AI drug discovery examples: what the verified public record shows in 2026

The pitches promise new medicines. The public record shows AI doing narrower, upstream work, and here are the examples TIN checked itself.

If you buy or fund life-science software, you have heard the pitch: an AI platform that discovers new drugs. It is hard to tell which of those claims rest on a published result and which rest on a slide.

The honest way to sort them is to ask what the model actually did, and who checked the output. Not “is it AI-powered”, but which step of a long pipeline it moved, and whether the result survived independent review.

Measured that way, the verified record is narrower than the marketing and more useful. In a 2020 Cell study, an MIT model screened more than 107 million molecules and returned a shortlist that led to a genuinely novel antibiotic [1]. That is a real example. It is also a discovery, not a medicine on a shelf.

What ai drug discovery means

AI drug discovery is the use of a machine-learning model to decide where scientists look next, by predicting which molecules are likely to be active or what shape a biological target takes. The model narrows the search; humans run the validation.

That definition matters because it draws the line the marketing blurs. A model that prioritises candidates is doing real work. It is not the same claim as a drug that an AI invented and proved.

How does ai drug discovery work

AI drug discovery works by training a model on known biology or chemistry, then applying it across a space far too large to test at the bench. The model scores or ranks; wet-lab experiments and, eventually, clinical trials decide.

The pattern is consistent across the credible examples. The model reads a huge candidate set, returns a small, ranked shortlist, and hands it back to people. The speed is real and the shortlist is real. The proof of a drug is still a trial.

The verified examples

These are the three examples TIN checked itself, against primary and independent sources, and wrote up as case files. Each one shows the same shape and a different target.

Halicin, an antibiotic from a 107-million-molecule screen. TIN’s case file on the MIT halicin study verified, against the peer-reviewed Cell paper and independent trade coverage, that a deep neural network screened more than 107 million molecules and that “23 compounds were chosen for further investigation, with eight displaying antibacterial activity against a range of pathogens” [1]. MIT reports the screen “took only three days” [3]. The verified result is a preclinical discovery: halicin cleared infections in mice, not a drug that has cleared trials.

AlphaFold, protein structure prediction. TIN’s case file on AlphaFold and CASP14 verified, against the independent Protein Structure Prediction Center’s official results, that AlphaFold ranked first in the CASP14 blind assessment with a summed z-score of 244.02 against 90.82 for the second-placed group. In 2024 the work won the Nobel Prize in Chemistry “for protein structure prediction” [4]. Knowing a target’s shape is upstream of designing a drug against it, which is why this sits in a discovery discussion at all.

AlphaFold3, and the reproducibility catch. TIN’s case file on AlphaFold3 records both the capability and the governance problem. Nature described the model as able to “predict not just the structures of protein complexes, but also when proteins interact with other kinds of molecule, including DNA and RNA” [5], the step closest to drug binding. But the code was withheld at publication, and only after an open letter did DeepMind release it, “six months after backlash”, under a noncommercial license [6]. A capability you cannot reproduce or use commercially is a weaker input to your pipeline than the headline suggests.

A citable comparison

Example	What the AI did	Verified result	Pipeline stage
Halicin (MIT, 2020)	Screened over 107 million molecules, ranked 23 candidates	8 confirmed antibacterial; halicin cleared infections in mice [1]	Candidate discovery, preclinical
AlphaFold (CASP14)	Predicted protein 3D structure blind	Ranked first, z-score 244.02 vs 90.82; Nobel 2024 [4]	Target understanding, upstream
AlphaFold3 (2024)	Predicted protein interactions with other molecules	Capability shipped; code released only after an open letter [5][6]	Interaction prediction, drug-adjacent

The column that should shape a buying decision is the last one. None of these examples is a finished drug. All of them move a step that used to be slow.

Will ai actually deliver new drugs

It may, and the honest answer today is that the public wins are upstream of the clinic. The strongest verified examples surface candidates and predict structures fast. The long, expensive part, proving a molecule is safe and effective in people, is not where the verified AI results sit.

That is not a reason to dismiss the field. It is a reason to price it correctly. A model that turns a three-year screening problem into a three-day one is worth paying for on its own terms, without pretending it has also run the trial.

The bottom line

When a vendor says “AI drug discovery”, ask which step the model verifiably did, and ask to see the published result that a third party could check. The examples that hold up, halicin and AlphaFold, are specific about the step and open about the evidence. The ones that fall apart are usually specific about the promise and vague about the proof. That test travels: it works for any AI claim, in any industry, where the outcome is supposed to be the product.

Sources

1. Chemistry World, Royal Society of Chemistry, “AI tool screens 107 million molecules, discovers potent new antibiotics”, 2020-02-21. https://www.chemistryworld.com/news/ai-tool-screens-107-million-molecules-discovers-potent-new-antibiotics/4011233.article
2. Cell (via Europe PMC), “A Deep Learning Approach to Antibiotic Discovery” (Stokes et al.), 2020-02-20. https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:32084340%20AND%20SRC:MED&resultType=core&format=json
3. MIT News, “Artificial intelligence yields new antibiotic”, 2020-02-20. https://news.mit.edu/2020/artificial-intelligence-identifies-new-antibiotic-0220
4. The Royal Swedish Academy of Sciences, “The Nobel Prize in Chemistry 2024”, 2024-10-09. https://www.nobelprize.org/prizes/chemistry/2024/summary/
5. Nature (news), “Major AlphaFold upgrade offers boost for drug discovery”, 2024-05-08. https://www.nature.com/articles/d41586-024-01463-0
6. Science, “Google DeepMind releases code behind its most advanced protein-prediction program”, 2024-11-11. https://www.science.org/content/article/google-deepmind-releases-code-behind-its-most-advanced-protein-prediction-program

Questions

What is AI drug discovery?

AI drug discovery is the use of machine-learning models to prioritise where scientists look, by predicting which molecules are likely to be active or what shape a protein target takes. The model narrows an enormous search space; humans still run the wet-lab and clinical validation.

What are real examples of AI drug discovery?

Real examples include halicin, an antibiotic an MIT model surfaced from a screen of more than 107 million molecules, and AlphaFold, which predicted protein structures accurately enough to win the 2024 Nobel Prize in Chemistry. Both did upstream work; neither is itself an approved medicine.

How does AI drug discovery work?

AI drug discovery works by training a model on known data, then applying it to score candidates or predict structures across a space too large to test by hand. The model ranks; the bench and the clinic decide.

Sources

1. Chemistry World, Royal Society of Chemistry, AI tool screens 107 million molecules, discovers potent new antibiotics , 2020-02-21
2. Cell (via Europe PMC), A Deep Learning Approach to Antibiotic Discovery (Stokes et al., Cell 2020) , 2020-02-20
3. MIT News, Artificial intelligence yields new antibiotic , 2020-02-20
4. The Royal Swedish Academy of Sciences, The Nobel Prize in Chemistry 2024 , 2024-10-09
5. Nature (news), Major AlphaFold upgrade offers boost for drug discovery , 2024-05-08
6. Science, Google DeepMind releases code behind its most advanced protein-prediction program , 2024-11-11

---

This is analysis, not a verified outcome. It carries no verification badge and never will. The proof lives in the case files, where every figure is checked against the public record and the method is printed on the page.