A structure-informed deep learning framework for modeling TCR-peptide-HLA interactions

This paper introduces StriMap, a structure-informed deep learning framework that accurately models TCR-peptide-HLA interactions to enable the identification of antigenic drivers in autoimmunity and the prioritization of targets for immunotherapy, as demonstrated by its successful discovery of validated molecular mimics linking ankylosing spondylitis and inflammatory bowel disease.

The Gist

Imagine your body's immune system as a highly sophisticated security force. Its job is to patrol your body, looking for intruders (like viruses or bacteria) or traitors (like cancer cells).

Here is how the security system works, using a simple analogy:

1. The HLA (The Bulletin Board): Think of your cells as buildings. Every building has a "bulletin board" (called an HLA) on its front door. This board displays tiny scraps of protein (peptides) found inside the building.
   • If the scrap is from a normal meal, the board says, "Everything is fine."
   • If the scrap is from a virus or a cancer cell, the board says, "Intruder alert!"
2. The TCR (The Security Guard): The T-cell Receptor (TCR) is the security guard patrolling the streets. The guard doesn't look at the whole building; they only look at the bulletin board.
3. The Interaction: If the guard sees a "bad" scrap on the board that matches their specific training, they sound the alarm and attack.

The Problem:
Scientists have gotten really good at predicting what gets put on the bulletin board (which scraps the cell displays). But predicting whether a specific security guard will actually recognize that scrap and attack is incredibly hard. It's like knowing what's on a flyer, but not knowing if a specific person will stop to read it.

The space of possibilities is massive. There are billions of possible guards and billions of possible flyers. Testing them all in a lab is too slow and expensive.

The Solution: StriMap
The authors of this paper built a new AI tool called StriMap (Structure-informed TRi-molecular Interaction Mapping).

Think of StriMap as a super-intelligent "Matchmaker" AI that doesn't just look at the text of the flyer and the resume of the guard. It looks at the shape and the chemistry of both.

• Old AI: Looked at the sequence of letters (like reading a book).
• StriMap: Looks at the 3D shape of the paper and the guard's hand, simulating how they physically fit together. It combines the "text" (sequence) with the "shape" (structure) and the "feel" (chemistry) to predict if they will lock together perfectly.

How They Tested It (The Two Big Wins)

The researchers tested this AI in two very different scenarios:

1. Fighting Cancer (The "Needle in a Haystack" Game)

In cancer, the body creates "mutant" scraps (neoepitopes) that shouldn't be there. Doctors want to find the one specific guard that can spot the one specific mutant scrap to build a personalized vaccine.

• The Test: They asked StriMap to find the right guards for known cancer mutations.
• The Result: StriMap was much better at finding the "needle in the haystack" than previous tools. It could rank the best candidates so high that doctors only need to test a few, saving time and money.

2. Solving Autoimmune Mysteries (The "Mimicry" Detective)

Sometimes, the security system gets confused. It sees a "bad" scrap from a harmless bacteria that looks exactly like a "bad" scrap from your own body. This causes the guard to attack your own tissues (Autoimmunity). This is called Molecular Mimicry.

• The Mystery: Ankylosing Spondylitis (AS) is a painful joint disease linked to a specific genetic marker (HLA-B27). Scientists suspected bacteria in the gut were the trigger, but they didn't know which bacteria or which protein.
• The Test: StriMap scanned 13 million bacterial protein scraps from gut bacteria. It asked: "Which of these 13 million scraps looks like a match for the AS-associated guards?"
• The Discovery: The AI picked out a few top candidates. The researchers tested them in a lab, and three of them actually worked, activating the immune cells.
• The Twist: One of these "bad" bacterial scraps was found to be very common in patients with Inflammatory Bowel Disease (IBD) as well. This suggests that AS and IBD might share a common bacterial trigger, offering a new clue for treating both diseases.

Why This Matters

Before StriMap, finding these matches was like trying to find a specific key in a pile of a billion keys by trying them one by one. StriMap is like a metal detector that instantly tells you which 10 keys are likely to fit the lock.

• For Cancer: It helps design better vaccines and therapies faster.
• For Autoimmunity: It helps us understand why our immune system attacks us, potentially leading to cures that stop the root cause rather than just treating symptoms.

In short, StriMap is a powerful new lens that helps us see the invisible handshake between our immune cells and the world around them, helping us fight disease more intelligently.

Technical Summary

1. Problem Statement

The interaction between T cell receptors (TCRs), peptides, and human leukocyte antigens (HLAs) is the fundamental mechanism of antigen-specific T cell immunity. While computational methods have advanced in predicting peptide-HLA (pHLA) presentation, accurately modeling the coupled TCR-pHLA recognition remains a significant bottleneck.

• Limitations of Current Approaches: Existing deep learning models often treat pHLA presentation and TCR recognition as independent tasks, relying heavily on sequence data while ignoring structural constraints. This leads to poor generalizability, especially in "zero-shot" scenarios (predicting interactions for unseen epitopes or TCRs) and fails to capture the conditional dependence where TCR recognition is contingent upon the peptide being presented by the HLA.
• Data Scarcity: Experimental data for TCR-pHLA interactions is sparse, biased, and lacks comprehensive negative examples, making it difficult to train models that can distinguish true binders from near-neighbor decoys.

2. Methodology: The StriMap Framework

The authors developed StriMap (Structure-informed TRi-molecular Interaction Mapping), a unified deep learning framework that jointly models pHLA presentation and TCR recognition.

