TransMHCII: a novel MHC-II binding prediction model built using a protein language model and an image classifier

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in immunology

Background:

• Deep learning models like AlphaFold2 have advanced protein structure prediction.
• Predicting biological properties from protein structures remains a challenge.
• Major histocompatibility complex class II (MHC-II) binding affinity is crucial for immune response.

Purpose of the Study:

• To develop a novel deep learning method for predicting peptide-MHC-II binding affinity.
• To explore a transfer learning approach by integrating protein language models with image classification architectures.
• To evaluate the performance of the proposed model against established prediction tools.

Main Methods:

• Extracted features from protein language models (PLMs) including ESM1b, ProtXLNet, and ProtT5-XL-UniRef.
• Utilized image classification models like EfficientNet and Vision Transformer (ViT-16) as backbones.
• Implemented a transfer learning strategy by combining PLM features with image model architectures.
• Trained and validated the final model, named TransMHCII, on peptide-MHC-II binding affinity data.

Main Results:

• The optimal pairing of PLMs and image classifiers resulted in the TransMHCII model.
• TransMHCII demonstrated superior performance compared to NetMHCIIpan 3.2 and NetMHCIIpan 4.0-BA.
• Performance improvements were observed across metrics including receiver operating characteristic area under the curve, balanced accuracy, and Jaccard scores.

Conclusions:

• The novel architecture integrating PLMs and image classifiers is effective for predicting MHC-II binding affinity.
• This approach offers a promising direction for developing deep learning models for complex biological problems.
• TransMHCII advances the field of immunoinformatics and peptide-MHC binding prediction.