AI Drug Discovery: what are the biggest unsolved problems?

Is AI drug discovery actually working today?

AI drug discovery is working today, but mainly at finding and designing promising drug candidates; we still cannot say that it reliably produces more successful medicines.

There is enough clinical evidence now to move beyond the old debate about whether AI can produce real drugs at all. An analysis published in Drug Discovery Today found that AI-discovered molecules had an estimated 80–90% success rate in Phase I, compared with historical industry ranges closer to 40–65%. That is unusually strong. Phase I largely tests whether a drug can enter humans with acceptable initial safety and pharmacokinetics, so AI appears increasingly good at producing compounds that behave like plausible medicines.

The advantage becomes much less obvious once efficacy enters the picture. The same analysis estimated Phase II success at around 40%, based on a much smaller sample, which was broadly in line with historical industry performance.

Insilico Medicine's rentosertib is currently the clearest example of an AI-originated program going further. AI was used both to identify TNIK as a target for idiopathic pulmonary fibrosis and to design the molecule. A randomized Phase IIa trial involving 71 patients found a mean 98.4 mL improvement in forced vital capacity after 12 weeks in the highest-dose group. Rentosertib has since moved into Phase III.

So AI drug discovery has crossed an important line. Companies are no longer showing only virtual molecules, mouse studies or computational benchmarks. Real AI-originated drugs are reaching patients and sometimes producing encouraging efficacy data. What we still lack is enough late-stage clinical evidence to know whether those successes happen more often than they do in conventional drug discovery.

What is the hardest problem AI drug discovery still cannot solve?

The hardest problem in AI drug discovery is choosing biological interventions that will actually change a human disease.

The pharmaceutical industry has never lacked chemical ideas. Its bigger problem is that most plausible biological hypotheses eventually turn out to be wrong, incomplete or clinically irrelevant. Roughly nine out of ten drugs entering clinical development still fail before approval, with insufficient efficacy and safety responsible for much of that attrition.

Human genetics gives us a useful measure of how much better target selection could matter. A large Nature analysis covering nearly 30,000 target-indication pairs found that mechanisms supported by human genetic evidence had about 2.6 times the probability of clinical success of unsupported mechanisms.

AI can search gene expression, genetics, scientific literature, proteins, single-cell data and disease networks at a scale no research team could handle manually. But finding an association is much easier than proving that changing it will help a patient. A gene may rise because it causes the disease, because the body is trying to fight the disease, or simply because another process upstream has changed.

The direction of treatment creates another layer of uncertainty. Knowing that a protein is involved in Alzheimer's disease does not tell us whether to inhibit it, activate it, alter it only in certain cells or intervene only at a particular disease stage.

Target discovery is getting faster. Biological truth remains stubbornly slow.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Why is human biology still so difficult for AI to predict?

Human biology remains difficult for AI because the answer changes with the context: the same drug can behave differently across cells, tissues, patients, doses and disease stages.

That is a very different learning problem from recognizing objects in images or predicting the next word in a sentence. A compound may inhibit a purified protein beautifully and then fail inside a cell. It may work in one cell line and disappear in primary human cells. A treatment that produces a striking effect in mice may do very little in people.

A recent Nature Reviews Drug Discovery assessment points to this conditional nature of life-science data as one reason AI has translated more slowly into drug R&D than into fields such as speech or image recognition. Biological measurements are full of dependencies that models can easily mistake for general rules.

The training problem is difficult too. Language models can learn from enormous amounts of existing human text. For many of the questions that matter most in drug discovery, the answer simply has not been generated yet. Researchers have to run the experiment, dose the animal or eventually treat patients before anyone knows what happens.

Drug discovery therefore has plenty of data in the broad sense—sequences, images, structures, assays and papers—but surprisingly little clean evidence showing what a new intervention will do in a real human disease.

Can AI drug discovery models trust their training data?

AI drug discovery has a serious data-quality problem because large biological datasets often combine experiments that were never designed to be compared directly.

Databases such as ChEMBL and PubChem contain huge quantities of chemical and biological measurements. The problem is that those measurements come from different laboratories, protocols, concentrations, assay technologies and research objectives. Two numbers sitting neatly beside each other in a database may represent rather different experiments.

Negative evidence is particularly weak. Scientists publish interesting positive findings more readily than compounds that did nothing. Pharmaceutical companies also hold decades of failed assays and abandoned programs that rarely become public. Yet those failures can be exactly what a model needs to learn where chemical or biological ideas stop working.

