In 1967, malaria was killing soldiers and civilians across Southeast Asia, and the parasite Plasmodium falciparum had learned to survive chloroquine. China launched a secret national programme, Project 523, to find a new antimalarial. In 1969 a young pharmacologist, Tu Youyou, was given charge of a team at the Academy of Traditional Chinese Medicine in Beijing. Her method was not a robotic screen. It was a library search of a different kind: her team collected more than 2,000 herbal recipes from classical texts and folk practitioners and tested hundreds of plant extracts against malaria in mice [1].
Sweet wormwood (Artemisia annua, qinghao) gave promising but inconsistent results. The breakthrough came from a single line in Ge Hong’s fourth-century Handbook of Prescriptions for Emergencies, which told the reader to soak a handful of qinghao in water, wring out the juice and drink it. Tu noticed that the recipe did not call for boiling. She reasoned that heat might be destroying the active principle, and switched to a low-temperature extraction with diethyl ether. In 1971, the neutral ether extract fully suppressed parasites in rodent and primate malaria models [1].
The molecule behind this effect, artemisinin, turned out to be chemically unusual: a sesquiterpene lactone carrying a 1,2,4-trioxane ring with an endoperoxide bridge. Inside the parasite, iron from digested haemoglobin is thought to cleave this peroxide bond, releasing carbon-centred radicals that damage parasite proteins. Artemisinin-based combination therapies are now first-line treatment for falciparum malaria worldwide, and in 2015 Tu Youyou shared the Nobel Prize in Physiology or Medicine for the discovery [2].
This story is the blueprint for the rest of this article. It contains every step of natural-product discovery: an information source (ancient texts), a hypothesis (heat destroys the active compound), an extraction method, a biological assay, structure elucidation and mechanism. What has changed in the past decade is that artificial intelligence (AI) can now perform the “reading” part of that process at a scale no human team could match, across genomes, proteomes, mass spectra and chemical space.
Why nature’s chemistry still matters
Natural products are molecules shaped by hundreds of millions of years of evolutionary selection. They are rich in stereocentres, ring systems and oxygen atoms, and they tend to occupy regions of chemical space that synthetic combinatorial libraries rarely reach. Newman and Cragg’s long-running analysis of approved drugs from 1981 to 2019 showed that natural products and their derivatives, mimics and pharmacophores remain a major source of new medicines, and that anti-infective drugs in particular remain heavily dependent on natural structures [3].
Yet the pharmaceutical industry largely stepped back from natural products in the 1990s and 2000s. The reasons were practical: repeated rediscovery of known compounds, difficult supply of rare organisms, slow structure elucidation and complex chemistry. Atanasov and colleagues argued in 2021 that new technologies, especially genome mining, advanced analytics and computational methods, are now removing many of these barriers [4].
The core problem: too many molecules, too little signal
A single plant leaf can contain thousands of specialised metabolites. The Human Metabolome Database (HMDB 5.0) lists more than 200,000 metabolite entries, most of them predicted rather than experimentally confirmed [5]. In a typical untargeted liquid chromatography–mass spectrometry (LC–MS) run of a biological sample, only a small fraction of detected features can be matched to a known structure. Researchers call the unannotated remainder the “dark metabolome”.
Classical bioactivity-guided fractionation deals with this by repeatedly splitting an extract and testing each fraction until a single active compound remains. It works, but it is slow, consumes material and often ends with a compound that is already known. AI changes the strategy. Instead of isolating first and asking questions later, machine-learning models can predict which sequences, spectra or gene clusters are likely to encode useful chemistry, so that laboratory effort is spent only on the most promising candidates.
Mining the chemistry of the human body
Antimicrobial peptides (AMPs) are short, usually cationic and amphipathic peptides that disrupt bacterial membranes. The classical view was that they come from dedicated immune genes. César de la Fuente’s group at the University of Pennsylvania tested a different idea: that functional antimicrobial fragments may be “encrypted” inside larger human proteins with no immune role, and released by proteolysis.
