Reproduce “UniCorn: A Unified Contrastive Learning Approach for Multi-view Molecular Representation Learning” on OpenResearch

About this paper

Shikun Feng, Yuyan Ni, Minghao Li, Yanwen Huang, Zhi-Ming Ma, Wei-Ying Ma, Yanyan Lan

UniCorn is a unified contrastive learning framework developed by researchers from Tsinghua University and Chinese Academy of Sciences that integrates 2D graph masking, 3D denoising, and 2D-3D cross-modal distillation to learn comprehensive molecular representations. It achieved state-of-the-art performance across 33 out of 38 quantum chemical, physicochemical, and biological prediction tasks, demonstrating a universal capability in molecular property prediction.

Key insights

• Reconstructive and contrastive losses can be theoretically unified, serving as mutual bounds and leading to representation clustering at different granularities within the learned embeddings.
• Different self-supervised learning methods induce distinct clustering patterns that correspond to the inductive biases of various molecular properties, clarifying why 3D denoising excels for quantum tasks and 2D masking for biological tasks.
• Integrating these diverse pre-training mechanisms into a single framework allows for learning comprehensive, multi-view molecular representations that capture fine-grained 3D, conformation-invariant, and substructure-based information simultaneously.

Summary by alphaXiv.