Title: Uncertainty-Calibrated Diffusion for Reliable 3D Molecular Graph Generation

Abstract:

Bayesian inference offers a rigorous approach to capturing epistemic uncertainty in neural networks by framing predictions as probability distributions instead of fixed outputs. However, diffusion models designed for 3D molecular graph generation face significant challenges due to the delicate geometric structures and stringent chemical rules they must obey. This sensitivity makes their inference processes highly vulnerable to miscalibrated uncertainty. A critical yet frequently ignored problem is the interaction between the epistemic uncertainty inherent in the learned denoiser and the aleatoric uncertainty deliberately introduced during the reverse diffusion phase. This interaction causes systematic variance inflation, creating a discrepancy between the actual distribution and the one simulated by the model. Such errors are especially harmful in high-precision molecular generation, where minor deviations can compromise chemical validity. In this study, we conduct both theoretical and empirical analyses to understand how epistemic uncertainty travels through diffusion inference and undermines sampling quality. Based on these insights, we introduce UCD (Uncertainty-Calibrated Diffusion), a straightforward but powerful technique that adjusts the reverse diffusion process to properly handle epistemic uncertainty. Comprehensive tests on standard 3D molecular benchmarks show that UCD consistently enhances sampling performance across various baseline methods, achieving new state-of-the-art results for 3D molecular diffusion. The source code is accessible at https://github.com/jiuguaiwf/UCD.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC