Benchmarking sequence-based and AlphaFold-based methods for pMHC-II binding core prediction: distinct strengths and...

Area of Science:

• Immunology
• Computational Biology
• Structural Biology

Background:

• Peptide-MHC class II (pMHC-II) interactions are vital for T-cell recognition and distinguishing self from non-self antigens.
• Accurate prediction of the pMHC-II binding core is essential for understanding T-cell receptor engagement and immune responses.
• High MHC-II polymorphism and peptide promiscuity necessitate computational methods for predicting pMHC-II interactions, with recent advances in AlphaFold-based structure prediction offering new avenues.

Purpose of the Study:

• To benchmark sequence-based and AlphaFold-based structure prediction methods for pMHC-II binding core prediction.
• To evaluate the performance of AlphaFold2 fine-tuned (AF-FT) and AlphaFold3 (AF3) against NetMHCIIpan-4.3 and DeepMHCII.
• To develop consensus strategies integrating sequence- and structure-based models for improved pMHC-II binding prediction.

Main Methods:

• A non-redundant dataset of 72 pMHC-II complexes was curated from the IMGT database, including negative controls.
• Performance evaluation of NetMHCIIpan-4.3, DeepMHCII, AF-FT, and AF3 using precision, recall, and F1 score.
• Development of consensus strategies combining AlphaFold-based and sequence-based models for predicting pMHC-II binding status.

Main Results:

• AlphaFold-based methods (AF3 and AF2-FT) demonstrated high recall for positive binders but frequently misclassified unbound peptides.
• NetMHCIIpan achieved the highest negative recall for non-binders but had lower positive recall.
• Consensus approaches combining AlphaFold-based binder prediction with NetMHCIIpan filtering significantly improved overall prediction precision.

Conclusions:

• AlphaFold-based and sequence-based methods possess complementary strengths for pMHC-II binding core prediction.
• AlphaFold-based methods excel at identifying positive binders, whereas NetMHCIIpan is superior for identifying non-binders.
• Future efforts should focus on enhancing AlphaFold-based models' prediction of unbound peptides and improving NetMHCIIpan's binding prediction accuracy.