Pretraining a foundation model for small-molecule natural products

2026-04-29 · Source: Nature Machine Intelligence · Field: Science & Research — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Health & Medical Research · Depth: Expert, long

Summary

Researchers developed NaFM, a foundation model specifically designed for natural products (NPs) to overcome limitations of existing deep learning methods in drug discovery. Unlike previous supervised learning models, NaFM employs a tailored pretraining strategy that incorporates contrastive learning and masked graph learning objectives. This approach emphasizes evolutionary information from molecular scaffolds while capturing side-chain details. NaFM achieved state-of-the-art results across various downstream tasks, including NP taxonomy classification, where it outperformed synthetic molecule-focused baselines, demonstrating its superior understanding of natural synthesis. The model also effectively captured evolutionary information at both gene and microbial levels and proved useful in virtual screening for identifying potential drug candidates. All datasets and source code are publicly available via figshare and Zenodo.

Key takeaway

For AI scientists and drug discovery researchers working with natural products, NaFM offers a significant advancement over conventional deep learning models. Its specialized pretraining approach, which accounts for natural product characteristics, provides more accurate representations for tasks like taxonomy classification and virtual screening. You should consider integrating NaFM into your workflows to enhance the discovery of novel drug candidates and gain deeper insights into natural synthesis pathways.

Key insights

NaFM is a foundation model for natural products, outperforming existing methods in drug discovery tasks.

Principles

• Tailor pretraining strategies to unique molecular properties.
• Incorporate evolutionary information for better molecular understanding.
• Combine contrastive and masked graph learning objectives.

Method

NaFM uses a scaffold-aware pretraining framework with contrastive learning and masked graph learning objectives to capture both evolutionary and side-chain information from natural product molecules.

In practice

• Use NaFM for improved natural product taxonomy classification.
• Apply NaFM to virtual screening for drug candidate identification.
• Leverage NaFM for biological source discrimination.

Topics

• Natural Product Foundation Model
• Drug Discovery
• Contrastive Learning
• Masked Graph Learning
• Molecular Scaffolds

Code references

• mwang87/NP-Classifier
• TomAIDD/NaFM-Official

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Learning the chemical language of natural products — The NaFM foundation model enhances natural product mining by learning their chemical language.
• Customized Amazon Nova models improve molecular-property prediction in drug discovery — Fine-tuned LLMs can unify complex molecular property prediction, outperforming or matching specialized GNNs.
• The Future Of Drug Discovery Is 4 Billion Years Old (Viswa Colluru, Founder & CEO at Enveda) — AI-driven exploration of nature's chemistry offers a potent, overlooked pathway for discovering first-in-class, affordable medicines.
• Learning Sparse Latent Predictive Foundation Model for Multimodal Neuroimaging — Neuro-JEPA is a sparse multimodal neuroimaging foundation model using a latent predictive objective and Mixture-of-Experts.
• A diffusion model conditioned on compound bioactivity profiles for generating high-content images — pDIFF uses bioactivity profiles to generate high-content images, improving virtual hit expansion for drug discovery.
• A Triple-Modal Contrastive Learning Framework with Sequence, Graph, and 3D Features for Drug-Target Interaction Prediction — TriMod-DTI uses triple-modal contrastive learning with 1D, 2D, and 3D features for superior drug-target interaction prediction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Nature Machine Intelligence.