Artificial intelligence is being applied to natural product drug discovery mainly in three places: sorting and classifying known compounds, finding the genes that make them, and predicting which molecules from nature might kill bacteria. The evidence so far consists of tools, databases, and laboratory results, not approved medicines. A 2025 deep learning study of archaeal proteins, for example, reported 80 synthesized peptides, 93 percent of them active in vitro, and one that worked in mice.
Key takeaways
- Natural products and their derivatives have a long record in medicine, especially in cancer and infection, according to a 2020 review of approvals from 1981 to September 2019.
- A 2021 review says technical barriers led industry to pull back from natural products from the 1990s, and that genome mining and better analytics are reviving interest.
- Open resources such as COCONUT, GNPS, and NPClassifier give machine learning models data and labels to work with.
- The archaeasin study mined 233 archaeal proteomes and found 12,623 candidate peptides, but its animal results are early-stage.
- No source reviewed here shows an AI-discovered natural product drug approved for patients.
Why natural products still matter
Newman and Cragg's review in the Journal of Natural Products, covering January 1981 to September 2019, counted approved drugs by origin. For cancer, it reports that 185 small molecules were approved from 1981 onward. Of these, 62 (33.5 percent) were natural products or direct derivatives, and adding 58 synthetic compounds built on natural product structures raised the figure to 64.9 percent. For the period from about 1946 to 1980, 40 of 75 small-molecule cancer drugs (53.3 percent) were natural products or derivatives. The authors conclude that natural product structures remain an active source of drugs.
That history explains the interest. It also sets a bar: any computational method has to find molecules that are not already known.
Why discovery slowed and what changed
A 2021 review in Nature Reviews Drug Discovery, by Atanasov and an international taskforce, describes the difficulty plainly. Screening, isolating, characterizing, and optimizing natural compounds is technically demanding, and these barriers contributed to a decline in pursuit by the pharmaceutical industry from the 1990s onward. The review points to improved analytical tools, genome mining and engineering, and advances in growing microbes as reasons interest has returned, particularly for antimicrobial resistance.
Computation fits into that picture as one of several tools, not the sole driver.
Where machine learning enters the workflow
| Step | Problem | Documented tool or study |
|---|---|---|
| Collect structures | Natural product data are scattered across databases | COCONUT, an open aggregation launched in 2021 and rebuilt as version 2.0 in 2025, with bulk download and community curation |
| Share spectra | Mass spectrometry data stay locked in papers | GNPS, an open knowledge base described in Nature Biotechnology in 2016 |
| Classify | Thousands of structures need labels | NPClassifier, a deep neural network that classifies structures from Morgan fingerprints (2021) |
| Find the genes | Biosynthetic gene clusters hide in genomes | antiSMASH, a genome mining tool in use since 2011; version 7 supports 81 cluster types |
| Predict activity | Which candidates are worth synthesizing | Deep learning mining of archaeal proteomes, 2025 |
The table lists published tools, not a ranking. The antiSMASH abstract describes it as a genome mining platform and does not call it AI, so its place here is as a source of the gene data that prediction models can use.
One worked example: peptides from archaea
In 2025 Marcelo Torres, Fangping Wan, and César de la Fuente-Nunez reported in Nature Microbiology that they had applied deep learning to 233 archaeal proteomes and flagged 12,623 molecules with potential antimicrobial activity. They called these peptides archaeasins. Of 80 that were synthesized, 93 percent showed activity in lab tests against bacteria that included Acinetobacter baumannii, Escherichia coli, and Staphylococcus aureus. One candidate, archaeasin-73, reduced A. baumannii in mouse infection models by an amount comparable to polymyxin B. The National Library of Medicine record for the study carries the abstract.
Several limits apply. The peptides come from proteins, not from the small molecules in the 2020 review, and mouse results rarely predict human outcomes. A candidate still faces toxicity testing, manufacturing questions, and clinical trials that take years.
What claims to question
- Is the result a prediction, a lab test, an animal study, or a human trial?
- Did the model find something new, or rediscover a known compound?
- Are the data and code open for others to repeat?
- Is the headline number a hit rate in a dish, not a drug?
This article gives general information about research. It is not medical advice and does not recommend any treatment or product.
FAQ
Has AI discovered an approved drug from a natural product?
None of the sources reviewed here reports one. The published results are databases, classifiers, and early laboratory findings.
Are natural products better drug sources than synthetic molecules?
The review data show a large share of approvals tied to natural product structures, especially in cancer and infection. They do not show that natural sources are better in every disease area.
What is a biosynthetic gene cluster?
It is a group of neighboring genes in a microbe's genome that together build a secondary metabolite, which is often a bioactive compound. Tools such as antiSMASH locate them.
Can the open databases be used for free?
COCONUT states it is openly accessible and offers bulk download in several formats. Each database sets its own license, so terms should be read before reuse.
Related reading
- AI in drug delivery
- testing what an AI product claims
- what measurements show about machine and natural intelligence
- a 1976 paper on AI and human nature
More in AI Basics.
Last verified: October 2026
Sources
- Newman D. J. and Cragg G. M., Natural Products as Sources of New Drugs over the Nearly Four Decades from 01/1981 to 09/2019, Journal of Natural Products, 2020
- Atanasov A. G. et al., Natural products in drug discovery: advances and opportunities, Nature Reviews Drug Discovery, 2021
- Sorokina M. et al., COCONUT online: Collection of Open Natural Products database, Journal of Cheminformatics, 2021
- COCONUT 2.0: a comprehensive overhaul and curation of the collection of open natural products database, Nucleic Acids Research, 2025
- Wang M. et al., Sharing and community curation of mass spectrometry data with GNPS, Nature Biotechnology, 2016
- Kim H. W. et al., NPClassifier, Journal of Natural Products, 2021
- Blin K. et al., antiSMASH 7.0, Nucleic Acids Research, 2023
- Torres M. D. T., Wan F., de la Fuente-Nunez C., Deep learning reveals antibiotics in the archaeal proteome, Nature Microbiology, 2025