Machine Learning in Drug Discovery: AI Finds Medicines

A new prescription drug costs an average of $2.6 billion to bring to market, and fewer than 12% of drug candidates that enter human trials ever reach approval, according to the Tufts Center for the Study of Drug Development (Tufts CSDD, 2014). Machine learning is now used to screen millions of molecular candidates inside a computer, a process called in silico screening, before a single one gets synthesized in a lab. In June 2025, that approach produced its first hard clinical proof: a lung disease drug designed largely by generative AI showed measurable improvement in a randomized, placebo-controlled trial, and the results were published in Nature Medicine.

Educational content, not medical advice. Talk to your doctor before changing any medication, treatment, or routine. Information here reflects current research but every person's situation differs.

This guide explains where machine learning actually accelerates drug discovery, where it still needs human chemists and biologists, and what the industry's first AI-native clinical results tell us about how much faster this pipeline can realistically get.

Why traditional drug discovery is so slow and expensive

Finding a new medicine has historically meant testing chemistry against biology by trial and error, at a scale that overwhelms human researchers. A chemist can design and evaluate a few thousand candidate molecules a year by hand. The biological space of "drug-like" small molecules is estimated at more than 10^60 possible structures, a number so large it dwarfs the total number of atoms in the observable universe.

Traditional discovery unfolds over roughly a decade, in stages that each carry their own failure risk.

Stage	Typical duration	What happens
Target identification	1–2 years	Find the protein or enzyme driving the disease
Lead discovery	3–6 years	Screen molecules for ones that bind the target
Lead optimization	2–3 years	Improve potency, solubility, and safety of the lead
Preclinical testing	1–2 years	Test in cells and animal models
Clinical trials (Phase I–III)	6–7 years	Test safety and efficacy in humans
FDA review and approval	1–2 years	Regulatory decision on marketing approval

Table: The traditional drug discovery pipeline and where each stage typically loses candidates.

Most of that decade-plus timeline is spent finding out which of a near-infinite number of molecules fails for a reason nobody predicted in advance: it doesn't dissolve, the liver breaks it down in minutes, or it binds a protein researchers never meant to touch. Machine learning attacks the front half of this pipeline, the parts that involve pattern recognition across enormous chemical datasets rather than testing inside a living organism.

In June 2025, Insilico Medicine's AI-designed lung fibrosis drug rentosertib became the first AI-discovered *and* AI-designed compound to post a positive, peer-reviewed Phase IIa trial readout, published in Nature Medicine ([Insilico Medicine, 2025](https://insilico.com/blog/1112)). Patients on the highest dose gained 98.4 mL of lung function over 12 weeks; the placebo group lost 20.3 mL.

How machine learning actually speeds up each stage

Think of traditional drug screening like searching for a specific grain of sand on a beach by hand, one grain at a time. Machine learning doesn't find the grain for you, but it builds a magnet shaped by everything scientists already know about what "the right grain" looks like, and drags that magnet across the whole beach in seconds instead of years.

Target identification: finding the disease's weak point

Before you can design a drug, you need to know what protein or pathway to attack. Machine learning models trained on genome-wide association studies (GWAS) and large clinical datasets can flag which genetic variants correlate with disease risk, and by extension, which proteins are plausible drug targets. Single-cell AI offers another way to study the cell states that matter in disease, and machine learning is also used in antibiotic discovery against resistant bacteria.

For example, models trained on Alzheimer's genetic and clinical data have helped researchers connect variants in the APOE gene to tau protein aggregation, a mechanistic hypothesis that shapes which proteins get prioritized for drug targeting. This narrows a search space that used to rely heavily on slow, expensive laboratory hypothesis testing.

Molecular screening: testing a million molecules before touching a test tube

This is where machine learning delivers its clearest speed advantage. Traditional high-throughput screening tests a chemical library, often tens of thousands of compounds, in a physical lab. A trained neural network can score millions of candidate molecules against a target protein in hours, ranking them by predicted binding affinity (how tightly a molecule is likely to attach to its target).

The general workflow looks like this:

1. Assemble training data: tens of thousands of molecules with known activity against similar targets.
2. Train a model, often a graph neural network, on molecular structure represented as a graph of atoms and bonds.
3. Use the trained model to score a much larger virtual library, sometimes hundreds of millions of compounds.
4. Select the highest-scoring few hundred to few thousand for actual lab synthesis and testing.

