Reproduce “Local-Global Multimodal Contrastive Learning for Molecular Property Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiayu Liu, Zhengyi Lu, Yunhong Liao, Chan Fan, Hou-biao Li
Researchers from the University of Electronic Science and Technology of China and Oakland University developed the Local–Global Multimodal Contrastive Learning (LGM-CL) framework for molecular property prediction. This framework integrates local and global structural information from molecular graphs with chemistry-aware textual semantics, achieving improved ROC-AUC scores on classification tasks (e.g., BACE 87.29 ± 0.19) and lower RMSE on regression tasks (e.g., ESOL 0.496 ± 0.025) across MoleculeNet benchmarks.
Key insights
- Explicitly integrating multi-scale structural information from molecular graphs with chemically informed textual semantics significantly enhances the comprehensiveness and accuracy of molecular representations.
- Self-supervised contrastive learning effectively aligns heterogeneous molecular views (graph and text) into a robust and transferable unified embedding space, reducing reliance on extensive labeled data.
- A precisely designed chemistry-aware prompting strategy can mitigate LLM hallucination, enabling the generation of valuable semantic augmentations that complement structural information for property prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.