PMGen

PMGen predicts peptide-MHC (pMHC) complex structures and designs optimized peptide sequences through an integrated framework for structure prediction and structure-guided design of variable-length peptides across MHC Class I and II.

Key Features:

• Structure prediction: PMGen enforces anchor constraints within AlphaFold2 through initial guess and template engineering strategies to achieve state-of-the-art structural fidelity without model fine-tuning.
• Peptide design: PMGen uses ProteinMPNN sampling on predicted backbones to generate higher affinity peptides while preserving parental 3D conformation.
• Cross-class coverage: PMGen handles variable-length peptides across both MHC Class I and MHC Class II.
• Binding-core identification: AlphaFold pLDDT scores enable sequence-independent binding-core identification.
• Structure recovery: PMGen can recover incorrectly predicted anchor positions and accurately captures mutation-induced conformational changes in neoantigen/wild-type pairs.

Scientific Applications:

• Immunotherapy design: PMGen enables structure-driven design of mimotopes and altered peptide ligands by exploiting spatial and biophysical insights from pMHC structures.
• Neoantigen analysis: PMGen accurately captures mutation-induced conformational changes in neoantigen/wild-type pairs.
• Machine learning training data generation: PMGen generates large-scale high-confidence pMHC structures to improve downstream models such as ProteinMPNN, as demonstrated by improving peptide sequence recovery from 0.14 to 0.64 on 85 unseen MHC-I alleles.

Methodology:

PMGen enforces anchor constraints within AlphaFold2 through two complementary strategies, initial guess and template engineering. For peptide design, PMGen applies ProteinMPNN sampling on predicted pMHC backbones.