IMGT/RobustpMHC: robust training for class-I MHC peptide binding prediction.

Area of Science:

• Immunoinformatics
• Computational Biology
• Machine Learning

Background:

• Accurate prediction of peptide-Major Histocompatibility Complex (MHC) class I binding is vital for vaccine development and immunotherapies.
• Current deep neural network approaches for peptide-MHC (pMHC) prediction have limitations, including reliance on pseudo-sequence extraction and poor generalization across datasets.

Purpose of the Study:

• To develop novel computational methods for more accurate and robust pMHC binding prediction.
• To address the limitations of existing pMHC prediction methods by utilizing full peptide and MHC sequences and leveraging unlabeled data.

Main Methods:

• Introduced PerceiverpMHC, a transformer-based architecture for learning representations from full peptide and MHC sequences.
• Developed IMGT/RobustpMHC, employing a self-supervised learning strategy on unlabeled data to enhance prediction robustness.
• Compiled CrystalIMGT, a crystallography-verified dataset, and developed a transfer learning pipeline to address distribution gaps.

Main Results:

• Demonstrated that neural architectures can effectively learn pMHC binding intricacies from full sequences.
• IMGT/RobustpMHC achieved over 6% improvement in binding prediction accuracy compared to state-of-the-art methods across eight diverse datasets.
• The transfer learning pipeline successfully mitigated distribution gaps presented by the CrystalIMGT dataset.

Conclusions:

• PerceiverpMHC and IMGT/RobustpMHC offer significant advancements in pMHC binding prediction accuracy and generalization.
• Self-supervised learning and transfer learning are effective strategies for improving robustness and addressing data distribution challenges in pMHC prediction.
• These methods have the potential to enhance the design of vaccines and immunotherapies.