Using an algorithm based on physicochemical properties of known AMPs, including sequence length, net charge and average hydrophobicity, they scanned the entire human proteome and identified 2,603 candidate encrypted peptide antibiotics. These were encoded in proteins with functions unrelated to the immune system. Selected peptides were active in vitro, and peptides from the same body region showed synergistic activity in two mouse infection models [6]. The work suggests that human tissues, such as the cardiovascular and digestive systems, carry a hidden layer of host-defence chemistry.
The same group extended the idea into the past. Using machine learning, Maasch and colleagues mined the proteomes of extinct hominins, including Neanderthals and Denisovans, for peptide antibiotics, an approach they called “molecular de-extinction”. Several resurrected peptides were synthesised and showed antibacterial activity, including in mouse models [7]. The scientific point is not nostalgia. Extinct sequences are a source of molecular diversity that current pathogens have never encountered.
The human body is not only human. Our microbes produce their own natural products. In 2014, Donia and colleagues systematically mined biosynthetic gene clusters in human-associated bacterial genomes and identified lactocillin, a thiopeptide antibiotic made by a vaginal Lactobacillus strain [8]. In 2024, Santos-Júnior and colleagues scaled this up with machine learning across 63,410 metagenomes and 87,920 prokaryotic genomes, building the AMPSphere catalogue of 863,498 non-redundant predicted peptides. Of 100 peptides that were synthesised and tested, 79 were active, and 63 acted against clinically relevant pathogens [9].
A hit rate of 79 per cent is striking. In traditional random screening, hit rates are often well below one per cent. This difference is the clearest quantitative argument for AI-guided prioritisation.
Mining plant chemistry
Plant metabolomics generates millions of tandem mass spectra (MS/MS). Each spectrum is a fragmentation fingerprint of a molecule, but turning that fingerprint into a structure is hard. Three computational advances have changed the field:
• Molecular networking (GNPS). The Global Natural Products Social Molecular Networking platform clusters spectra by similarity, so that structurally related molecules group together and known compounds can be quickly recognised and set aside (dereplication) [10].
• Fingerprint prediction (SIRIUS / CSI:FingerID). These tools use fragmentation trees and machine learning to predict a molecular formula and a structural fingerprint from a spectrum, then search structure databases [11].
• Compound-class prediction (CANOPUS). Deep learning assigns unknown spectra to chemical classes, such as alkaloids or flavonoids, even when no reference structure exists. This makes the dark metabolome partly readable [12].
In practice, a researcher studying a medicinal plant can now map its chemical space, flag clusters that do not match any known compound and send only those features for isolation and testing.
Specialised metabolites are made by enzymes, and the genes for those enzymes are often clustered. Tools such as antiSMASH detect and annotate biosynthetic gene clusters in microbial and, increasingly, plant genomes [13]. Plant pathways are harder because the genes are often scattered across the genome. Here, co-expression analysis and machine learning help to link genes that switch on together under the same conditions.
Example 4: The Taxol pathway. Paclitaxel (Taxol), a frontline anticancer drug, is still mainly produced from its precursor baccatin III extracted from yew (Taxus) trees. For decades, several enzymes in the pathway remained unknown. In 2025, McClune and colleagues at Stanford used a strategy called multiplexed perturbation × single nuclei (mpXsn), profiling yew cell states across tissues and stress conditions to group co-regulated genes. This resolved new genes, including FoTO1, a protein that solved a long-standing bottleneck in the first oxidation step. Together, 17 genes enabled de novo production of baccatin III in Nicotiana benthamiana leaves at levels comparable to yew needles [14].
Example 5: QS-21 in yeast. QS-21 is a complex saponin adjuvant from the soapbark tree (Quillaja saponaria) used in modern vaccines. In 2024, Liu and colleagues reconstituted its full biosynthesis in engineered yeast, a route that could reduce dependence on bark harvesting [15].
Structure prediction tools such as AlphaFold now provide reliable three-dimensional models for most plant enzymes, which helps scientists guess substrate pockets and reaction types for uncharacterised genes [16].
