AI Drug Discovery

Why Traditional Drug Development Is So Slow and Costly

Traditional drug development runs on a long chain of expensive, uncertain experiments. A single approved medicine can take ten to fifteen years from first idea to pharmacy shelf. Scientists must find a biological target, design molecules, and test them in cells, animals, and people. Each stage discards most candidates, so the survivors carry the cost of every failure before them. The widely cited figure of about USD 2.6 billion per approved drug captures how brutal this attrition really is. That estimate, traced through the federal review of clinical trial costs, reflects years of dead ends. The price is not just money, it is the diseases left untreated while teams chase long odds.

The failure rate is the core of the problem, not a footnote. Around 90 percent of drug candidates that enter human trials never reach approval. Many collapse in Phase II, where a promising idea finally meets real patients and disappoints. A 2025 RAND analysis even put median research and development cost near USD 708 million, far below the headline number. That gap shows how contested these economics are, yet every estimate agrees the waste is enormous. Whichever figure holds, the message for the industry stays the same. The earliest decisions, which target and which molecule, shape almost everything that follows.

Most of that risk is locked in before a drug ever reaches a patient. Choosing the wrong target wastes a decade and billions on a molecule that was doomed from the start. Designing molecules by hand and intuition leaves vast regions of chemistry unexplored. Testing them one batch at a time turns discovery into a slow, serial grind. This is precisely the bottleneck that artificial intelligence aims to break. By learning from past data, models can flag weak targets and weak molecules early. Our overview of drug discovery and development using AI shows why front-loading these decisions matters so much.

How the AI Discovery Pipeline Works End to End

Building on that bottleneck, AI-driven discovery reorganizes the early pipeline into a tighter, data-driven loop. The process starts with data, the messy record of genes, proteins, diseases, and known drugs. Models learn patterns in that data to suggest which biological target might drive a disease. From target to candidate, AI tries to replace blind trial and error with informed prediction at every step. Generative systems then design molecules likely to hit that target safely. Other models score how each molecule will bind, dissolve, and clear the body. Only the strongest candidates move into real laboratory and clinical testing.

The defining feature is the loop, not any single clever model. Predictions feed lab experiments, and experimental results feed back to retrain the models. This lab-in-the-loop cycle sharpens each round of suggestions with fresh evidence. Insilico Medicine famously ran this loop to move a target to Phase II in roughly eighteen months, a pace the critical review of AI small-molecule discovery documents in detail. That speed comes from compressing design cycles, not from skipping the science. Each stage still demands chemistry, biology, and careful human judgment. The payoff is fewer wasted experiments and faster learning from the ones that run.

No part of this pipeline works without trustworthy, well-organized data. Garbage data produces confident but wrong predictions that waste real lab time. Teams spend heavily to clean, label, and standardize their biological datasets first. The quality of that foundation often decides whether the whole loop succeeds. Public resources and shared molecular datasets now make this groundwork far easier. The piece on AI-ready molecular datasets explains why clean chemical data underpins every later step. With that base in place, the first real decision is which target to pursue.

Finding New Drug Targets With Machine Learning

Turning to the first decision, target discovery is where AI can save the most time and money. A target is usually a protein or gene whose behavior drives a disease. Choosing the right one is the single most consequential call in the entire pipeline. Machine learning models scan genomics, proteomics, and clinical data to rank which targets most plausibly cause a disease. They surface patterns no human could hold in mind across millions of data points. Insilico’s PandaOmics engine used this approach to prioritize a protein called TNIK for lung fibrosis. That target selection set up the entire rentosertib program that followed.

The strength of AI here is breadth, the ability to weigh evidence at scale. Models integrate gene expression, genetic associations, and text from millions of research papers. They can connect a faint genetic signal to a disease mechanism buried in old literature. This wide view helps teams find novel targets that competitors have overlooked. It also flags targets likely to fail, sparing years of doomed effort. Work on AI in genomics and genetic analysis shows how these signals are mined. Strong target picks still need wet-lab confirmation before any molecule is designed.

