AI Drug Discovery: How Machine Learning Works

Summary

Machine learning is changing how new medicines are found, but not yet how many are approved. In drug discovery, machine learning models now propose disease targets, design molecules and screen millions of virtual compounds before a laboratory experiment starts. More than 173 AI-discovered programs were in clinical development in early 2026, and Insilico Medicine's rentosertib became the most advanced, with a Phase III trial announced in July 2026. Yet trackers still report no fully approved AI-discovered drug, and only about 6.7% of drugs entering Phase 1 reach the market. This guide explains the traditional pipeline, where AI enters each stage, what large language models and quantum computing really add, which companies lead, and how the FDA and EMA expect AI evidence to be documented. Market estimates differ sharply, which shows how much definitions matter in this young field.

Table of Contents

1. What Is AI-Driven Drug Discovery?

2. The Drug Discovery and Development Process

3. Target Identification with Machine Learning

4. Compound Screening and Molecule Design

5. LLMs and AI Tools for Drug Discovery

6. Drug Discovery and Quantum Computing

7. Top AI Drug Discovery Companies and Platforms

8. Regulation, Limits and the Road Ahead

9. Common Mistakes to Avoid

10. FAQ

11. Conclusion

1. What Is AI-Driven Drug Discovery?

AI-driven drug discovery is the use of machine learning models, trained on chemical, biological and clinical data, to find disease targets, design or rank candidate molecules and predict how those molecules will behave before anyone runs an experiment. You will also see it called AI-powered drug discovery. The idea is simple: instead of testing millions of compounds blindly, a model learns patterns from past results and points scientists toward the few experiments most likely to pay off.

Two distinctions keep the topic honest. The first is between discovery and development. Discovery covers finding a target and a molecule that acts on it. Development covers preclinical safety work, human trials and regulatory review. Most AI progress so far sits in discovery, while most cost and failure sit in development. The second distinction is between AI-assisted and AI-discovered. A drug whose target or structure came largely from an AI platform is AI-discovered, while a drug that only used AI for some analysis along the way is AI-assisted. Industry trackers often count both, which inflates headline numbers.

The money follows the promise. Research and Markets estimates the global AI in drug discovery market at 5.09 billion US dollars in 2026, growing to 17.56 billion by 2031, a 28.1% compound annual growth rate. A second report from the same publisher puts the market at about 3.1 billion dollars in 2025 and 43.9 billion by 2035. The gap reflects what each report counts, such as platforms alone or platforms plus services. Treat these as vendor forecasts that show direction rather than precise sizes. The second report also says machine learning held 82.6% of the market in 2025 and software 67.9%, which matches what researchers see in practice: the field runs on models and data rather than on new laboratory hardware.

Three forces made this possible. Protein structure prediction, led by AlphaFold, gave models a map of biology. Larger public and proprietary datasets gave them examples to learn from. And language models gave scientists a way to read the literature and plan work at scale. The sections below follow the pipeline from the first question, what should we target, to the last, what will regulators accept.

2. The Drug Discovery and Development Process

To see where AI helps, you need the baseline. The classic drug discovery and development process has five broad steps. Scientists first identify and validate a biological target, usually a protein involved in a disease. They then look for hits, molecules that bind the target, by screening large compound libraries. Hits become leads through repeated chemical tweaking, and leads are optimized for potency, safety and how the body absorbs and clears them. The best candidates then enter preclinical testing, followed by three phases of human trials and a regulatory review.

Every step is slow and every step loses candidates. The most useful recent number comes from a Citeline analysis of 2014 to 2023 data, reported by MedPath: only 6.7% of drugs entering Phase 1 ultimately reach the market, down from 10.4% in Citeline's 2014 estimate. Phase 2 is the biggest bottleneck, with a 28% success rate, while Phase 3 succeeds 55% of the time. In plain terms, most molecules that look promising in the lab fail in patients, usually for lack of efficacy or for safety reasons that earlier tests did not predict.