Reviews of public protein-ligand datasets have repeatedly found fragmented measurements, inconsistent experimental protocols and shortages of reliable inactive-compound data. More rows therefore do not automatically create a better training set.

This helps explain why proprietary experimental platforms have become strategically important. Recursion has generated large cellular imaging datasets rather than relying entirely on public biology. Other AI drug companies are combining external data with internal synthesis, screening or disease-model experiments.

These days, owning the feedback loop between prediction and experiment can be more valuable than simply owning another large model.

Can AI predict a drug when it has never seen anything similar before?

AI still struggles when a drug, target or biological system falls far outside its training experience, which is exactly where many valuable discoveries begin.

Random train-test splits can make this weakness easy to miss. Closely related molecules or proteins may appear in both the training and test sets, allowing an algorithm to achieve an impressive score by learning patterns that are already familiar.

Real drug discovery is harsher. Researchers want a model to predict what will happen with a new chemical scaffold, an unusual protein pocket or a target for which little experimental data exist.

A recent Nature Reviews Drug Discovery critique makes the problem especially clear: chemical space is so vast that genuinely novel drug predictions routinely become out-of-distribution predictions. Excellent performance near familiar chemistry can therefore tell us surprisingly little about performance on the molecule a medicinal chemist actually wants to invent next.

Prospective testing is the cleanest answer. Freeze the model, give it an unseen problem, record its predictions before the experiments happen and then see whether the molecules work.

Until AI drug discovery systems perform well repeatedly under those conditions, benchmark accuracy will continue to look more reassuring than real discovery feels.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Have AlphaFold and protein AI basically solved the structure problem?

AlphaFold has transformed structural biology, but protein AI still cannot tell drug hunters everything they need to know about how a molecule will behave around a target.

AlphaFold made useful structural predictions available for an enormous number of proteins that previously lacked experimentally determined structures. AlphaFold 3 went further by modeling interactions involving proteins, DNA, RNA, ligands and other molecular components.

Drug discovery adds complications that a static structure cannot capture. Proteins move. Binding pockets appear and disappear. Water molecules and ions change interactions. A ligand itself can push a protein into a different conformation. Membranes, post-translational modifications and other molecular partners can alter what happens inside a cell.

More importantly, medicinal chemists rarely need only a yes-or-no prediction of binding. They may have 300 related compounds and need to know whether one binds at 10 nanomolar while another binds at 500 nanomolar, whether that difference survives in a cell and whether an apparently stronger binder creates another liability elsewhere.

Recent evaluations of AlphaFold 3 show impressive protein-ligand structure prediction while still finding limitations around major conformational changes and affinity ranking.

Protein AI has removed a huge structural-information bottleneck. The harder problem now is predicting the dynamic molecular physics that separates an interesting structure from a good drug.

Can generative AI make drugs that medicinal chemists would actually want?

Generative AI can already make credible molecules, but designing a molecule that looks good computationally is easier than designing one that survives an entire drug program.

A useful compound has to satisfy several constraints at once. It needs potency against the target, selectivity, solubility, permeability, appropriate metabolism and enough exposure in the right tissue. It also needs acceptable safety, chemical stability and a realistic manufacturing route.

Those properties fight each other constantly. Increasing lipophilicity can improve membrane penetration while hurting solubility and increasing off-target binding. A chemical change that improves potency may worsen metabolism. Fixing one toxicity issue can damage exposure.

That is why the strongest generative systems today are increasingly built around multi-parameter optimization and repeated experimental feedback rather than unrestricted molecular invention.

The early clinical record also gives us a useful clue. AI-discovered molecules have so far performed unusually well in Phase I in the first systematic analysis of the field. That result is still based on a limited sample, but it is consistent with AI becoming genuinely useful at producing compounds with strong drug-like properties.

The remaining test is whether these systems can repeatedly produce better compounds with fewer design cycles than strong medicinal-chemistry teams would have produced otherwise.

Can AI-generated molecules actually be manufactured?

AI-generated molecules are still sometimes much easier to draw than to make, so synthesizability remains a practical constraint on generative drug design.

A molecular generator can roam through chemical space without worrying much about whether a chemist can buy the starting materials, control the reaction, purify the product or manufacture kilograms of it later. That freedom can produce clever structures that score well computationally and become painful as soon as they reach the lab.