Designing New Molecules With Generative AI

Building on a chosen target, generative AI tackles the next hard problem, inventing the molecule itself. Chemical space is almost unimaginably large, with more possible drug-like molecules than atoms in the solar system. Searching it by hand is hopeless, so chemists historically explored tiny, familiar corners. Generative models flip this around by proposing brand new molecules designed to fit a specific target pocket. They learn the grammar of viable chemistry from millions of known compounds. The model then generates novel structures predicted to bind tightly and behave safely. Insilico’s Chemistry42 engine used this method to design the inhibitor behind rentosertib.

The newest systems design molecules directly inside the three-dimensional binding site. Structure-based generative models grow a candidate atom by atom within the protein cavity. This approach optimizes many properties at once, not just binding strength. Teams can ask for potency, solubility, and safety in a single design request. The review of generative AI in drug discovery describes this shift as a true paradigm change. It moves the field from screening existing libraries to inventing fit-for-purpose chemistry. The catch is that a designed molecule on screen still has to be synthesized in a lab.

Generative design is powerful, yet it is not magic and it makes mistakes. Models sometimes propose molecules that are unstable, toxic, or impossible to manufacture. Chemists must filter these suggestions and reject the ones that defy practical chemistry. The best teams treat generated molecules as strong starting points, not finished drugs. Synthetic accessibility scoring helps screen out structures no chemist could build. This human filtering step keeps the loop grounded in real-world feasibility. Once a candidate looks promising, its target binding must be checked against an accurate protein structure.

The economics of generative design keep improving as the models and tools mature. Cloud compute now lets a small team generate and rank thousands of candidate molecules in a single day. Open-source generative frameworks have lowered the cost of entry for academic groups. That access means novel chemistry is no longer the preserve of a few large firms. Shared public benchmarks also let teams compare model quality on the same yardstick. The result is faster and cheaper design loops that still depend on careful human review. None of this removes the need to synthesize and test the winning molecules.

Predicting Protein Structures With AlphaFold

Beyond designing molecules, AI also cracked one of biology’s oldest puzzles. Knowing a protein’s three-dimensional shape is essential, because structure determines how a drug can bind it. For fifty years, mapping a single structure could take a scientist months or even years of painstaking lab work. DeepMind’s AlphaFold changed that overnight by predicting protein structures in seconds rather than years. The breakthrough gave researchers reliable shapes for nearly every known protein at once. That foundation reshaped how AI-driven discovery teams approach previously invisible targets. It turned a rare, slow measurement into an everyday computational resource.

The latest version pushed the science far past folding single proteins. AlphaFold 3 predicts the structure and interactions of proteins, DNA, RNA, and small molecules together, a leap the DeepMind announcement of AlphaFold 3 describes in detail. Modeling how a drug and a protein actually meet is the heart of rational design. The model handles most of the molecule types stored in the Protein Data Bank. Researchers can access much of that power through the public AlphaFold Server. This turns a slow experimental step into a fast, scalable computation for thousands of teams. The connection between folding and discovery appears in our piece on protein folding and computation.

Structure prediction is powerful, yet it is not a complete answer on its own. A predicted shape is a static snapshot, while real proteins flex and shift constantly. Binding also depends on water, motion, and subtle chemistry a single structure cannot capture. Experimental methods like cryo-electron microscopy still validate the most critical structures. Used wisely, AlphaFold accelerates the start of design rather than guaranteeing a finished drug. The smartest teams treat its predictions as strong hypotheses, not final truth. With a structure and candidate molecules in hand, the next task is ranking them fast.

Virtual Screening and Hit Identification

With that structure in hand, virtual screening ranks huge libraries of molecules by computer. Instead of testing millions of compounds in a lab, teams score them digitally first. The model predicts how strongly each compound binds the chosen target. This computational triage can shrink a library of millions down to a few dozen promising hits. Those best-scoring compounds, called hits, then move forward to real experimental testing. The savings in time, money, and lab materials at this stage are substantial. Screening that once took months can now run in days on shared hardware.