Cost is just as contested. You will often read that a new drug takes 10 to 15 years and more than a billion dollars. That framing is a historical benchmark rather than a current measurement, and the best recent evidence is more nuanced. A January 2025 RAND study of 38 drugs approved by the FDA in 2019, using company spending from 2014 to 2019, found a median direct R&D cost of 150 million dollars per drug but a mean of 369 million. The gap exists because a few extremely expensive programs pull the average up. RAND's numbers cover direct R&D spending, so they are not directly comparable with the larger figures that include failures and the cost of capital.

This is the opportunity for machine learning. Even small improvements in who gets picked for the next experiment compound across thousands of decisions. If a model can steer a team away from a doomed target before the money is spent, the saving is far larger than any speedup in the lab itself. That logic explains why investors, pharmaceutical companies and regulators now treat AI as a serious part of the pipeline rather than a curiosity.

3. Target Identification with Machine Learning

Target identification in drug discovery is the first and arguably the riskiest decision: choose the wrong protein and every later step is wasted. Machine learning attacks this problem by combining sources no single scientist can read in full. Models ingest multi-omics data, which measures genes, proteins and metabolites together, along with millions of published papers, patents and clinical records. They then rank which proteins are most likely to drive a disease and are also realistic to drug. The data fusion principle is similar to the one we describe in our guide to multi-modal AI fusion, where combining different signals reveals patterns that none shows alone.

The clearest real-world example is rentosertib. Insilico Medicine's PandaOmics platform identified TNIK as a target in idiopathic pulmonary fibrosis, a progressive scarring disease of the lungs, and the company's generative chemistry platform then designed a molecule against it, according to Bio-IT World's reporting. That end-to-end story, from AI-chosen target to AI-designed molecule to human trial, is why the drug is so closely watched.

Structure prediction is the second pillar. Once you know a target, you need its three-dimensional shape to design something that fits. AlphaFold 3, released in 2024, extended structure prediction from single proteins to proteins interacting with other molecules, including potential drugs. In February 2026, Isomorphic Labs, the drug discovery spin-off of Google DeepMind, announced a more powerful successor. Nature reported that the company is keeping the model proprietary, which leaves open-source developers guessing how to match it. That detail matters for science policy: the strongest tools may sit behind company walls, which makes independent validation harder.

Be careful with speed claims. You will see statements that AI cuts target identification time by half or more. Those figures come mostly from companies and secondary sources, and we found no independent, industry-wide study that confirms them. A fair summary is that machine learning widens the search and reorders priorities faster than manual review, but whether the chosen targets succeed in patients is only answered years later in the clinic. Rentosertib's target, for example, took from identification to a Phase III announcement in July 2026 to be tested, and even that drug still has to prove itself in a larger trial.

4. Compound Screening and Molecule Design

Once a target is chosen, the hunt for a molecule begins. Compound screening in drug discovery has traditionally meant testing hundreds of thousands of real compounds in robotic assays. Machine learning adds two shortcuts. Virtual screening scores enormous digital libraries against a protein structure so that only the most promising few thousand are ever made and tested. Generative chemistry goes further and invents new molecules that match a wish list of properties, such as strong binding, low toxicity and a shape that the body can absorb.

How well a molecule binds depends on the chemical interactions between its atoms and the target pocket, including covalent bonds, ionic attractions and weaker forces. If those ideas are new to you, our explainer on ionic versus covalent bonds covers the basics that underlie every docking score. Models also predict how a compound will be absorbed, distributed, metabolized and excreted, which is why the field talks about ADMET and DMPK. Catching a poor profile on a computer is far cheaper than discovering it in animals.

Rentosertib again provides the most detailed public evidence. In a Phase IIa trial published in Nature Medicine on 3 June 2025, Insilico reported results from 71 patients with idiopathic pulmonary fibrosis across 22 sites in China. The study was randomized, double-blind and placebo-controlled and lasted 12 weeks. It met its primary safety endpoint, and lung function, measured as forced vital capacity, improved in a dose-dependent way. Patients on 60 milligrams once daily gained a mean of 98.4 milliliters, while the placebo group lost 20.3 milliliters.

Those numbers are encouraging, but the lead investigator himself noted that each group was small and that larger studies are needed. That caution applies to the whole field. A model that designs a good molecule on paper still has to pass biology, manufacturing and safety, none of which a computer can fully simulate. AI shortens the path to a promising candidate, yet the final proof remains a human trial. Think of generative design as a very fast first draft, not a finished product.

