BERTMHC

BERTMHC predicts peptide binding and antigen presentation by Major Histocompatibility Complex (MHC) class II molecules using a transformer neural network pretrained on protein sequences and a multiple-instance learning framework.

Key Features:

• Transformer Neural Network Model: Implements a transformer architecture to model sequential protein/peptide information for MHC class II binding and presentation prediction.
• Self-Supervised Pretraining: Leverages self-supervised pretraining on a large corpus of protein sequences to improve representation learning and generalization.
• Multiple Instance Learning (MIL) Framework: Applies MIL to deconvolve mass spectrometry data where multiple MHC alleles may present the same peptide, enabling allele-level assignment of presentations.
• Prediction Targets: Predicts both peptide–MHC class II binding affinities and mass spectrometry-derived peptide presentation.
• Implementation: Provided as a Python-based implementation for model training and inference.

Scientific Applications:

• Immunotherapy and Therapeutic Vaccine Design: Supports selection of class II epitopes to design vaccines that elicit sustained immune responses against cancer-associated genetic alterations.
• Epitope Discovery from Mass Spectrometry: Facilitates assignment and analysis of peptides identified by mass spectrometry to specific MHC class II alleles for antigen presentation studies.

Methodology:

The approach combines self-supervised pretraining of a transformer model on large protein sequence datasets with a multiple-instance learning framework to address peptide assignment in mass spectrometry-derived MHC class II presentation data.