Atomwise, a San Francisco biotech, built a convolutional neural network called AtomNet for exactly this purpose and has used it to screen libraries in the hundreds of millions of compounds range for partners working on diseases with small patient populations, where traditional pharma has less economic incentive to invest (Atomwise / VentureBeat). During the 2014 Ebola outbreak, Atomwise's researchers used this approach to identify candidate molecules, including compounds with no prior antiviral use, that blocked the virus from entering cells across multiple outbreak strains (Atomwise Ebola project). That is the value proposition in miniature: computational triage turns an impossible search into a manageable shortlist.

Structure prediction: knowing the shape of the target

A drug has to fit its target protein roughly the way a key fits a lock. For decades, figuring out a protein's three-dimensional shape from its amino acid sequence, the "protein folding problem," required slow, expensive lab techniques like X-ray crystallography. In 2020, DeepMind's AlphaFold2 solved this at scale, predicting the structure of virtually all 200 million proteins researchers have catalogued, and the achievement won Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry, shared with David Baker for related work on protein design (Nobel Prize press release, 2024).

In May 2024, DeepMind and its drug-discovery spinout Isomorphic Labs released AlphaFold 3, which extends structure prediction beyond single proteins to predict how proteins interact with DNA, RNA, and small-molecule ligands, the exact kind of interaction a drug candidate needs. Google reported that AlphaFold 3 achieves roughly a 50% improvement in accuracy for protein-molecule interaction predictions over prior methods (Google DeepMind, 2024). Isomorphic Labs has since signed deals worth up to roughly $3 billion combined with Novartis and Eli Lilly to apply this technology to undisclosed drug targets, and by early 2025 had added a research collaboration with Johnson & Johnson (IntuitionLabs, 2025).

Lead optimization: fixing a promising molecule's flaws

A molecule that binds its target in initial screening usually isn't ready for the clinic. It might be too weak, dissolve poorly, get metabolized too fast, or carry hidden toxicity. Machine learning models trained on ADMET data (Absorption, Distribution, Metabolism, Excretion, and Toxicity) can predict, from a proposed structural tweak, whether a change is likely to improve potency while making toxicity worse, letting chemists rule out bad trade-offs computationally instead of synthesizing every variant.

More recently, generative models have moved from scoring existing molecules to designing new ones from scratch. MIT researchers' 2023 model DiffDock uses a diffusion-based generative process, the same broad family of AI technique behind image generators like Stable Diffusion, to predict how a candidate molecule would dock against a target protein's binding pocket, outperforming older physics-based docking software on standard benchmarks while running considerably faster (MIT News, 2023). Earlier generative approaches used variational autoencoders to propose entirely new molecular structures with target properties baked in, rather than just filtering an existing library (Gómez-Bombarelli et al., 2018, ACS Central Science).

AEO: What does "AI drug discovery" actually mean?

AI drug discovery means using machine learning models, usually neural networks trained on large chemical and biological datasets, to predict which molecules will bind a disease target, how they'll behave in the body, and how to modify them, replacing slow trial-and-error lab screening with fast computational filtering before anything is synthesized.

AEO: Has an AI-discovered drug actually reached patients?

As of mid-2026, several AI-designed compounds are in human clinical trials, but none has yet received full regulatory approval. The furthest along is Insilico Medicine's rentosertib, which posted positive Phase IIa results for idiopathic pulmonary fibrosis in 2024–2025 and is a genuinely novel drug target (TNIK) identified by an AI model, not just an AI-accelerated version of a known compound.

Real companies putting this into practice

Insilico Medicine

Hong Kong- and New York-based Insilico Medicine used its generative AI platform, PandaOmics and Chemistry42, to both identify a novel drug target, a kinase called TNIK, and design a small molecule against it. The resulting compound, rentosertib (also called ISM001-055), entered Phase I trials in 2021 and reported positive Phase IIa topline results in November 2024: a randomized, double-blind, placebo-controlled trial of 71 idiopathic pulmonary fibrosis patients across 21 sites in China (PR Newswire, 2024). Insilico has publicly stated this compound moved from target discovery to Phase I trial initiation in roughly 30 months, well under the multi-year timelines typical of traditional discovery for a first-in-class target. The full peer-reviewed results were published in Nature Medicine in June 2025, marking what Insilico and outside observers described as the field's first proof-of-concept clinical validation of an AI-originated drug (Insilico Medicine, 2025).