The practical takeaway for readers is to separate three claims that are often blurred: that AI can propose a molecule quickly, that the molecule is better than a conventionally found one, and that it will become an approved medicine. The first is well supported, the second is plausible but rarely shown head to head, and the third is still unproven.

5. LLMs and AI Tools for Drug Discovery

Large language models are the newest layer. Unlike structure predictors, which handle molecules directly, LLMs for drug discovery mostly work with text, code and tools. They read papers, summarize evidence, write analysis scripts, plan experiments and call specialized software. The result is closer to a research assistant than to a molecule designer, and it is why many of the AI tools for drug discovery launched in 2025 and 2026 are described as agents or workbenches.

The big model developers have all moved in. OpenAI announced GPT-Rosalind, a reasoning model for biology, drug discovery and translational medicine. According to heise, it can access more than 50 scientific tools and data sources through a life sciences plug-in for Codex, but access is limited to qualified U.S. enterprise customers under a trusted access program. OpenAI had already announced a collaboration with Sanofi and Formation Bio in May 2024 to fine-tune models on Sanofi data. Google's work runs mainly through Isomorphic Labs, as described above, where AlphaFold-derived models are applied to small molecules and antibodies.

Anthropic launched Claude for Life Sciences in October 2025 and followed with Claude Science, a beta workbench for drug discovery scientists. Pharmaceutical Technology reported on 1 July 2026 that it runs locally or on a remote machine, connects tools such as PubMed, Jupyter and R, and includes more than 60 functions for areas like genomics and cheminformatics. The same report says Anthropic is starting an internal program to find treatments for neglected diseases, partly to learn how to build better tools for pharma. For readers evaluating these products, the common pattern is orchestration: the model coordinates existing databases and software rather than replacing them.

LLMs bring real risks. They can state false facts with confidence, and a hallucinated citation or an invented binding result in a drug program is costly. They also inherit biases from the literature they read, which favors positive results. The safest uses today keep a human reviewer in the loop and ask the model to show its sources and code. Regulators are moving the same way, as the final section explains.

It is also worth asking what these tools change for smaller teams. A university lab or a small biotech without a dedicated bioinformatics group can now query genomic databases, run standard analyses and draft protocols through a conversation. That lowers the entry cost of exploring a new disease, which may matter most for neglected conditions that large companies ignore. Whether lower cost translates into more approved drugs is a separate question, because the expensive part, running trials, is unchanged. In short, LLMs speed up the thinking around the experiments, not the experiments themselves.

6. Drug Discovery and Quantum Computing

Quantum computing is often presented as the next leap for drug discovery, because molecules are quantum systems and classical computers struggle to simulate them exactly. The pitch is that a large enough quantum machine could calculate the electronic behavior of a drug and its target far more accurately than today's approximations. The reality in 2026 is a research milestone, not a product.

The most striking result so far is a hybrid one. In May 2026, RIKEN announced that Cleveland Clinic, RIKEN and IBM had modeled a protein complex of 12,635 atoms, described as the largest protein system simulated with quantum computers. The team also reached the finals of the 2026 ACM Gordon Bell Prize, according to The Quantum Insider. Hybrid means that a quantum processor handles the hardest part of the calculation while ordinary supercomputers do the rest, so the achievement shows a workflow rather than a machine that outperforms classical methods on drug problems.

That distinction is important. Independent summaries of the field note that current quantum systems are not fault-tolerant, that no quantum-derived drug candidate has completed clinical development, and that Google's October 2025 verifiable quantum advantage result was a physics benchmark rather than a drug discovery demonstration. IBM's public roadmap targets quantum advantage for commercially relevant problems by the end of 2026 and fault-tolerant machines by 2029. Roadmaps are promises, not results, and they have slipped before in this industry.

Where might quantum computing help first? Likely in narrow, expensive calculations, such as modeling how a drug binds to a metal-containing enzyme or estimating energies that classical methods approximate poorly. Those results would then feed back into the machine learning models described earlier. Some researchers also explore quantum generative models for designing small molecules, but these remain at the proof-of-concept stage.

