Enhancing protein immunogenicity prediction via uncertainty weighted deep ensemble.

Area of Science:

• Computational Biology
• Machine Learning
• Immunology

Background:

• Machine learning (ML) models have advanced protein immunogenicity prediction for natural proteins.
• Current ML models lack reliable uncertainty estimates, limiting their practical application.
• Predictive uncertainty is crucial for robust immunogenicity assessment in therapeutic development.

Purpose of the Study:

• To develop an uncertainty-aware method for protein immunogenicity prediction.
• To enhance the reliability and practical utility of ML models in this domain.
• To integrate predictive uncertainty into ML models for improved performance.

Main Methods:

• Introduction of DUNE (Deep Uncertainty-weighted Ensemble), a novel ML integration method.
• Ensemble of probabilistic member models to incorporate uncertainty estimates.
• Evaluation of DUNE against deterministic and probabilistic ensemble strategies.

Main Results:

• DUNE significantly enhances protein immunogenicity predictive performance by incorporating uncertainty.
• The DUNE method outperforms existing deterministic single learners and ensemble strategies.
• DUNE provides a more reliable and robust framework for immunogenicity prediction.

Conclusions:

• DUNE enables uncertainty-aware protein immunogenicity prediction, improving trustworthiness.
• The method enhances the practical utility of ML models in therapeutic antigen design.
• Incorporating uncertainty estimates is key to advancing data-driven immunogenicity prediction.