How Scientists Use AI to Discover New Drugs

Bringing one new drug to market costs an estimated $2.6 billion and takes 10 to 15 years, with roughly 90% of candidates failing somewhere between the lab and the pharmacy shelf (Tufts Center for the Study of Drug Development, 2016; DiMasi et al., Journal of Health Economics, 2016). Machine learning is now used at nearly every early stage of that pipeline — to predict how molecules behave, to design new compounds from scratch, and to flag which candidates are worth testing first. What AI has not yet done is fix the slowest, most expensive part of the process: clinical trials in real patients, still governed by biology, safety, and regulation, not compute.

This article walks through where AI genuinely accelerates drug discovery today, where the evidence is thin, and why "AI discovered a drug" headlines usually describe a molecule that is still years away from your pharmacy.

Key takeaways

• AI's clearest, most verified win is structural biology: DeepMind's AlphaFold — a neural network that predicts a protein's 3D shape from its amino acid sequence — won its creators, Demis Hassabis and John Jumper, a share of the 2024 Nobel Prize in Chemistry (Nobel Prize Committee, 2024).
• Machine learning speeds up the front end of drug discovery — target identification, virtual screening, and molecule design — by narrowing millions of candidates down to a shortlist worth synthesizing.
• AI has not meaningfully shortened the overall drug approval timeline. Clinical trials, which test safety and efficacy in real humans, remain the dominant bottleneck, and phase 2/3 success rates are still around 10% to 15% of candidates that enter human testing (Wong et al., Biostatistics, 2019).
• Several AI-designed molecules have reached human clinical trials (companies including Insilico Medicine, Recursion Pharmaceuticals, and Exscientia), but as of mid-2026 none has completed the full path to FDA approval as a fully AI-originated drug.
• The honest summary: AI is compressing months out of the discovery phase, not years out of the whole pipeline.

The five stages of drug discovery — and where AI actually fits

Before judging what AI has and has not changed, it helps to see the whole pipeline. Drug discovery is not one process; it is a sequence of increasingly expensive filters, each designed to kill bad candidates before they reach human trials. Target identification — finding the biological molecule (usually a protein) that a drug should act on — is a hypothesis-generation exercise: which malfunctioning gene, enzyme, or receptor plausibly explains a disease?

Once a target exists, lead compound discovery searches for a molecule that binds to that target with the right effect. Historically this meant physically testing thousands of compounds in a lab, one plate at a time. Preclinical testing then checks whether a promising lead is safe and effective in cells and animal models before it ever touches a human. Clinical trials are the multi-year, multi-phase test in real patients: phase 1 for safety, phase 2 for early efficacy, phase 3 for large-scale confirmation. Regulatory review is the final gate, where an agency like the U.S. Food and Drug Administration (FDA) examines the full trial dataset before granting approval.

AlphaFold has predicted structures for more than 200 million proteins — nearly every protein known to science — and has been used by more than two million researchers in 190 countries since its 2020/2021 release (Nobel Prize Committee, 2024; Google DeepMind, 2024).

The table below maps where AI tools are genuinely doing work today, as distinct from where the marketing gets ahead of the science.

Stage	What happens	Typical duration	AI's real role today
Target identification	Find the protein or gene driving a disease	1–3 years	Predicting protein structure (AlphaFold) and mining genomic/literature data to rank candidate targets
Lead discovery & screening	Find molecules that bind the target	2–4 years	Virtual screening of compound libraries; generative models designing novel molecules computationally
Preclinical testing	Test safety/efficacy in cells and animals	1–2 years	Predicting toxicity and off-target effects before physical synthesis, reducing wasted lab work
Clinical trials (phases 1–3)	Test in human patients	6–8 years	Patient recruitment matching, trial-site optimization, synthetic control arms in limited cases — the bottleneck AI has barely touched
Regulatory review & approval	Agency evaluates full data package	1–2 years	Minimal; this stage is procedural and legal, not computational

Discovery and screening are where AI genuinely compresses time; clinical trials, roughly 60% to 70% of total pipeline duration, remain largely unchanged (Wong et al., 2019).

