ACME: pan-specific peptide-MHC class I binding prediction through attention-based deep neural networks

Area of Science:

• Immunoinformatics
• Computational Biology
• Bioinformatics

Background:

• Accurate prediction of peptide-MHC binding is crucial for developing effective cancer vaccines.
• Current algorithms require improvement in prediction accuracy and reduction of false positives.

Purpose of the Study:

• To develop a novel, pan-specific algorithm for predicting peptide-MHC class I binding affinities.
• To improve the precision of binding affinity predictions, especially for unseen MHC alleles.

Main Methods:

• Attention-based Convolutional Neural Networks (ACME) were developed.
• The algorithm integrates convolutional neural networks with an attention mechanism.
• ACME was tested for its ability to predict binding affinities for various MHC class I molecules.

Main Results:

• ACME significantly outperforms existing state-of-the-art methods, increasing the Pearson correlation coefficient by up to 23 percentage points.
• The algorithm demonstrates accuracy even for novel MHC alleles not present in the training data.
• ACME's ability to identify strong-binding peptides was experimentally validated.

Conclusions:

• ACME provides a powerful and practical tool for studying peptide-MHC class I interactions.
• The interpretability of ACME offers insights into peptide-MHC binding preferences.
• ACME is available as open-source software, facilitating its use in research.