From discovery to design
Example 6: Halicin. In 2020, Stokes and colleagues trained a graph neural network on the growth-inhibition data of about 2,300 molecules against Escherichia coli. The model then screened a drug-repurposing library and more than 100 million molecules in silico. It identified halicin, a compound originally studied for diabetes, which killed many drug-resistant bacteria by disrupting the proton motive force [17]. In 2024, the same group used explainable deep learning to identify a new structural class of antibiotics active against MRSA [18].
Example 7: Generative design. In 2025, Krishnan and colleagues went beyond screening. Generative models designed more than 36 million candidate molecules, either by growing from a chemical fragment or freely from scratch. After computational filtering, synthesis and testing, two lead compounds emerged: NG1, active against drug-resistant Neisseria gonorrhoeae, and DN1, active against MRSA in a mouse skin-infection model. Both appear to work through membrane disruption and are structurally distinct from existing antibiotics [19].
These designed molecules are not natural products. But the models that create them are trained on chemistry that evolution discovered first. Natural products teach the models what bioactive chemistry looks like.
Revisiting artemisinin with today’s toolkit
Imagine Tu Youyou’s project in 2026. A language model could read thousands of classical and regional texts and rank plants by how often, and how specifically, they are linked to fever cycles. Untargeted LC–MS of the top candidates, followed by molecular networking, would flag Artemisia annua as carrying an unusual cluster of sesquiterpenoids. Fingerprint prediction would highlight a peroxide-containing scaffold with no close match in common drug libraries. Co-expression analysis of glandular trichome transcriptomes would point to amorpha-4,11-diene synthase and its partner oxidases. In fact, this last step has already happened: the core artemisinic acid pathway was later engineered into yeast, giving a semi-synthetic supply route.
The human insight in the real story, that boiling destroys the active compound, would still matter. AI accelerates reading and ranking. It does not replace the chemist who notices that a recipe says “soak” and not “boil”.
Limits, risks and open questions
• Data bias. Models learn from public datasets that over-represent well-studied organisms and assay types. Rare plants and under-studied human tissues remain poorly covered.
• Prediction is not validation. In silico hits must still pass synthesis, cytotoxicity, pharmacokinetics, animal studies and clinical trials. Most will fail.
• Membrane-active compounds. Many AI-discovered antimicrobials act on membranes, which raises questions of selectivity against human cells and in vivo stability.
• Access and benefit sharing. Traditional knowledge and biodiversity used in discovery are subject to the Nagoya Protocol and, in India, the Biological Diversity Act, 2002. Ethical discovery must respect these rules.
• Dual-use concerns. Generative chemistry models can, in principle, be misused. Responsible development needs safeguards.
Nature’s chemistry is a vast library, written in genes, proteins and metabolites, and mostly unread.
For most of history, scientists read it one book at a time, as Tu Youyou did with her 2,000 recipes.
AI now lets us scan whole shelves: the human proteome, the global microbiome, the transcriptome of a yew tree and the spectra of an entire medicinal plant.
The early results are real: validated peptides from our own tissues, reconstructed plant pathways and new antibiotic scaffolds. The next decade will show how many of these reach patients.
The most productive future will likely combine machine scale with the human curiosity that first read an old recipe carefully.
Disclaimer
This article is for general educational and informational purposes only. It is not medical advice and must not be used for diagnosis, treatment or self-medication. Many compounds discussed here are experimental and have been tested only in laboratory or animal studies; they are not approved medicines. Herbal products and plant extracts can be toxic or interact with prescribed drugs. Always consult a qualified healthcare professional before using any medicine or herbal remedy. The descriptions of research are summaries of published studies and may not reflect later findings. The author and publisher make no representation as to the completeness or accuracy of the information and accept no liability for any loss or harm arising from its use. Mention of any product, organisation or tool does not imply endorsement.
References
1. Tu Y. The discovery of artemisinin (qinghaosu) and gifts from Chinese medicine. Nat Med. 2011;17(10):1217–1220. doi:10.1038/nm.2471
2. The Nobel Assembly at Karolinska Institutet. The Nobel Prize in Physiology or Medicine 2015 (William C. Campbell, Satoshi Ōmura, Tu Youyou). NobelPrize.org; 2015.