The sensible reading for a general audience is to watch, not to bet. If you read a headline claiming that quantum computers have discovered a drug, check whether it describes a simulation of a protein, a software benchmark or an actual molecule that entered a trial. Today only the first two exist. Classical machine learning, running on ordinary graphics processors, remains the engine behind every AI-discovered drug now in human testing.

7. Top AI Drug Discovery Companies and Platforms

How many AI drug discovery companies are there? No authoritative count exists, because the answer depends on whether you include software vendors, contract research firms and traditional pharmaceutical groups with AI teams. The Research and Markets forecast names NVIDIA, Schrodinger, Recursion, Insilico Medicine, Google, Microsoft, Tempus AI, Illumina, XtalPi and Iktos among the leading participants, which shows how broad the definition is. Below we focus on the ones with clinical or funding milestones you can check.

Insilico Medicine is the most advanced on the clinical side. Its rentosertib reached Phase IIa with published results in 2025, and in July 2026 Bio-IT World reported a Phase III trial: a randomized, double-blind, placebo-controlled study of 320 patients running 52 weeks. Isomorphic Labs raised 600 million dollars in March 2025 and has drug design collaborations with Eli Lilly and Novartis, with potential milestones of up to 1.7 billion and 1.2 billion dollars respectively, according to industry analyses. Recursion, which merged with Exscientia, and Xaira Therapeutics, which launched with 1 billion dollars of initial funding, round out the best-funded names. These deal values are potential payments tied to milestones, not money received.

The clinical scoreboard is the better test than funding. A tracker published by Healthcare Discovery in April 2026 counted more than 173 AI-discovered programs in clinical development, with roughly 94 in Phase I, 56 in Phase II and 15 in Phase III. Those phases add up to about 165, so the breakdown is approximate. A 2024 BCG analysis published in Drug Discovery Today, summarized by HT World, reported Phase I success of 80 to 90% for AI-discovered molecules against 40 to 65% historically, but only about 40% in Phase II, in line with industry averages, and from a limited sample.

The pattern is clear: AI seems to help molecules clear early safety hurdles, while efficacy in patients remains the hard part. As of mid-October 2026, industry trackers report no fully approved AI-discovered drug. One June 2026 article projects a first FDA approval of an AI-discovered drug in 2026 or 2027, but that is a forecast, not a fact.

When comparing drug discovery platforms, look for three things: a clinical-stage asset, published data in a peer-reviewed journal, and a partner willing to pay upfront. Platforms with all three deserve more attention than those with a large funding round alone. Our guide to molar mass calculations shows the kind of basic quantitative literacy that helps when reading company data tables.

8. Regulation, Limits and the Road Ahead

Regulators have moved from watching to writing rules. In January 2025 the FDA published draft guidance on using artificial intelligence to support regulatory decision making for drugs and biological products. On 14 January 2026 the FDA and the EMA jointly set out ten principles for good AI practice across the medicines lifecycle. The FDA's version lists them as: keep humans at the center of design, take a risk-based approach, follow established standards, define a clear context of use, draw on multidisciplinary expertise, govern and document data well, apply sound model design practices, assess performance based on risk, manage the model across its life cycle, and communicate clear and essential information.

Neither agency calls these principles binding. The EMA describes them as broad guidance that will underpin future AI guidance, and says EU guideline development is already underway. For companies, the message is practical: document where your training data came from, state exactly what each model is used for, and keep monitoring it after launch.

The limits of the technology are equally clear. Data quality and provenance come first, because a model trained on biased or noisy assays produces confident but wrong predictions. Validation comes second. The idea of context of use means that a model proven for one task, such as ranking compounds, cannot be assumed to work for another, such as predicting patient response. Reproducibility is third, since proprietary models and changing training data make independent replication difficult, as the Isomorphic example showed. Finally there is clinical translation: despite large pipelines, the field still lacks a fully approved product, and biology remains the main bottleneck.

What should readers expect next? Over the next two years the key tests are concrete. Rentosertib's Phase III will show whether an AI-discovered drug can deliver benefit in a large trial. The 15 or so Phase III programs will show whether early success rates carry through. Regulators will turn principles into detailed guidance. And market forecasts will either be validated by approved products or revised downward. Our guide to voice biomarkers and AI shows a parallel story, where promising signals still need rigorous clinical validation before they change care.

