Systematic benchmarking of molecular foundation models for natural-product annotation and biosynthetic inference

Abstract Molecular foundation models offer a potential route to improve natural-product annotation and support biosynthetic inference, but their utility across these tasks remains insufficiently characterized. We benchmarked eight foundation-model representations against Morgan fingerprints across five tasks covering natural-product-likeness, natural product chemical class prediction, PMN and KEGG pathway association, and PlantCyc reaction-neighbor prediction. Models were evaluated under random and scaffold-based splits using both full data and sample-efficiency settings. In the full data setting, MolFormer achieved the highest mean AUROC for natural product likeness prediction, while Morgan fingerprints were strongest across other tasks. In the sample efficiency analyses, pretrained foundation model representations showed advantages when labeled data were scarce across several tasks. Even in the PMN pathway task, where foundation models did not outperform Morgan fingerprints individually, their embeddings provided complementary predictive information, as shown by error pattern and ensemble analyses. These results show that molecular foundation models are particularly useful for tasks related to global structural characteristics and when labeled training data are limited. This benchmark extends the evaluation of molecular foundation models beyond conventional drug-property tasks and provides a practical reference for selecting and developing representations for natural-product annotation and biosynthetic inference.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11357474/v1
Primary Topic
Computational Drug Discovery Methods
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Systematic benchmarking of molecular foundation models for natural-product annotation and biosynthetic inference

Fan Lin, Shu Xing, Yuqiang Han, Dong Wei
6 links
Research Square
Computational Drug Discovery Methods
preprint

Systematic benchmarking of molecular foundation models for natural-product annotation and biosynthetic inference

Fan Lin, Shu Xing, Yuqiang Han, Dong Wei
preprint en

Abstract

Abstract Molecular foundation models offer a potential route to improve natural-product annotation and support biosynthetic inference, but their utility across these tasks remains insufficiently characterized. We benchmarked eight foundation-model representations against Morgan fingerprints across five tasks covering natural-product-likeness, natural product chemical class prediction, PMN and KEGG pathway association, and PlantCyc reaction-neighbor prediction. Models were evaluated under random and scaffold-based splits using both full data and sample-efficiency settings. In the full data setting, MolFormer achieved the highest mean AUROC for natural product likeness prediction, while Morgan fingerprints were strongest across other tasks. In the sample efficiency analyses, pretrained foundation model representations showed advantages when labeled data were scarce across several tasks. Even in the PMN pathway task, where foundation models did not outperform Morgan fingerprints individually, their embeddings provided complementary predictive information, as shown by error pattern and ensemble analyses. These results show that molecular foundation models are particularly useful for tasks related to global structural characteristics and when labeled training data are limited. This benchmark extends the evaluation of molecular foundation models beyond conventional drug-property tasks and provides a practical reference for selecting and developing representations for natural-product annotation and biosynthetic inference.

Research Square
Jilin University (CN), Chinese Academy of Sciences (CN), Fuyao University of Science and Technology (CN)
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon