Improving Molecular Property Prediction with Score-Based Diffusion Models

We explore diffusion-based training as a flexible approach to molecular property prediction in settings shaped by chemical, biophysical, and latent heterogeneity. To address this, we build on score-based generative diffusion models and introduce a conditional diffusion module for predictive tasks, equipped with a joint training and sampling guidance scheme and an optional gated residual mechanism that regulates feature- vs target-driven information flow across noise levels. The module acts as a plug-in learning block for existing molecular encoders, making it complementary to advances in representation learning. Across diverse feature extractors and graph neural networks, diffusion-based training improves over conventional maximum-likelihood objectives on small-molecule regression benchmarks and matches or exceeds strong baselines on multitask regression and classification datasets. It also produces well-calibrated predictive uncertainty with near-perfect prediction interval coverage, enables confidence-guided ranking, and performs well in challenging settings including activity cliffs and class-imbalanced multitask prediction. Overall, these results position diffusion as a promising representation-agnostic approach for effective molecular property prediction.