BERTMHC: improved MHC-peptide class II interaction prediction with transformer and multiple instance learning.

Area of Science:

• Genomics and Bioinformatics
• Immunology
• Computational Biology

Background:

• Advancements in cancer genomics enable targeted therapeutic vaccine development.
• Accurate prediction of peptide-major histocompatibility complex (MHC) interactions is vital for vaccine efficacy.
• Current prediction models for MHC class II presentation lag behind MHC class I.

Purpose of the Study:

• To develop an advanced computational model for predicting peptide binding and presentation to MHC alleles.
• To improve the accuracy of predicting MHC class II epitopes for enhanced therapeutic vaccine design.

Main Methods:

• Developed a transformer neural network model utilizing self-supervised pretraining on protein sequences.
• Implemented a multiple instance learning (MIL) framework to deconvolute mass spectrometry data for MHC presentation.
• Evaluated model performance against state-of-the-art methods using peptide and MHC sequence data.

Main Results:

• Self-supervised pretraining significantly boosted model performance on prediction tasks.
• The combined pretraining and MIL approach outperformed existing models for both MHC binding and cell surface presentation predictions.
• The model demonstrates superior accuracy in predicting peptide presentation by MHC class II alleles.

Conclusions:

• The novel transformer-based model with pretraining and MIL significantly advances MHC class II presentation prediction.
• This improved prediction capability is crucial for the rational design of highly specific and effective cancer therapeutic vaccines.
• The developed tools are publicly available to facilitate further research and development in cancer immunotherapy.