Reusability Report: Evaluating the performance of a meta-learning foundation model on predicting the antibacterial activity of natural products

2026-02-12 · Source: Nature Machine Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Data Science & Analytics · Depth: Expert, extended

Summary

A reusability report published in Nature Machine Intelligence on February 12, 2026, evaluates ActFound, a meta-learning foundation model, for predicting the antibacterial activity of natural products (NPs). Developed by Feng et al., ActFound uses pairwise learning and meta-learning to enable fine-tuning with minimal new data. This study fine-tuned ActFound on an antibacterial NPs dataset, which typically lacks the large, labeled data required by traditional deep learning methods. While ActFound did not achieve the same accuracy on this dataset as reported for other cross-domain tasks in its original publication, it demonstrated comparable or superior performance against other state-of-the-art models, particularly in low-shot settings (e.g., 8-16 fine-tuning compounds). The reduced performance is attributed to the high diversity and dissimilarity of compounds within the NPs dataset, which limits the effectiveness of ActFound's pairwise learning function.

Key takeaway

For AI Researchers and Scientists working on drug discovery with limited bioactivity data, ActFound offers a viable framework, especially for datasets rich in structure-activity relationships. Your decision to adopt ActFound should consider the chemical diversity of your target dataset; highly dissimilar compounds may reduce its accuracy, despite its superior performance in low-shot settings compared to other models.

Key insights

ActFound, a meta-learning foundation model, shows promise for bioactivity prediction with limited data, especially for similar compounds.

Principles

• Foundation models adapt quickly to new tasks with minimal data.
• Pairwise learning benefits datasets with similar compound bioactivities.

Method

ActFound was fine-tuned on an antibacterial NPs dataset using kNN-MAML, treating each bacterial strain as an assay. Performance was evaluated across 8-128 few-shot settings and compared to other meta-learning models.

In practice

• Use ActFound for bioactivity prediction with scarce labeled data.
• Prioritize datasets with high compound similarity for optimal ActFound performance.

Topics

• Meta-learning
• Foundation Models
• Bioactivity Prediction
• Natural Products
• Few-shot Learning

Code references

• BFeng14/ActFound
• caitlinbutt04/Actfound_reusability_report

Best for: AI Researcher, AI Scientist, Research Scientist

Related on AIssential

• A world model for proteins is here — Biohub's open-source protein "world model" (ESMFold2, ESM Atlas) accelerates drug discovery by mapping, predicting, and designing proteins.
• Author Correction: Foundation model of neural activity predicts response to new stimulus types — Accurate documentation of model architectures is critical for reproducibility in neural activity prediction.
• Benchmarking LLM Agents on Meta-Analysis Articles from Nature Portfolio — LLM agents critically fail at criterion-based screening in scientific meta-analysis, despite high retrieval recall.
• A Systematic Evaluation of Molecular Mixture Behavior Prediction — Evaluating molecular mixture ML models requires decomposing error into pure-compound and non-ideal interaction components, moving beyond absolute accuracy.
• Exascale Multi-Task Graph Foundation Models for Imbalanced, Multi-Fidelity Atomistic Data — Exascale graph foundation models accelerate materials discovery by screening billions of structures in seconds.
• A Causal Foundation Model for Structure and Outcome Prediction — TabPFN-CFM is a causal foundation model predicting structure and outcomes, supporting Pearl's Causal Hierarchy queries.

Open in AIssential →

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