking score. Performance was generally better for approach (1) over (2) although it varied by the dataset and the performance metric. The computational cost for prediction increases linearly with the number of input seeds, and thus there is a tradeoff between increased accuracy and increased complexity with leveraging additional structures.Classification

[0153] From the curated feature space, classification models was trained to distinguish between binding complexes and non-binding complexes. For the within-data set classification, 5-fold cross validation was performed, where the data was split into 5 folds and was further split each fold into 80% train and 20% test using scikit-leam’s StratifiedKFold function70. Several regression models was then trained including Ridge regression35(i.e. LogisticClassifier with L2 regularization) and XGBoost36as implemented by scikit-leam using the training data, and assessed predictive accuracy on the 20% held out test set. The final ROC curves are the average across the 5 folds. Other architectures / penalty paradigms was tested, and they achieved similar performance, with slight variability depending on the dataset. Ridge regression was chosen for the bulk of the analyses since it can help alleviate multicollinear features (which is likely an issue given the feature construction process) and showed the best performance on the generalization tasks. For the across-data set classification, the full feature set was trained on one dataset and then assess predictive accuracy on the other.Feature Importance Analysis-53- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0154] Once the classification models are trained, feature importance analysis was conducted using scikit-learn’s permutation importance function, which measures how much shuffling a feature worsens model performance70,71. Min-max normalization was then applied to these feature importances to transform their values into the range between 0 and 1. The feature space was clustered by taking the matrix of TCR:pMHC pairs vs. features and using the nearest-neighbors based Leiden clustering, as implemented in scanpy72.LARGE-LANGUAGE MODEL (BERTie)Architecture

[0155] BERTie model consists of a BERT encoder and a Multilayer Perceptron (MLP) layer. For each CDR3P-epitope pair a [CLS] token was prepended and concatenate the two sequences together adding a [SEP] token between them and at the end. The sequences are padded on the right to maximum length within each batch. The BERT encoder takes in the tokenized sequence <t0, tl, t2, ... tT>, where tO, tl, ... represent amino acids in CDR3P or epitope or the special characters, and T denotes the index of the last token in the combined sequence. The model then generates the hidden representation from the last layer for each token of the input sequence: fE: ti — hi, Vi G {0, 1, 2 ... T}. The hidden state was then passed of the [CLS] token hCLS to the Multi-Layer Perceptron (MLP) classifier layer which makes a prediction fD: hCLS — [0,1], where 1 indicates that the pair is predicted to be valid (binder) or 0 - invalid (non-binder). A weighted binary cross-entropy loss evaluates how close the predictions are to the truth and is back propagated through the model to update weight parameters in all previous layers in the direction that would decrease this loss. BERTie versions was developed where tokens are single amino acids and combinations of 2 or 3 amino acids. Where BERTie ensemble was referred to, predictions of BERTie models based on each encoding approach are ensembled (averaged) to arrive at the ensemble prediction.Datasets

[0156] BERTie is trained and evaluated on three publicly available datasets:Adaptive Biotechnologies’ ImmuneCODE database containing putative SARS-CoV-2-specific TCR sequences and their corresponding epitopes (MIRA)49; VDJdb: a collection of -54- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282paired TCR sequences and epitopes from previously published studies, largely in infectious diseases39; and McPAS-TCR50, a manually curated database of TCR sequences associated with various pathologies and antigens.

[0157] ImmuneCODE / MIRA49contains 143,938 unique CDR3P sequences from TCRs of over 1,400 subjects exposed to or infected with SARS-CoV-2 virus. In a series of multiplexed experiments based on MIRA assay technology49, these TCRs were co-cultured with various segments of SARS-CoV-2 viral peptidome and were found to pair with 363 unique SARS-CoV-2 epitopes. VDJdb39version 2022-03-30 is a compilation of 51,257 sequences of TCR CDR3s from different studies and 227 epitopes known to be recognized by these TCRs. McPAS50contains over 20,000 TCR sequences matching over 300 unique epitopes with minimal overlap with VDJdb. The three datasets were deduplicated, and only CDR3s of TCRP chains for human MHC class I examples were used as input. The described datasets only contain positive TCR-epitope pairs. Negative examples was generated by creating permutations of TCR and epitope sequences to create pairs not seen in the dataset (invalid / non-binder pairs).

[0158] BERTie was trained on 80% of the combined dataset and tested on the remaining 20%; however, where indicated in the main text, another version of BERTie was then retrained on all of the human MHC class I data.

[0159] Data Table 1: enfold Features for Predicting Binding.NAME FULL NAME DESCRIPTIONpLDDT predicted local distance per-atom confidence score to estimate the difference test reliability of a predicted structurePAE predicted alignment matrix of expected errors (in Angstroms) for error each residue if the predicted structure is aligned to the true structureiPAE Interface / interaction subset of the PAE matrix over regions of predicted alignment interface between chains in the complex error-55- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282pTM predicted template estimate of the accuracy of a prediction modeling score relative to the true structure ipTM interface predicted estimate of the accuracy of interfaces relative template modeling score to the true structurecontact probs Contact Probabilities matrix of interchain contact probabilities (<8A) between representative atoms from eachresidue derived from AF3's distogram head RMSD root mean square root mean square deviation (RMSD)deviation between residues in a predicted target structure and a predicted reference structure, where the reference is a known binding complex; pairwise RMSD between residues of pairs of predicted structures for the same complex (set of TCR:pMHC pr protein:MHC sequences), whereas predicted structures are aligned overall or on one chain and RMSD is calculated between other chains of the complexes. The resulting set of RMSD features is then aggregated for the ensemble by calculating the mean, minimum, maximum, and standard deviation across all pairs of predicted structures of the ensembleEmbeddings Protein language model Sequence embeddings generated with protein embeddings language models such as ESM2, ESM-C, Prot-Bert, TCR-Bert, and others, aggregated via mean-pooling or PCA (principal component analysis) or other aggregation methods. Embeddings are generated for individual chains as well as pairs of chains-56- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282(example: TCR beta + peptide, peptide + MHC) and the entire complex. Embedding- related features may include residuals between a chain’s embedding as part of the complex and that chain’s separateembedding.REFERENCES

[0160] 1. Waldman, A. D., Fritz, J. M. & Lenardo, M. J. A guide to cancer immunotherapy: from T cell basic science to clinical practice. Nat. Rev. Immunol. 20, 651— 668 (2020).

[0161] 2. Arstila, T. P. et al. A Direct Estimate of the Human aP T Cell Receptor Diversity. Science 286, 958-961 (1999).

[0162] 3. Ke§mir, C., Borghans, J. A. M. & de Boer, R. J. Diversity of Human aP T Cell Receptors. Science 288, 1135-1135 (2000).

[0163] 4. Cooper, M. D. & Alder, M. N. The evolution of adaptive immune systems. Cell 124, 815-822 (2006).

[0164] 5. Hacohen, N., Fritsch, E. F., Carter, T. A., Lander, E. S. & Wu, C. J.Getting personal with neoantigen-based therapeutic cancer vaccines. Cancer immunology research 1, 11-15 (2013).

[0165] 6. Schumacher, T. N. & Schreiber, R. D. Neoantigens in cancer immunotherapy. Science 348, 69-74 (2015).

[0166] 7. Roudko, V., Greenbaum, B. & Bhardwaj, N. Computational prediction and validation of tumor-associated neoantigens. Frontiers in Immunology 11, 27 (2020).

[0167] 8. Dash, P. et al. Quantifiable predictive features define epitope-specific T cell receptor repertoires. Nature 547, 89-93 (2017).-57- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0168] 9. Glanville, J. et al. Identifying specificity groups in the T cell receptor repertoire. Nature 547, 94-98 (2017).

[0169] 10. Karnaukhov, V. K. et al. Structure-based prediction of T cell receptor recognition of unseen epitopes using TCRen. Nat. Comput. Sci. 4, 510-521 (2024).

[0170] 11. Perez, M. A. S. et al. Predicting antigen-specificities of orphan T cell receptors from cancer patients with TCRpcDist. Adv. Sci. (Weinh.) 11, e2405949 (2024).

[0171] 12. Wang, A. et al. RACER-m leverages structural features for sparse T cell specificity prediction. Sci. Adv. 10, eadl0161 (2024).

[0172] 13. Yin, R. et al. TCRmodel2: high-resolution modeling of T cell receptor recognition using deep learning. Nucleic Acids Res. 51, W569-W576 (2023).

[0173] 14. Wu, F. et al. Fast and accurate modeling of TCR-peptide-MHC complexes using tFold-TCR. bioRxiv 2025.01.12.632367 (2025)doi:10.1101 / 2025.01.12.632367.

[0174] 15. Nielsen, M. et al. Lessons learned from the IMMREP23 TCR-epitope prediction challenge. Immunoinformatics (Amst.) 16, 100045 (2024).

[0175] 16. Deng, L. et al. Performance comparison of TCR-pMHC prediction tools reveals a strong data dependency. Front. Immunol. 14, 1128326 (2023).

[0176] 17. Luksza, M. et al. A neoantigen fitness model predicts tumour response to checkpoint blockade immunotherapy. Nature 551, 517-520 (2017).

[0177] 18. Luksza, M. et al. Neoantigen quality predicts immunoediting in survivors of pancreatic cancer. Nature 606, 389-395 (2022).

[0178] 19. Rojas, L. A. et al. Personalized RNA neoantigen vaccines stimulate T cells in pancreatic cancer. Nature 618, 144-150 (2023).-58- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0179] 20. Moris, P. et al. Current challenges for unseen-epitope TCR interaction prediction and a new perspective derived from image classification. Briefings in bioinformatics 22, bbaa318 (2021).

[0180] 21. McMaster, B., Thorpe, C., Ogg, G., Deane, C. M. & Koohy, H. Can AlphaFold’s breakthrough in protein structure help decode the fundamental principles of adaptive cellular immunity? Nat. Methods 21, 766-776 (2024).

[0181] 22. Bradley, P. Structure-based prediction of T cell receptor:peptide-MHC interactions. Elife 12, e82813 (2023).

[0182] 23. Zhu, T. et al. Accurate structure prediction of immune proteins using parameter-efficient transfer learning. bioRxiv (2024) doi: 10.1101 / 2024.11.13.621715.

[0183] 24. Wu, D. et al. Structural characterization and AlphaFold modeling of human T cell receptor recognition of NRAS cancer neoantigens. Sci. Adv. 10, eadq6150 (2024).

[0184] 25. Culka, M. et al. Predicting specificity of TCR-pMHC interactions using machine learning and biophysical models. bioRxiv 2025.04.04.647165 (2025) doi:10.1101 / 2025.04.04.647165.

[0185] 26. Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493-500 (2024).

[0186] 27. Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123-1130 (2023).

[0187] 28. Baek, M. etal. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871-876 (2021).

[0188] 29. Wu, R. etal. High-resolution de novo structure prediction from primary sequence. bioRxiv 2022.07.21.500999 (2022) doi: 10.1101 / 2022.07.21.500999.

[0189] 30. Bennett, N. R. et al. Improving de novo protein binder design with deep learning. Nat. Commun. 14, 2625 (2023).-59- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0190] 31. Johansen, K. H. et al. De novo designed pMHC binders facilitate T cell induced killing of cancer cells. bioRxiv (2024) doi: 10.1101 / 2024.11.27.624796.

[0191] 32. Householder, K. D. et al. De novo design and structure of a peptidecentric TCR mimic binding module. bioRxivorg (2024) doi: 10.1101 / 2024.12.16.628822.(0192] 33. Liu, B. et al. Design of high specificity binders for peptide-MHC-I complexes. bioRxivorg 2024.11.28.625793 (2024) doi:10.1101 / 2024.11.28.625793.

[0193] 34. Visani, G. M. et al. T-cell receptor specificity landscape revealed through de novo peptide design. arXiv [q-bio.QM] (2025).

[0194] 35. Hoerl, A. E. & Kennard, R. W. Ridge regression: Applications to nonorthogonal problems. Technometrics 12, 69-82 (1970).

[0195] 36. Chen, T. & Guestrin, C. Xgboost: A scalable tree boosting system, in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining 785-794 (2016).

[0196] 37. Dens, C., Laukens, K., Bittremieux, W. & Meysman, P. The pitfalls of negative data bias for the T-cell epitope specificity challenge. Nature Machine Intelligence 5, 1060-1062 (2023).

[0197] 38. Wallner, B. AFsample: improving multimer prediction with AlphaFold using massive sampling. Bioinformatics 39, (2023).

[0198] 39. Shugay, M. etal. VDJdb: a curated database of T-cell receptor sequences with known antigen specificity. Nucleic Acids Res. 46, D419-D427 (2018).

[0199] 40. Andreatta, M. & Nielsen, M. Gapped sequence alignment using artificial neural networks: application to the MHC class I system. Bioinformatics 32, 511-517 (2016).

[0200] 41. Reynisson, B., Alvarez, B., Paul, S., Peters, B. & Nielsen, M.NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen-60- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data. Nucleic acids research 48, W449-W454 (2020).

[0201] 42. Motmaen, A. etal. Peptide-binding specificity prediction using finetuned protein structure prediction networks. Proc. Natl. Acad. Sci. U. S. A. 120, e2216697120 (2023).

[0202] 43. Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) (eds. Burstein, J., Doran, C. & Solorio, T.) 4171-4186 (Association for Computational Linguistics, Minneapolis, Minnesota, 2019).

[0203] 44. Meynard-Piganeau, B., Feinauer, C., Weigt, M., Walczak, A. M. & Mora, T. TULIP: A transformer-based unsupervised language model for interacting peptides and T cell receptors that generalizes to unseen epitopes. Proc. Natl. Acad. Sci. U. S. A. 121, e2316401121 (2024).

[0204] 45. Cai, M., Bang, S., Zhang, P. & Lee, H. ATM-TCR: TCR-Epitope binding affinity prediction using a multi-head self-attention model. Front. Immunol. 13, 893247 (2022).

[0205] 46. Jensen, M. F. & Nielsen, M. NetTCR 2.2 - Improved TCR specificity predictions by combining pan- and peptide-specific training strategies, loss-scaling and integration of sequence similarity. (2024) doi:10.7554 / elife.93934.2.

[0206] 47. Springer, I., Tickotsky, N. & Louzoun, Y. Contribution of T cell receptor alpha and beta CDR3, MHC typing, V and J genes to peptide binding prediction. Front. Immunol. 12, 664514 (2021).

[0207] 48. wwPDB consortium. Protein Data Bank: the single global archive for 3D macromolecular structure data. Nucleic Acids Res. 47, D520-D528 (2019).-61- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0208] 49. Nolan, S. et al. A large-scale database of T-cell receptor beta sequences and binding associations from natural and synthetic exposure to SARS-CoV-2. Frontiers in Immunology 16, 1488851 (2025).

[0209] 50. Tickotsky, N., Sagiv, T., Prilusky, J., Shifrut, E. & Friedman, N.McPAS-TCR: a manually curated catalogue of pathology-associated T cell receptor sequences. Bioinformatics 33, 2924-2929 (2017).

[0210] 51. Vita, R. et al. The Immune Epitope Database (IEDB): 2024 update. Nucleic Acids Res. 53, D436-D443 (2025).

[0211] 52. Roney, J. P. & Ovchinnikov, S. State-of-the- Art Estimation of Protein Model Accuracy Using AlphaFold. Phys. Rev. Lett. 129, 238101 (2022).

[0212] 53. Sasse, A. etal. Reconstructing the sequence specificities of RNA-binding proteins across eukaryotes. bioRxiv (2024).

[0213] 54. Mayer, A. et al. How different are self and nonself? arXiv [q-bio.CB] (2022).

[0214] 55. Mallein, B., Paparella, F., Schertzer, E. & Talyigas, Z. Selection of the fittest or selection of the luckiest: the emergence of Goodhart’s law in evolution. arXiv [q-bio.PE] (2025).

[0215] 56. Koltun, V. & Hafner, D. The h-index is no longer an effective correlate of scientific reputation. PLoS One 16, e0253397 (2021).