Modern screening blends physics-based docking with fast machine learning predictions. Docking simulates how a molecule physically fits into a protein’s binding pocket. Learned models, trained on known active compounds, then rerank those poses for accuracy. This hybrid keeps the rigor of physics while gaining the speed of pattern recognition. Teams now screen virtual libraries of billions of compounds that were never even synthesized. Clean chemical data, as covered in work on data challenges in life science AI, underpins reliable screening results. Even a strong hit, though, can still fail on safety once tested.

Predicting Toxicity, Safety, and ADMET Properties

Beyond binding, a real drug must survive the body without harming it. Most candidates fail not because they miss the target but because they are unsafe. AI models now predict ADMET properties, how a drug is absorbed, distributed, metabolized, excreted, and how toxic it might be. These predictions flag dangerous molecules long before they reach an animal or a person. A model can warn that a compound will damage the liver or fail to clear the body. Catching these flaws early saves enormous downstream cost and protects trial participants. Safety prediction is where AI quietly prevents some of the most expensive failures.

The honest limitation is that toxicity is fiendishly hard to predict reliably. Biology is complex, and a model trained on past drugs can miss novel dangers. False confidence here is risky, since a missed toxicity signal can harm patients. Teams therefore treat ADMET predictions as filters, not final safety verdicts. Regulators still require real animal and human safety data before approval. The critical review of real-world AI outcomes stresses that early-stage speed does not guarantee clinical safety. AI narrows the field of candidates, but the clinic remains the final judge.

The cost of a missed safety signal explains why teams stay cautious here. A toxic surprise late in development can erase years of work and huge budgets. Models help by flagging the riskiest molecules among thousands of candidates early. They cannot, though, capture every rare or long-term effect that only appears in people. Regulators therefore still demand staged animal and human safety studies before any approval. The honest framing treats these predictions as a first filter, not a final verdict. Used that way, safety models quietly remove many doomed candidates before they reach a patient.

Repurposing Existing Drugs With AI

Shifting from new molecules, AI also finds fresh uses for drugs that already exist. Approved drugs already cleared safety testing, so repurposing one skips years of early risk. Models compare disease signatures against the known effects of thousands of existing medicines. A strong match suggests an old drug might treat a new condition. Drug repurposing with AI can move a candidate toward patients far faster because its safety profile is already known. This approach drew intense interest during global disease outbreaks when speed mattered most. It offers a lower-cost path to treatments for rare and neglected diseases.

The appeal of repurposing is speed and lower cost, but the science stays demanding. A drug safe at one dose for one disease may fail at the dose a new disease needs. Many AI-suggested repurposing hits still collapse in proper clinical testing. The signals that look strong in data can reflect noise rather than real biology. Teams must validate every repurposing idea in cells, animals, and trials. Research on new treatments for incurable diseases shows both the promise and the caution. Repurposing widens the search, yet it does not bypass the need for evidence.

Repurposing also reshapes the economics of treating small patient populations. Rare diseases often lack the market size to justify a billion-dollar program. An existing and already-manufactured drug can lower that financial barrier dramatically. AI can scan for these matches across the entire approved pharmacopeia at once. That breadth surfaces candidates a human team would never have time to check. The result is a practical bridge between data-rich models and patients who have few options. Beyond molecules, AI is now reaching into how trials themselves are run.

AI in Clinical Trial Design and Patient Selection

Beyond the molecule itself, AI now influences the clinical trials that decide a drug’s fate. Trials are the most expensive part of development and the most common place to fail. Choosing the wrong patients can sink a drug that actually works in the right ones. AI helps design smarter trials by identifying patients most likely to respond based on their biology. Models analyze genetic and clinical data to define sharper enrollment criteria. This precision can shrink trial size, cut cost, and speed clear answers. It also reduces the number of patients exposed to a drug that may not help them.

The same models can predict where a trial might stumble before it starts. They forecast enrollment bottlenecks, dropout risk, and likely site performance. AI also mines real-world health records to find eligible patients faster. These gains matter because a delayed trial burns money every single day. Yet biased or incomplete data can steer trials toward narrow, unrepresentative groups. Work on predicting cancer drug response shows how patient selection can sharpen outcomes. Trial design gains from AI, but only when the underlying data is fair and complete.

