Bayesian Uncertainty-Guided Fidelity Fusion for Bioactivity Prediction

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Accurate molecular bioactivity prediction is crucial for rational drug design.
• Challenges include data scarcity and imbalanced labels in molecular property prediction tasks.

Purpose of the Study:

• To develop a data-efficient Bayesian framework for molecular property prediction.
• To integrate classification-to-regression knowledge fusion and uncertainty quantification.
• To improve molecular modeling for drug discovery in resource-limited settings.

Main Methods:

• Proposed the Bayesian Class-Attentive Transformer Network (BCATNet).
• BCATNet fuses classification data priors into a Bayesian regression task using cross-token attention.
• Employed uncertainty quantification and active learning for data efficiency.

Main Results:

• BCATNet demonstrated superior performance and robustness under reduced regression supervision compared to baseline models.
• Bayesian uncertainty estimates correlated with prediction errors and enabled risk stratification.
• Uncertainty-driven active learning strategies achieved optimal performance on benchmark tasks.

Conclusions:

• BCATNet offers a generalizable paradigm for uncertainty-aware molecular modeling by bridging classification and regression.
• The framework provides a principled approach for reliable, interpretable, and resource-efficient drug discovery.
• Explicit classification-to-regression knowledge fusion is a competitive alternative to generic molecular pretraining.