USMPep

USMPep predicts peptide binding affinity to major histocompatibility complex (MHC) molecules to support identification of neoepitopes for immunotherapy and personalized cancer treatment.

Key Features:

• RNN architecture: Achieves state-of-the-art performance in MHC class I binding prediction using a simple recurrent neural network (RNN) as a single generic model with a unified set of hyperparameters.
• Versatility across peptide lengths: Handles peptides of varying lengths without relying on heuristics.
• Benchmark performance: Demonstrates competitive predictive performance on benchmark datasets including IEDB and a recent HPV dataset.
• Ensembling and pretraining: Accuracy can be improved through ensembling multiple regressors and language model pretraining.
• MHC class II applicability: Shows solid performance in predicting MHC class II binding affinities despite limited training data.

Scientific Applications:

• Neoepitope identification for immunotherapy: Predicts peptide–MHC interactions to aid selection of neoepitopes for personalized cancer vaccines and other immunotherapeutic strategies.
• MHC class II prediction support: Provides binding-affinity predictions for MHC class II to extend applications beyond MHC class I studies.

Methodology:

Training a recurrent neural network on peptide sequences to predict binding affinities to MHC molecules using a universal single-model architecture with unified hyperparameters; performance can be enhanced by ensembling regressors and language model pretraining and handles peptide length variation without heuristics.