The Data and Infrastructure Behind AI Drug Discovery

Given the scale of these models, a vast and expensive machinery of data and computing sits beneath them. This work is only as good as the biological data it learns from. High-quality, well-labeled data is the scarce resource that decides which programs actually succeed. Companies invest heavily in generating their own experimental data at industrial scale. Recursion, for example, built automated labs that produce enormous proprietary datasets. That data advantage, not just clever algorithms, often separates leaders from followers. Without it, even the best model learns from noise and produces unreliable predictions.

Generating that data requires robotics, cloud computing, and serious capital. Self-driving labs run thousands of experiments with minimal human intervention. Each experiment feeds the models that design the next round of molecules. This fusion of wet-lab automation and computing is the real engine of the field. Sharing and standardizing data, as discussed in work on artificial intelligence in healthcare, lowers the barrier for smaller teams. Public databases and partnerships help spread access beyond the largest players. Even with great data and tools, the field still runs into hard, unsolved limits.

Risks and Hard Limits of AI in Drug Discovery

Stepping back from the wins, the field carries real risks worth stating plainly. The biggest is the clinical translation gap, the chasm between fast discovery and slow clinical proof. AI can accelerate the design of a molecule, but it cannot guarantee that molecule will work in patients. Exscientia’s DSP-1181, the first AI-designed drug in trials, was discontinued after Phase I. BenevolentAI’s BEN-2293 met safety endpoints yet missed efficacy in atopic dermatitis. These setbacks, catalogued in the critical review of AI discovery outcomes, are sobering. Speed in discovery does not erase the stubborn biology of disease.

Data problems sit close behind the translation gap as a core risk. Models trained on biased or thin data produce confident but misleading predictions. A molecule that looks perfect in silico can fail the moment it meets messy biology. Models can also hallucinate molecules that are toxic or impossible to synthesize. Validation in the lab remains expensive, slow, and absolutely unavoidable. Even market leaders feel this, as Recursion’s 2025 pipeline cuts showed when early data disappointed. The report on Recursion trimming its pipeline made the limits visible.

There is also a quieter risk of overpromising and inflated expectations. Headlines about AI curing disease can outrun the slow reality of clinical science. That hype distorts investment and can erode public trust when timelines slip. Demis Hassabis himself pushed Isomorphic’s first trials from 2025 to late 2026. Honest framing matters, because patients and investors deserve realistic timelines. The technology is genuinely powerful, yet it is one tool inside a long process. Keeping expectations grounded is part of using AI responsibly, which leads directly into ethics.

Concentration of capability is a quieter risk that also deserves attention. A handful of well-funded labs hold the best data, the strongest models, and the deepest talent. That imbalance could narrow which diseases attract serious investment and effort. Conditions with small patient markets may still be overlooked despite the new tools. Open data and shared infrastructure can spread capability more evenly across the field. Policy choices, not just technology, will decide how widely the benefits actually reach. The field works best when many teams, not just a few giants, can contribute.

Ethics, Bias, and Trust in AI-Discovered Medicines

Building on those risks, ethics shapes whether AI medicines earn public trust. Bias in training data is the first concern, because biased data yields biased drugs. If datasets underrepresent some populations, models may design drugs that serve them poorly. A medicine optimized on narrow data can widen, rather than close, existing health gaps. Diverse, representative biological data is therefore an ethical requirement, not a technical nicety. Transparency matters too, since clinicians must understand why a model chose a candidate. Black-box predictions are hard to trust when human lives are at stake.

A darker ethical concern is the dual-use nature of generative chemistry. The same models that design medicines could, in principle, design harmful molecules. Researchers have shown that flipping a model’s objective can generate toxic compounds. Responsible labs build guardrails, access controls, and oversight to prevent misuse. Accountability is the final piece, since someone must own the outcomes of AI decisions. Broader debates in ethical concerns in AI healthcare apply directly to drug design. Trust is earned through fairness, transparency, and clear human responsibility for every choice.

Regulation and Intellectual Property Questions

Beyond ethics, regulation and ownership raise questions the law is still answering. Regulators must decide how to evaluate drugs designed largely by algorithms. The core safety and efficacy standards stay the same, but the evidence trail changes. No regulator yet approves a drug because an AI designed it, the molecule must still pass full clinical trials. Agencies are now drafting guidance on how AI evidence fits the approval process. Companies need clarity to plan multi-year programs around shifting expectations. The piece on FDA approval and regulation of AI healthcare tools traces this evolving landscape.