Researchers have responded by building synthetic reality into the generation process. Some systems restrict molecules to known building blocks and reaction types. Others include synthetic-accessibility scores or run retrosynthesis software before candidates are selected.

Each approach has drawbacks. Simple accessibility scores can miss real synthetic problems. Full retrosynthetic planning is much heavier computationally. Restricting models to familiar reactions makes the molecules easier to produce but narrows the chemical space that AI can explore.

For drug companies, extreme novelty is rarely the goal anyway. A more useful AI molecule is one that explores chemistry humans might have missed while remaining practical enough to synthesize, test and eventually manufacture.

What an AI model may optimize	What a drug program actually needs
Predicted potency	Reproducible experimental potency
Molecular novelty	Useful and defensible novelty
Synthetic-accessibility score	A practical synthesis route
Individual property scores	Several acceptable properties at once
A virtual molecule	A stable, scalable physical compound

Why can't AI predict drug toxicity reliably yet?

AI still cannot predict drug toxicity reliably because safety problems can emerge through too many different mechanisms, tissues and time scales for one model to learn cleanly.

Some risks are becoming easier to screen computationally. Models can estimate hERG liability, metabolic stability, reactive chemistry and interactions with known off-target proteins. Medicinal chemists already use these predictions to remove obvious liabilities earlier.

The hardest safety failures look very different. A toxic metabolite may appear only after prolonged exposure. Weak interactions with several proteins may combine into a clinically relevant problem. Liver injury can be idiosyncratic. Immune toxicity depends heavily on patient biology. Chronic effects may appear months after treatment begins.

The available data are biased too. Dangerous molecules discovered early never reach clinical trials, while rare late toxicities may be represented by only a handful of cases. Models consequently have fewer useful examples precisely where prediction could save the most money and patients.

The likely path forward combines chemistry models with human genetics, cellular imaging, organoids, organ-on-chip systems and clinical data. A molecular structure alone rarely contains enough information to predict everything a human body might do with it.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Can AI predict whether a drug will actually work in patients?

AI still cannot reliably predict whether a drug will work in patients, and that is where the field's biggest claims are currently being tested.

A drug has to survive a long chain of conditions. The target must genuinely affect the disease. Researchers must modulate it in the right direction. The molecule has to reach the relevant tissue at sufficient concentration. The effect must last long enough to help without producing unacceptable toxicity. The patients enrolled in a trial must also have biology that still depends on that mechanism.

One weak link can kill the whole program.

Clinical results make the problem unusually visible. The first systematic review of AI-discovered compounds reported around 80–90% Phase I success but only about 40% Phase II success, although the Phase II sample was small. Once researchers have to show that the biological intervention actually helps patients, much of the apparent advantage disappears.

Rentosertib is therefore worth watching closely. Its randomized Phase IIa result was good enough to move the idiopathic pulmonary fibrosis program into Phase III, making it one of the field's most advanced tests of an AI-discovered target and molecule. More recently, researchers analyzing the same clinical program also reported changes in proteomic aging-clock measures, an interesting secondary finding but nowhere near proof that the drug slows human aging.

For now, the important result remains the lung-disease trial. If several AI-originated programs eventually reproduce this kind of efficacy in larger Phase III trials, we will have much stronger evidence that AI is improving more than molecule design.

Why do drugs that work in cells and mice still fail in humans?

AI cannot fully fix the translation problem because many laboratory disease models reproduce only a small part of what actually happens inside a patient.

Cell lines are simplified biological systems. Mouse models are different species. Even sophisticated engineered disease models usually capture only selected mechanisms from conditions that may involve immune cells, aging, genetics, metabolism, environmental exposure and years of progression.

A treatment can therefore produce a perfectly reproducible result in an experimental model while telling us little about the people who will eventually receive it.

This is one reason human genetics has become so valuable for target discovery: genetic evidence comes from humans and can sometimes show whether lifelong changes in a biological pathway alter disease risk. Organoids, induced pluripotent stem cells, patient-derived tissues and richer phenotypic assays try to close the same gap from another direction.

AI can extract more information from these systems than researchers could manually. It can recognize subtle cellular phenotypes, combine different data types and identify patterns across thousands of perturbations. But better analysis cannot rescue an experiment that poorly represents the disease.

We still need better models of patients, not simply better models trained on laboratory data.

Is the wet lab becoming AI drug discovery's new bottleneck?

