Enhancing protein immunogenicity prediction via uncertainty weighted deep ensemble

Recent advances in machine learning (ML) have significantly improved data-driven protein immunogenicity prediction. Although these ML models perform well on natural proteins, their practical utility is limited by the absence of reliable uncertainty estimates for their predictions. The objective of this work is to develop a method that incorporates predictive uncertainty to achieve uncertainty-aware protein immunogenicity prediction. We introduce DUNE (Deep UNcertainty-weighted Ensemble), a novel method designed to integrate predictive uncertainty into ML models. DUNE is built to enhance prediction performance by incorporating uncertainty estimates from probabilistic member models in an ensemble into the final prediction. Experimental results demonstrate that incorporating uncertainty estimates through DUNE significantly enhances protein immunogenicity predictive performance. Our proposed DUNE method outperforms existing deterministic single learners and various other deterministic and probabilistic ensemble-based classification strategies. DUNE provides a more reliable and robust framework for protein immunogenicity prediction. By achieving uncertainty-aware prediction, DUNE can improve the trustworthiness and practical utility of ML models in therapeutic antigen design.