Intellectual property is the second open question with real money at stake. If an AI designs a molecule, who owns the resulting patent and invention. Patent systems were built around human inventors, not generative models. Unclear ownership can chill the investment that funds expensive clinical trials. Regulatory and filing workflows are themselves being reshaped, as covered in work on AI for regulatory filings and pharma factories. Courts and patent offices are only beginning to address these disputes. Until the rules settle, legal uncertainty will shadow even the most promising programs.

Data governance adds a third layer of complexity to the legal picture. Patient data that trains these models is sensitive and tightly regulated across regions. Companies must prove they handled consent, privacy, and security correctly at every step. A weak data trail can stall an otherwise promising approval submission badly. Clear documentation of how a model reached its conclusions is becoming an expectation. Teams that build that paper trail early will move through review far more smoothly. Good governance is now a competitive advantage, not just a box to tick.

The Future of AI in Drug Discovery

Looking ahead, the field is moving toward greater autonomy and integration. The next wave is agentic AI, systems that plan and run discovery campaigns with less human steering. Self-driving labs that design, test, and redesign molecules in a continuous loop are the clear direction of travel. Foundation models trained on biology promise to generalize across many diseases at once. Isomorphic Labs aims to start its first AI-designed-drug trials by the end of 2026. Its roughly USD 3 billion in deals with Novartis and Eli Lilly signals serious industry conviction. The field is shifting from isolated tools to fully integrated discovery engines.

Money and validation are flowing toward this vision at a striking pace. Isomorphic raised USD 600 million in March 2025 to scale its drug design engine. The plans for those trials are detailed in coverage of Isomorphic preparing AI-designed drug trials. The 2025 rentosertib readout gave the whole field its first real clinical proof point. Analysts expect the next decade to bring the first drugs discovered end to end by AI. Trends explored in future trends in AI-powered healthcare point the same way. Success now depends less on algorithms and more on patient data, trials, and trust.

How to Implement an AI Drug Discovery Project

In practice, a small team can begin an AI-driven discovery project using mostly open tools and public data. The sequence below is a realistic starting path, not a shortcut to a finished and approved drug. Each of the 5 steps pairs one concrete action with the scientific judgment it still demands. The goal is to compress the early stages while keeping every result grounded in real evidence.

Step 1 – Define the disease and target hypothesis

Start by naming the 1 disease and the biological target you believe drives it. A sharp hypothesis focuses every later modeling choice and saves weeks of wasted compute. Gather the genetic, expression, and literature evidence that supports your chosen target first. Write down clearly why this target is druggable and what success would actually look like. Review the roughly 90 percent of programs that fail so you avoid a known dead end. A target picked on weak evidence cannot be rescued by any amount of clever modeling later. This framing step costs almost nothing, yet it shapes the entire multi-year project ahead.

Step 2 – Assemble and clean your data

Collect public datasets on your target, related proteins, and at least a few thousand known compounds. Standardize every molecule format so your tools can read them without silent parsing errors. Clean, deduplicate, and validate each record before you train anything on the data. Poor data quality is the single most common reason that early projects quietly fail. Aim for 1 trustworthy, well-labeled table rather than 10 messy ones stitched together. Public resources like ChEMBL and shared molecular datasets give a strong, free starting point. A simple validation script catches malformed structures before they waste hours of model time.

Step 3 – Get a reliable protein structure

Obtain a 3 dimensional structure for your target protein before you design any molecules. Check the Protein Data Bank first for an experimental structure of that exact target. When none of the structures exist, predict one using a model such as AlphaFold. A good structure lets you design and dock molecules against a real binding pocket. Always confirm the per-residue confidence scores before you trust any region of the model. Treat a low-confidence loop as a hypothesis, not as settled fact about the protein. This 1 step turns an invisible target into something your tools can actually work against.

