The standard pitch for AI in drug discovery goes like this: the old way took a decade and billions of dollars, the new way takes months, and the revolution is imminent. I have watched a version of this pitch at every conference I have attended for five years, and it misdescribes the problem badly enough to be worth unpacking properly.
The bottleneck in pharmaceutical development is not finding molecules. It never was.
Where Drugs Actually Fail
Roughly nine out of ten compounds that enter human trials never reach approval. Most of those failures happen in Phase II and Phase III — not because the molecule could not be found or synthesised, but because it did not work in patients, or it worked and was unsafe, or the biological hypothesis it was built on turned out to be wrong.
A widely cited industry estimate puts the capitalised cost of bringing one approved drug to market at around 2.6 billion dollars, a figure that is disputed in both directions but is uncontroversially large. The reason it is large is that it includes the cost of all the failures. You pay for the nine that died to get the one that lived.
This matters for evaluating AI's contribution, because the discovery stage — target identification through lead optimisation — is a comparatively small fraction of both the time and the money. Compress it to zero and the overall timeline shortens by a few years at most. Everything downstream is governed by how long it takes to dose humans safely and observe what happens, and no amount of computation changes that.
So when you see a claim that AI has cut development from fifteen years to three, check which stages are being measured. Almost always it is time from project start to a nominated preclinical candidate, which is real and useful and not the same thing as bringing a drug to market.
What AI Genuinely Changed
With that framing in place, the actual advances are substantial.
Protein structure prediction is the clearest win in computational biology in my lifetime. AlphaFold2 solved a problem that had resisted fifty years of effort, and the associated database released predicted structures for essentially the entire known protein universe. Structures that would each have taken a crystallographer months or years are now a lookup.
It is worth being precise about what this does and does not give you. AlphaFold predicts a static folded structure with a confidence score, and it does so extremely well for ordered domains. It is less informative about conformational dynamics, about disordered regions, and about how a small molecule will actually bind and with what affinity — which is the thing a medicinal chemist needs. Later systems extend to complexes and ligands with genuine improvements. But a predicted structure is the start of a drug discovery problem, not the end of one, and the field's occasional suggestion that structure prediction solved drug design is not a view held by anyone who does the chemistry.
Virtual screening at a scale that was previously impossible. Physical high-throughput screening runs perhaps hundreds of thousands to low millions of compounds and is expensive. Machine-learned scoring functions let you triage libraries of billions of enumerable molecules computationally, then test only the interesting fraction in the lab. The models are imperfect — enrichment over random is what you get, not certainty — but reordering a billion-compound library so that the plausible candidates come first is a real capability that did not exist a decade ago.
Generative molecular design. Rather than searching an existing library, train a model on chemical space and have it propose structures satisfying multiple constraints at once: binding to the target, synthesisable, soluble, not obviously toxic, not infringing someone's patent. Multi-objective optimisation over chemistry is well suited to generative models, and the molecules produced are novel and chemically sensible. Whether they are good drugs is decided in animals and then in humans, which brings us back to the bottleneck.
Synthesis planning. Retrosynthesis — working backwards from a target molecule to purchasable starting materials — is a search problem with a huge branching factor and well-characterised chemistry rules. This is a domain where machine learning has been quietly and undramatically useful for years.
What the Clinic Has Shown So Far
This is the section most articles on this subject skip, and it is the only one that answers the question.
As things stand, no drug whose discovery is attributed to AI has completed the full path to regulatory approval. Several are in clinical trials, and the results so far are mixed in exactly the way you would expect from early-stage drug development.
The first widely publicised AI-originated candidate to enter human trials, a compound for obsessive-compulsive disorder developed in a pharma partnership, was discontinued during Phase I. That is unremarkable — most Phase I compounds are discontinued — but it received far less coverage than its entry into trials had, which tells you something about how this field is reported.
On the more encouraging side, Insilico Medicine's candidate for idiopathic pulmonary fibrosis, where both the target and the molecule came out of their generative platform, reported positive Phase 2a results. This is a meaningful milestone: an AI-nominated target and an AI-designed molecule showing a signal in patients. It is also a small, early trial in a disease with few options, and Phase 2a signals fail to replicate in Phase III with depressing regularity. Treat it as genuine evidence that the approach can produce clinical candidates, and not yet as evidence that it produces better ones.
The commercial side has been correspondingly sober. Several of the best-known AI-first drug discovery companies have consolidated, narrowed their pipelines, or shifted towards partnership models where large pharmaceutical companies carry the clinical risk. Recursion and Exscientia, two of the most prominent, merged in 2024. This is what an industry looks like when the technology is real and the timelines are longer than the funding assumed.
The Problems That Do Not Go Away
Data is the binding constraint, and not in the way people assume. There is a lot of chemical and biological data and most of it is unusable for training: measured under incompatible assay conditions, skewed towards well-studied targets, and — critically — missing the negatives. Failed experiments are rarely published. A model trained on the literature learns from a heavily filtered record of what worked, which is the textbook setup for overconfidence.
In-silico to in-vivo translation remains the hard part. A model can rank compounds by predicted binding affinity quite well. Predicting what a compound does in a living organism — metabolism, off-target effects, tissue distribution, the immune response, toxicity that emerges only after chronic dosing — is a vastly harder problem with far less training data, because the only source of that data is expensive animal and human experiments.
Target validation is where most of the value would be, and it is the least tractable. If AI could tell you reliably which biological targets are worth pursuing, it would eliminate the most expensive category of failure: the beautifully engineered molecule that hits its target perfectly in a disease where hitting that target does not help. Work on this using genomic and clinical evidence is promising and early. It is also, unlike molecular generation, hard to demonstrate in a press release, which is roughly why you hear less about it.
How I Would Read the News
When you encounter a claim about AI in drug discovery, three questions separate substance from marketing.
Which stage does the speedup apply to — discovery, or the whole development path? Almost always discovery, which is not the bottleneck.
Is the endpoint computational or clinical? A predicted structure, a docking score, a generated molecule, and a patient outcome are four wildly different levels of evidence, and they get reported in the same sentence constantly.
What happened to the ones that failed? Pipelines get announced loudly and discontinued quietly. The base rate for clinical attrition applies to AI-designed compounds too, and so far there is no published evidence that it applies to them any less.
None of this is an argument that the field is hype. The tools are genuinely powerful and I use versions of them in my own work. But the honest statement is that AI has substantially improved one part of a long pipeline whose failure modes live mostly elsewhere. That is a real contribution. It is not a revolution yet, and calling it one before a single approval has cleared is how a promising field spends credibility it will need later.
Tags
Taresh Sharan
support@sharaninitiatives.com