Multimodal Representation Learning for Exploring Natural Product Chemical Space
Year: 2026
Type: preprint
Abstract: Natural products represent a major source of therapeutic molecules, yet their structural diversity and limited biological annotations pose substantial challenges for computational exploration and discovery. Here, we develop a multimodal molecular representation framework that integrates chemical language, molecular fingerprints, physicochemical descriptors, and molecular graphs into a unified latent space for natural product discovery. Using structurally out-of-distribution molecular spaces, we show that molecular representation and sampling strategy jointly determine which regions of chemical space are explored. Multimodal pretraining captured chemically informative representations that enabled exploration of structurally unseen natural products, while comparisons with singlemodality representations showed that representation advantages depended on the discovery objective rather than being universally superior. The pretrained representation was transferable across anti-inflammatory and antioxidant activity predictions, and its integration with active learning enabled data-efficient exploration of biologically active molecules. Prospective evaluation of structurally unseen natural products further connected model-guided exploration with experimental antioxidant activity. These findings establish molecular representation and sampling as coupled determinants of chemical-space exploration and show
Source: ChemRxiv
Institutions: Khon Kaen University, Can Tho University
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7213883386