Explainable Model Layers in TCR-pMHC Binding

• Explainable Model Layers (TCR-EML) are a novel architectural design incorporating transformer-based cross-attention and contact prototypes to provide residue-level interpretability for TCR-pMHC binding predictions.
• The framework uses hierarchical feature fusion and differentiable binarization to model biochemical interactions, achieving superior ROC-AUC performance and robust generalization on unseen epitopes.
• By aligning deep learning outputs with experimental binding maps, TCR-EML enhances model trustworthiness and aids applications in vaccine design, cancer immunotherapy, and autoimmune research.

Explainable Model Layers (TCR-EML) are architectural components within machine learning models, particularly transformer-based predictors for T-cell receptor (TCR) and peptide–MHC (pMHC) binding, that provide interpretable residue-level explanations for classification decisions. Unlike post-hoc attribution techniques, TCR-EML models incorporate structural and biochemical mechanisms directly into the layer design, enabling native transparency of how particular amino acid residues and their interactions drive predicted TCR-pMHC binding outcomes. This approach allows for both highly accurate predictions and biologically meaningful rationale, addressing a critical limitation in the application of deep learning to immunological problems.

1. Layer Architecture and Fusion Mechanisms

TCR-EML is structured around two core architectural blocks:

• Feature Enhancement and Fusion (FEF) Block: Multiple cross-attention layers systematically integrate protein LLM embeddings of CDR3α, CDR3β, and peptide sequences. Initial cross-attention is performed to fuse CDR3a and CDR3b representations:

$$E_{a \rightarrow b} = A(E_a, E_b), \quad E_{b \rightarrow a} = A(E_b, E_a)$$

Subsequently, peptide embeddings $E_e$ are fused with these cross-fused TCR representations through further cross-attention operations. This hierarchical fusion enables invariant feature interaction across TCR chains and between TCR and peptide, capturing biologically relevant context for downstream contact modeling.

• Contact Prototype Layers: These layers generate residue-by-residue contact scores using a domain-informed similarity measure:

$$S = \frac{E_1 \cdot E_2^T}{\|E_1\| \|E_2\|} \cdot \tau$$

where $\tau$ is a trainable temperature parameter. This similarity score reflects the inverse of the estimated contact distance, emphasizing direct biological interpretability. The contact scores $S$ are then thresholded using a differentiable binarization:

$M_i = \sigma\left[(S - t_i) \cdot N\right]$

where $\sigma(\cdot)$ is the sigmoid, $t_i$ is the threshold, and $N$ is a normalizing constant.

The composite scores are aggregated as:

Component	Formula	Interpretation
Similarity score	$$S = \frac{E_1 \cdot E_2^T}{\|E_1\| \|E_2\|} \cdot \tau$$	Contact likelihood per residue
Binarized contact	$M_i = \sigma[(S - t_i) \cdot N]$	Contact presence indicator
Output score	$\hat{y} = \frac{w_{a,e} + w_{b,e}}{2}$	Overall binding prediction

These contact prototypes allow inspection of the model's "reasoning" through direct correspondence with residue-level contact maps and experimentally verified binding interfaces.

2. Integration of Biochemical Binding Mechanisms

TCR-EML explicates TCR-pMHC binding by emulating domain knowledge of physical residue contacts:

• Similarity scores between fused TCR and peptide embeddings model the likelihood of structural proximity.
• Thresholded and aggregated similarity values approximate the binary nature of residue contacts observed in crystallographic structures.
• The use of temperature scaling ($\tau$) and differentiable binarization ($M_i$) ensures smooth optimization while retaining strict correspondence to biochemical constraints.

This mechanistic fidelity enables quantification of binding events in terms directly relatable to experimentally determined contact regions, providing not just predictions but interpretable maps consistent with underlying molecular interactions.

3. Predictive Accuracy and Generalization

Empirical evaluation with various protein LLM backbones demonstrates:

• Substantially increased ROC-AUC for TCR-pMHC binding prediction compared to linear classifier baselines and other transformer models, e.g., achieving up to 99.9% ROC-AUC with ProteinBERT on top-100 epitopes.
• Consistent performance gains across expanded peptide test sets and strong generalization to unseen epitope sequences.
• The improvement margin ranges from 4% to 20% over non-explainable baselines, indicating that embedding explainability in the layer design does not compromise but enhances predictive capability.

These results substantiate the utility of explainable model layers in maintaining performance while introducing transparent, residue-level interpretation.

4. Formal Evaluation of Explainability

Explainability is quantitatively assessed using the Binding Region Hit Rate (BRHR), based on the TCR-XAI benchmark. BRHR is computed as the fraction of top-k residues identified by the contact prototype that overlap with experimental binding regions (usually defined by residue proximity in crystal structures):

• BRHR values consistently exceed 0.71 on key peptide-to-CDR3 interactions.
• High overlap signifies that the model's intermediate explanations closely match the true physical interface responsible for recognition.

This level of explanatory power facilitates validation, scientific scrutiny, and clinical translation, marking a shift from opaque predictions to biochemically rational outputs.

5. Applications in Immunology and Translational Medicine

The mechanistic transparency enabled by TCR-EML supports several critical applications:

• Vaccine Design: The mapping of residue importance informs the selection of epitopes likely to elicit robust T cell responses.
• Cancer Immunotherapy: Precise determination of TCR binding hot spots permits the design of T cell therapies targeting neoantigens.
• Autoimmune Disease Research: Identifying the binding determinants provides insight into the molecular basis of self-reactivity.

In each application, the ability to directly relate predictions to the physical interactions underlying immune recognition is essential for model trustworthiness, hypothesis generation, and therapeutic intervention.

6. Relation to Existing Methods and Paradigm

Relative to post-hoc explainability and attention-based attribution, TCR-EML provides native, domain-embedded explanation:

• The model's prototype layers differ from token-based attribution by computing direct residue-residue similarity using PLM features and then matching these to biochemical contact maps.
• Transparency is embedded at the layer level, enabling post-training inspection without recourse to external interpretation algorithms.
• The approach represents an "explain-by-design" paradigm shift for biological sequence modeling, not previously adopted in TCR-pMHC binding prediction.

These distinctions formalize a rigorous path from raw sequence data to interpretable immunological inference.

7. Significance and Future Directions

TCR-EML provides a robust framework unifying predictive performance and model interpretability, with the following implications:

• Bridges the gap between state-of-the-art deep learning and mechanistically reliable biological modeling.
• Promotes scientific discovery through residue-resolved explanations, facilitating hypothesis testing and experimental validation.
• Opens avenues for expanded domain application: integrating prototype layers into other protein interaction prediction tasks or extending to non-sequence modalities.

A plausible implication is that future explainable architectures will continue to draw on mechanistic principles, bringing machine learning outputs into tighter alignment with experimental and clinical practice.