[0216] 57. Dill, K. A. & Chan, H. S. From Levinthal to pathways to funnels. Nat. Struct. Mol. Biol. 4, 10-19 (1997).

[0217] 58. Arteaga, G. Y., Schon, T. B. & Pielawski, N. Hallucination detection in LLMs: Fast and memory-efficient fine-tuned models. arXiv [cs.LG] (2024).

[0218] 59. Ahdritz, G. etal. OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Nat. Methods 21, 1514-1524 (2024).-62- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0219] 60. Heather, J. M. etal. Stitchr: stitching coding TCR nucleotide sequences from V / J / CDR3 information. Nucleic acids research 50, e68-e68 (2022).

[0220] 61. Uniprot-Consortium. UniProt : the universal protein knowledgebase in 2021. Nucleic Acids Research 49, 480-489 (2021).

[0221] 62. Robinson, J., Barker, D. J. & Marsh, S. G. E. 25 years of the IPD- IMGT / HLA database. HI A 103, el 5549 (2024).

[0222] 63. Morcos, F. etal. Direct-coupling analysis of residue coevolution captures native contacts across many protein families. Proc. Natl. Acad. Sci. U. S. A. 108, E1293-301 (2011).

[0223] 64. Glogl, M. etal. Target-conditioned diffusion generates potent TNFR superfamily antagonists and agonists. Science 386, 1154-1161 (2024).

[0224] 65. Muratspahic, E. etal. De novo design of miniprotein agonists and antagonists targeting G protein-coupled receptors. bioRxiv 2025.03.23.644666 (2025) doi : 10.1101 / 2025.03.23.644666.

[0225] 66. Yin, R., Feng, B. Y., Varshney, A. & Pierce, B. G. Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants. Protein Science 31, e4379 (2022).

[0226] 67. Cotet, T.-S. etal. Crowdsourced protein design: Lessons from the Adaptyv EGFR binder competition. bioRxiv (2025) doi: 10.1101 / 2025.04.17.648362.

[0227] 68. Cock, P. J. A. et al. Biopython: freely available Python tools for computational molecular biology and bioinformatics. Bioinformatics 25, 1422 (2009).

[0228] 69. Hamelryck, T. & Manderick, B. PDB file parser and structure class implemented in Python. Bioinformatics 19, 2308-2310 (2003).

[0229] 70. Pedregosa, F. et al. Scikit-learn: Machine learning in Python, the Journal of machine Learning research 12, 2825-2830 (2011).-63- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0230] 71. Breiman, L. Random forests. Machine learning 45, 5-32 (2001).

[0231] 72. Wolf, F. A., Angerer, P. & Theis, F. J. SCANTY: large-scale single-cell gene expression data analysis. Genome biology 19, 1-5 (2018).Example 2: Use of enFoldX for peptide:MHC Binding

[0232] We will apply enFoldX pipeline and method to the peptide:MHC binding problem, including MHC Class I and Class II. Briefly, we will run sequentially the following four steps: structure look-up via MSA and template search for the MHC allele, structure prediction via diffusion, feature extraction from predicted structures and ensembles of predicted structures, training a classifier on available peptide:MHC binding data, and applying it to new peptide:MHC combinations to predict binding. The binding prediction (inference) also includes the MSA and template search for the complex we infer for, structure prediction, feature extraction, and then passing the features to the classifier and evaluating whether the predicted probability of binding is above or below a binding threshold.

[0233] The classes of features defined for the peptide:MHC (Class I and Class II) binding prediction are the same as for TCR:pMHC and include Confidence metrics (pLDDT, PAE, iPAE, pTM, ipTM), Contact probabilities, and RMSD. The pairwise interaction metrics (iPAE, ipTM, and contact probabilities) are defined between each two of the three components: (1) for peptide:MHC Class I: MHC a chain, peptide, and optionally MHC p2-microglobulin chain and (2) for peptide:MHC Class II: MHC a chain, peptide, and optionally MHC P chain. In addition, we will define and optionally include in the classifier a set of these classes of features for subsets of amino acids at the critical interaction interfaces, such as the MHC cleft and the peptide. (This is akin to CDR3-specific features for TCR:pMHC prediction).

[0234] Moreover, we will apply both the “best predicted” structure as well as the structure ensemble approach to the peptide:MHC (Class I and Class II) predictions. We will train the peptide:MHC classifier on datasets including but not limited to IEDB and IEDB-EL, PRIDE, SysteMIC Atlas, MassIVE, HLA Ligand Atlas, NetMHCpan datasets, MHCFlurry datasets, MHCNuggets, as well as other immunopeptidomic datasets.-64- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282Example 3: System and Methods for Determining Values Predicting Probabilities of Protein-Protein Interactions in T-Cell Receptor (TCR)-Major Histocompatibility Complex (MHC)-Peptide Complexes

[0235] Referring now to FIG. 14, depicted is a block diagram of a system 100 for determining values predicting probabilities of protein-protein interactions in T-cell receptor (TCR)-major histocompatibility complex (MHC)-peptide complexes or peptide-MHC complexes. In an overview, the system 100 can include at least one data processing system 105, at least one computing device 110, at least one sequencing device 120, and at least one database 170, communicably coupled via at least one network 115. The data processing system 105 can include at least one data aggregator 125, at least one model trainer 130, at least one fold constructor 135, at least one feature extractor 140, at least one binding classifier 145, and an output evaluator 150, among others. The data processing system 105 can include at least one folding model 155, at least one protein language model 160, and at least one classifier model 165, among others. The system 100 may be used to implement the functionalities detailed herein Examples 1 and 2. Each component can be implemented using hardware, including one or more processors coupled with memory, or a combination of hardware and software (e.g., as described in Example 4).

[0236] The data processing system 105 can be any computing system including one or more processors coupled with memory and software to execute instructions herein. The data processing system 105 can be capable of performing the various processes and tasks described herein. The data processing system 105 can be in communication with the components of the system 100 (e.g., computing device 110 or the sequencing device 120). The data processing system 105 can be situated, located, or otherwise associated with at least one server group. The server group can correspond to a data center, a branch office, or a site at which one or more servers corresponding to the data processing system 105 are situated.

[0237] The data processing system 105 can include or execute any number of modules, processes, components, or subcomponents to perform the various processes and tasks described herein. The data aggregator 125 can retrieve sequences corresponding to an MHC polypeptide, a peptide epitope, a TCR-a polypeptide, a TCR-P polypeptide, and -65- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282training data, among others. The model trainer 130 can initialize, train, or otherwise establish any one of the machine learning (ML) models, such as the classifier model 165. The fold constructor 135 can create a plurality of candidate quaternary structures using one or more input sequences and the folding model 155. The feature extractor 140 can determine one or more features for each candidate quaternary structure, including confidence metrics, structure metrics, root mean square deviation (RMSD) metrics, embeddings, or aggregated values thereof. The binding classifier 145 can provide the generated features as input to the classifier model 165 to determine at least one value indicating interaction, binding probability, or classification for a candidate quaternary structure. The output evaluator 150 can provide output for provision based on the results from the classifier model 165.

[0238] The folding model 155 can be any type of structure prediction algorithm or model to generate one or more candidate quaternary structures from one or more biomolecular sequences. The folding model 155 can be, for example, AlphaFold (e.g., AlphaFold 3), RoseTTAFold, OpenFold, Boltz-1, OmegaFold, or ESMFold, among others. In general, the folding model 155 can include at least one input, at least one output, and a set of weights or parameters relating the input with the output. The input can include one or more sequences corresponding to an MHC polypeptide, a peptide epitope, a TCR-a polypeptide, or a TCR-P polypeptide. The output can include one or more candidate quaternary structures and, in some embodiments, one or more scores indicating predicted structural accuracy.

[0239] The protein language model 160 can be any type of machine learning algorithm or model to generate embeddings for latent representations from sequence inputs of candidate structures. In some embodiments, the protein language model 160 can include protein language models, such as ESM2, ESM-C, Prot-Bert, TCR-Bert, among others. The sequence embeddings can be generated for an MHC polypeptide, a peptide epitope, a TCR-a polypeptide, or a TCR-P polypeptide, among others, or any combination thereof. The resulting embeddings can be aggregated using mean-pooling, principal component analysis (PC A), or other aggregation methods, and embedding-related features can further include residuals between a chain embedding as part of the complex and a separate embedding for-66- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282that chain alone. In general, the input to the protein language model 160 can include one or more sequences for candidate quaternary structures. The output can include one or more embeddings or learned feature representations for use by the classifier model 165.

[0240] The classifier model 165 can be any type of machine learning algorithm or model configured to determine a value indicating a probability of interaction, binding likelihood, or binding classification for a candidate peptide-MHC complex or TCR-peptide-MHC complex. The classifier model 165 can be, for example, a clustering algorithm, a support vector machine (SVM), a Naive Bayesian classifier, a decision tree, a random forest classifier, a logistic regression model, a ridge regression model, a gradient boosting model, or a deep learning based model (e.g., a convolutional neural network (CNN), a recurrent neural network (RNN), a feed forward neural network (FNN), a long-short term memory (LSTM) network, an autoencoder, a transformer model, or a multilayer perceptron), among others. In general, the classifier model 165 can include at least one input, at least one output, and a set of weights relating the input with the output. The input can include the features (e.g., one or more confidence metrics, one or more structure metrics, one or more RMSD metrics, one or more embeddings, or aggregated values) for one or more candidate quaternary structures. The output can include at least one value indicating probability of interaction, binding likelihood, or classification as binding or non-binding for the candidate quaternary structure.

[0241] The sequencing device 120 may be any device to perform genetic sequencing on a gene segment (e.g., deoxyribonucleic acid (DNA) or ribonucleic acid (RNA) sample) in a sample taken from a subject and generate sequencing datasets using the genetic sequencing. The genetic sequencing carried out may be a high throughput, massively parallel sequencing technique (sometimes herein referred to as next-generation sequencing), such as whole genome sequencing (WGS), pyrosequencing, Reversible dye-terminator sequencing, SOLiD sequencing, Ion semiconductor sequencing, and Helioscope single molecule sequencing, among others. In some embodiments, the sequencing device 120 may generate sequencing data for any one or more of: a MHC, peptide epitope, TCR-a, or TCR-P, among others. Using the genetic sequencing data, the sequencing device 120 may generate genomic dataset profiling the subject. In generating the genomic dataset, the-67- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282sequencing device 120 may execute read alignment, variant calling, and gene expression quantification, among others. The sequencing device 120 may use the gene sequencing to generate a sequencing dataset. The sequencing dataset may be maintained using one or more files according to a format (e.g., FASTQ, BAM, SAM, BCL, or VCF formats).

[0242] The computing device 110 can be any device comprising of one or more processors coupled with memory and software and capable of providing an output projection image. The computing device 110 can be any device comprising of one or more processors coupled with memory and software and capable of providing an output to a user. The computing device 110 can be associated with a clinician, researcher, technician, administrator, or other user. The computing device 110 can include at least one display to present information from the data processing system 105. The computing device 110 can be in communication with the data processing system 105, the sequencing device 120, and the database 170 via the network 115.

[0243] The database 170 (sometimes herein referred to as a data repository) can store and maintain various resources and data associated with the data processing system 105, the computing device 110, or the sequencing device 120. The database 170 can include software to manage the collection of data, such as a database management system (DBMS), query engine, storage engine, or transaction manager, among others. The database 170 can store, maintain, or house data associated with the system 100 using one or more data structures. For example, the database 170 can include training datasets to train the classifier model 165, one or more biomolecular sequences, candidate quaternary structures, extracted features, model parameters, output values, threshold values, and associations between predicted values and candidate quaternary structures.

[0244] Referring now to FIGs. 15A and 15B, depicted are block diagrams of a process 200 for training classifier models in the system 100 for determining values predicting probabilities of protein-protein interactions. Starting with FIG .15A, under the process 200, the data aggregator 125 can access the database 170 to retrieve, obtain, or otherwise receive a training dataset 205 to train the classifier model 165. The classifier model 165 may be trained to predict protein-protein interactions or binding probabilities for peptide-major histocompatibility (pMHC) complex or a T-cell receptor (TCR) peptide- -68- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282major histocompatibility complex (TCR:pMHC) complex. The training dataset 205 can include a set of examples. Each example can be associated with at least one complex structure 202.

[0245] The example may be derived from a previous sample from a mammal subject, such as a human or a mouse, among others. The mammal subject may be at risk of or diagnosed with a pathogen or cancer. The example may be retrieved from the database 170 (e.g., storing experimental data or genomic sequencing data). The cancer may include, for example, at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof. The pathogen may include, for example, viral pathogens, such as Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), among others.

[0246] In some embodiments, the complex structure 202 may correspond to a peptide-major histocompatibility (pMHC) complex. The complex structure 202 may include at least one peptide epitope 204 and at least one MHC 206. The peptide epitope 204 may originate from at least one of a pathogen, a self-antigen, or cancer in the subject. The peptide epitope 204 may include at least one of a class I epitope or a class II epitope. In some embodiments, the class I epitope may have a length of 8-11 amino acids. In some embodiments, the class II epitope may have a length of 12-25 amino acids. The MHC 206 may include at least one of an MHC class I or an MHC class II. The MHC polypeptide sequence may include at least one of HLA-A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP, among others. The MHC class I may include an alpha chain including an al domain, a2 domain, and a3 domain. In some embodiments, the MHC class I may also-69- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282include a P2-microglobulin chain. The MHC class II may include an alpha chain having an al domain and a2 domain. In some embodiments, the MHC class II may also include a beta chain having an pi domain and P2 domain.

[0247] In some embodiments, the complex structure 202 may correspond to a T-cell receptor (TCR) peptide-major histocompatibility complex (TCR:pMHC) complex. The complex structure 202 may include at least one TCR 208, in addition to the peptide epitope 204 and the MHC 206. The TCR 208 may include a TCRa polypeptide and a TCRP polypeptide that are isolated or obtained from at least one of CD8+ T cell or a CD4+ T cell, among others. In some embodiments, the complex structure 202 may include at least one T-cell 212 from which the TCR 208 is isolated or obtained. The T-cell 212 may include, for example, at least one of CD8+ T cell or a CD4+ T cell. In some embodiments, T-cell 212 may include the TCR 208 to be bound with the peptide epitope 204 and the MHC 206. The T-cell 212 can be specific for, cross-reactive with, or non-reactive to a given peptide epitope 204 presented by the MHC 206. In some embodiments, the complex structure 202 may include at least one antigen presenting cell (APC) 210. The APC 210 may include, for example, a dendritic cell, macrophage, B cell, tumor cell, infected cell, engineered cell, or another cell presenting the peptide epitope 204 and the MHC 206.

[0248] With retrieval of the training data 205, the data aggregator 125 can extract, select, or otherwise identify at least one example from the training data 205. In the training dataset 205, the example can include a respective set of sequences 215A-N (hereinafter referred to as sequences 215) and at least one corresponding label 220 for the complex structure 202. The set of sequences 215 may include one or more of: a first sequence corresponding to a polypeptide of the MHC 206; a second sequence corresponding to the peptide epitope 204; a third sequence corresponding to the TCR-a polypeptide of the TCR 208; or the fourth sequence corresponding to the TCR-P polypeptide of the TCR 208, among others. The set of sequences 215 may be generated using next-generation sequencing (NGS) (e.g., by the sequencing device 120). When training the classifier model 165 for peptide-major histocompatibility (pMHC) complex, the set of sequences 215 may include the first sequence corresponding to the MHC 206 and the second sequence corresponding to the peptide epitope 204. When training the classifier model 165 for-70- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282TCR:pMHC complex, the set of sequences 215 may include the first sequence corresponding to a polypeptide of the MHC 206; the second sequence corresponding to the peptide epitope 204; the third sequence corresponding to the TCR-a polypeptide of the TCR 208; and a fourth sequence corresponding to the TCR-P polypeptide of the TCR 208.