Core Architecture

• Sequence and Structure Feature Extractor (SSFE): A multimodal encoder that processes input sequences (Peptide, HLA, TCR $\alpha$/$\beta$ chains) by fusing three distinct feature channels:
  1. Physicochemical Features: Residue-level properties (hydrophobicity, charge, volume, etc.) derived from AAindex.
  2. Sequence Context: Evolutionary embeddings from pretrained protein language models (e.g., ESM2, ProtT5).
  3. Structural Geometry: 3D structural features predicted by ESMFold (backbone torsion angles, solvent accessibility, contact counts) processed through an Equivariant Graph Neural Network (EGNN) to encode spatial relationships invariant to rotation and translation.
• Interaction Modeling: The framework employs Bilinear Attention Networks (BAN) to model pairwise residue-residue interactions between molecular components (e.g., peptide vs. HLA pocket, TCR vs. pHLA complex). This allows for fine-grained modeling of the binding interface without constructing full outer-product tensors.
• Coupled Architecture: StriMap uses a hierarchical approach:
  1. First, it models pHLA presentation.
  2. Second, it conditions TCR recognition on the learned pHLA landscape. This enforces biological consistency, ensuring the model understands that TCRs only recognize peptides that are successfully presented.

Training Innovations

• Dynamic Negative Sampling (DNS): To address the scarcity of validated negative pairs, the model regenerates shuffled negative pairs (mismatched TCR-pHLA) at every training epoch, expanding the effective negative space.
• Mutation-Aware Hard Negatives: For cancer applications, the model generates "hard negatives" by introducing small mutations (1–3 amino acids) to positive peptides or TCR CDR3 regions. This forces the model to learn subtle distinctions between mutant neoantigens and wild-type self-peptides.
• Loss Function: Utilizes Focal Loss to handle class imbalance and prioritize hard-to-classify examples.

3. Key Contributions

1. Unified Tri-molecular Modeling: StriMap is the first framework to jointly model the entire TCR-pHLA recognition cascade, integrating structural, physicochemical, and sequence-context features.
2. State-of-the-Art Performance: It achieves superior performance across multiple benchmarks for both pHLA presentation and TCR-pHLA interaction prediction, outperforming leading tools like NetMHCpan, NetTCR, and deepAntigen.
3. Generalizability: The framework demonstrates robust performance in out-of-distribution (OOD) settings, including zero-shot prediction of unseen epitopes and cross-patient inference.
4. Dual Application Utility: The authors validated the framework in two distinct clinical domains:
   • Cancer Immunotherapy: Prioritizing neoepitopes for vaccines and TCRs for adoptive cell therapy.
   • Autoimmunity: Discovering microbial molecular mimics driving ankylosing spondylitis (AS).

4. Key Results

Benchmarking Performance

• pHLA Prediction: StriMap achieved top-tier AUPRC and AUROC scores across ID (in-distribution), DS (distribution-shifted), and OOD (out-of-distribution) benchmarks. It successfully captured sub-allelic heterogeneity (e.g., distinct binding motifs within HLA-B*27:05).
• TCR-pHLA Prediction: In the IMMREP23 challenge and other independent datasets, StriMap outperformed NetTCR and MixTCRpred. Ablation studies confirmed that removing structural modules or dynamic negative sampling significantly degraded performance.
• Vaccine/TCR Prioritization: In a melanoma dataset, StriMap's "joint score" (combining pHLA and TCR scores) significantly improved the ranking of experimentally validated immunogenic neoepitopes compared to pHLA-only models.

Case Study: Cancer (TP53 Neoantigens)

• StriMap was tested on a zero-shot setting for TP53 mutations. Using mutation-aware negative sampling, it successfully ranked cognate TCRs for mutant peptides higher than baseline methods (TAPIR, NetTCR-2.0), even when neither the peptide nor TCR was seen during training.
• It effectively disentangled whether immunogenicity changes were driven by altered HLA presentation or altered TCR binding.

Case Study: Autoimmunity (Ankylosing Spondylitis)

• Screening: StriMap screened 13 million peptides derived from 43,241 proteins of gut bacteria linked to AS.
• Discovery: It identified candidate molecular mimics (e.g., peptides from Streptococcus and Akkermansia muciniphila) predicted to bind HLA-B*27:05 and activate TRBV9+ T cells.
• Experimental Validation:
  • Three top-ranked bacterial peptides (ARVMALMPF, PRIMMISMK, GRILALVPK) were experimentally validated to activate AS8.4 TCR-expressing T cells in a Jurkat reporter assay.
  • ARVMALMPF (from Streptococcus) was found to be significantly enriched in patients with Inflammatory Bowel Disease (IBD), suggesting a shared microbial trigger between AS and IBD.

5. Significance and Impact

• Rational Immunotherapy Design: StriMap provides a generalizable tool for selecting neoepitopes for personalized cancer vaccines and identifying TCRs for adoptive cell therapy, addressing the critical need to prioritize a small number of candidates from vast libraries.
• Mechanistic Insight into Autoimmunity: The study bridges the gap between microbial genomics and autoimmune pathogenesis, providing a computational pipeline to identify specific bacterial triggers for diseases like AS and IBD.
• Open Access: The authors released the codebase and a web portal (www.strimap.com), enabling the community to run predictions and fine-tune models on user-specific data, fostering iterative improvement through "lab-in-the-loop" validation.

In summary, StriMap represents a paradigm shift from independent sequence-based prediction to a structure-informed, coupled deep learning framework, significantly advancing the ability to decode the "language" of T cell immunity for therapeutic applications.