Large-Scale Structure-Based Prediction of Stable Peptide Binding to Class I HLAs Using Random Forests

Area of Science:

• Immunoinformatics
• Structural Biology
• Computational Biology

Background:

• Predicting peptide-HLA (pHLA) binding is crucial for immunotherapy development.
• Current machine learning predictors primarily rely on sequence data.
• Exploring structure-based prediction models offers potential for improved generalization, especially for underrepresented HLA alleles.

Purpose of the Study:

• To develop and validate a structure-based pan-allele model for predicting stable pHLA binding.
• To leverage a large dataset of modeled pHLA structures for training.
• To assess the generalizability and interpretability of the structure-based approach.

Main Methods:

• Utilized APE-Gen to model over 150,000 pHLA structures, creating the largest dataset of its kind.
• Extracted residue-residue distance features from modeled pHLA structures.
• Trained a random forest model using these structure-based features for binding prediction.

Main Results:

• The structure-based model achieved competitive Area Under the Receiver Operating Characteristic (AUROC) values in leave-one-allele-out validation.
• The model demonstrated strong performance using significantly less data compared to sequence-based methods.
• An interpretation analysis revealed how model predictions are composed, aligning with chemical intuition.

Conclusions:

• Structure-based prediction models can achieve generalizable and interpretable pHLA binding predictions.
• This approach offers a promising alternative to sequence-based methods, particularly for alleles with limited experimental data.
• The developed model represents a significant advancement in leveraging structural information for immunoinformatics applications.