3. Newman DJ, Cragg GM. Natural products as sources of new drugs over the nearly four decades from 01/1981 to 09/2019. J Nat Prod. 2020;83(3):770–803. doi:10.1021/acs.jnatprod.9b01285
4. Atanasov AG, Zotchev SB, Dirsch VM, et al. Natural products in drug discovery: advances and opportunities. Nat Rev Drug Discov. 2021;20(3):200–216. doi:10.1038/s41573-020-00114-z
5. Wishart DS, Guo A, Oler E, et al. HMDB 5.0: the Human Metabolome Database for 2022. Nucleic Acids Res. 2022;50(D1):D622–D631. doi:10.1093/nar/gkab1062
6. Torres MDT, Melo MCR, Crescenzi O, Notomista E, de la Fuente-Nunez C. Mining for encrypted peptide antibiotics in the human proteome. Nat Biomed Eng. 2022;6(1):67–75. doi:10.1038/s41551-021-00801-1
7. Maasch JRMA, Torres MDT, Melo MCR, de la Fuente-Nunez C. Molecular de-extinction of ancient antimicrobial peptides enabled by machine learning. Cell Host Microbe. 2023;31(8):1260–1274. doi:10.1016/j.chom.2023.07.001
8. Donia MS, Cimermancic P, Schulze CJ, et al. A systematic analysis of biosynthetic gene clusters in the human microbiome reveals a common family of antibiotics. Cell. 2014;158(6):1402–1414. doi:10.1016/j.cell.2014.08.032
9. Santos-Júnior CD, Torres MDT, Duan Y, et al. Discovery of antimicrobial peptides in the global microbiome with machine learning. Cell. 2024;187(14):3761–3778. doi:10.1016/j.cell.2024.05.013
10. Wang M, Carver JJ, Phelan VV, et al. Sharing and community curation of mass spectrometry data with Global Natural Products Social Molecular Networking. Nat Biotechnol. 2016;34(8):828–837. doi:10.1038/nbt.3597
11. Dührkop K, Fleischauer M, Ludwig M, et al. SIRIUS 4: a rapid tool for turning tandem mass spectra into metabolite structure information. Nat Methods. 2019;16(4):299–302. doi:10.1038/s41592-019-0344-8
12. Dührkop K, Nothias LF, Fleischauer M, et al. Systematic classification of unknown metabolites using high-resolution fragmentation mass spectra. Nat Biotechnol. 2021;39(4):462–471. doi:10.1038/s41587-020-0740-8
13. Blin K, Shaw S, Augustijn HE, et al. antiSMASH 7.0: new and improved predictions for detection, regulation, chemical structures and visualisation. Nucleic Acids Res. 2023;51(W1):W46–W50. doi:10.1093/nar/gkad344
14. McClune CJ, Liu JC-T, Wick C, et al. Discovery of FoTO1 and Taxol genes enables biosynthesis of baccatin III. Nature. 2025;643(8071):582–592. doi:10.1038/s41586-025-09090-z
15. Liu Y, Zhao X, Gan F, et al. Complete biosynthesis of QS-21 in engineered yeast. Nature. 2024;629(8013):937–944. doi:10.1038/s41586-024-07345-9
16. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583–589. doi:10.1038/s41586-021-03819-2
17. Stokes JM, Yang K, Swanson K, et al. A deep learning approach to antibiotic discovery. Cell. 2020;180(4):688–702. doi:10.1016/j.cell.2020.01.021
18. Wong F, Zheng EJ, Valeri JA, et al. Discovery of a structural class of antibiotics with explainable deep learning. Nature. 2024;626(7997):177–185. doi:10.1038/s41586-023-06887-8
19. Krishnan A, Anahtar MN, Valeri JA, et al. A generative deep learning approach to de novo antibiotic design. Cell. 2025 (published online 14 Aug 2025). doi:10.1016/j.cell.2025.07.033
","deck":"","published_at":"2026-10-02T13:51:10.794Z","title":"Reading Nature’s Chemical Library with Machines","url":"https://kandulamahesh.com/insights/reading-nature-s-chemical-library-with-machines"}