Energy-Guided Denoising Contrastive Learning for Molecular Property Prediction.

Self-supervised coordinate denoising presents a promising approach to learning 3D molecular representations for predicting molecular properties. However, existing methods often rely on random perturbations, which overlook the heterogeneous chemical environments and may lead to physically unreasonable structures. To address this limitation, we propose an energy-guided denoising contrastive learning framework with an adaptive noise generator. This generator produces structure-aware, atom-specific perturbations conditioned on local chemical environments, guiding the encoder to learn a meaningful representation of the molecular potential energy surface (PES). Our pretraining framework integrates three complementary components: (i) equivariant denoising, (ii) contrastive alignment between original and perturbed conformations, and (iii) self-supervised energy-gap prediction. Extensive evaluations demonstrate that our method achieves state-of-the-art performance across several molecular property prediction benchmarks. The case studies show that our method is interpretable and has learned the relationship between chemical and energy information. The source code is available at https://github.com/wangjx22/EDCL.