USMPep: universal sequence models for major histocompatibility complex binding affinity prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Immunology

Background:

• Immunotherapy offers personalized cancer treatment, but predicting peptide-Major Histocompatibility Complex (MHC) binding remains a challenge.
• Existing MHC binding prediction algorithms often require complex training, specific peptide lengths, or heuristic methods.

Purpose of the Study:

• To develop a simplified yet effective algorithm for MHC binding prediction.
• To evaluate the performance of a recurrent neural network for MHC class I and II binding prediction.

Main Methods:

• Utilized a simple recurrent neural network architecture (USMPep).
• Trained and evaluated the model on benchmark datasets including IEDB and a recent HPV dataset.
• Assessed performance on both MHC class I and class II binding prediction tasks.

Main Results:

• USMPep achieved state-of-the-art results on MHC class I binding prediction using a single architecture and hyperparameters.
• The model demonstrated competitive performance even when trained from scratch.
• USMPep showed solid performance for MHC class II binding prediction despite limited training data.

Conclusions:

• Competitive MHC binding affinity prediction is achievable with standard neural network architectures and training procedures.
• The developed method avoids reliance on heuristics, offering a more robust approach.