Leveraging attention-based deep multiple instance and multiple task learning for improved neoepitope identification.

Area of Science:

• Computational biology
• Immunology
• Artificial intelligence

Background:

• Predicting major histocompatibility complex class I (MHC class I) neoepitopes is vital for personalized cancer immunotherapy.
• Existing prediction methods face challenges with multi-allele ligand presentation and accurate neoepitope identification.

Purpose of the Study:

• To introduce NeoMHCI, a novel deep learning model designed for precise MHC class I neoepitope identification.
• To enhance the prediction accuracy of ligand presentation across multiple MHC class I alleles and improve neoepitope prioritization.

Main Methods:

• Developed NeoMHCI, integrating attention-based multiple instance learning (MIL) and multi-task learning.
• Utilized MIL for high-quality peptide embeddings across diverse MHC class I molecules.
• Employed fine-tuning to enhance immunogenicity prioritization.

Main Results:

• NeoMHCI demonstrated superior performance on benchmark datasets compared to existing methods.
• Achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.948 for multi-allele ligand presentation prediction.
• Obtained the highest top-5 accuracy (42.3%) in neoepitope recognition.

Conclusions:

• NeoMHCI offers a significant advancement in predicting MHC class I neoepitopes.
• The model shows strong potential for developing personalized cancer vaccines and immunotherapies.