[0249] In some embodiments, the data aggregator 125 may identify one or more of: (i) a first input file identifying the first sequence corresponding to the polypeptide of the MHC 206, (ii) a second input file identifying the second sequence corresponding to the peptide epitope 204, (iii) a third input file identifying the third sequence corresponding to the TCR-a polypeptide of the TCR 208, or (iv) a fourth input file identifying the fourth sequence corresponding to the TCR-P polypeptide of the TCR 208, among others. The input files may be of any suitable format, such as FASTQ, BAM, SAM, BCL, or VCF formats. When training the classifier model 165 for pMHC complexes, the data aggregator 125 may identify the first input file identifying the first sequence corresponding to the MHC 206 and the second input file identifying the second sequence corresponding to the peptide epitope 204. When training the classifier model 165 for TCR:pMHC complexes, the data aggregator 125 may identify the first input file identifying the first sequence corresponding to the polypeptide of the MHC 206, the second input file identifying the second sequence corresponding to the peptide epitope 204, the third input file identifying the third sequence corresponding to the TCR-a polypeptide of the TCR 208 and the fourth input file identifying the fourth sequence corresponding to the TCR-P polypeptide of the TCR 208.

[0250] In addition, the label 220 may identify or indicate the complex structure 202 associated with the example as one of binding or non-binding. When training the classifier model 165 for pMHC complexes, the label 220 may identify or indicate the pMHC complex corresponding to the complex structure 202 as one of binding or non-binding. When training the classifier model 165 for TCR:pMHC complexes, the label 220 may may identify or indicate the TCR:pMHC complex corresponding to the complex structure 202 as one of binding or non-binding. The label 220 can be derived or measured from an experimentally measured binding affinity, effective affinity, monovalent affinity, cellular activation readout, cytotoxicity readout, expansion readout, peptide presentation readout, or-71- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282assay result, among others. The label 220 may be generated or manually created based on the results of the experimental tests.

[0251] With retrieval, the fold constructor 135 may generate or create a set of candidate quaternary structures 230A-N (hereinafter generally referred to as candidate quaternary structures 230 using the set of sequences 215. Any number of candidate quaternary structures 230 may be created, ranging for example between 5 and 1,000. In some embodiments, the fold constructor 135 may create the set of candidate quaternary structures 230 using the set of input files. In some embodiments, the fold constructor 135 may generate a set of output files corresponding to the set of candidate quaternary structures 230, using the set of input files. Each output file may define or include a respective candidate quaternary structure 230. When training the classifier model 165 for pMHC complexes, each candidate quaternary structure 230 may define or include (i) a respective MHC structure and (ii) a respective peptide epitope structure for forming the pMHC complex. When training the classifier model 165 for TCR:pMHC complexes, each candidate quaternary structure 230 may define or include (i) a respective MHC structure, (ii) a respective peptide epitope structure, (iii) a respective TCR-a structure, and (iv) a respective TCR-P structure, for the TCR:pMHC complex.

[0252] In some embodiments, the fold constructor 135 may apply or execute the folding model 155 using the set of sequences 215 or the set of input files. To execute, the fold constructor 135 may feed or provide the set of sequences 215 (or the set of input files defining the sequences 215) as input to the folding model 155. Based on executing the folding model 155, the fold constructor 135 may generate the set of candidate quaternary structures 230. In some embodiments, the fold constructor 135 may perform pre-processing on the set of sequences 215, prior to input to the folding model 155. For example, the fold constructor 135 may perform formatting, deduplication, or sequence validation for one or more polypeptides defined by the set of sequences 215. In some embodiments, the fold constructor 135 may perform template-search and multiple sequence alignment (MSA) on the set of sequences 215, in conjunction with input into the folding model 155. The template search may be used to identify reference sequences from the database 170 corresponding to the set of sequences 215. The MSA may be used to identify evolutionary-72- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282variation and conservation across sequence positions in the set of sequences 215. The information obtained through template search and MSA may be provided by the fold constructor 135 as input to the folding model 155.

[0253] Moving onto FIG. 15B, for each candidate quaternary structure 230, the feature extractor 140 may calculate, determine, or otherwise generate a set of features 235A-N (hereinafter generally referred to as features 235). Each set of features 235 may correspond to a respective candidate quaternary structure 230. The set of features 235 may identify or include two or more of: (i) one or more confidence metrics, (ii) one or more structure metrics, (iii) one or more root mean square deviation (RMSD) metrics, or (iv) a set of embedding, among others. In some embodiments, the set of features 235 for each candidate quaternary structure 230 may be maintained as part of the training dataset 205 for use by the model trainer 130. In some embodiments, the feature extractor 140 may generate the set of features 235 from the respective candidate quaternary structure 230.

[0254] In some embodiments, the confidence metrics in the set of features 235 may include one or more of: (i) a predicted local distance difference test (pLDDT) metric indicating a predicted local reliability of at least a portion of the candidate quaternary structure 230; (ii) a predicted alignment error (PAE) metric corresponding to a matrix of estimated positional errors for each residue of the candidate quaternary structure 230 to a reference true structure; (iii) an interface predicted alignment error (iPAE) metric corresponding to a subset of the matrix of estimated position errors over regions of interface between chains in the candidate quaternary structure 230; (iv) a predicted template modeling (pTM) metric corresponding to an estimate of accuracy of predicted candidate quaternary structure 230 relative to the reference true structure; or (v) an interface predicted template modeling (ipTM) metric corresponding to an estimate of accuracy of interfaces of the predicted candidate quaternary structure 230 relative to interfaces of the reference true structure, among others. The feature extractor 140 may determine the confidence metrics from confidence based on executing the folding model 155 using the set of sequences 215 in generating the corresponding candidate quaternary structure 230.

[0255] In some embodiments, the structural metrics in the set of features 235 may indicate probabilities of contact between amino acid residues represented in the-73- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282corresponding candidate quaternary structure 230. When training the classifier model 165 for pMHC complexes, the structure metrics may indicate a probability of contact between at least one amino acid residue of the MHC structure and at least one amino acid residue of the peptide epitope structure. When training the classifier model 165 for TCR:pMHC complexes, the structure metrics may indicate a probability of contact between two or more of (i) at least one MHC amino acid residue of the MHC structure, (ii) at least one peptide epitope amino acid residue of the peptide epitope structure, (iii) at least one TCR-a amino acid residue of the TCR-a structure, (iv) at least one TCR-P amino acid residue of the TCR-P structure, (v) at least one amino acid residue in a complementarity determining region 3 (CDR3) portion of the TCR-a structure, or (vi) at least one amino acid residue in a CDR3 portion of the TCR-P structure. In some embodiments, the feature extractor 140 may determine the structure metrics from contact probabilities or distance distributions output by the folding model 155.

[0256] In some embodiments, the RMSD metrics in the set of features 235 may indicate a measure of structural difference between positions of atoms or residues of the candidate quaternary structure 230 to another quaternary structure. In some embodiments, the feature extractor 140 may calculate or determine the RMSD metric based on a deviation between a corresponding candidate quaternary structure 230 and at least one reference quaternary structure. The reference quaternary structure may correspond to a complex (e.g., pMHC or TCR:pMHC complex) that was previously identified as binding (e.g., through experimental measurements). For example, the feature extractor 140 may align the candidate quaternary structure 230 with a reference quaternary structure and may calculate the RMSD metric between corresponding residues, atoms, chains, or substructures based on a comparison. The alignment may be performed on all chains of the complex structure 202 or on a subset of chains. When training the classifier model 165 for pMHC complexes, the RMSD metric may be determined based on a deviation between the candidate quaternary structure 230 and at least one reference quaternary structure for the pMHC complex. When training the classifier model 165 for TCR:pMHC complexes, the RMSD metric may be determined based on a deviation between the candidate quaternary structure 230 and at least one reference quaternary structure for the TCR:pMHC complex.-74- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0257] In some embodiments, the feature extractor 140 may calculate or determine the RMSD metric based on a set of deviations. Each deviation may be between any two of the set of candidate structures 230 generated using the set of sequences 215 for the same complex structure 202. Each pair of candidate quaternary structures 230 may be aligned overall, aligned on one or more selected chains, or aligned on one or more subcomplexes, and the feature extractor 140 may determine the RMSD values for one or more other chains, residues, or interface regions. For example, the feature extractor 140 may align two candidate quaternary structures 230 on the MHC chain and calculate RMSD for the peptide epitope, align on the pMHC portion and calculate RMSD for the TCR-a or TCR-P chain, or align on one TCR chain and calculate RMSD on the other TCR chain or on a CDR3 region. The feature extractor 140 may aggregate the resulting set of pairwise RMSD values to determine one or more RMSD metrics, such as a mean, minimum, maximum, standard deviation, or range across the pairwise RMSD values.

[0258] In some embodiments, the set of embeddings in the set of features 235 may be determined or generated based on executing the protein language model 160 using the set of sequences 215 for the complex structure 202. The set of embeddings may correspond to a lower-dimensional representation of latent features in the sequence information for the complex structure 202. In processing the set of sequences 215 using the protein language model 160, the feature extractor 140 may perform mean pooling, principal component analysis, among others, and may determine residual features among the sequences. When training the classifier model 165 for pMHC complexes, the set of embeddings may include embeddings for the MHC 206 and the peptide epitope 204 for the pMHC complex. When training the classifier model 165 for TCR:pMHC complexes, the set of embeddings may include embeddings for the MHC 206, the peptide epitope 204, the TCR-a polypeptide of the TCR 208, and the TCR-P polypeptide of the TCR 208 for the TCR:pMHC complex.

[0259] With generation of the set of features 235, the feature extractor 140 may perform a structural ensemble approach or a best structure approach. Under the structural ensemble approach, the feature extractor 140 may aggregate the corresponding set of features 235 across the set of candidate quaternary structures 230. In aggregating, the feature extractor 140 may determine a mean, standard deviation, minimum, maximum, or-75- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282range, for each feature 235 across the candidate quaternary structures 230. Based on the feature extractor 140 may generate an aggregated set of features 235 (sometimes herein referred to as an ensemble of features). Each feature 235 may include an aggregation of the features 235 across the candidate quaternary structures 230. The aggregated set of features 235 may be used to further evaluate the complex structure 202.

[0260] Under the best structure approach, the feature extractor 140 may identify or select at least one candidate quaternary structure 230 from the set of candidate quaternary structures 230. From executing the folding model 155 using the set of sequences 215, the feature extractor 140 may calculate, determine, or otherwise generate a respective score for each candidate quaternary structure 230. The score may indicate a prediction accuracy of the corresponding candidate quaternary structure 230. Based on the scores, the feature extractor 140 may select the candidate quaternary structure 230 from the set of candidate quaternary structures 230. For instance, the feature extractor 140 may select the candidate quaternary structure 230 with the highest accuracy score. With the selection, the feature extractor 140 may identify the set of features 325 corresponding to the candidate quaternary structure 230. The set of features 235 corresponding to the candidate quaternary structure 230 (e.g., with the best score) may be used to further evaluate the complex structure 202.

[0261] The model trainer 130 may initialize, train, and establish the classifier model 165 using the training dataset 205. In some embodiments, the model trainer 130 may instantiate the classifier model 165 with a set of weights or parameters having random, predefined, or default values. To train, the model trainer 130 may apply or execute the classifier model 165 using the set of features 235 (e.g., the aggregated set of features 235 across the set of candidate quaternary structure 230 or the set of features 235 corresponding to the selected candidate quaternary structure 230). The model trainer 130 may feed or provide the set of features 235 as input to the classifier model 165. In feeding, the model trainer 130 may process the set of features 235 using the set of weights of the classifier model 165. Based on execution of the classifier model 165, the model trainer 130 may calculate, generate, or otherwise determine at least one value 240. The value 240 may indicate a probability of protein-protein interaction or a probability of binding in the complex structure 202. When training the classifier model 165 for pMHC complexes, the-76- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282value 240 may indicate a probability of interaction between the MHC 206 and the peptide epitope 204 in the pMHC complex. When training the classifier model 165 for TCR:pMHC complexes, the value 240 may indicate a probability of interaction among the MHC 206, the peptide epitope 204, the TCR-a polypeptide of the TCR 208, and the TCR-P polypeptide of the TCR 208 in the TCR:pMHC complex.

[0262] In some embodiments, the model trainer 130 may determine or generate a classification for the complex structure 202 based on the value 240. To generate, the model trainer 130 may compare the value 240 with a threshold value. The threshold value may define a value at which to identify the complex structure 202 corresponding to the input as to the classifier model 165 as predicted to be binding or non-binding. The threshold value may be predefined at a set value or may be determined from experimental measurements. If the value 240 satisfies (e.g., greater than or equal to) the threshold value, the model trainer 130 may determine the classification of the complex structure 202 as binding. When training the classifier model 165 for pMHC complexes, the model trainer 130 may determine the classification indicating the pMHC complex as predicted to be binding.When training the classifier model 165 for TCR:pMHC complexes, the model trainer 130 may determine the classification indicating the TCR:pMHC complex as predicted to be binding. If the value 240 does not satisfy (e.g., is less than) the threshold value, the model trainer 130 may determine the classification of the complex structure 202 as non-binding. When training the classifier model 165 for pMHC complexes, the model trainer 130 may determine the classification indicating the pMHC complex as predicted to be non-binding. When training the classifier model 165 for TCR:pMHC complexes, the model trainer 130 may determine the classification indicating the TCR:pMHC complex as predicted to be nonbinding.

[0263] With the generation of the value 240 or the classification, the model trainer 130 may determine, calculate, or otherwise generate at least one loss metric 245 based on a comparison between the value 240 and the label 225. The loss metric 245 may indicate a degree of deviation of the output of the classifier model 165 from the label 225. The loss metric 245 may be determined in accordance with one or more loss functions, such as a norm loss, mean squared error, quadratic loss, cross-entropy loss, logistic loss, or another-77- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282supervised learning loss function. Using the loss metric 245, the model trainer 130 may modify, change, or otherwise update at least one of the set of weights of the classifier model 165. The updating may be in accordance with a back propagation or optimization function, depending on the type of classifier model 165. The optimization may include, for example, stochastic gradient descent, adaptive moment estimation (Adam), adaptive gradient methods, coordinate descent, boosting updates, or another optimization routine. By updating the classifier model 165 using the loss metric 245, the value 240 and the resulting predicted classification generated by the classifier model 165 may more accurately correspond to the labels 225 of the training examples. The model trainer 130 may iteratively train the classifier model 165 until convergence. Upon convergence, the model trainer 130 may store and maintain the trained classifier model 165 for use in inference on new pMHC complexes or new TCR:pMHC complexes.

[0264] Referring now to FIG. 16, depicted is a block diagram of a process 300 for creating candidate complex structures in the system 100 for determining values predicting probabilities of protein-protein interactions. Under the process 300, the data aggregator 125 receive, obtain, or otherwise retrieve a set of sequences 320A-N (hereinafter generally referred to as sequences 320) from a database 170 or a sequencing device 120. The set of sequences 320 may correspond to at least one candidate complex structure 302. In some embodiments, the set of sequences 320 may be retrieved from the database 170, such as a database storing experimental data, immunopeptidome data, genomic sequencing data, or previously acquired sequence data. In some embodiments, the set of sequences 320 may be newly generated by the sequencing device 120. For example, the sequencing device 120 may perform sequencing (e.g., next-generation sequencing) on a sample from a mammal subject 315, such as a human (e.g., as depicted) or a mouse, among others.

[0265] The mammal subject 315 may be at risk of or diagnosed with a pathogen or cancer. The cancer may include, for example, at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin's disease, intestinal cancer, kidney-78- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non-Hodgkin's lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof. The pathogen may include, for example, viral pathogens such as severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), among others.

[0266] In some embodiments, the candidate complex structure 302 may correspond to a peptide-major histocompatibility (pMHC) complex. The candidate complex structure 302 may include at least one peptide epitope 304 and at least one MHC 306. The peptide epitope 304 may originate from at least one of a pathogen, a self-antigen, or cancer in the subject 315. The peptide epitope 304 may include at least one of a class I epitope or a class II epitope. In some embodiments, the class I epitope may have a length of 8-11 amino acids. In some embodiments, the class II epitope may have a length of 12-25 amino acids. The MHC 306 may include at least one of an MHC class I or an MHC class II. The MHC polypeptide sequence may include at least one of HL A- A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP, among others. The MHC class I may include an alpha chain including an al domain, a2 domain, and a3 domain. In some embodiments, the MHC class I may also include a p2-microglobulin chain. The MHC class II may include an alpha chain having a al domain and a2 domain. In some embodiments, the MHC class II may also include a beta chain having a pi domain and P2 domain.

