Uncertainty Weighted Deep Ensemble to Enhance Protein Property Prediction

NeurIPS 2025 Workshop FPI Submission80 Authors

Track: Main Track

Keywords: Uncertainty quantification, protein property prediction, deep ensemble

Abstract: Recent advances in machine learning (ML) have led to significant improvements in data-driven protein property prediction. While these ML models have demonstrated strong prediction performance on natural proteins, their practical utility still remains limited due to the absence of reliable uncertainty estimates for their predictions. In this work, we present **DUNE** (**D**eep **UN**certainty-weighted **E**nsemble), a method for incorporating predictive uncertainty into ML models, to achieve uncertainty-aware protein property prediction. We demonstrate how incorporating uncertainty estimates can enhance the overall predictive performance across three property prediction tasks; immunogenicity, toxicity and allergenicity. Experimental results show that our proposed DUNE outperforms existing ensemble based classification strategies.

Submission Number: 80