How AlphaFold changed structural biology — and why that matters for drugs

For 50 years, figuring out the 3D shape of a protein from its sequence of amino acids was one of biology's hardest unsolved problems, described by researchers as "the protein folding problem." Shape matters because a protein's function comes from its 3D structure. A protein — a folded chain of amino acids that carries out most of the work inside a cell — only binds to a drug molecule if the drug's shape fits its structure the way a specific key fits a specific lock.

Before AlphaFold, determining a single protein structure experimentally, using methods like X-ray crystallography, could take a PhD student years and cost tens of thousands of dollars. DeepMind's AlphaFold2, released in 2020 and detailed in Nature in 2021, used a neural network trained on known protein structures to predict shapes computationally, often within hours and with accuracy competitive with experimental methods (Jumper et al., Nature, 2021). In October 2024, the Nobel Committee awarded Demis Hassabis and John Jumper a share of the Nobel Prize in Chemistry for this work, alongside David Baker for related computational protein design (Nobel Prize Committee, 2024).

Think of it like the difference between hand-sculpting a lock's shape from scratch versus using a 3D scanner. Both eventually produce an accurate model, but the scanner does it thousands of times faster and lets you check far more locks before committing to a key. That is what AlphaFold gave structural biologists: a way to see a target protein's shape before spending years crystallizing it, which is a prerequisite for designing a molecule that fits it.

What did AlphaFold actually solve, and what did it not solve?

AlphaFold predicts a static protein shape with high accuracy for most proteins, work that earned its creators the 2024 Nobel Prize in Chemistry. It does not predict how a protein moves, how it binds to a specific drug molecule, or whether that binding will be safe in a living human body — those questions still require laboratory and clinical testing (Jumper et al., Nature, 2021).

Where AI genuinely accelerates the early pipeline

Two techniques do most of the real work in AI-assisted drug discovery today, and it helps to separate them clearly, because they get conflated in press coverage.

Virtual screening uses machine learning models to predict, computationally, how well each molecule in a library, sometimes containing billions of candidate compounds, will bind to a target protein. Instead of physically synthesizing and testing every molecule, researchers rank candidates by predicted binding affinity and only synthesize the small fraction that scores well. This is a filtering step, similar to how a spam filter does not read every word of every email closely; it triages first, then a human looks closely only at what passed the filter.

Generative molecule design goes a step further: rather than searching an existing library, a generative model proposes entirely new molecular structures optimized for a target, similar to how a generative image model proposes a new image from a text prompt rather than retrieving an existing photo. Companies like Insilico Medicine have used this approach to design novel drug candidates and move them into human trials faster than traditional discovery would allow, with Insilico reporting an experimental fibrosis drug candidate that went from target identification to a phase 1-ready molecule in under 18 months, versus a typical multi-year timeline for that stage (Insilico Medicine, published research and company disclosures, 2021–2023). It is worth noting that reaching phase 1 readiness measures speed to a testable candidate, not proof that the drug works or is safe in humans.

Other real, published applications include AI models that predict a molecule's toxicity before it is synthesized, reducing wasted lab work on compounds likely to fail early safety checks, and models that predict how a drug will be absorbed, distributed, metabolized, and excreted by the body (a set of properties researchers shorthand as ADMET).

None of these tools replace a wet-lab experiment. A model's prediction that a molecule will bind a target, or that it will be non-toxic, is a hypothesis that still must be confirmed in cells, animals, and eventually humans. The AI narrows the search; it does not certify the answer.

Why AI has not shortened the overall pipeline yet

This is the part of the story that gets flattened in headlines, and where SciHub101 wants to be precise rather than exciting. Even if every candidate a company puts into human trials today were discovered twice as fast using AI, the total time to an approved drug would barely move, because discovery is a relatively small slice of the total timeline. Clinical trials, roughly six to eight years across three phases, account for the majority of total development time, and that duration is set by biology and regulation, not computation (Wong et al., Biostatistics, 2019).