The honest summary is that AI has already changed how candidates are found, and it has not yet changed how many reach patients. That second change is what the next few years will decide.

Common Mistakes to Avoid

Confusing "AI-discovered" with "AI-approved." Hundreds of programs are in trials, but trackers report no fully approved AI-discovered drug yet. Check the phase before repeating a success story.

Comparing market forecasts without checking definitions. The 3.1 billion and 5.09 billion dollar estimates above come from different reports with different scopes. Always note what is counted and who produced the number.

Quoting the "10 to 15 years and billions of dollars" line as current fact. It is a historical benchmark. The recent RAND study found a median direct R&D cost of 150 million dollars, with a mean of 369 million.

Assuming LLMs design molecules by themselves. Today's language models mostly orchestrate tools, read literature and write code. Molecule design relies on specialized generative and structure models.

Treating quantum computing as ready. The 12,635-atom simulation is a real milestone, yet no quantum-derived drug candidate has completed clinical development.

Ignoring context of use. A model validated for one decision cannot be assumed valid for another, which is the point regulators stress most.

Frequently Asked Questions

1. How does AI drug discovery work?

AI drug discovery works by training models on chemical, biological and clinical data so they can predict which targets, molecules and experiments are most likely to succeed. A typical workflow starts with a model ranking disease targets from omics data and literature. Structure models then predict how candidate molecules fit the target, and generative models propose new compounds that meet a property wish list. Other models forecast absorption, toxicity and metabolism. Scientists make and test the best candidates, and the new results are fed back to improve the model. The loop repeats until a lead is ready for preclinical work. Humans stay in charge of every decision that reaches a patient.

2. How is AI being used in drug discovery and development?

AI is used at nearly every stage. In discovery it ranks targets, screens virtual libraries and designs molecules. In development it can help choose patients, analyze trial data and monitor safety signals, though regulators ask that such uses be documented and validated. Language models read papers, write code and plan experiments, as in Anthropic's Claude Science workbench and OpenAI's GPT-Rosalind. Structure models such as those from Isomorphic Labs predict protein and ligand interactions. The strongest public evidence is Insilico Medicine's rentosertib, whose target and molecule came from AI platforms and which has a published Phase IIa result and an announced Phase III trial.

3. How can AI speed up drug discovery?

AI speeds up discovery mainly by narrowing the search before costly experiments. Virtual screening scores huge digital libraries so that only the best few thousand compounds are made. Prediction models flag poor safety or absorption profiles early, and generative design proposes candidates that already match the target profile. Language models cut the time spent reading literature. Claims of specific time savings, such as halving target identification, mostly come from companies and lack independent confirmation. The more reliable evidence is that AI-discovered molecules show higher Phase I success, 80 to 90% in a 2024 BCG analysis, though Phase II success looks similar to the industry average.

4. What is AI-powered drug discovery?

AI-powered drug discovery is another name for AI-driven drug discovery: using machine learning and related methods to find targets, design molecules and predict their behavior. It covers a spectrum, from AI-assisted projects that use models for a few analyses to AI-discovered programs where a platform produced the target or structure. Platforms from companies such as Insilico Medicine, Recursion and Isomorphic Labs fall under this heading, as do research tools from major AI developers. The term does not mean an AI works alone. Human scientists choose targets, run experiments, interpret results and take responsibility for clinical decisions, and regulators expect that oversight to be documented.

5. Why is drug discovery so hard?

Drug discovery is hard because biology is complex and tests in the lab often fail to predict results in patients. A molecule must bind its target, avoid dangerous side effects, reach the right tissue, stay in the body long enough and be manufacturable. Citeline data from 2014 to 2023 shows that only 6.7% of drugs entering Phase 1 reach the market, with Phase 2 succeeding only 28% of the time. Most failures come from lack of efficacy or unexpected safety problems in humans. Diseases like idiopathic pulmonary fibrosis also lack good models, so researchers cannot know a target is right until a trial is run.