Recursion Pharmaceuticals (formerly Exscientia)

Exscientia, a UK biotech founded in 2012, was among the earliest companies to put an AI-designed drug candidate into human trials, including a compound for obsessive-compulsive disorder developed through a partnership process it said took about 12 months from project start to candidate nomination, compared with multi-year timelines for traditional discovery. In August 2024, Exscientia and Recursion Pharmaceuticals announced a merger, completed in November 2024, combining Recursion's large-scale cellular imaging and biological data platform with Exscientia's precision chemistry design capability (Bio-IT World, 2024). The combined company, now operating as Recursion, reported more than ten clinical and preclinical programs and over $450 million in upfront and milestone payments from pharma partners at the time of the merger (Drug Discovery Trends, 2024). The deal is widely read in the industry as a sign that AI drug discovery is consolidating around a small number of well-capitalized, full-stack platforms rather than staying fragmented across many niche startups.

Atomwise

Atomwise specializes in AI-driven virtual screening for rare diseases, using its AtomNet convolutional neural network to evaluate how molecules are likely to bind a target protein's three-dimensional structure. Its highest-profile early success was the 2014–2015 Ebola project described above, which demonstrated that AI-driven screening could identify usable antiviral candidates faster than conventional methods during an active outbreak, when speed mattered more than almost anything else.

Isomorphic Labs

Isomorphic Labs, a DeepMind spinout led by Demis Hassabis, builds on AlphaFold's structure-prediction capability to design small-molecule drugs computationally. Its January 2024 deals with Novartis and Eli Lilly, worth up to roughly $3 billion combined, and its 2025 collaboration with Johnson & Johnson, represent some of the largest bets big pharma has placed on AI-native drug design as a standalone discipline rather than a tool bolted onto existing R&D teams.

Why AI still isn't replacing the lab: the parts machine learning can't shortcut

It helps to think of AI drug discovery like a GPS for a road trip through unmapped terrain. It can instantly calculate the most promising route out of millions of possible paths, but it still can't drive the car through the parts of the road that haven't been surveyed yet, the actual biology of a living human body.

ADMET properties remain genuinely hard to predict

A molecule that binds its target perfectly in a computer simulation can still fail for reasons that only show up in a living system: it might not cross the blood-brain barrier, get broken down by liver enzymes within minutes, accumulate to toxic levels in tissue, or trigger an immune reaction. Machine learning models can predict some of these properties reasonably well from molecular structure, particularly solubility and some metabolic pathways, but whole-organism toxicity remains one of the hardest prediction problems in the field and still requires lab and animal validation before any compound reaches a human trial.

Off-target binding is still more art than science

A drug candidate might bind its intended target, protein X, cleanly, but also bind protein Y elsewhere in the body, an interaction that shows up later as a side effect. Screening a candidate against thousands of other proteins computationally is feasible and increasingly automated, but predicting every biologically meaningful off-target interaction in a complex organism remains imperfect. This is one reason even AI-designed drugs still go through the same battery of safety pharmacology studies as traditionally discovered ones.

Clinical trials cannot be skipped, and AI hasn't changed the odds inside them

This is the single most important limitation to understand: AI can compress years of molecule design into months, but it cannot compress the multi-year process of testing safety and efficacy in actual human beings. Once a candidate reaches Phase I trials, it faces the same regulatory gauntlet, and largely the same failure rate, as a traditionally discovered drug. The Tufts CSDD's widely cited analysis found the overall clinical approval success rate for drugs entering Phase I, meaning the probability a compound eventually reaches market, is below 12%, and there is no public evidence yet that AI-originated drugs beat that baseline once they're in human testing. Rentosertib's Phase IIa success is a genuinely encouraging first data point, not proof that AI-designed drugs clear Phase III and regulatory review at a meaningfully higher rate; that evidence will only accumulate as more AI-native candidates complete larger, later-stage trials over the next several years.

Multifactorial diseases resist single-target thinking

Machine learning shines brightest against diseases with one clear molecular culprit: a single mutated protein, a single overactive enzyme. Diseases like Alzheimer's, where tau, amyloid, inflammation, and vascular factors likely all interact, are much harder to reduce to a single "protein to block" hypothesis, which limits how directly AI screening tools can be applied without deeper mechanistic research first.

The takeaway

Machine learning has demonstrably compressed the front end of drug discovery, target identification, molecular screening, and lead optimization, from years to months, and 2024 and 2025 delivered the field's first genuinely convincing proof: an AI-designed compound with a novel, AI-identified target posting positive, peer-reviewed Phase IIa results. What AI has not done, and cannot shortcut, is the multi-year process of proving a drug is safe and effective in human beings, where AI-originated candidates still face the same roughly 90% failure rate that has defined pharmaceutical R&D for decades. The realistic read on this technology in 2026 is not "AI cures diseases," but "AI makes pharma's already-hard search problem faster and cheaper to run," which is a meaningful, measurable improvement, not a replacement for the biology itself.