[0267] In some embodiments, the candidate complex structure 302 may correspond to a T-cell receptor (TCR)-peptide-major histocompatibility complex (TCR:pMHC) complex. The candidate complex structure 302 may include at least one TCR 308, in addition to the peptide epitope 304 and the MHC 306. The TCR 308 may include a TCRa polypeptide and a TCRP polypeptide that are isolated or obtained from at least one of a CD8+ T cell or a CD4+ T cell, among others. In some embodiments, the candidate complex structure 302 may include at least one T-cell 312 from which the TCR 308 is isolated or obtained. The T-cell 312 may include, for example, at least one of a CD8+ T cell or a CD4+ T cell. In some embodiments, the T-cell 312 may include the TCR 308 to be-79- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282bound with the peptide epitope 304 and the MHC 306. The T-cell 312 can be specific for, cross-reactive with, or non-reactive to a given peptide epitope 304 presented by the MHC 306. In some embodiments, the candidate complex structure 302 may include at least one antigen presenting cell (APC) 310. The APC 310 may include, for example, a dendritic cell, macrophage, B cell, tumor cell, infected cell, engineered cell, or another cell presenting the peptide epitope 304 and the MHC 306.

[0268] With retrieval of the sequences 320, the data aggregator 125 can extract, select, or otherwise identify a set of sequences corresponding to the candidate complex structure 302. The set of sequences 320 may include one or more of a first sequence corresponding to a polypeptide of the MHC 306; a second sequence corresponding to the peptide epitope 304; a third sequence corresponding to the TCR-a polypeptide of the TCR 308; or a fourth sequence corresponding to the TCR-P polypeptide of the TCR 308, among others. The set of sequences 320 may be generated using next-generation sequencing (NGS), such as by the sequencing device 120, or may be retrieved from the database 170. When performing inference for a pMHC complex, the set of sequences 320 may include the first sequence corresponding to the MHC 306 and the second sequence corresponding to the peptide epitope 304. When performing inference for a TCR:pMHC complex, the set of sequences 320 may include the first sequence corresponding to a polypeptide of the MHC 306, the second sequence corresponding to the peptide epitope 304, the third sequence corresponding to the TCR-a polypeptide of the TCR 308, and the fourth sequence corresponding to the TCR-P polypeptide of the TCR 308.

[0269] In some embodiments, the data aggregator 125 may identify one or more of (i) a first input file identifying the first sequence corresponding to the polypeptide of the MHC 306, (ii) a second input file identifying the second sequence corresponding to the peptide epitope 304, (iii) a third input file identifying the third sequence corresponding to the TCR-a polypeptide of the TCR 308, or (iv) a fourth input file identifying the fourth sequence corresponding to the TCR-P polypeptide of the TCR 308, among others. The input files may be of any suitable format, such as FASTQ, BAM, SAM, BCL, or VCF formats. When performing inference for pMHC complexes, the data aggregator 125 may identify the first input file identifying the first sequence corresponding to the MHC 306 and-80- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282the second input file identifying the second sequence corresponding to the peptide epitope 304. When performing inference for TCR:pMHC complexes, the data aggregator 125 may identify the first input file identifying the first sequence corresponding to the polypeptide of the MHC 306, the second input file identifying the second sequence corresponding to the peptide epitope 304, the third input file identifying the third sequence corresponding to the TCR-a polypeptide of the TCR 308, and the fourth input file identifying the fourth sequence corresponding to the TCR-P polypeptide of the TCR 308.

[0270] Using the set of sequences 320, the fold constructor 135 may generate or create a set of candidate quaternary structures 325A-N (hereinafter generally referred to as candidate quaternary structures 325). Any number of candidate quaternary structures 325 may be created, ranging for example between 5 and 1,000. In some embodiments, the fold constructor 135 may create the set of candidate quaternary structures 325 using the set of input files. In some embodiments, the fold constructor 135 may generate a set of output files corresponding to the set of candidate quaternary structures 325, using the set of input files. Each output file may define or include a respective candidate quaternary structure 325. When performing inference for pMHC complexes, each candidate quaternary structure 325 may define or include (i) a respective MHC structure and (ii) a respective peptide epitope structure for forming the pMHC complex. When performing inference for TCR:pMHC complexes, each candidate quaternary structure 325 may define or include (i) a respective MHC structure, (ii) a respective peptide epitope structure, (iii) a respective TCR-a structure, and (iv) a respective TCR-P structure, for the TCR:pMHC complex.

[0271] In some embodiments, the fold constructor 135 may apply or execute the folding model 155 using the set of sequences 320 or the set of input files. To execute, the fold constructor 135 may feed or provide the set of sequences 320, or the set of input files defining the sequences 320, as input to the folding model 155. Based on executing the folding model 155, the fold constructor 135 may generate the set of candidate quaternary structures 325. In some embodiments, the fold constructor 135 may perform preprocessing on the set of sequences 320 prior to input to the folding model 155. For example, the fold constructor 135 may perform formatting, deduplication, or sequence validation for one or more polypeptides defined by the set of sequences 320. In some embodiments, the fold-81- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282constructor 135 may perform template-search and multiple sequence alignment (MSA) on the set of sequences 320 in conjunction with input into the folding model 155. The template search may be used to identify reference sequences from the database 170 corresponding to the set of sequences 320. The MSA may be used to identify evolutionary variation and conservation across sequence positions in the set of sequences 320. The information obtained through template search and MSA may be provided by the fold constructor 135 as input to the folding model 155.

[0272] Referring now to FIG. 17, depicted is a block diagram of a process 400 for executing classifier models in the system 100 for determining values predicting probabilities of protein-protein interactions. Under the process 400, for each candidate quaternary structure 325, the feature extractor 140 may calculate, determine, or otherwise generate a corresponding set of features 405A-N (hereinafter generally referred to as features 405). Each set of features 405 may correspond to a respective candidate quaternary structure 325. The set of features 405 may include or identify two or more of (i) one or more confidence metrics, (ii) one or more structure metrics, (iii) one or more root mean square deviation (RMSD) metrics, or (iv) a set of embeddings, among others. In some embodiments, the feature extractor 140 may generate the set of features 405 from the respective candidate quaternary structure 325 for subsequent evaluation by the classifier model 165.

[0273] In some embodiments, the confidence metrics in the set of features 405 may include one or more of (i) a predicted local distance difference test (pLDDT) metric indicating a predicted local reliability of at least a portion of the candidate quaternary structure 325; (ii) a predicted alignment error (PAE) metric corresponding to estimated positional errors for residues of the candidate quaternary structure 325 relative to a reference true structure; (iii) an interface predicted alignment error (iPAE) metric corresponding to a subset of the estimated positional errors over interface regions between chains in the candidate quaternary structure 325; (iv) a predicted template modeling (pTM) metric corresponding to an estimate of overall accuracy of the predicted candidate quaternary structure 325 relative to a reference true structure; or (v) an interface predicted template modeling (ipTM) metric corresponding to an estimate of accuracy of interfaces of the predicted candidate quaternary structure 325 relative to interfaces of a reference true-82- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282structure, among others. The feature extractor 140 may determine the confidence metrics from confidence outputs generated by the folding model 155 in generating the corresponding candidate quaternary structure 325.

[0274] In some embodiments, the structural metrics in the set of features 405 may indicate probabilities of contact between amino acid residues represented in the corresponding candidate quaternary structure 325. When performing inference for pMHC complexes, the structure metrics may indicate a probability of contact between at least one amino acid residue of the MHC 306 and at least one amino acid residue of the peptide epitope 304. When performing inference for TCR:pMHC complexes, the structure metrics may indicate a probability of contact between two or more of: (i) at least one MHC amino acid residue of the MHC 306, (ii) at least one peptide epitope amino acid residue of the peptide epitope 304, (iii) at least one TCR-a amino acid residue of the TCR 308, (iv) at least one TCR-P amino acid residue of the TCR 308, (v) at least one amino acid residue in a complementarity determining region 3 (CDR3) portion of the TCR-a structure of the TCR 308, or (vi) at least one amino acid residue in a CDR3 portion of the TCR-P structure of the TCR 308. In some embodiments, the feature extractor 140 may determine the structure metrics from contact probabilities or distance distributions output by the folding model 155.

[0275] In some embodiments, the RMSD metrics in the set of features 405 may indicate a measure of structural difference between positions of atoms or residues of the candidate quaternary structure 325 and another quaternary structure. In some embodiments, the feature extractor 140 may calculate or determine an RMSD metric based on a deviation between a corresponding candidate quaternary structure 325 and at least one reference quaternary structure. The reference quaternary structure may correspond to a complex, such as a pMHC or TCR:pMHC complex, previously identified as binding through experimental measurements. For example, the feature extractor 140 may align the candidate quaternary structure 325 with a reference quaternary structure and may calculate the RMSD metric between corresponding residues, atoms, chains, or substructures based on the alignment. The alignment may be performed on all chains of the candidate complex structure 302 or on a subset of chains. When performing inference for pMHC complexes, the RMSD metric may be determined based on a deviation between the candidate quaternary structure 325 and-83- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282at least one reference quaternary structure for the pMHC complex including the peptide epitope 304 and the MHC 306. When performing inference for TCR:pMHC complexes, the RMSD metric may be determined based on a deviation between the candidate quaternary structure 325 and at least one reference quaternary structure for the TCR:pMHC complex including the peptide epitope 304, the MHC 306, and the TCR 308.

[0276] In some embodiments, the feature extractor 140 may calculate or determine an RMSD metric based on a set of deviations among the candidate quaternary structures 325. Each deviation may be between any two of the set of candidate quaternary structures 325 generated using the set of sequences 320 for the same candidate complex structure 302. Each pair of candidate quaternary structures 325 may be aligned overall, aligned on one or more selected chains, or aligned on one or more subcomplexes, and the feature extractor 140 may determine an RMSD value for one or more other chains, residues, or interface regions. For example, the feature extractor 140 may align two candidate quaternary structures 325 on the MHC 306 and calculate RMSD for the peptide epitope 304, align on the pMHC portion and calculate RMSD for the TCR 308, or align on one chain of the TCR 308 and calculate RMSD on another chain of the TCR 308 or on a CDR3 region thereof. The feature extractor 140 may aggregate the resulting set of pairwise RMSD values to determine one or more RMSD metrics, such as a mean, minimum, maximum, standard deviation, or range across the pairwise RMSD values.

[0277] In some embodiments, the set of embeddings in the set of features 405 may be determined or generated based on executing the protein language model 160 using the new sequences 320 corresponding to the candidate complex structure 302. The set of embeddings may correspond to a lower-dimensional representation of latent features in the sequence information for the candidate complex structure 302. In processing the new sequences 320 using the protein language model 160, the feature extractor 140 may perform mean pooling, principal component analysis, or another aggregation technique, and may determine residual features among the sequences. When performing inference for pMHC complexes, the set of embeddings may include embeddings for the MHC 306 and the peptide epitope 304. When performing inference for TCR:pMHC complexes, the set of-84- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282embeddings may include embeddings for the MHC 306, the peptide epitope 304, the TCR-a polypeptide of the TCR 308, and the TCR-P polypeptide of the TCR 308.

[0278] With generation of the set of features 405, the feature extractor 140 may perform either a structural ensemble approach or a best structure approach. Under the structural ensemble approach, the feature extractor 140 may aggregate corresponding sets of features 405 across the set of candidate quaternary structures 325. In aggregating, the feature extractor 140 may determine a mean, standard deviation, minimum, maximum, or range for each feature across the candidate quaternary structures 325. Based thereon, the feature extractor 140 may generate an aggregated set of features 405, sometimes herein referred to as an ensemble 410 of features 405. Each feature 405 in the ensemble 410 may include an aggregation of corresponding features across the candidate quaternary structures 325. The aggregated set of features 405 may be used to further evaluate the candidate complex structure 302.

[0279] Under the best structure approach, the feature extractor 140 may identify or select at least one candidate quaternary structure 325 from the set of candidate quaternary structures 325. From executing the folding model 155 using the new sequences 320A-N, the feature extractor 140 may calculate, determine, or otherwise obtain a respective score for each candidate quaternary structure 325. The score may indicate a predicted accuracy of the corresponding candidate quaternary structure 325. Based on the scores, the feature extractor 140 may select one candidate quaternary structure 325 from the set of candidate quaternary structures 325. For instance, the feature extractor 140 may select the candidate quaternary structure 325 with the highest accuracy score. With the selection, the feature extractor 140 may identify the set of features 405 corresponding to the selected candidate quaternary structure 325. The set of features 405 corresponding to the selected candidate quaternary structure 325 may be used to further evaluate the candidate complex structure 302.

[0280] The binding classifier 145 may apply or execute the classifier model 165 using the set of features 405 (e.g., the aggregated set of features 405 across the set of candidate quaternary structure 325 or the set of features 405 corresponding to the selected candidate quaternary structure 325). The binding classifier 145 may feed or provide the set -85- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282of features 405 as input to the classifier model 165. In feeding, the binding classifier 145 may process the set of features 405 using the set of weights of the classifier model 165. Based on execution of the classifier model 165, the binding classifier 145 may calculate, generate, or otherwise determine at least one value 420. The value 420 may indicate a probability of protein-protein interaction or a probability of binding in the complex structure 202. When performing inference for pMHC complexes, the value 420 may indicate a probability of interaction between the MHC 306 and the peptide epitope 304 in the pMHC complex. When performing inference for TCR:pMHC complexes, the value 420 may indicate a probability of interaction among the MHC 306, the peptide epitope 304, the TCR-a polypeptide of the TCR 308, and the TCR-P polypeptide of the TCR 308 in the TCR:pMHC complex.

[0281] In some embodiments, the binding classifier 145 may determine or generate a classification for the complex structure 202 based on the value 420. To generate, the binding classifier 145 may compare the value 420 with a threshold value. The threshold value may define a value at which to identify the complex structure 202 corresponding to the input as to the classifier model 165 as predicted to be binding or non-binding. The threshold value may be predefined at a set value or may be determined from experimental measurements. If the value 420 does not satisfy (e.g., is less than) the threshold value, the binding classifier 145 may determine the classification of the complex structure 202 as nonbinding. When performing inference for pMHC complexes, the binding classifier 145 may determine the classification indicating the pMHC complex as predicted to be non-binding. When performing inference for TCR:pMHC complexes, binding classifier 145 may determine the classification indicating the TCR:pMHC complex as predicted to be nonbinding. If the value 420 satisfies (e.g., greater than or equal to) the threshold value, the binding classifier 145 may determine the classification of the complex structure 202 as binding. When performing inference for pMHC complexes, the binding classifier 145 may determine the classification indicating the pMHC complex as predicted to be binding.When performing inference for TCR:pMHC complexes, the binding classifier 145 may determine the classification indicating the TCR:pMHC complex as predicted to be binding.-86- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0282] Referring now to FIG. 18, depicted is a block diagram of a process 500 for providing outputs in the system 100 for determining values predicting probabilities of protein-protein interactions. Under the process 500, the output evaluator 150 may store and maintain an association between one or more of: the set of sequences 320, the set of candidate quaternary structure 325, the value 420, and the classification, using one or more data structures on the database 180. The one or more data structures can include an array, a linked list, a stack, a tree, and a hash table, among others. The data structures may be maintained as one or more files, such as Extensible Markup Language (XML), JavaScript Object Notation (JSON), structured query language (SQL), comma-separated-value (CSV), among others. For example, the data structure can be a hash table where the sequences 320 for the complex structure 302 are the key to the hash table and the value 420 or the classification. The hash table can include a plurality of keys corresponding to a plurality of candidate complexes mapped to a plurality of values indicating predicted binding or nonbinding.

