A modular protein language modelling approach to immunogenicity prediction.

Area of Science:

• Computational biology
• Immunoinformatics
• Personalized medicine

Background:

• Predicting neoantigen immunogenicity is vital for developing personalized cancer vaccines.
• Limited training data and low reactivity rates pose significant challenges in this prediction task.

Purpose of the Study:

• To introduce ImmugenX, a novel modular protein language modeling approach for predicting CD8+ T-cell epitope immunogenicity.
• To evaluate ImmugenX's performance against state-of-the-art models in predicting pMHC interactions and immunogenicity.

Main Methods:

• Developed ImmugenX, a modular framework with pMHC and optional TCR encoding modules.
• Trained the pMHC encoding module on prediction tasks including binding affinity, eluted ligand prediction, and stability.
• Utilized context-specific immunogenicity prediction head modules.

Main Results:

• ImmugenX's encoding module demonstrated comparable or superior performance on pMHC binding affinity, eluted ligand prediction, and stability tasks.
• ImmugenX significantly outperformed existing models in pMHC immunogenicity prediction (AUC=0.619, AP=0.514), achieving a 7% increase in average precision.
• Integrating T-cell receptor (TCR) context further enhanced immunogenicity prediction performance.

Conclusions:

• ImmugenX offers a robust and improved approach to neoantigen immunogenicity prediction for CD8+ epitopes.
• The model's interpretability analysis identified weaknesses in current models and potential biases in public datasets.
• ImmugenX advances the development of effective personalized immunotherapies.