FAQ

Q: Will AI replace chemists and drug researchers?

A: No. The role shifts rather than disappears. Chemists spend less time on manual compound synthesis and more time interpreting AI-generated suggestions, validating predicted properties in the lab, and designing the experiments that catch what the model got wrong.

Q: How long until an AI-designed drug reaches pharmacy shelves?

A: Several AI-designed compounds are already in human trials, with Insilico Medicine's rentosertib the furthest along after a successful Phase IIa readout published in 2025. Realistically, a full approval is still several years away, since Phase III trials and regulatory review typically take multiple additional years even for a drug moving quickly through Phase II.

Q: Can AI design a drug for any disease?

A: It works best for diseases with one clear molecular target, like a single overactive protein. It's much harder to apply to multifactorial diseases such as Alzheimer's, where several biological pathways interact and no single "block this protein" hypothesis fully explains the disease.

Q: How much cheaper is AI-assisted drug discovery than the traditional process?

A: There's no single audited industry-wide figure yet, since most AI-native drugs haven't finished the full pipeline. What's documented is that the traditional average cost is about $2.6 billion per approved drug, and AI-driven platforms report compressing the early discovery timeline, from target identification to a clinical-ready candidate, from several years down to as little as 12–30 months in specific, publicized cases like Exscientia's and Insilico's.

Q: What's the difference between AI-assisted screening and AI-generated molecule design?

A: AI-assisted screening ranks and filters molecules that already exist in a chemical library. AI-generated design, using models like variational autoencoders or diffusion models such as DiffDock, proposes brand-new molecular structures from scratch that were never previously synthesized, then predicts their properties before anyone makes them in a lab.

Q: Did AlphaFold win a Nobel Prize?

A: Yes. Demis Hassabis and John Jumper won half of the 2024 Nobel Prize in Chemistry for developing AlphaFold2, which predicts a protein's 3D structure from its amino acid sequence. David Baker won the other half for his work on computational protein design.

Q: What was the first clinical trial success for an AI-designed drug?

A: Insilico Medicine's rentosertib, targeting a kinase called TNIK for idiopathic pulmonary fibrosis, reported positive Phase IIa results in November 2024, with the peer-reviewed data published in Nature Medicine in June 2025. It's cited as the field's first published proof-of-concept that an AI-discovered target and an AI-designed molecule can succeed together in a randomized human trial.

Q: Do off-target effects and side effects still happen with AI-designed drugs?

A: Yes. AI can screen a candidate against many known proteins computationally, but it cannot fully predict how a molecule behaves inside a complex living system. AI-designed drugs go through the same preclinical toxicology and clinical safety monitoring as any other drug candidate.

Q: Why did Exscientia merge with Recursion Pharmaceuticals?

A: The two companies combined in November 2024 to pair Recursion's large-scale biological and cellular imaging data with Exscientia's AI-driven chemistry design tools, creating what both companies described as a more complete, "full-stack" AI drug discovery platform than either could offer alone.

Sources

1. Tufts Center for the Study of Drug Development, cost-to-develop estimate ($2.6 billion) and Phase I approval success rate (below 12%): appliedclinicaltrialsonline.com
2. Insilico Medicine, ISM001-055/rentosertib Phase IIa topline results, November 2024: PR Newswire
3. Insilico Medicine, first Phase IIa readout and Nature Medicine publication, 2025: insilico.com/blog/1112
4. Recursion and Exscientia merger completion, November 2024: Bio-IT World
5. Recursion-Exscientia merger deal terms and pipeline size: Drug Discovery Trends
6. Atomwise, AI-driven virtual screening platform (AtomNet): VentureBeat
7. Atomwise, 2014 Ebola antiviral screening project: O'Reilly
8. Google DeepMind and Isomorphic Labs, AlphaFold 3 release and accuracy claims, May 2024: blog.google
9. Isomorphic Labs, pharma partnership deal values and pipeline, 2024–2025: IntuitionLabs
10. The Nobel Prize in Chemistry 2024, official press release (Baker, Hassabis, Jumper): NobelPrize.org
11. MIT News, DiffDock diffusion model for molecular docking, March 2023: news.mit.edu
12. Gómez-Bombarelli et al., "Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules," ACS Central Science, 2018: pubs.acs.org

---

Read next:

• The AlphaFold Protein Folding Breakthrough
• The CRISPR Gene Editing Breakthrough
• Complete AI Guide 2026
• Complete Machine Learning Guide 2026