We are excited to announce the publication of our latest research titled "Explainable Uncertainty Quantifications for Deep Learning-Based Molecular Property Prediction," in Journal of Cheminformatics. This work introduces an innovative method for quantifying uncertainties in deep learning models applied to molecular property predictions. The approach brings transparency and explainability to predictions, making it possible to identify the factors driving uncertainty, thereby improving the reliability of machine learning models in chemical research.
Key Findings:
- Atom-Based Uncertainty Quantification: The study introduces a novel method that quantifies uncertainty at the atomic level, allowing researchers to trace the source of uncertainties back to individual atoms within molecules. This approach provides a deeper chemical insight into model predictions, highlighting which atoms contribute most to prediction errors.
- Separating Aleatoric and Epistemic Uncertainties: The research distinguishes between aleatoric (data-related) and epistemic (model-related) uncertainties, offering a comprehensive framework to assess both types. This distinction is crucial in determining whether more data is needed (epistemic uncertainty) or if the inherent noise in the data limits prediction accuracy (aleatoric uncertainty).
- Post-Hoc Calibration of Uncertainty Estimates: The study also presents a post-hoc calibration method to refine the uncertainty estimates produced by ensemble models, improving their confidence intervals. This approach reduces overestimation of uncertainties and enhances the overall predictive accuracy of deep learning models in molecular property prediction.
This groundbreaking research addresses a critical challenge in machine learning for chemistry by making predictions more interpretable and reliable, particularly in areas where high-quality data is scarce. The explainable uncertainty quantification framework proposed in this work is expected to accelerate advancements in drug discovery, materials science, and chemical engineering.
For more details, please access the full publication here.