MHCRoBERTa

MHCRoBERTa predicts interactions between peptides and type I MHC molecules to estimate peptide–MHC class I binding affinities for neoantigen identification.

Key Features:

• RoBERTa pre-training approach: A RoBERTa-based pre-training strategy is tailored for predicting interactions between type I MHC molecules and peptides.
• Transfer learning with label-agnostic protein sequences: The model leverages transfer learning using label-agnostic protein sequences to improve prediction accuracy.
• Performance metrics: Benchmark results report a Spearman rank correlation coefficient (SRCC) of 0.785 and an area under the curve (AUC) of 0.817, with the SRCC 14.3% higher than NetMHCpan3.0 and 3% higher than MHCflurry.
• IC50 prediction improvement: Demonstrates enhanced prediction of IC50 values relevant to peptide–MHC binding affinity assessments.
• Visualization of multi-head self-attention: Provides visualization of multi-head self-attention across layers and heads to assess learned syntax and semantics for prediction tasks.

Scientific Applications:

• Cancer immunotherapy and neoantigen identification: Used to screen and prioritize peptide candidates for neoantigen discovery and personalized cancer vaccine development.

Methodology:

RoBERTa pre-training tailored for peptide–MHC class I interaction prediction, transfer learning using label-agnostic protein sequences, and visualization of multi-head self-attention across layers and heads.