GeoPMB: An Interface-Aware Geometric Deep Learning Framework for Peptide-MHCI Binding Prediction with Evolutionary...

Area of Science:

• Computational Biology
• Immunoinformatics
• Structural Bioinformatics

Background:

• Peptide-Major Histocompatibility Complex class I (pMHCI) binding is crucial for adaptive immunity and therapeutic development.
• Experimental methods for identifying immunogenic peptides are resource-intensive and time-consuming.
• Existing computational predictors have limitations in generalization and capturing 3D structural details.

Purpose of the Study:

• To develop a novel computational framework, GeoPMB, for accurate prediction of pMHCI binding.
• To address the limitations of existing sequence-based and structure-based prediction methods.
• To enhance the identification of immunogenic peptides for vaccine and immunotherapy design.

Main Methods:

• Integration of geometric deep learning with pretrained protein language models (PLMs).
• Utilizing PLMs for evolutionary context and semantic feature extraction from sequences.
• Employing a geometric graph network to model spatial dependencies and physicochemical features of the pMHCI complex.

Main Results:

• GeoPMB significantly outperforms state-of-the-art methods in pMHCI binding specificity and affinity prediction.
• Demonstrated superior generalization capabilities, particularly for underrepresented or rare alleles.
• Achieved high performance in antibody-antigen docking pose ranking, showcasing versatility.

Conclusions:

• GeoPMB offers a powerful, structurally aware approach for precision immunology.
• The framework enhances the prediction of immunogenic peptides, facilitating therapeutic design.
• GeoPMB represents a significant advancement in computational prediction for immunology and drug discovery.