[0283] The output evaluator 150 may identify or determine whether the complex structure 302 is to be selected for provision or excluded from provision based on the value 420 (or the classification). There may be a set of candidate complex (e.g., the candidate complex structure 302) that were evaluated using the classifier model 165. If the value 420 does not satisfy (e.g., less than) the threshold value, the output evaluator 150 may determine that the candidate complex structure 302 is non-binding. When the candidate complex structure 302 is determined as non-binding, the output evaluator 150 may exclude the candidate complex structure 302 from provision. In some embodiments, the output evaluator 150 may store and maintain an indication of the candidate complex structure 302 as excluded from provision in the database 180. If the value 420 satisfies (e.g., greater than or equal to) the threshold value, the output evaluator 150 may determine that the candidate complex structure 302 is binding. When the candidate complex structure 302 is determined as binding, the output evaluator 150 may select the candidate complex structure 302 for provision. In some embodiments, the output evaluator 150 may store and maintain an indication of the candidate complex structure 302 as selected for provision in the database 180.-87- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0284] In some embodiments, the output evaluator 150 may identify or determine whether the subject 315 is a candidate for therapy based on the value 420 (or the classification). The determination may be performed when the sample used to evaluate is obtained from the sample of the subject 315. The therapy may include, for example, a neoantigen vaccine to alleviate or address the cancer in the subject 315. The neoantigen vaccine may include at least a portion of the complex structure 302 (e.g., the peptide epitope 304). The neoantigen vaccine may trigger T-cell responses to suppress or alleviate presence of cancer in the subject 315. When the candidate complex structure 302 is determined as binding, the output evaluator 150 may identify the subject 315 as a candidate for therapy. When the candidate complex structure 302 is determined as non-binding, the output evaluator 150 may identify the subject 315 as a non-candidate for therapy.

[0285] The output evaluator 150 may create or generate at least one output 505 based on the results of the evaluation using the classification model 165. The output 505 may be generated based on any one or more of: the set of sequences 320, the set of candidate quaternary structure 325, the value 420, the classification indicating the complex structure 302 as binding or non-binding, the indication of the candidate complex structure 302 as selected or excluded for provision, or the identification of the subject 315 as a candidate or non-candidate for therapy, among others. With the generation, the output evaluator 150 may send, transmit, or otherwise provide the output 505 to the computing device 110. The output 505 can be provided as a system call (e.g., to cause the computing device 110 to render a message), a notification, an alert, a text message, or electronic mail (e-mail), among other messages in a human-readable format.

[0286] The computing device 110 may retrieve, obtain, or otherwise receive the output 505 from the data processing system 105. Upon receipt, the computing device 110 may render, display, or otherwise present the output 505 via a user interface. The information presented via the user interface may include any one or more of: the set of sequences 320, the set of candidate quaternary structure 325, the value 420, the classification indicating the complex structure 302 as binding or non-binding, the indication of the candidate complex structure 302 as selected or excluded for provision, or the identification of the subject 315 as a candidate or non-candidate for therapy, among others.-88- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0287] The provision of the output 505 may be used in a number of settings in an environment 510. The environment 510 can correspond to a clinical, laboratory, manufacturing, pharmacy, or research setting in which the information provided by the output 505 may be used. For example, the environment 510 can include instruments, devices, personnel, and materials for synthesis of peptides or proteins, recombinant expression, purification, refolding, assay-based validation, vaccine formulation, and administration to a subject 315 . In some embodiments, the environment 510 can comprise a hospital, clinic, diagnostic laboratory, cell therapy facility, biomanufacturing facility, or another setting in which the selected complex structure 302 is created, validated, or used.

[0288] In some embodiments, at least one protein 520 may be created or synthesized using the set of sequences 320 for the selected complex structure 302. For example, the set of sequences 320 may be used to produce the protein 520 corresponding to the MHC 306 or the peptide epitope 304, that may be assembled into the selected complex structure 302. The MHC 306 and the peptide epitope 304 may become bound to the TCR-a polypeptide and TCR-P polypeptide of the TCR 308 e.g., of the subject 315). The protein 520 may be created by recombinant expression of one or more polypeptides corresponding to the MHC 306, along with folding and purification. The peptide epitope 304 may be produced (e.g., using solid-phase peptide synthesis) and then loaded into the MHC 306 to form the pMHC complex for the protein 520. For example, the MHC 306 can be expressed as a class I or class II MHC molecule, optionally with an accessory chain such as p2-microglobulin for class I embodiments, and the peptide epitope 304 can be loaded into the MHC 306 to form a pMHC complex. In some embodiments, the TCR 308 may be created using gene synthesis and assembly, and may be added to the pMHC complex to form the TCR:pMHC complex.

[0289] In some embodiments, at least one binding metric 515 may be calculated, determined, or measured from the protein 520 of the selected complex structure 302. The binding metric 515 may include at least one of a monovalent affinity metric or an effective affinity metric, among others. Measurement of the binding metric 515 can include subjecting the protein 520 or the assembled selected complex structure 302 to one or more binding assays, stability assays, presentation assays, or other analytical measurements configured to quantify interaction among components of the selected complex structure 302.-89- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282For example, when the selected complex structure 302 corresponds to a pMHC complex, the binding metric 515 may quantify interaction between the MHC 306 and the peptide epitope 304. When the selected complex structure 302 corresponds to a TCR:pMHC complex, the binding metric 515 may quantify interaction among the peptide epitope 304, the MHC 306, and the TCR 308. The measured binding metric 515 may be compared to a threshold (e.g., an expected value) to determine that the selected complex structure 302 is a validated binding complex.

[0290] In some embodiments, the subject 315 may be administered with a therapeutically effective amount of at least one vaccine 525. The subject 315 may have been identified as a candidate for the administration of therapy. The vaccine 525 the peptide epitope 304 corresponding to the peptide epitope structure of the complex structure 302. In some embodiments, the vaccine 525 may be a neoantigen vaccine or a pathogenspecific vaccine e.g., a vaccine targeting a bacterial, viral or fungal antigen) created using the sequence of the epitope 304 identified from the selected complex structure 302. The creation of the vaccine 525 can include synthesizing a peptide corresponding to the peptide epitope 304 or nucleic acid construct encoding the peptide epitope, and formulating the peptide with vaccine components, such as an adjuvant, carrier, delivery vehicle, nanoparticle, lipid formulation 304. The generated vaccine 525 can then be administered to the subject 315 to induce or trigger an immune response (e.g., against the pathogen-derived or cancer-associated epitope represented by the selected complex structure 302). In some embodiments, the vaccine may be simultaneously, sequentially, or separately administered to the subject along with an additional therapeutic agent. Examples of therapeutic agents include, but are not limited to, an anticancer agent, a therapeutic antibody, an antiviral agent, an antimicrobial agent, etc. Examples of anticancer agents include, but are not limited to, 13-cis-Retinoic Acid, 2-Amino-6-Mercaptopurine, 2-CdA, 2-Chlorodeoxyadenosine, 5-fluorouracil, 6-Thioguanine, 6-Mercaptopurine, Accutane, Actinomycin-D, Adriamycin, Adrucil, Agrylin, Ala-Cort, Aldesleukin, Alemtuzumab, Alitretinoin, Alkaban-AQ, Alkeran, All-transretinoic acid, Alpha interferon, Altretamine, Amethopterin, Amifostine, Aminoglutethimide, Anagrelide, Anandron, Anastrozole, Arabinosylcytosine, Aranesp, Aredia, Arimidex, Aromasin, Arsenic trioxide, Asparaginase, ATRA, Avastin, BCG, BCNU, Bevacizumab, Bexarotene, Bicalutamide, BiCNU,-90- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282Blenoxane, Bleomycin, Bortezomib, Busulfan, Busulfex, C225, Calcium Leucovorin, Campath, Camptosar, Camptothecin-11, Capecitabine, Carac, Carboplatin, Carmustine, Carmustine wafer, Casodex, CCNU, CDDP, CeeNU, Cerubidine, cetuximab, Chlorambucil, Cisplatin, Citrovorum Factor, Cladribine, Cortisone, Cosmegen, CPT-11, Cyclophosphamide, Cytadren, Cytarabine, Cytarabine liposomal, Cytosar-U, Cytoxan, Dacarbazine, Dactinomycin, Darbepoetin alfa, Daunomycin, Daunorubicin, Daunorubicin hydrochloride, Daunorubicin liposomal, DaunoXome, Decadron, Delta-Cortef, Deltasone, Denileukin diftitox, DepoCyt, Dexamethasone, Dexamethasone acetate, Dexamethasone sodium phosphate, Dexasone, Dexrazoxane, DHAD, DIC, Diodex, Docetaxel, Doxil, Doxorubicin, Doxorubicin liposomal, Droxia, DTIC, DTIC-Dome, Duralone, Efudex, Eligard, Ellence, Eloxatin, Elspar, Emcyt, Epirubicin, Epoetin alfa, Erbitux, Erwinia L-asparaginase, Estramustine, Ethyol, Etopophos, Etoposide, Etoposide phosphate, Eulexin, Evista, Exemestane, Fareston, Faslodex, Femara, Filgrastim, Floxuridine, Fludara, Fludarabine, Fluoroplex, Fluorouracil, Fluorouracil (cream), Fluoxymesterone, Flutamide, Folinic Acid, FUDR, Fulvestrant, G-CSF, Gefitinib, Gemcitabine, Gemtuzumab ozogamicin, Gemzar, Gleevec, Lupron, Lupron Depot, Matulane, Maxidex, Mechlorethamine, -Mechlorethamine Hydrochlorine, Medralone, Medrol, Megace, Megestrol, Megestrol Acetate, Melphalan, Mercaptopurine, Mesna, Mesnex, Methotrexate, Methotrexate Sodium, Methylprednisolone, Mylocel, Letrozole, Neosar, Neulasta, Neumega, Neupogen, Nilandron, Nilutamide, Nitrogen Mustard, Novaldex, Novantrone, Octreotide, Octreotide acetate, Oncospar, Oncovin, Ontak, Onxal, Oprevelkin, Orapred, Orasone, Oxaliplatin, Paclitaxel, Pamidronate, Panretin, Paraplatin, Pediapred, PEG Interferon, Pegaspargase, Pegfilgrastim, PEG-INTRON, PEG-L-asparaginase, Phenylalanine Mustard, Platinol, Platinol-AQ, Prednisolone, Prednisone, Prelone, Procarbazine, PROCRIT, Proleukin, Prolifeprospan 20 with Carmustine implant, Purinethol, Raloxifene, Rheumatrex, Rituxan, Rituximab, Roveron-A (interferon alfa-2a), Rubex, Rubidomycin hydrochloride, Sandostatin, Sandostatin LAR, Sargramostim, Solu-Cortef, Solu-Medrol, STI-571, Streptozocin, Tamoxifen, Targretin, Taxol, Taxotere, Temodar, Temozolomide, Teniposide, TESPA, Thalidomide, Thalomid, TheraCys, Thioguanine, Thioguanine Tabloid, Thiophosphoamide, Thioplex, Thiotepa, TICE, Toposar, Topotecan, Toremifene, Trastuzumab, Tretinoin, Trexall, Trisenox, TSP A, VCR,-91- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282Velban, Velcade, VePesid, Vesanoid, Viadur, Vinblastine, Vinblastine Sulfate, Vincasar Pfs, Vincristine, Vinorelbine, Vinorelbine tartrate, VLB, VP- 16, Vumon, Xeloda, Zanosar, Zevalin, Zinecard, Zoladex, Zoledronic acid, Zometa, Gliadel wafer, Glivec, GM-CSF, Goserelin, granulocyte colony stimulating factor, Halotestin, Herceptin, Hexadrol, Hexalen, Hexamethylmelamine, HMM, Hycamtin, Hydrea, Hydrocort Acetate, Hydrocortisone, Hydrocortisone sodium phosphate, Hydrocortisone sodium succinate, Hydrocortone phosphate, Hydroxyurea, Ibritumomab, Ibritumomab Tiuxetan, Idamycin, Idarubicin, Ifex, IFN-alpha, Ifosfamide, IL 2, IL-11, Imatinib mesylate, Imidazole Carboxamide, Interferon alfa, Interferon Alfa-2b (PEG conjugate), Interleukin 2, Interleukin-11, Intron A (interferon alfa-2b), Leucovorin, Leukeran, Leukine, Leuprolide, Leurocri stine, Leustatin, Liposomal Ara-C, Liquid Pred, Lomustine, L-PAM, L-Sarcolysin, Meticorten, Mitomycin, Mitomycin-C, Mitoxantrone, M-Prednisol, MTC, MTX, Mustargen, Mustine, Mutamycin, Myleran, Iressa, Irinotecan, Isotretinoin, Kidrolase, Lanacort, L-asparaginase, and LCR.

[0291] Examples of antiviral agents include, but are not limited to, ganciclovir, azidothymidine (AZT), lamivudine (3TC), etc.

