When Multimodal Fusion Fails: Contrastive Alignment as a Necessary Stabilizer for TCR-Peptide Binding Prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Multimodal learning in biology often struggles with imperfect auxiliary data.
• TCR-peptide binding prediction faces challenges with noisy, structure-derived residue graphs.
• Naive fusion of sequence and graph data can degrade performance, especially with limited supervision.

Purpose of the Study:

• To develop a method for effectively integrating imperfect structural information into multimodal learning for biological predictions.
• To improve the stability and generalization of TCR-peptide binding prediction models.
• To address the limitations of naive data fusion in the presence of noisy biological data.

Main Methods:

• Introduced TRACE, a framework using parallel sequence (ESM-2) and graph (GNN) towers.
• Applied CLIP-style intra-entity contrastive alignment to encourage modality consistency.
• Evaluated using a leakage-controlled TCHard RN protocol with pair-disjoint splits and audited methodology.

Main Results:

• TRACE achieved the best mean AUROC (0.578±0.033) among matched baselines.
• Intra-entity alignment stabilized training, providing consistent gains and robustness to graph corruption.
• The method prevented performance collapse under limited supervision, outperforming unconstrained fusion.

Conclusions:

• Effective integration and careful evaluation of modalities are crucial for multimodal learning in biology.
• TRACE offers a stable and effective approach for leveraging imperfect structural data alongside sequence information.
• The developed protocol provides a reliable benchmark for evaluating multimodal biological prediction models.