Predicting HLA class II antigen presentation through integrated deep learning.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Accurate prediction of antigen presentation by human leukocyte antigen (HLA) class II molecules is crucial for vaccine development and cancer immunotherapies.
• Current computational methods face limitations due to insufficient training data and algorithmic constraints.

Purpose of the Study:

• To develop MARIA (major histocompatibility complex analysis with recurrent integrated architecture), a novel computational tool for predicting antigen presentation by HLA class II molecules.
• To improve the accuracy of predicting immunogenic epitopes for therapeutic applications.

Main Methods:

• MARIA utilizes a multimodal recurrent neural network architecture.
• The model is trained on diverse data including in vitro binding measurements, mass spectrometry-identified peptide HLA ligand sequences, antigen gene expression levels, and protease cleavage signatures.
• Leverages an improved machine learning framework for enhanced predictive capabilities.

Main Results:

• MARIA demonstrated superior performance compared to existing methods, with an area under the curve ranging from 0.89 to 0.92 in validation datasets.
• Peptides identified with high MARIA scores were more likely to elicit strong CD4+ T cell responses in independent cancer neoantigen studies.
• The tool successfully identified immunogenic epitopes in diverse cancer and autoimmune disease contexts.

Conclusions:

• MARIA represents a significant advancement in predicting antigen presentation by HLA class II molecules.
• The tool enables the identification of immunogenic epitopes for developing effective vaccines and cancer immunotherapies.
• MARIA has broad applicability in identifying targets for diverse cancers and autoimmune diseases.