[0292] Examples of antibacterial agents, include but are not limited to, acedapsone; acetosulfone sodium; alamecin; alexidine; amdinocillin; amdinocillin pivoxil; amicycline; amifloxacin; amifloxacin mesylate; amikacin; amikacin sulfate; aminosalicylic acid; aminosalicylate sodium; amoxicillin; amphomycin; ampicillin; ampicillin sodium; apalcillin sodium; apramycin; aspartocin; astromicin sulfate; avilamycin; avoparcin; azithromycin; azlocillin; azlocillin sodium; bacampicillin hydrochloride; bacitracin; bacitracin methylene disalicylate; bacitracin zinc; bambermycins; benzoylpas calcium; berythromycin; betamicin sulfate; biapenem; biniramycin; biphenamine hydrochloride; bispyrithione magsulfex; butikacin; butirosin sulfate; capreomycin sulfate; carbadox; carbenicillin disodium; carbenicillin indanyl sodium; carbenicillin phenyl sodium; carbenicillin potassium; carumonam sodium; cefaclor; cefadroxil; cefamandole; cefamandole nafate; cefamandole sodium; cefaparole; cefatrizine; cefazaflur sodium; cefazolin; cefazolin sodium; cefbuperazone; cefdinir; cefepime; cefepime hydrochloride; cefetecol; cefixime; cefmenoxime hydrochloride; cefmetazole; cefmetazole sodium; cefonicid monosodium; cefonicid sodium; cefoperazone sodium; ceforanide; cefotaxime sodium; cefotetan;-92- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282cefotetan disodium; cefotiam hydrochloride; cefoxitin; cefoxitin sodium; cefpimizole; cefpimizole sodium; cefpiramide; cefpiramide sodium; cefpirome sulfate; cefpodoxime proxetil; cefprozil; cefroxadine; cefsulodin sodium; ceftazidime; ceftibuten; ceftizoxime sodium; ceftriaxone sodium; cefuroxime; cefuroxime axetil; cefuroxime pivoxetil; cefuroxime sodium; cephacetrile sodium; cephalexin; cephalexin hydrochloride; cephaloglycin; cephaloridine; cephalothin sodium; cephapirin sodium; cephradine; cetocycline hydrochloride; cetophenicol; chloramphenicol; chloramphenicol palmitate; chloramphenicol pantothenate complex; chloramphenicol sodium succinate; chlorhexidine phosphanilate; chloroxylenol; chlortetracycline bisulfate; chlortetracycline hydrochloride; cinoxacin; ciprofloxacin; ciprofloxacin hydrochloride; cirolemycin; clarithromycin; clinafloxacin hydrochloride; clindamycin; clindamycin hydrochloride; clindamycin palmitate hydrochloride; clindamycin phosphate; clofazimine; cioxacillin benzathine; cioxacillin sodium; cloxyquin; colistimethate sodium; colistin sulfate; coumermycin; coumermycin sodium; cyclacillin; cycloserine; dalfopristin; dapsone; daptomycin; demeclocycline; demeclocycline hydrochloride; demecycline; denofungin; diaveridine; dicloxacillin; dicloxacillin sodium; dihydrostreptomycin sulfate; dipyrithione; dirithromycin; doxycycline; doxycycline calcium; doxycycline fosfatex; doxycycline hyclate; droxacin sodium; enoxacin; epicillin; epitetracycline hydrochloride; erythromycin; erythromycin acistrate; erythromycin estolate; erythromycin ethyl succinate; erythromycin gluceptate; erythromycin lactobionate; erythromycin propionate; erythromycin stearate; ethambutol hydrochloride; ethionamide; fleroxacin; floxacillin; fludalanine; flumequine; fosfomycin; fosfomycin tromethamine; fumoxicillin; furazolium chloride; furazolium tartrate; fusidate sodium; fusidic acid; gentamicin sulfate; gloximonam; gramicidin; haloprogin; hetacillin; hetacillin potassium; hexedine; ibafloxacin; imipenem; isoconazole; isepamicin; isoniazid; josamycin; kanamycin sulfate; kitasamycin; levofuraltadone; levopropylcillin potassium; lexithromycin; lincomycin; lincomycin hydrochloride; lomefloxacin; Lomefloxacin hydrochloride; lomefloxacin mesylate; loracarbef; mafenide; meclocycline; meclocycline sulfosalicylate; megalomicin potassium phosphate; mequidox; meropenem; methacycline; methacycline hydrochloride; methenamine; methenamine hippurate; methenamine mandelate; methicillin sodium; metioprim; metronidazole hydrochloride; metronidazole phosphate; mezlocillin; mezlocillin sodium; minocycline;-93- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282minocycline hydrochloride; mirincamycin hydrochloride; monensin; monensin sodiumr; nafcillin sodium; nalidixate sodium; nalidixic acid; natainycin; nebramycin; neomycin palmitate; neomycin sulfate; neomycin undecylenate; netilmicin sulfate; neutramycin; nifuiradene; nifuraldezone; nifuratel; nifuratrone; nifurdazil; nifurimide; nifiupirinol; nifurquinazol; nifurthiazole; nitrocycline; nitrofurantoin; nitromide; norfloxacin; novobiocin sodium; ofloxacin; onnetoprim; oxacillin; oxacillin sodium; oximonam; oximonam sodium; oxolinic acid; oxytetracycline; oxytetracycline calcium; oxytetracycline hydrochloride; paldimycin; parachlorophenol; paulomycin; pefloxacin; pefloxacin mesylate; penamecillin; penicillin G benzathine; penicillin G potassium; penicillin G procaine; penicillin G sodium; penicillin V; penicillin V benzathine; penicillin V hydrabamine; penicillin V potassium; pentizidone sodium; phenyl aminosalicylate; piperacillin sodium; pirbenicillin sodium; piridicillin sodium; pirlimycin hydrochloride; pivampicillin hydrochloride; pivampicillin pamoate; pivampicillin probenate; polymyxin B sulfate; porfiromycin; propikacin; pyrazinamide; pyrithione zinc; quindecamine acetate; quinupristin; racephenicol; ramoplanin; ranimycin; relomycin; repromicin; rifabutin; rifametane; rifamexil; rifamide; rifampin; rifapentine; rifaximin; rolitetracycline; rolitetracycline nitrate; rosaramicin; rosaramicin butyrate; rosaramicin propionate; rosaramicin sodium phosphate; rosaramicin stearate; rosoxacin; roxarsone; roxithromycin; sancycline; sanfetrinem sodium; sarmoxicillin; sarpicillin; scopafungin; sisomicin; sisomicin sulfate; sparfloxacin; spectinomycin hydrochloride; spiramycin; stallimycin hydrochloride; steffimycin; streptomycin sulfate; streptonicozid; sulfabenz; sulfabenzamide; sulfacetamide; sulfacetamide sodium; sulfacytine; sulfadiazine; sulfadiazine sodium; sulfadoxine; sulfalene; sulfamerazine; sulfameter; sulfamethazine; sulfamethizole; sulfamethoxazole; sulfamonomethoxine; sulfamoxole; sulfanilate zinc; sulfanitran; sulfasalazine; sulfasomizole; sulfathiazole; sulfazamet; sulfisoxazole; sulfisoxazole acetyl; sulfisboxazole diolamine; sulfomyxin; sulopenem; sultamricillin; suncillin sodium; talampicillin hydrochloride; teicoplanin; temafloxacin hydrochloride; temocillin; tetracycline; tetracycline hydrochloride; tetracycline phosphate complex; tetroxoprim; thi amphenicol; thiphencillin potassium; ticarcillin cresyl sodium; ticarcillin disodium; ticarcillin monosodium; ticlatone; tiodonium chloride; tobramycin; tobramycin sulfate; tosufloxacin;-94- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282trimethoprim; trimethoprim sulfate; trisulfapyrimidines; troleandomycin; trospectomycin sulfate; tyrothricin; vancomycin; vancomycin hydrochloride; virginiamycin; or zorbamycin.

[0293] In one use case, the prediction of binding between a T-cell receptor (TCR) and a peptide-major histocompatibility complex (pMHC) can be used for immune monitoring in a subject, such as a subject receiving a neoantigen vaccine. For example, a computing system may determine that TCRs in a subject’s immune repertoire are predicted to bind one or more specific peptide epitopes included in the vaccine. By analyzing TCR sequence data obtained from the subject over time, the computing system can identify repertoire expansion associated with predicted recognition of particular vaccine epitopes and thereby monitor an immune response to the administered vaccine. The predicted binding information can be used to assess whether there is an immune response to the vaccine. The resultant analysis may be used to support decisions regarding continued monitoring, repeat vaccine administration, booster administration, or inclusion of additional peptide epitopes in a subsequent vaccine formulation. Comparison of predicted TCR responses with subsequent tumor sequencing can identify loss of a previously targeted peptide epitope, which may indicate immunoediting and support selection of replacement or additional epitopes for later treatment.

[0294] In another use case, the prediction of TCR:pMHC binding can be used to identify self-reactive TCRs associated with autoimmunity or immune-mediated pathology. For example, the framework may be used to evaluate whether a TCR is predicted to bind an MHC complex presenting a self-peptide. The computing system may identify TCRs that contribute to autoimmune disease, inflammatory disease, tissue-specific immune damage, or other undesired immune responses. When a self-reactive TCR is identified, the information can be used to guide selective depletion, inhibition, blockade, or modulation of cells expressing that TCR. Such interventions can include, for example, targeted antibodies, engineered cell therapies such as CAR-T cells, receptor-specific depletion agents, immune-modulating biologies, or other targeted therapeutic modalities. Thus, the prediction framework can support both diagnostic and therapeutic workflows for identifying and addressing pathogenic self-reactive T-cell populations.-95- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0295] Referring now to FIG. 19, depicted is a flow diagram of a method 600 of determining values predicting probabilities of protein-protein interactions in T-cell receptor (TCR)- peptide-major histocompatibility complex (MHC) (TCR:pMHC) complexes or peptide-MHC (pMHC) complexes. The method 600 can be implemented by any components detailed herein, such as the system 100 or the system 700. Under the method 600, a computing system may retrieve sequence files corresponding to components of a complex (605). The computing system may create a set of candidate sequences (610). The computing system may generate a set of features for each structure (615). The computing system may generate an ensemble of features based on aggregating the features across the candidate structures (620). The computing system may execute a classification model using the ensemble of features (625). The computing system may determine a value indicating a binding probability for the complex based on execution (630). The computing system may determine whether the value satisfies a threshold (635). If the value satisfies (e.g., is greater than or equal to) the threshold, the computing system may select the complex for provision (640). Otherwise, if the value does not satisfy (e.g., is less than) the threshold, the computing system may exclude the complex from provision (645). The computing system may generate an output based on the results (650). The computing system may provide the output for presentation (655).Example 4: Computing and Network Environment

[0296] Various operations described herein can be implemented on computer systems. FIG. 20 shows a simplified block diagram of a representative server system 700, client computing system 714, and network 726 usable to implement certain embodiments of the present disclosure. In various embodiments, server system 700 or similar systems can implement services or servers described herein or portions thereof. Client computing system 714 or similar systems can implement clients described herein. The system 100 described herein can be similar to the server system 700. Server system 700 can have a modular design that incorporates a number of modules 702 (e.g., blades in a blade server embodiment); while two modules 702 are shown, any number can be provided. Each module 702 can include processing unit(s) 704 and local storage 706.-96- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0297] Processing unit(s) 704 can include a single processor, which can have one or more cores, or multiple processors. In some embodiments, processing unit(s) 704 can include a general-purpose primary processor as well as one or more special-purpose coprocessors such as graphics processors, digital signal processors, or the like. In some embodiments, some or all processing units 704 can be implemented using customized circuits, such as application specific integrated circuits (ASICs) or field programmable gate arrays (FPGAs). In some embodiments, such integrated circuits execute instructions that are stored on the circuit itself. In other embodiments, processing unit(s) 704 can execute instructions stored in local storage 706. Any type of processors in any combination can be included in processing unit(s) 704.

[0298] Local storage 706 can include volatile storage media (e.g., DRAM, SRAM, SDRAM, or the like) and / or non-volatile storage media (e.g., magnetic or optical disk, flash memory, or the like). Storage media incorporated in local storage 706 can be fixed, removable or upgradeable as desired. Local storage 706 can be physically or logically divided into various subunits such as a system memory, a read-only memory (ROM), and a permanent storage device. The system memory can be a read-and-write memory device or a volatile read-and-write memory, such as dynamic random-access memory. The system memory can store some or all of the instructions and data that processing unit(s) 704 need at runtime. The ROM can store static data and instructions that are needed by processing unit(s) 704. The permanent storage device can be a non-volatile read-and-write memory device that can store instructions and data even when module 702 is powered down. The term “storage medium” as used herein includes any medium in which data can be stored indefinitely (subject to overwriting, electrical disturbance, power loss, or the like) and does not include carrier waves and transitory electronic signals propagating wirelessly or over wired connections.

[0299] In some embodiments, local storage 706 can store one or more software programs to be executed by processing unit(s) 704, such as an operating system and / or programs implementing various server functions such as functions of the system 100 or any other system described herein, or any other server(s) associated with system 100 or any other system described herein.-97- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0300] Software” refers generally to sequences of instructions that, when executed by processing unit(s) 704 cause server system 700 (or portions thereof) to perform various operations, thus defining one or more specific machine embodiments that execute and perform the operations of the software programs. The instructions can be stored as firmware residing in read-only memory and / or program code stored in non-volatile storage media that can be read into volatile working memory for execution by processing unit(s) 704. Software can be implemented as a single program or a collection of separate programs or program modules that interact as desired. From local storage 706 (or non-local storage described below), processing unit(s) 704 can retrieve program instructions to execute and data to process in order to execute various operations described above.

[0301] In some server systems 700, multiple modules 702 can be interconnected via a bus or other interconnect 708, forming a local area network that supports communication between modules 702 and other components of server system 700. Interconnect 708 can be implemented using various technologies including server racks, hubs, routers, etc.

[0302] A wide area network (WAN) interface 710 can provide data communication capability between the local area network (interconnect 708) and the network 726, such as the Internet. Technologies can be used, including wired (e.g., Ethernet, IEEE 702.3 standards) and / or wireless technologies (e.g., Wi-Fi, IEEE 702.11 standards).

[0303] In some embodiments, local storage 706 is intended to provide working memory for processing unit(s) 704, providing fast access to programs and / or data to be processed while reducing traffic on interconnect 708. Storage for larger quantities of data can be provided on the local area network by one or more mass storage subsystems 712 that can be connected to interconnect 708. Mass storage subsystem 712 can be based on magnetic, optical, semiconductor, or other data storage media. Direct attached storage, storage area networks, network-attached storage, and the like can be used. Any data stores or other collections of data described herein as being produced, consumed, or maintained by a service or server can be stored in mass storage subsystem 712. In some embodiments, additional data storage resources may be accessible via WAN interface 710 (potentially with increased latency).-98- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0304] Server system 700 can operate in response to requests received via WAN interface 710. For example, one of modules 702 can implement a supervisory function and assign discrete tasks to other modules 702 in response to received requests. Work allocation techniques can be used. As requests are processed, results can be returned to the requester via WAN interface 710. Such operation can generally be automated. Further, in some embodiments, WAN interface 710 can connect multiple server systems 700 to each other, providing scalable systems capable of managing high volumes of activity. Other techniques for managing server systems and server farms (collections of server systems that cooperate) can be used, including dynamic resource allocation and reallocation.

[0305] Server system 700 can interact with various user-owned or user-operated devices via a wide-area network such as the Internet. An example of a user-operated device is shown in FIG. 20 as client computing system 714. Client computing system 714 can be implemented, for example, as a consumer device such as a smartphone, other mobile phone, tablet computer, wearable computing device (e.g., smart watch, eyeglasses), desktop computer, laptop computer, and so on.

[0306] For example, client computing system 714 can communicate via WAN interface 710. Client computing system 714 can include computer components such as processing unit(s) 716, storage device 718, network interface 720, user input device 722, and user output device 724. Client computing system 714 can be a computing device implemented in a variety of form factors, such as a desktop computer, laptop computer, tablet computer, smartphone, other mobile computing device, wearable computing device, or the like.

[0307] Processing unit(s) 716 and storage device 718 can be similar to processing unit(s) 704 and local storage 706 described above. Suitable devices can be selected based on the demands to be placed on client computing system 714; for example, client computing system 714 can be implemented as a “thin” client with limited processing capability or as a high-powered computing device. Client computing system 714 can be provisioned with program code executable by processing unit(s) 716 to enable various interactions with server system 700.-99- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0308] Network interface 720 can provide a connection to the network 726, such as a wide area network (e.g., the Internet) to which WAN interface 710 of server system 700 is also connected. In various embodiments, network interface 720 can include a wired interface (e.g., Ethernet) and / or a wireless interface implementing various RF data communication standards such as Wi-Fi, Bluetooth, or cellular data network standards (e.g., 3G, 4G, LTE, etc ).

[0309] User input device 722 can include any device (or devices) via which a user can provide signals to client computing system 714. The client computing system 714 can interpret the signals as indicative of particular user requests or information. In various embodiments, user input device 722 can include any or all of a keyboard, touch pad, touch screen, mouse or other pointing device, scroll wheel, click wheel, dial, button, switch, keypad, microphone, and so on.

[0310] User output device 724 can include any device via which client computing system 714 can provide information to a user. For example, user output device 724 can include a display to display images generated by or delivered to client computing system 714. The display can incorporate various image generation technologies, e.g., a liquid crystal display (LCD), light-emitting diode (LED) including organic light-emitting diodes (OLED), projection system, cathode ray tube (CRT), or the like, together with supporting electronics (e.g., digital-to-analog or analog-to-digital converters, signal processors, or the like). Some embodiments can include a device such as a touchscreen that function as both input and output device. In some embodiments, other user output devices 724 can be provided in addition to or instead of a display. Examples include indicator lights, speakers, tactile “display” devices, printers, and so on.

[0311] Some embodiments include electronic components, such as microprocessors, storage and memory that store computer program instructions in a computer-readable storage medium. Many of the features described in this specification can be implemented as processes that are specified as a set of program instructions encoded on a computer-readable storage medium. When these program instructions are executed by one or more processing units, they cause the processing unit(s) to perform various operation indicated in the program instructions. Examples of program instructions or computer code include machine -100- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282code, such as is produced by a compiler, and files including higher-level code that are executed by a computer, an electronic component, or a microprocessor using an interpreter. Through suitable programming, processing unit(s) 704 and 716 can provide various functionality for server system 700 and client computing system 714, including any of the functionality described herein as being performed by a server or client, or other functionality.

[0312] It will be appreciated that server system 700 and client computing system 714 are illustrative and that variations and modifications are possible. Computer systems used in connection with embodiments of the present disclosure can have other capabilities not specifically described here. Further, while server system 700 and client computing system 714 are described with reference to particular blocks, it is to be understood that these blocks are defined for convenience of description and are not intended to imply a particular physical arrangement of component parts. For instance, different blocks can be but need not be located in the same facility, in the same server rack, or on the same motherboard. Further, the blocks need not correspond to physically distinct components. Blocks can be configured to perform various operations, e.g., by programming a processor or providing appropriate control circuitry, and various blocks might or might not be reconfigurable depending on how the initial configuration is obtained. Embodiments of the present disclosure can be realized in a variety of apparatus including electronic devices implemented using any combination of circuitry and software.