Step 4 – Run virtual screening or generative design

With a structure ready, screen a compound library or generate brand new molecules. Docking tools score how each of the millions of candidates fits the target pocket. Generative tools instead propose fresh molecules shaped directly for that 1 binding site. Rank the output by predicted binding, drug-likeness, and synthetic accessibility together. Always filter generated molecules for synthesizability before celebrating a strong docking score. Keep only a short list of 20 to 50 credible candidates for the next stage. This narrowing is exactly where AI saves the most laboratory time and money.

Step 5 – Predict safety and validate in the lab

Run ADMET and toxicity predictions on your shortlisted candidates as the next filter. Discard molecules flagged for likely toxicity, poor absorption, or fast clearance early. Treat these predictions as filters, never as a substitute for real laboratory testing. Synthesize the surviving 5 or 10 molecules and test them in cells and animals. Feed every experimental result back to retrain and sharpen your models each round. This lab-in-the-loop cycle is the real heart of any credible discovery program. Expect many rounds, since even strong candidates can fail once they meet messy biology.

Key Insights on AI in Drug Discovery

• Insilico’s rentosertib raised lung function by a mean 98.4 mL in the Phase IIa readout, the first AI-designed-drug proof point.
• One approved medicine is often priced near 2.6 billion dollars, yet a 2025 cost review cites a far lower 708 million figure.
• About 90 percent of trial candidates fail, a rate the analysis of clinical failure ties to gaps that AI cannot fully close.
• Insilico reached Phase II in roughly 18 months, a pace the critical review credits to its PandaOmics and Chemistry42 engines.
• Isomorphic Labs raised 600 million dollars in 2025, and its trial plans now target a late-2026 clinical start.
• Recursion paused or cut programs in 2025 after weak data, a retrenchment the pipeline coverage ties to six active projects.
• AlphaFold 3 models proteins with DNA, RNA, and small molecules, a leap the DeepMind announcement spans across the Protein Data Bank.
• The AI drug discovery market sat near 6 billion dollars in 2025, with the 2034 forecast projecting strong growth.

Taken together, these numbers tell a story of real progress shadowed by stubborn limits. The 2025 rentosertib readout proves an AI-designed drug can help patients, not just impress in data. Speed records and billion-dollar deals show genuine industry conviction behind the approach. Yet the 90 percent failure rate and Recursion’s cuts prove biology still resists prediction. The honest read is that AI compresses discovery while the clinic stays the final, unforgiving judge. Progress now hinges on data quality, trial design, and trust as much as on better models.

How AI Compares to Traditional Drug Discovery

Stepping back from the science, the contrast between the two approaches comes into sharp focus. Traditional methods lean on human intuition and slow, serial experiments at every turn. The AI-driven path front-loads prediction to narrow the options before costly lab work begins. The table below sets the two side by side across the dimensions that matter most. Both paths still converge on the same clinical trials that ultimately decide success.

Dimension	Traditional Drug Discovery	AI Drug Discovery
Target identification	Manual hypothesis from focused lab studies	Models rank targets across genomic and literature data at scale
Molecule design	Chemist intuition exploring familiar chemistry	Generative models invent novel structures for a target pocket
Early-stage speed	Several years to a clinical candidate	As fast as roughly 18 months in leading programs
Screening scale	Thousands of physical compounds tested	Billions of virtual compounds scored computationally
Cost driver	Serial lab experiments and high attrition	Data generation, compute, and validation infrastructure
Clinical success	About 10 percent of trial entrants approved	Same trials and odds once a candidate reaches humans
Key risk	Slow, costly dead ends discovered late	Confident wrong predictions from biased or thin data
Human role	Drives every design and decision directly	Validates, filters, and interprets model suggestions

AI Drug Discovery in Practice Across the Industry

Insilico Medicine’s Rentosertib for Lung Fibrosis

Among the clearest wins, Insilico Medicine deployed its Pharma.AI platform to find a target and design a drug for idiopathic pulmonary fibrosis. PandaOmics prioritized the protein TNIK, and Chemistry42 generated the inhibitor that became rentosertib. In a Phase IIa trial, the 60 mg arm drove an increase of 98.4 mL in lung function versus a 20.3 mL decline on placebo, per Insilico’s readout. The program reached Phase II in roughly eighteen months, saving years against a normal timeline. The limitation is real, since Phase IIa is small and a larger trial must still confirm efficacy and long-term safety. Even this landmark remains an early signal, not a finished and approved medicine. It is the clearest proof so far that AI-designed drugs can genuinely help patients.

