Multimodal learning on heterogeneous subgraphs and LLMs representation for MHC-peptide binding affinity prediction

Area of Science:

• Computational biology
• Immunoinformatics
• Machine learning

Background:

• Accurate prediction of Major Histocompatibility Complex (MHC)-peptide binding affinity is critical for developing effective immunotherapies.
• Current computational methods face challenges in simultaneously modeling functional semantics, evolutionary constraints, and structural dynamics of polymorphic residues.
• There is a need for advanced models that can capture complex interactions between MHC molecules and peptides.

Purpose of the Study:

• To develop a novel computational model, the Contrast learning-based Multi-feature Heterogeneous Subgraph model (CMHS), for enhanced MHC-peptide binding affinity prediction.
• To integrate diverse data representations, including sequence and structural information, for a more comprehensive understanding of MHC-peptide interactions.
• To establish a new benchmark in predicting hypervariable immune interactions.

Main Methods:

• Utilized LoRA fine-tuning with ESM2 and BLOSUM50 for MHC-exclusive sequence representation, capturing functional dependencies and conserved residues.
• Employed a biophysics-guided heterogeneous graph network for structural representation, incorporating a novel trainable Gaussian noise layer guided by crystallographic B-factors.
• Implemented a three-stage message-passing framework with subgraph aggregation and extraction, followed by contrastive learning to align sequence and graph representation spaces.

Main Results:

• The CMHS model demonstrated significant improvements in prediction accuracy across 16 HLA allele benchmarks.
• Achieved an average Spearman rank correlation coefficient (SRCC) improvement of 8.7% compared to existing methods.
• Showcased an average Area Under the Curve (AUC) improvement of 7.6%, indicating enhanced predictive performance.

Conclusions:

• The CMHS model represents a new paradigm for predicting hypervariable immune interactions, particularly MHC-peptide binding affinity.
• The integration of sequence and structural data through contrastive learning significantly enhances prediction capabilities.
• This approach offers a promising direction for advancing immunotherapeutic development through improved computational modeling.