[0313] While the disclosure has been described with respect to specific embodiments, one skilled in the art will recognize that numerous modifications are possible. Embodiments of the disclosure can be realized using a variety of computer systems and communication technologies including but not limited to the specific examples described herein. Embodiments of the present disclosure can be realized using any combination of dedicated components and / or programmable processors and / or other programmable devices. The various processes described herein can be implemented on the same processor or different processors in any combination. Where components are described as being configured to perform certain operations, such configuration can be accomplished, e.g., by designing electronic circuits to perform the operation, by programming programmable-101- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282electronic circuits (such as microprocessors) to perform the operation, or any combination thereof. Further, while the embodiments described above may make reference to specific hardware and software components, those skilled in the art will appreciate that different combinations of hardware and / or software components may also be used and that particular operations described as being implemented in hardware might also be implemented in software or vice versa.

[0314] Computer programs incorporating various features of the present disclosure may be encoded and stored on various computer-readable storage media; suitable media include magnetic disk or tape, optical storage media such as compact disk (CD) or DVD (digital versatile disk), flash memory, and other non-transitory media. Computer-readable media encoded with the program code may be packaged with a compatible electronic device, or the program code may be provided separately from electronic devices (e.g., via Internet download or as a separately packaged computer-readable storage medium).

[0315] Thus, although the disclosure has been described with respect to specific embodiments, it will be appreciated that the disclosure is intended to cover all modifications and equivalents within the scope of the following claims.EQUIVALENTS

[0316] The present technology is not to be limited in terms of the particular embodiments described in this application, which are intended as single illustrations of individual aspects of the present technology. Many modifications and variations of this present technology can be made without departing from its spirit and scope, as will be apparent to those skilled in the art. Functionally equivalent methods and apparatuses within the scope of the present technology, in addition to those enumerated herein, will be apparent to those skilled in the art from the foregoing descriptions. Such modifications and variations are intended to fall within the scope of the present technology. It is to be understood that this present technology is not limited to particular methods, reagents, compounds compositions or biological systems, which can, of course, vary. It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only, and is not intended to be limiting.-102- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282

[0317] In addition, where features or aspects of the disclosure are described in terms of Markush groups, those skilled in the art will recognize that the disclosure is also thereby described in terms of any individual member or subgroup of members of the Markush group.

[0318] As will be understood by one skilled in the art, for any and all purposes, particularly in terms of providing a written description, all ranges disclosed herein also encompass any and all possible subranges and combinations of subranges thereof. Any listed range can be easily recognized as sufficiently describing and enabling the same range being broken down into at least equal halves, thirds, quarters, fifths, tenths, etc. As a nonlimiting example, each range discussed herein can be readily broken down into a lower third, middle third and upper third, etc. As will also be understood by one skilled in the art all language such as “up to,” “at least,” “greater than,” “less than,” and the like, include the number recited and refer to ranges which can be subsequently broken down into subranges as discussed above. Finally, as will be understood by one skilled in the art, a range includes each individual member. Thus, for example, a group having 1-3 cells refers to groups having 1, 2, or 3 cells. Similarly, a group having 1-5 cells refers to groups having 1, 2, 3, 4, or 5 cells, and so forth.

[0319] All patents, patent applications, provisional applications, and publications referred to or cited herein are incorporated by reference in their entirety, including all figures and tables, to the extent they are not inconsistent with the explicit teachings of this specification.-103- 4906-5272-8748.2

Claims

Atty. Dkt. No.: 115872-3282CLAIMS1. A method of determining values indicating probabilities of protein-protein interactions in peptide-major histocompatibility complex (MHC) (pMHC) complexes, comprising:retrieving, by one or more processors, (i) a first sequence corresponding to an MHC polypeptide and (ii) a second sequence corresponding to a peptide epitope that binds to the MHC polypeptide to form a pMHC complex;creating, by the one or more processors, using the first sequence and the second sequence, a plurality of candidate quaternary structures, each of the plurality of candidate quaternary structures comprising (i) a respective MHC structure and (ii) a respective peptide epitope structure for forming the pMHC complex;generating, by the one or more processors, for at least one candidate quaternary structure of the plurality of candidate quaternary structures, a plurality of features comprising two or more of (i) a confidence metric, (ii) a structure metric, or (iii) a root mean square deviation (RMSD) metric;providing, by the one or more processors, the plurality of features as input to a classifier model to determine a value indicating a probability of interaction between the MHC and the peptide epitope in the pMHC complex; andstoring, by the one or more processors, using one or more data structures, an association between the respective value and the pMHC complex.

2. The method of claim 1, further comprising:determining, by the one or more processors, that the pMHC complex is non-binding based on the value not satisfying a threshold value; andexcluding, by the one or more processors, the pMHC complex from provision.

3. The method of claim 1, further comprising:determining, by the one or more processors, that the pMHC complex is binding based on the value satisfying a threshold value; andselecting, by the one or more processors, from a plurality of candidate pMHC complexes, the pMHC complex for provision.-104- 4906-5272-8748.2Atty. Dkt. No.: 115872-32824. The method of claim 3, further comprising generating the pMHC complex using an MHC and a peptide epitope synthesized in accordance with a plurality of sequences defining the pMHC complex selected from the plurality of candidate pMHC complexes.

5. The method of claim 4, further comprising measuring a binding affinity metric from the pMHC complex to validate as a binding complex, wherein the binding affinity comprises at least one of a monovalent affinity metric or an effective affinity metric.

6. The method of claim 3, further comprising administering a subject at risk of cancer or diagnosed with cancer, with neoantigen vaccine comprising the peptide epitope.

7. The method of claim 6, wherein the cancer comprises at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof.

8. The method of claim 1, wherein the first sequence corresponds to at least one of a wildtype MHC polypeptide or a mutant type MHC polypeptide,wherein the second sequence corresponds to a wildtype type peptide epitope or a mutant type peptide epitope, orwherein the first sequence and the second sequence are derived from at least one mammal comprising one or more of a human or a mouse.-105- 4906-5272-8748.2Atty. Dkt. No.: 115872-32829. The method of claim 1, wherein the confidence metric comprising at least one of (i) a predicted local distance difference test (pLDDT) metric, (ii) a predicted alignment error (PAE) metric, (iii) an interface predicted alignment error (iPAE), (iv) a predicted template modeling (pTM) metric, or (v) an interface predicted template modeling (ipTM) metric.

10. The method of claim 1, wherein the structure metric indicates a probability of contact between at least one amino acid residue of the MHC structure and at least one amino acid residue of the peptide epitope structure.

11. The method of claim 1, wherein the RMSD metric is determined based on a deviation between the candidate quaternary structure and at least one reference quaternary structure for the pMHC complex.

12. The method of claim 1, wherein the RMSD metric is determined based on an aggregation of a plurality of deviations, each of the plurality of deviations between two candidate quaternary structures of a plurality of candidate quaternary structures.

13. The method of claim 1, wherein the plurality of features further comprises a plurality of embeddings generated based on executing a protein language model using a plurality of sequences corresponding to the pMHC complex.

14. The method of claim 1, wherein retrieving the first sequence and the second sequence further comprises retrieving, from one or more databases or from a sequencing device, (i) a first input file identifying the first sequence corresponding to the MHC polypeptide sequence and (ii) a second input file identifying the second sequence corresponding to the peptide epitope sequence,wherein creating the plurality of candidate quaternary structures further comprises generating, using the first input file and the second input file, a plurality of output files corresponding to the plurality of candidate quaternary structures.-106- 4906-5272-8748.2Atty. Dkt. No.: 115872-328215. The method of claim 1, wherein the classifier model is established using a plurality of training examples, each of the plurality of training examples comprising:(i) a respective example plurality of features comprising two or more of an example confidence metric, an example structure metric, or an example RMSD metric, for an example pMHC complex, and(ii) a respective label indicating the example pMHC complex as one of binding or non-binding.

16. The method of claim 1, wherein creating the plurality of candidate quaternary structures further comprises executing a folding model using the first sequence and the second sequence, and further comprising:generating, by the one or more processors, based on executing the folding model, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective score indicating a respective accuracy of the candidate quaternary structure; and selecting, by the one or more processors, from the plurality of candidate quaternary structures, the at least one candidate quaternary structure based on the respective score indicating the respective accuracy.

17. The method of claim 16, wherein the folding model comprises at least one of AlphaFold, RoseTTAFold, OpenFold, Boltz-1, OmegaFold, orESMFold.

18. The method of claim 1, further comprising determining, by the one or more processors, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective plurality of features comprising two or more: (i) a respective confidence metric, (ii) a respective structure metric, (iii) a respective root mean square deviation (RMSD) metric, or (iv) a plurality of embeddings;wherein generating the plurality of features further comprises generating the plurality of features based on an aggregation of the respective plurality of features for each candidate quaternary structure of the plurality of candidate quaternary structures, wherein providing the input to the classifier model further comprises determining the value indicating the indicating the probability of interaction within the pMHC complex.-107- 4906-5272-8748.2Atty. Dkt. No.: 115872-328219. The method of claim 1, wherein the MHC corresponding to the first sequence comprises at least one of an MHC class I or an MHC class II,wherein the MHC polypeptide sequence comprises at least one of HLA-A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP,wherein the MHC class I comprises an alpha chain including al domain, a2 domain, and a3 domain, and optionally a p2-microglobulin chain, andwherein the MHC class II comprises an alpha chain having al domain and a2 domain and optionally a beta chain having an pi domain and P2 domain.

20. The method of claim 1, wherein the peptide epitope corresponding to the second sequence originates from at least one of a pathogen, a self-antigen, or cancer in a subject, andwherein the peptide epitope comprises at least one of a class I epitope or a class II epitope, and wherein the class I epitope has a length of 8-11 amino acids, and wherein the class II epitope has a length of 12-25 amino acids.

21. A method of determining values indicating probability of protein-protein interactions in T-cell receptor (TCR)-peptide-major histocompatibility complex (MHC) (TCR:pMHC) complexes, comprising:retrieving, by one or more processors, (i) a first sequence corresponding to an MHC polypeptide and (ii) a second sequence corresponding to a peptide epitope that binds to the MHC polypeptide to form an peptide-MHC (pMHC) complex; (iii) a third sequence corresponding to a TCR-a polypeptide and (iv) a fourth sequence corresponding to a TCR-P polypeptide that binds to the TCR-a polypeptide to form a TCR, wherein the TCR binds with the pMHC complex to form a TCR-peptide-MHC complex;creating, by the one or more processors, using the first sequence, the second sequence, the third sequence, and the fourth sequence, a plurality of candidate quaternary structures for the to form the TCR-peptide-MHC complex, each of the plurality of quaternary structures comprising: (i) a respective MHC structure, (ii) a respective peptide epitope structure, (iii) a respective TCR-a structure, and (iv) a respective TCR-P structure;-108- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282generating, by the one or more processors, for at least one candidate quaternary structure of the plurality of candidate quaternary structures, a plurality of features comprising two or more of a confidence metric, a structure metric, or a root square mean deviation (RMSD) metric;determining, by the one or more processors, based on providing the plurality of features as input to a classifier model, at least one value indicating interaction among the MHC structure, the peptide epitope structure, the TCR-a structure, and the TCR-P structure of the at least one candidate quaternary structure for the TCR:pMHC complex; and storing, by the one or more processors, using one or more data structures, an association between the value and the at least one candidate quaternary structure for the TCR:pMHC complex.

22. The method of claim 21, further comprising:determining, by the one or more processors, that the TCR:pMHC complex is nonbinding based on the value not satisfying a threshold value; andexcluding, by the one or more processors, the TCR:pMHC complex from provision.

23. The method of claim 21, further comprising:determining, by the one or more processors, that the TCR:pMHC complex is binding, based on the value satisfying a threshold; andselecting, by the one or more processors, from a plurality of candidate TCR:pMHC complexes, the at least one TCR:pMHC complex for provision.

24. The method of claim 23, further comprising generating the TCR:pMHC complex using a peptide epitope, an MHC molecule, and a TCR synthesized in accordance with a plurality sequences defining the TCR:pMHC complex selected from the plurality of candidate TCR:pMHC complexes.

25. The method of claim 24, further comprising measuring a binding affinity metric from the TCR:pMHC complex to validate as a binding complex, wherein the binding affinity metric comprises at least one of a monovalent affinity or effective affinity.-109- 4906-5272-8748.2Atty. Dkt. No.: 115872-328226. The method of claim 23, further comprising administering a subject at risk of cancer or diagnosed with cancer, with neoantigen vaccine comprising the peptide epitope corresponding to the peptide epitope structure of the TCR:pMHC complex.

27. The method of claim 26, wherein the cancer comprises at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof.

28. The method of claim 21, wherein the first sequence corresponds to at least one of a wildtype MHC polypeptide or a mutant type MHC polypeptide,wherein the second sequence corresponds to a wildtype type peptide epitope or a mutant type peptide epitope,wherein the first sequence, the second sequence, the third sequence, or the fourth sequence are derived from at least one mammal comprising one or more of a human or a mouse.

29. The method of claim 21, wherein the confidence metric comprises at least one of (i) a predicted local distance difference test (pLDDT) metric, (ii) a predicted alignment error (PAE) metric, (iii) an interface predicted alignment error (iPAE), (iv) a predicted template modeling (pTM) metric, or (v) an interface predicted template modeling (ipTM) metric.-110- 4906-5272-8748.2Atty. Dkt. No.: 115872-328230. The method of claim 21, wherein the structure metric indicates a probability of contact between two or more of: (i) at least one MHC amino acid residue of the MHC structure, (ii) at least one peptide epitope amino acid residue of the peptide epitope structure, (iii) at least one TCR-a amino acid residue of the TCR-a structure, (iv) at least one TCR- P amino acid residue of the TCR-P structure, (v) at least one amino acid residue in the Complementarity Determining Region 3 (CDR3) portion of the TCR-a structure, or (vi) at least one amino acid residue in the CDR3 portion TCR-P structure.

31. The method of claim 21, wherein the RMSD metric is determined based on a deviation between the candidate quaternary structure and a reference quaternary structure for the TCR:pMHC complex, orwherein the RMSD metric is determined based on an aggregation of a plurality of deviations, each of the plurality of deviations between two candidate quaternary structures of a plurality of candidate quaternary structures.

32. The method of claim 21, wherein the respective plurality of features for each candidate TCR:pMHC complex further comprises a respective plurality of embeddings generated based on executing a protein language model using a plurality of sequences corresponding to the TCR:pMHC complex.

33. The method of claim 21, wherein retrieving further comprises retrieving, from one or more databases or from a sequencing device, (i) a first input file identifying the first sequence corresponding to the MHC polypeptide, (ii) a second input file identifying the second sequence corresponding to the peptide epitope, (iii) a third input file identifying the third sequence corresponding to the TCR-a polypeptide and (iv) a fourth input file identifying the fourth sequence corresponding to the TCR-P polypeptide,wherein creating the plurality of candidate quaternary structures further comprises generating, using the first input file, the second input file, the third input file, and the fourth input file, a plurality of output files corresponding to the plurality of candidate quaternary structures.-Ill- 4906-5272-8748.2Atty. Dkt. No.: 115872-328234. The method of claim 21, wherein the classifier model is established using a plurality of training examples, each of the plurality of training examples comprising:(i) a respective example plurality of features comprising two or more of a respective example confidence metric, a respective example structure metric, a respective example root mean square deviation (RMSD) metric, for an example TCR:pMHC complex(ii) a respective label indicating the example TCR:pMHC complex as one of binding or non-binding.