DeepMind’s AlphaFold Reshaping Structural Biology

DeepMind built AlphaFold to predict protein structures and then released a database of hundreds of millions of them. Researchers worldwide now pull accurate shapes in seconds for targets that once required years of lab work, an enormous time saving. AlphaFold 3 extended this to model proteins binding DNA, RNA, and small molecules together, a capability the DeepMind AlphaFold 3 announcement details across most Protein Data Bank molecule types. The outcome shows up as a sharp increase in productivity, with millions of researchers using the predictions for drug and enzyme work. The limitation is that predicted structures are static snapshots, while real proteins flex and shift constantly. Critical targets still require experimental validation by methods like cryo-electron microscopy. AlphaFold accelerates the start of design rather than replacing the wet lab entirely.

Recursion’s Industrial-Scale Discovery Engine

Recursion built automated laboratories that run millions of experiments to generate proprietary biological data at industrial scale. The company trained models on that data to map cell biology and propose drug candidates across many diseases. After merging with Exscientia in 2024, it pushed a broad pipeline of programs into the clinic. The reality check came in 2025, when weak early data forced a reduction to roughly six active programs, a retrenchment the report on its pipeline trim documents in detail. That decrease of several programs shows scale and data alone do not guarantee clinical wins. The limitation is the same translation gap that still humbles the whole field. Recursion’s story is both real ambition and a sober lesson in managing expectations.

Lessons From Real AI Drug Discovery Programs

Case Study: Isomorphic Labs and the Bet on AlphaFold

Among the most watched bets, Isomorphic Labs faced a clear problem, that even powerful structure prediction does not by itself produce approved drugs. The DeepMind spinout set out to turn AlphaFold’s modeling into a full drug design engine. It built that engine and signed research collaborations with Novartis and Eli Lilly worth roughly USD 3 billion in potential value. In March 2025 it raised USD 600 million to scale the platform and staff up for clinical work. The measurable signal of conviction is large, with billions committed before a single drug enters trials. The limitation is timing, since CEO Demis Hassabis pushed the first human trials from 2025 to late 2026, a slip the coverage of its trial preparations records. The lesson is that capital and technology still bend to the slow pace of clinical science. Even the best-funded AI lab cannot skip the trials that decide a drug’s fate.

Case Study: Exscientia’s DSP-1181 and the First Trial

Exscientia confronted the problem of proving that an AI-designed molecule could even reach human testing. The company developed and deployed DSP-1181, a candidate for obsessive-compulsive disorder, using its design platform. In 2020 it became one of the first AI-designed small molecules to enter a Phase 1 clinical trial. The molecule cleared Phase 1 with a favorable safety profile, a genuine milestone that drew millions in investor attention. The hard impact came next, because development was discontinued and the drug never advanced to Phase 2. That outcome, documented in the critical review of AI small-molecule outcomes, exposed the translation gap early. The limitation was stark, since speed and safety in Phase 1 did not translate into clinical success. The lesson is that entering trials fast is not the same as winning them.

Case Study: BenevolentAI’s BEN-2293 in Atopic Dermatitis

BenevolentAI faced the challenge of solving eczema with a candidate its knowledge-graph platform helped identify. The company developed and advanced BEN-2293, a topical treatment for atopic dermatitis, into a controlled clinical trial. The drug met its primary safety endpoints, showing it was well tolerated in patients. The measurable disappointment was efficacy, because the trial showed no statistically significant benefit over placebo. That setback, noted in the review of real-world AI drug outcomes, triggered a painful restructuring that cut a large percent of staff. The company closed offices and pivoted back to early-stage discovery partnerships to survive. The limitation is the recurring theme, that a model can surface a plausible target yet still miss in the clinic. The lesson is that AI sharpens discovery while late-stage biology remains the ultimate test.

Frequently Asked Questions About AI in Drug Discovery