A phase 3 trial has to enroll enough patients, over a long enough period, to detect a real treatment effect against the noise of natural disease variation and placebo response. That takes time regardless of how the drug candidate was designed. Regulatory agencies like the FDA also require a fixed body of safety and efficacy evidence before approval; no algorithm shortens the months required to observe whether a drug causes rare side effects that only appear after sustained exposure.

Independent industry analysis published in 2025 found that while AI-driven discovery methods have increased the number of candidate molecules entering preclinical and early trial phases, funding and effort devoted specifically to the trial phase has not kept pace, with clinical trial funding's share of total drug development investment falling from about 6.9% to 4.8% between 2020 and 2025 (industry venture-funding analysis, 2025). More candidates entering an already-strained trial system does not, by itself, produce more approved drugs faster; it can even create a backlog.

There are narrower, real efficiency gains inside trials. The FDA's 2025 guidance validating broader use of synthetic control arms, statistical comparison groups built from historical or real-world patient data instead of a live placebo arm, has in some trial designs reduced enrollment needs and shortened timelines by an estimated six to nine months (FDA guidance and industry reporting, 2025). AI-assisted pre-screening of electronic health records has also improved patient enrollment speed in some trials by up to roughly 25% (industry clinical-operations reporting, 2025). These are real, measurable gains, but they trim months off a process that still runs for years, not years off it.

Has any AI-discovered drug been approved by the FDA?

As of mid-2026, no drug whose discovery and design were led primarily by an AI system has completed the full path to FDA approval. Several AI-originated candidates, including molecules from Insilico Medicine and Exscientia, have entered human clinical trials, but trials themselves, not discovery, remain the long pole in the tent (company trial disclosures, 2021–2025).

What people get wrong about AI and drug discovery

The most common misconception is treating "AI discovered a new drug" as equivalent to "a new drug exists." What usually happened is that AI helped identify or design a candidate molecule that is now entering, or partway through, years of testing that AI does not accelerate. That candidate has roughly a 10% to 15% chance of eventually becoming an approved drug, the same base rate as candidates discovered by traditional methods once they reach human trials (Wong et al., 2019). AI changes which molecules get proposed and tested, not the odds each one clears trials.

A second common error is conflating AlphaFold's structure predictions with drug efficacy predictions. Knowing a protein's shape tells researchers where a drug might dock, similar to knowing the shape of a lock tells you what kind of key might fit. It does not tell you whether inserting that key is safe, whether the body will break the key down before it works, or whether turning that particular lock actually treats the disease. Those are separate, harder questions that still require laboratory and clinical evidence.

A third confusion is treating every "AI-powered" biotech company's discovery claims as independently verified. Some of the most credible published results, like AlphaFold's structural predictions, have been extensively validated against experimental data and peer-reviewed in journals like Nature (Jumper et al., 2021). Others, particularly claims about how fast a specific generative model found a specific candidate, come primarily from company press releases and should be read with the understanding that "reached phase 1" is an early, not final, milestone. This distinction matters for a YMYL-adjacent topic: readers should not conclude that any specific AI-associated drug or company represents a proven treatment. None of the tools or companies discussed here should be read as an endorsement or a claim of therapeutic efficacy; drug safety and effectiveness are established only through completed clinical trials and regulatory review.

How this connects to the AI techniques underneath

The models used in drug discovery are built on the same core machine learning ideas used across the field. Virtual screening and toxicity prediction models are typically trained the same way gradient descent finds better answers for any neural network — by iteratively adjusting the model's internal parameters to reduce prediction error on known molecule-target pairs. AlphaFold itself is a specialized neural network architecture, and the deeper mechanics of how it predicts structure are covered in the AlphaFold protein folding breakthrough.

It is also worth understanding the biological tool that AI-designed drugs increasingly work alongside: gene editing. Some AI-discovered candidates target diseases caused by specific gene mutations, and researchers increasingly pair computational drug design with tools like DNA sequences in living cells with letter-level accuracy.">CRISPR, covered in depth in the CRISPR gene editing breakthrough, to correct the underlying genetic cause rather than only treating symptoms with a small-molecule drug.