35. The method of claim 21, wherein creating the plurality of candidate quaternary structures further comprises executing a folding model using the first sequence, the second sequence, the third sequence, and the fourth sequence, and further comprising:generating, by the one or more processors, based on executing the folding model, for each candidate TCR:pMHC complex a plurality of candidate quaternary structures, each with a respective score indicating a respective accuracy of the candidate quaternary structure as determined by the folding model; andselecting, by the one or more processors, from the plurality of candidate quaternary structures, the at least one candidate quaternary structure based on the respective score indicating the respective accuracy36. The method of claim 35, wherein the folding model comprises at least one of AlphaFold, RoseTTAFold, OpenFold, Boltz-1, OmegaFold, ESMFold.

37. The method of claim 21, further comprising determining, by the one or more processors, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective plurality of features comprising two or more: (i) a respective confidence metric, (ii) a respective structure metric, (iii) a respective root mean square deviation (RMSD) metric, or (v) a respective plurality of embeddings,wherein generating the plurality of features further comprises generating the plurality of features based on an aggregation of the respective plurality of features for each candidate quaternary structure of the plurality of candidate quaternary structures,-112- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282wherein providing the input to the classifier model further comprises determining the value indicating the probability of interaction within the TCR:pMHC quaternary structure.

38. The method of claim 21, wherein the TCR comprises a TCRa polypeptide and a TCRP polypeptide that are isolated from at least one of CD8+ T cell or a CD4+ T cell, and the third sequence corresponding to TCR-a polypeptide and the fourth sequence corresponding to the TCR-P polypeptide are determined by next-generation sequencing.

39. The method of claim 21, wherein the MHC corresponding to the first sequence comprises at least one of an MHC class I or an MHC class II,wherein the MHC polypeptide comprises at least one of HLA-A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP,wherein the MHC class I comprises an alpha chain including al domain, a2 domain, and a3 domain, and optionally a P2-microglobulin chain, andwherein the MHC class II comprises an alpha chain having al domain and a2 domain and optionally a beta chain having an pi domain and P2 domain.

40. The method of claim 21, wherein the peptide epitope corresponding to the second sequence originates from at least one of a pathogen, a self-antigen, or cancer in a subject, andwherein the peptide epitope comprises at least one of a class I epitope or a class II epitope, and wherein the class I epitope has a length of 8-11 amino acids, and wherein the class II epitope has a length of 12-25 amino acids.

41. A system for determining values indicating probabilities of protein-protein interactions in peptide-major histocompatibility complex (MHC) (pMHC) complexes, comprising: one or more processors coupled with memory, configured to:retrieve (i) a first sequence corresponding to an MHC polypeptide and (ii) a second sequence corresponding to a peptide epitope that binds to the MHC polypeptide to form a pMHC complex;-113- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282create, using the first sequence and the second sequence, a plurality of candidate quaternary structures, each of the plurality of candidate quaternary structures comprising (i) a respective MHC structure and (ii) a respective peptide epitope structure for forming the pMHC complex;generate, for at least one candidate quaternary structure of the plurality of candidate quaternary structures, a plurality of features comprising two or more of (i) a confidence metric, (ii) a structure metric, or (iii) a root mean square deviation (RMSD) metric;provide the plurality of features as input to a classifier model to determine a value indicating a probability of interaction between the MHC and the peptide epitope in the pMHC complex; andstore, using one or more data structures, an association between the respective value and the pMHC complex.

42. The system of claim 41, wherein the one or more processors are further configured to determine that the pMHC complex is non-binding based on the value not satisfying a threshold value; and exclude the pMHC complex from provision.

43. The system of claim 41, wherein the one or more processors are further configured to determine that the pMHC complex is binding based on the value satisfying a threshold value; andselect, from a plurality of candidate pMHC complexes, the pMHC complex for provision.

44. The system of claim 41, wherein the pMHC complex is generated using an MHC and a peptide epitope synthesized in accordance with a plurality of sequences defining the pMHC complex selected from the plurality of candidate pMHC complexes.

45. The system of claim 41, wherein a binding affinity metric is measured from the pMHC complex to validate as a binding complex, wherein the binding affinity comprises at least one of a monovalent affinity metric or an effective affinity metric.-114- 4906-5272-8748.2Atty. Dkt. No.: 115872-328246. The system of claim 41, wherein a subject at risk of cancer or diagnosed with cancer is administered with neoantigen vaccine comprising the peptide epitope.

47. The system of claim 46, wherein the cancer comprises at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof.

48. The system of claim 41, wherein the first sequence corresponds to at least one of a wildtype MHC polypeptide or a mutant type MHC polypeptide,wherein the second sequence corresponds to a wildtype type peptide epitope or a mutant type peptide epitope, orwherein the first sequence and the second sequence are derived from at least one mammal comprising one or more of a human or a mouse.

49. The system of claim 41, wherein the confidence metric comprising at least one of (i) a predicted local distance difference test (pLDDT) metric, (ii) a predicted alignment error (PAE) metric, (iii) an interface predicted alignment error (iPAE), (iv) a predicted template modeling (pTM) metric, or (v) an interface predicted template modeling (ipTM) metric.

50. The system of claim 41, wherein the structure metric indicates a probability of contact between at least one amino acid residue of the MHC structure and at least one amino acid residue of the peptide epitope structure.-115- 4906-5272-8748.2Atty. Dkt. No.: 115872-328251. The system of claim 41, wherein the RMSD metric is determined based on a deviation between the candidate quaternary structure and at least one reference quaternary structure for the pMHC complex.

52. The system of claim 41, wherein the RMSD metric is determined based on an aggregation of a plurality of deviations, each of the plurality of deviations between two candidate quaternary structures of a plurality of candidate quaternary structures.

53. The system of claim 41, wherein the plurality of features further comprises a plurality of embeddings generated based on executing a protein language model using a plurality of sequences corresponding to the pMHC complex.

54. The system of claim 41, wherein the one or more processors are further configured to:retrieve, from one or more databases or from a sequencing device, (i) a first input file identifying the first sequence corresponding to the MHC polypeptide sequence and (ii) a second input file identifying the second sequence corresponding to the peptide epitope sequence; andgenerate, using the first input file and the second input file, a plurality of output files corresponding to the plurality of candidate quaternary structures.

55. The system of claim 41, wherein the classifier model is established using a plurality of training examples, each of the plurality of training examples comprising:(i) a respective example plurality of features comprising two or more of an example confidence metric, an example structure metric, or an example RMSD metric, for an example pMHC complex, and(ii) a respective label indicating the example pMHC complex as one of binding or non-binding.

56. The system of claim 41, wherein the one or more processors are further configured to:-116- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282execute a folding model using the first sequence and the second sequence to create the plurality of candidate quaternary structures;generate, based on executing the folding model, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective score indicating a respective accuracy of the candidate quaternary structure; andselect, from the plurality of candidate quaternary structures, the at least one candidate quaternary structure based on the respective score indicating the respective accuracy.

57. The system of claim 56, wherein the folding model comprises at least one of AlphaFold, RoseTTAFold, OpenFold, Boltz-1, OmegaFold, orESMFold.

58. The system of claim 41, wherein the one or more processors are further configured to:determine, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective plurality of features comprising two or more: (i) a respective confidence metric, (ii) a respective structure metric, (iii) a respective root mean square deviation (RMSD) metric, or (iv) a plurality of embeddings;generate the plurality of features based on an aggregation of the respective plurality of features for each candidate quaternary structure of the plurality of candidate quaternary structures; andprovide the plurality of features to determine the value indicating the indicating the probability of interaction within the pMHC complex.

59. The system of claim 41, wherein the MHC corresponding to the first sequence comprises at least one of an MHC class I or an MHC class II,wherein the MHC polypeptide comprises at least one of HLA-A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP,wherein the MHC class I comprises an alpha chain including al domain, a2 domain, and a3 domain, and optionally a p2-microglobulin chain, andwherein the MHC class II comprises an alpha chain having al domain and a2 domain and optionally a beta chain having an pi domain and P2 domain.-117- 4906-5272-8748.2Atty. Dkt. No.: 115872-328260. The system of claim 41, wherein the peptide epitope corresponding to the second sequence originates from at least one of a pathogen, a self-antigen, or cancer in a subject, andwherein the peptide epitope comprises at least one of a class I epitope or a class II epitope, and wherein the class I epitope has a length of 8-11 amino acids, and wherein the class II epitope has a length of 12-25 amino acids.

61. A system for determining values indicating probability of protein-protein interactions in T-cell receptor (TCR)-peptide-major histocompatibility complex (MHC) (TCR:pMHC) complexes, comprisingone or more processors coupled with memory, configured to:retrieve (i) a first sequence corresponding to an MHC polypeptide and (ii) a second sequence corresponding to a peptide epitope that binds to the MHC polypeptide to form an peptide-MHC (pMHC) complex; (iii) a third sequence corresponding to a TCR-a polypeptide and (iv) a fourth sequence corresponding to a TCR-P polypeptide that binds to the TCR-a polypeptide to form a TCR, wherein the TCR binds with the pMHC complex to form a TCR-peptide-MHC complex;create, using the first sequence, the second sequence, the third sequence, and the fourth sequence, a plurality of candidate quaternary structures for the to form the TCR-peptide-MHC complex, each of the plurality of quaternary structures comprising: (i) a respective MHC structure, (ii) a respective peptide epitope structure, (iii) a respective TCR-a structure, and (iv) a respective TCR-P structure;generate, for at least one candidate quaternary structure of the plurality of candidate quaternary structures, a plurality of features comprising two or more of a confidence metric, a structure metric, or a root square mean deviation (RMSD) metric;determine, based on providing the plurality of features as input to a classifier model, at least one value indicating interaction among the MHC structure, the peptide epitope structure, the TCR-a structure, and the TCR-P structure of the at least one candidate quaternary structure for the TCR:pMHC complex; andstore, using one or more data structures, an association between the value and the at least one candidate quaternary structure for the TCR:pMHC complex.-118- 4906-5272-8748.2Atty. Dkt. No.: 115872-328262. The system of claim 61, wherein the one or more processors are further configured to determine that the TCR:pMHC complex is non-binding based on the value not satisfying a threshold value; and exclude the TCR:pMHC complex from provision.

63. The system of claim 61, wherein the one or more processors are further configured to:determine that the TCR:pMHC complex is binding, based on the value satisfying a threshold; andselect, from a plurality of candidate TCR:pMHC complexes, the at least one TCR:pMHC complex for provision.

64. The system of claim 63, wherein the TCR:pMHC complex is generated using a peptide epitope, an MHC molecule, and a TCR synthesized in accordance with a plurality sequences defining the TCR:pMHC complex selected from the plurality of candidate TCR:pMHC complexes.

65. The system of claim 64, wherein a binding affinity metric is measured from the TCR:pMHC complex to validate as a binding complex, wherein the binding affinity metric comprises at least one of a monovalent affinity or effective affinity.

66. The system of claim 65, wherein a subject at risk of cancer or diagnosed with cancer is administered with neoantigen vaccine comprising the peptide epitope corresponding to the peptide epitope structure of the TCR:pMHC complex.

67. The system of claim 66, wherein the cancer comprises at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin’s-119- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof.

68. The system of claim 61, wherein the first sequence corresponds to at least one of a wildtype MHC polypeptide or a mutant type MHC polypeptide,wherein the second sequence corresponds to a wildtype type peptide epitope or a mutant type peptide epitope,wherein the first sequence, the second sequence, the third sequence, or the fourth sequence are derived from at least one mammal comprising one or more of a human or a mouse.

69. The system of claim 61, wherein the confidence metric comprises at least one of (i) a predicted local distance difference test (pLDDT) metric, (ii) a predicted alignment error (PAE) metric, (iii) an interface predicted alignment error (iPAE), (iv) a predicted template modeling (pTM) metric, or (v) an interface predicted template modeling (ipTM) metric.

70. The system of claim 61, wherein the structure metric indicates a probability of contact between two or more of: (i) at least one MHC amino acid residue of the MHC structure, (ii) at least one peptide epitope amino acid residue of the peptide epitope structure, (iii) at least one TCR-a amino acid residue of the TCR-a structure, (iv) at least one TCR-P amino acid residue of the TCR-P structure, (v) at least one amino acid residue in the Complementarity Determining Region 3 (CDR3) portion of the TCR-a structure, or (vi) at least one amino acid residue in the CDR3 portion TCR-P structure.

71. The system of claim 61, wherein the RMSD metric is determined based on a deviation between the candidate quaternary structure and a reference quaternary structure for the TCR:pMHC complex, or-120- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282wherein the RMSD metric is determined based on an aggregation of a plurality of deviations, each of the plurality of deviations between two candidate quaternary structures of a plurality of candidate quaternary structures.

72. The system of claim 61, wherein the plurality of features further comprises a plurality of embeddings generated based on executing a protein language model using a plurality of sequences corresponding to the TCR:pMHC complex.

73. The system of claim 61, wherein the one or more processors are further configured to:retrieve, from one or more databases further comprises retrieving, from one or more databases or from a sequencing device, (i) a first input file identifying the first sequence corresponding to the MHC polypeptide, (ii) a second input file identifying the second sequence corresponding to the peptide epitope, (iii) a third input file identifying the third sequence corresponding to the TCR-a polypeptide and (iv) a fourth input file identifying the fourth sequence corresponding to the TCR-P polypeptide; andgenerate, using the first input file, the second input file, the third input file, and the fourth input file, a plurality of output files corresponding to the plurality of candidate quaternary structures.

74. The system of claim 61, wherein the classifier model is established using a plurality of training examples, each of the plurality of training examples comprising:(i) a respective example plurality of features comprising two or more of a respective example confidence metric, a respective example structure metric, a respective example root mean square deviation (RMSD) metric, for an example TCR:pMHC complex, and(ii) a respective label indicating the example TCR:pMHC complex as one of binding or non-binding.

75. The system of claim 61, wherein the one or more processors are further configured to:execute a folding model using the first sequence, the second sequence, the third sequence, and the fourth sequence to create the plurality of candidate quaternary structures;-121- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282generate, based on executing the folding model, for each candidate TCR:pMHC complex a plurality of candidate quaternary structures, each with a respective score indicating a respective accuracy of the candidate quaternary structure as determined by the folding model; andselect, from the plurality of candidate quaternary structures, the at least one candidate quaternary structure based on the respective score indicating the respective accuracy.

76. The system of claim 75, wherein the folding model comprises at least one of AlphaFold, RoseTTAFold, OpenFold, Boltz-1, OmegaFold, ESMFold.

77. The system of claim 61, wherein the one or more processors are further configured to:determine, by the one or more processors, for each candidate quaternary structure of the plurality of candidate quaternary structures, a respective plurality of features comprising two or more: (i) a respective confidence metric, (ii) a respective structure metric, (iii) a respective root mean square deviation (RMSD) metric, or (v) a respective plurality of embeddings;generate the plurality of features based on an aggregation of the respective plurality of features for each candidate quaternary structure of the plurality of candidate quaternary structures; andprovide the input to the classifier model to determine the value indicating the probability of interaction within the TCR:pMHC quaternary structure.

78. The system of claim 61, wherein the TCR comprises a TCRa polypeptide and a TCRP polypeptide that are isolated from at least one of CD8+ T cell or a CD4+ T cell, and the third sequence corresponding to TCR-a polypeptide and the fourth sequence corresponding to the TCR-P polypeptide are determined by next-generation sequencing.

79. The system of claim 61, wherein the MHC corresponding to the first sequence comprises at least one of an MHC class I or an MHC class II,-122- 4906-5272-8748.2Atty. Dkt. No.: 115872-3282wherein the MHC polypeptide comprises at least one of HLA-A, HLA-B, HLA-C, HLA-DR, HLA-DQ, or HLA-DP,wherein the MHC class I comprises an alpha chain including al domain, a2 domain, and a3 domain, and optionally a p2-microglobulin chain, andwherein the MHC class II comprises an alpha chain having al domain and a2 domain and optionally a beta chain having an pi domain and P2 domain.

80. The system of claim 61, wherein the peptide epitope corresponding to the second sequence originates from at least one of a pathogen, a self-antigen, or cancer in a subject, andwherein the peptide epitope comprises at least one of a class I epitope or a class II epitope, and wherein the class I epitope has a length of 8-11 amino acids, and wherein the class II epitope has a length of 12-25 amino acids.-123- 4906-5272-8748.2