Yes. AI can generate hypotheses far faster than laboratories can test them, so experimental capacity is becoming one of the main constraints on AI drug discovery.

A generative model can propose thousands of molecules in minutes. Scientists still have to synthesize compounds, purify them, confirm their structures and run biological assays. Some experiments take hours, others take weeks, and animal or advanced disease-model studies can take much longer.

That imbalance is pushing AI drug companies toward automated design-make-test-learn loops. Recursion built much of its platform around high-throughput cellular experiments. Automated synthesis and screening systems can now feed results back into models that choose the next compounds to test. Researchers are also developing so-called self-driving laboratories in which software selects experiments and robotic systems carry them out.

The value here is speed. Optimization loops can increasingly run automatically when the objective is clear, while scientists still have to step in when an unexpected phenotype appears, an assay stops making biological sense or the whole target hypothesis needs to be reconsidered.

AI is therefore creating more demand for high-quality experiments. Once generating another hypothesis becomes cheap, the expensive part is finding out which hypothesis survives contact with biology.

Is AI making drug discovery cheaper yet?

AI is already cutting the cost and time of some discovery tasks, but there is still no convincing industry-wide evidence that it has dramatically reduced the cost of producing an approved medicine.

Virtual screening can replace some physical screening. Better predictions can reduce the number of compounds synthesized. Automated chemistry can shorten medicinal-chemistry cycles. A company that reaches a development candidate after making hundreds rather than thousands of molecules has saved real time and money.

Clinical failure overwhelms many of those savings. A drug that was discovered unusually cheaply can still burn tens or hundreds of millions of dollars if it fails after entering patients. Large clinical trials still need sites, investigators, manufacturing, monitoring and substantial numbers of participants.

The economics therefore depend heavily on whether AI eventually improves success rates, especially in Phase II and Phase III. Cutting early discovery cost by half would be useful. Preventing even a modest share of expensive late-stage failures could be far more valuable.

Pharma companies are already paying to find out. Isomorphic Labs' original collaborations with Eli Lilly and Novartis included $82.5 million in combined upfront payments and potential economics approaching $3 billion before royalties. Novartis later expanded its relationship with Isomorphic from three drug programs to as many as six, while Johnson & Johnson added another multi-target collaboration.

Those multibillion-dollar headline figures are mostly contingent milestones, so they should not be treated as money already earned. The more interesting evidence is the repeated behavior: large pharmaceutical companies are expanding some AI partnerships after getting an initial look at the technology.

That tells us AI discovery has become commercially credible. It still does not tell us how much the eventual approved drug will cost.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

How do we know AI actually deserves credit when a drug succeeds?

We often cannot tell how much credit AI deserves for a successful drug because modern programs combine algorithms, experimental platforms, medicinal chemists, existing biological knowledge and ordinary human judgment.

Imagine that an AI-assisted program reaches clinical development in two years rather than four. The model may have proposed better molecules. Automated chemistry may have compressed each design cycle. The target may simply have had unusually strong biology from the beginning. The company may also have prioritized the program more aggressively than comparable projects.

Success stories make this particularly difficult. A company has every reason to highlight the role of its platform when a candidate works, while a failed program can disappear quietly into ordinary biotech attrition.

The current generation of companies makes attribution even harder because the technologies are becoming integrated. Recursion's combination with Exscientia brought together large-scale phenotypic biology, computational chemistry and automated experimentation. Isomorphic Labs works directly with major pharmaceutical companies rather than developing every downstream step itself.

A better test would compare similar programs using different discovery workflows and measure the number of compounds made, experiments performed, months spent, money spent and candidates that eventually succeed.

Without those comparisons, “AI-discovered drug” remains a useful description but a weak measure of how much AI actually changed the outcome.

Are AI drug discovery benchmarks testing the wrong thing?

Many AI drug discovery benchmarks still reward good predictions on old data when the real job is making better decisions about experiments nobody has run yet.

This is one of the clearest conclusions from a recent Nature Reviews Drug Discovery assessment of the field. The authors argue that benchmarking needs to move away from validating models and toward testing whether those models improve actual decisions.

The distinction sounds small but changes the standard completely. A model can raise an accuracy score from 82% to 86% without altering a single compound that researchers choose to make. Similar molecules can leak across training and test sets. Teams can also tune algorithms against the same public benchmarks for years until performance on those datasets says little about genuinely unseen problems.