The takeaway

AI has made the discovery and design phase of drug development faster and cheaper in specific, well-documented ways, most clearly through AlphaFold's protein structure predictions and through virtual screening and generative molecule design that narrow enormous candidate pools down to a manageable shortlist. What it has not done, as of 2026, is shorten the years-long clinical trial process that determines whether a candidate is actually safe and effective, and that gap is why headlines about "AI-discovered drugs" describe promising starting points rather than finished treatments. The realistic picture is a pipeline getting faster at one end and staying exactly as slow, and exactly as necessary, at the other.

---

Read next:

• The AlphaFold Protein Folding Breakthrough
• The CRISPR Gene Editing Breakthrough
• How Gradient Descent Finds Better Answers

FAQ

Q: Did AI actually discover a drug that's now on the market? As of mid-2026, no FDA-approved drug has been fully attributed to AI-led discovery and design. Several AI-originated candidates from companies like Insilico Medicine and Exscientia have reached human clinical trials, but none has completed the multi-year trial and approval process yet.

Q: What did AlphaFold actually win the Nobel Prize for? Demis Hassabis and John Jumper won a share of the 2024 Nobel Prize in Chemistry for AlphaFold, an AI system that predicts a protein's 3D structure from its amino acid sequence, solving a problem structural biologists had worked on for roughly 50 years (Nobel Prize Committee, 2024).

Q: Has AI made new drugs cheaper to develop? AI has reduced costs in the discovery phase by cutting the number of compounds that must be physically synthesized and tested. It has not reduced the dominant cost driver, which is the multi-year clinical trial process, so total development cost estimates around $2.6 billion have not measurably dropped industry-wide.

Q: Why hasn't AI sped up clinical trials the way it sped up drug discovery? Clinical trials test safety and efficacy in real patients over months or years, a timeline set by biology, statistics, and regulatory requirements, not by computation. AI has improved narrower trial operations, like patient recruitment matching and synthetic control arms, but has not compressed the core multi-phase testing timeline.

Q: Can AI find new uses for drugs that are already approved?

Yes, and this is one of the more mature, lower-risk applications of AI in the field, known as drug repurposing: using machine learning to predict whether a drug already approved for one condition might also treat a different disease. Because a repurposed drug already carries an established human safety profile from its original approval, this approach can skip much of the early safety testing that new molecules require, meaningfully shortening the path to clinical trials for the new indication (industry and academic reviews of AI-driven repurposing, 2024–2025).

The technique typically works by mapping relationships between drugs, genes, proteins, and diseases into a large network, sometimes called a knowledge graph, and training a model to predict plausible new drug-disease connections based on patterns in that network. Research groups have used this approach to screen large numbers of approved compounds against thousands of disease indications simultaneously; one published framework mapped patient-derived genetic data onto protein interaction networks and flagged existing cancer drugs, including several with well-established safety records, as promising repurposing candidates for other tumor types (academic drug-repurposing research, published in peer-reviewed bioinformatics literature). A newer graph-based model called TxGNN has been developed specifically to predict repurposing candidates across more than 17,000 diseases, including rare conditions with few or no existing approved treatments (published graph neural network research, Nature Medicine, 2024).

Drug repurposing itself long predates AI. Sildenafil, originally developed and tested as a treatment for hypertension and angina, was repurposed after clinical trial observations revealed a different effect and became the drug marketed as Viagra. What AI has changed is the scale and speed of the search: instead of relying on incidental clinical observations or a researcher's hunch, machine learning models can now systematically scan an approved-drug library against a much broader set of disease targets before any new trial begins. As with drug discovery more broadly, a predicted repurposing candidate is still a hypothesis, not a proven treatment, and requires its own clinical trials confirming safety and efficacy for the new indication before regulators will approve it.

Q: Is it safe to take a drug that was "designed by AI"? Any drug reaching a pharmacy, regardless of how it was discovered or designed, must pass the same FDA safety and efficacy review as drugs discovered through traditional methods. The discovery method does not change the regulatory bar a drug must clear before approval.