Systematically benchmarking peptide-MHC binding predictors: From synthetic to naturally processed epitopes.

Area of Science:

• Immunoinformatics
• Computational Biology
• Machine Learning in Immunology

Background:

• Numerous machine learning (ML) predictors exist for T-cell epitope identification based on MHC binding affinities.
• Selecting optimal tools is challenging due to evolving data and methods.
• Reliability of predictors on newly generated, naturally processed epitopes remains unevaluated.

Purpose of the Study:

• To benchmark existing MHC-binding predictors using novel, untested synthetic and naturally processed epitope data.
• To evaluate the performance of various ML and structural modeling approaches.
• To provide an unbiased view for establishing best practices in T-cell epitope prediction.

Main Methods:

• Benchmarking 18 MHC-binding predictors on 32 HLA class I and 24 HLA class II alleles.
• Utilizing blind test sets of synthetic and naturally processed epitopes.
• Comparing artificial neural network (ANN), regression-based ML, and structural modeling methods.

Main Results:

• ANN-based predictors (mhcflurry, nn_align) outperformed others for MHC class I and II predictions (AUC=0.911).
• NetMHCpan4 showed comparable performance; customized mhcflurry achieved similar accuracy.
• Predictors trained on elution data (NetMHCpan4, MixMHCpred) performed better on naturally processed ligands.

Conclusions:

• ANN approaches are superior for MHC-binding prediction.
• mhcflurry and nn_align are top performers for specific MHC classes.
• Predictor performance varies across HLA types and datasets; elution data improves accuracy for natural ligands.