Prospective experiments are much harder to game. Give several systems the same new target or chemical series, lock the predictions before testing and then measure which approach finds useful compounds faster or with fewer experiments.

Today, there are still far too few independent, repeated prospective comparisons of that kind. This is one reason impressive AI drug-discovery papers continue to accumulate faster than convincing evidence of better clinical productivity.

Type of evidence	How convincing is it?
Better retrospective benchmark	Weak
Prediction on carefully separated unseen data	Useful
Successful prospective laboratory prediction	Strong
Repeated improvement across real drug programs	Very strong
Higher clinical success	Decisive

Can regulators actually trust AI in drug development?

Regulators already accept AI throughout drug development, but companies have to prove that each model is reliable enough for the decision it influences.

The FDA had received more than 500 drug and biological product submissions containing AI components between 2016 and 2023, covering nonclinical research, clinical development, manufacturing and post-marketing work. AI in regulated drug development is therefore already routine enough that the question has shifted from permission to proof.

The FDA's recent approach centers on context of use. A model that helps scientists choose exploratory laboratory experiments does not carry the same risk as one used to support a regulatory conclusion about efficacy, safety or product quality. Higher-stakes uses need stronger evidence that the model performs reliably for that specific purpose.

Fast-changing AI creates an awkward fit with pharmaceutical development. A model may be retrained, its data may change or a third-party provider may release a new version, while regulated research depends heavily on reproducibility and documentation.

Drug companies consequently need version control, traceable training and validation data, performance monitoring and clear limits around when a model should be trusted.

Regulation currently looks more like an engineering constraint than a fundamental blocker. Models can be used, but “the AI said so” will never replace evidence.

What would finally prove that AI has transformed drug discovery?

AI drug discovery will have genuinely transformed pharma when AI-originated programs succeed in patients more often, not when models generate molecules faster or win another benchmark.

We already have substantial proof at the technical level. AI can predict protein structures, search huge chemical spaces, analyze cellular images, generate molecules and help prioritize experiments.

We also have growing proof at the program level. Dozens of AI-associated candidates have reached clinical development, and some have advanced beyond Phase I. Rentosertib has now entered Phase III, which gives the field a particularly useful real-world test.

The missing evidence is statistical. As seen above, the first clinical analysis found unusually high Phase I success for AI-discovered molecules while Phase II performance looked much closer to historical norms. A recent Nature Reviews Drug Discovery review reached a similarly cautious conclusion after examining the broader field: technical progress has been extensive, yet evidence of clinically relevant impact remains limited.

That can change surprisingly quickly once enough programs mature. We would want dozens of comparable AI-originated candidates reaching later clinical stages, with outcomes adjusted for therapeutic area, modality, target novelty and strength of prior biological evidence. Otherwise, AI companies could look unusually successful simply because they chose easier chemistry or better-validated targets.

Clinical success is the scoreboard that eventually settles the argument.

What AI claims to improve	What would actually prove it
Better molecules	Prospective experimental superiority
Faster discovery	Shorter comparable drug programs
Better target selection	Higher efficacy success in Phase II
Fewer expensive mistakes	Higher Phase III and approval rates
Lower R&D cost	Lower cost per approved drug

If you want more recent data on this point, please see our latest AI in drug discovery market report.

So what are the biggest unsolved problems in AI drug discovery today?

The biggest unsolved problems in AI drug discovery today are understanding causal human biology, predicting unfamiliar chemistry and biology, anticipating toxicity and efficacy, and proving that better computational predictions translate into more successful drugs.

Molecule generation has moved surprisingly far. Protein structure prediction has moved even further. Automated laboratories are getting faster. Pharmaceutical companies are spending real money on AI partnerships, and an AI-discovered target and molecule has now reached Phase III.

The difficult part starts when a clean computational problem becomes messy human biology. We still struggle to know which disease mechanisms are causal, whether a target should be activated or inhibited, which experimental model actually represents patients and whether a compound that looks excellent before the clinic will produce a useful effect once people receive it.

Recent evidence makes the dividing line clearer than it was a few years ago. AI seems increasingly capable of improving the quality and speed of early drug discovery. We do not yet have enough evidence that it improves the probability of creating an effective medicine.

That is the real frontier now. If AI starts lifting Phase II and Phase III success across enough independent programs, drug discovery economics could change dramatically. Until then, AI has become a powerful discovery tool without solving the reason drug development has always been so difficult: biology keeps getting the final vote.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.