6. How long does drug discovery and development take?

The commonly quoted range is 10 to 15 years from target to approval, but that is a historical benchmark rather than a figure we could confirm from a current primary source. Duration varies by disease, trial size and regulatory pathway. Rentosertib shows how long even a fast program takes: an AI platform identified its target, a Phase IIa trial was published in June 2025, and a 52-week Phase III trial of 320 patients was announced in July 2026, with approval still ahead. AI aims to shorten the early discovery phase, but human trials set the pace for most of the timeline, so total time may fall less than headlines imply.

7. How much does drug discovery cost?

Estimates depend heavily on method. A January 2025 RAND study of 38 drugs approved by the FDA in 2019 found a median direct R&D cost of 150 million dollars and a mean of 369 million, using company spending from 2014 to 2019. The mean is higher because a few extremely costly programs skew the average. Much larger figures in the billions are widely quoted, but they typically include the cost of failed projects and of capital, which RAND's direct figure does not. Costs are also rising: Citeline data cited by MedPath shows falling success rates, which raises the cost per approved drug. Check which definition any number uses.

8. How many AI drug discovery companies are there?

There is no authoritative count. The answer changes with the definition: pure AI biotechs, software vendors, contract research firms and large pharmaceutical companies with internal AI teams all appear in different lists. A Research and Markets report published on 12 October 2026 names NVIDIA, Schrodinger, Recursion, Insilico Medicine, Google, Microsoft, Tempus AI, Illumina, XtalPi and Iktos among the leading participants. A secondary tracker counted more than 173 AI-discovered programs in clinical development in early 2026, which is a count of drugs rather than companies. For a reliable figure, look for a stated methodology before you cite any total.

Conclusion

Machine learning has become a normal part of drug discovery, from ranking targets to designing molecules and drafting research plans. The evidence so far is encouraging at the early stages: AI-discovered molecules appear to clear Phase I safety hurdles more often, and rentosertib has reached Phase III. Yet the hardest question, whether these drugs help patients more reliably than conventionally found ones, is still open, and no fully approved AI-discovered drug exists as of October 2026. Large language models make research faster, quantum computing remains a promising experiment, and regulators have started to set shared rules. If you follow one thing, follow the Phase III results and the first approval decision. Those outcomes will separate lasting progress from hype, and we will update this guide as they arrive.

Sources

1. Research and Markets via GlobeNewswire, Global AI in Drug Discovery Market Forecast to 2031, 12 October 2026.

2. Research and Markets via GlobeNewswire, AI in Drug Discovery Market, 2026 to 2035, 12 October 2026.

3. RAND, Typical Cost of Developing a New Drug Is Skewed by Few High-Cost Outliers, 7 January 2025.

4. MedPath, Why Drug Development Costs Are Rising, citing Citeline/Biomedtracker data for 2014 to 2023.

5. Insilico Medicine, Nature Medicine publication of Phase IIa rentosertib results, 3 June 2025.

6. Bio-IT World, Insilico Medicine Launches Phase III Trial for AI-Developed Drug, 8 July 2026.

7. Nature, Isomorphic Labs announces a more powerful model, 19 February 2026.

8. EMA, EMA and FDA set common principles for AI in medicine development, 14 January 2026.

9. FDA, Guiding Principles of Good AI Practice in Drug Development, 14 January 2026.

10. Pharmaceutical Technology, Anthropic launches Claude Science for drug discovery, 1 July 2026.

11. heise, OpenAI introduces GPT-Rosalind.

12. RIKEN, Cleveland Clinic, RIKEN and IBM quantum protein simulation, 5 May 2026.

13. Healthcare Discovery, The AI Drug Discovery Pipeline Reaches Its Moment of Truth in 2026, 9 April 2026.

14. HT World, AI-discovered drugs achieved higher success rate than those by humans, summarizing a BCG study in Drug Discovery Today.

MAGAZINE

© 2025. All rights reserved.

Disclaimer: While browsing our website, you may encounter affiliate links, as well as direct links to our own platforms where you can purchase our digital illustrations, products featuring our designs, or other science-related products.If you click on an affiliate link and make a purchase, we may earn a small commission at no additional cost to you. We only recommend products and services that we genuinely believe in and that align with our brand and values. Thank you for your support—it helps us continue creating content!

Privacy policy

Écrivez votre texte ici ...