We introduce GeoPMB, an interface-aware geometric deep learning framework designed for robust peptide-MHCI binding prediction by integrating evolutionary insights from protein language models with fine-grained geometric constraints. GeoPMB mitigates the overfitting common in sequence-based methods while capturing intricate geometric details often overlooked by coarse structure-based methods, enabling superior generalization to rare alleles and versatile application in antibody-antigen docking conformation ranking.
Required dependencies and versions:
- python 3.8.16
- pytorch 1.13.1
- biopython 1.79
- sentencepiece 0.1.99
- transformers 4.39.3
- numpy 1.23.5
- torch-geometric 2.3.0
- pytorch-scatter 2.1.1
- pytorch-cluster 1.6.1
- Peptide-MHCI Datasets: The two benchmark pMHCI datasets used in our study are available in
/data. - Antibody-Antigen Data: All antibody-antigen pairs are available in
/data, and complex structures can be downloaded from SAbDab.
GeoPMB requires 3D structural inputs. We utilize PANDORA to generate reliable 3D structures for peptide-MHCI pairs. Please ensure PANDORA is installed and configured correctly.
Convert PDB structures into geometric graphs and extract evolutionary features. The implementation details are located in /src/dataset/feature_utils.py.
Inference commands using trained models (located in /models):
python src/test.py \
--dataset_path ./examples \
--feature_path ./examples/Feature \
--output_path ./outputs