AI Drug Discovery: What It Has Actually Found

A new medicine takes somewhere around a decade to reach a pharmacy shelf. Most estimates put the cost of getting one there north of two billion dollars, and roughly nine out of ten candidates that reach human trials fail anyway.

That failure rate is the reason drug companies have poured money into AI drug discovery for a decade. The promise was always the same: find better candidates sooner, and stop wasting years on molecules that were never going to work.

For most of that decade the promise was all there was. Now there are results to look at, a supercomputer built specifically for the job, and a deal worth up to $2.75 billion riding on the idea. Here is what AI drug discovery genuinely does, what it has found, and the test it has not passed yet.

Key Takeaways

• AI does not invent medicines. It narrows an impossibly large search space down to a shortlist chemists can actually test
• One recent model screened 4.6 million compounds in hours to predict hydrogen positions, a detail that decides whether a molecule fits its target
• Eli Lilly built LillyPod, a pharmaceutical AI supercomputer, and signed a deal with Insilico Medicine worth up to $2.75 billion
• Around half of biotech firms adopting these tools report faster time-to-target, and 42% report better accuracy and hit rates
• None of it counts until an AI-derived drug clears Phase III, which is where the real verdict lands

What AI drug discovery actually does

The popular version of this story has a computer inventing a cure. That is not the work.

The number of small molecules that could theoretically be drug-like runs into the tens of billions at minimum. Estimates for the wider chemical space climb past anything a laboratory could ever synthesise, let alone test. A pharmaceutical screening programme might physically test a few million compounds over months at significant cost.

The real job is triage. Models trained on known chemistry and protein structures predict which molecules are worth making, which will bind to a target, and which will fall apart in the body. Chemists still make them. Biologists still test them. The models decide what goes on the bench.

If the concept of a model learning patterns from data is new to you, our explainer on how machine learning works covers the mechanism without the jargon.

Why binding prediction is so hard

A drug works by physically fitting into a pocket on a protein, like a key in a lock. Predicting that fit means modelling how atoms arrange themselves, and hydrogen atoms are the difficulty. They are tiny, they move, and standard imaging often cannot see where they sit.

Get the hydrogens wrong and your simulation of how a molecule docks is wrong too. Entire programmes have chased candidates that looked excellent in software and did nothing in a cell.

The result labs are pointing at now

In September, researchers published work in which a model scanned 4.6 million compounds in hours to predict hydrogen positions in drug-like molecules, sharpening how accurately those molecules can be modelled inside a protein binding site.

Put that against the older approach. Determining hydrogen placement by traditional computational chemistry is slow enough that teams do it for a handful of promising molecules, not millions. Running it across a library of 4.6 million changes what the question even is. You stop asking which of my twelve candidates is best and start asking which twelve of my millions are worth having.

A second strand of work published in Science Advances tackled the time problem in molecular simulation, using generative models to bridge from femtosecond to nanosecond time steps. In plain terms, simulating how a molecule behaves normally advances in steps so small that covering any useful stretch of time takes enormous compute. Skipping ahead accurately compresses simulations that ran for weeks into something far shorter.

Where the money is going

Spending tells you how seriously the industry takes this, and the spending is no longer experimental.

Eli Lilly inaugurated LillyPod, described as the first NVIDIA DGX SuperPOD built with DGX B300 systems for a pharmaceutical company. It is a dedicated machine for drug discovery, genomics and clinical development, not rented cloud capacity borrowed between other jobs.

The same company signed an agreement with Insilico Medicine valued at up to $2.75 billion, granting Lilly worldwide rights to develop preclinical candidates discovered through Insilico's generative platform. A deal that size is not a pilot. It is a bet that the molecules coming out the other end are worth more than the traditional pipeline they would otherwise fund.

This is the pattern across AI adoption generally. The organisations getting value are the ones rebuilding a workflow around the technology rather than bolting it onto the side, which is the same split that separates the winners from the companies shipping AI features without a use case.

What the adoption numbers actually show

Reported outcomes from biotech firms using these tools are encouraging and carefully worded, which is itself informative.

Roughly half of those adopting AI in biotech report faster time-to-target, meaning they identify which protein to aim at sooner. About 42% report an uplift in accuracy and hit rates, meaning more of the molecules they test do something useful.

Notice what neither figure claims. Neither says a drug reached patients faster. Both measure the earliest stage of a pipeline, the part before human trials, which is also the cheapest part and the part where failure costs least. That is genuine progress at the front of a very long process.

A worked example of what that saves

Say a discovery programme normally screens 2 million compounds to find 200 worth serious study, and that costs a year and a large fraction of a preclinical budget. Improving hit rate by 42% means the same 200 promising molecules emerge from a much smaller physical screen, or the same screen yields closer to 280.

Against a ten-year timeline, shaving months off the first year is real but modest. Against a 90% clinical failure rate, improving which candidates enter trials is worth vastly more, because a Phase III failure can burn hundreds of millions in a single result.

The test AI drug discovery has not passed

Every figure above describes finding candidates. None describes a medicine that works in people.

The industry is now entering the stretch where AI-derived candidates reach late-stage trials in numbers. Phase III is the stage that decides whether a drug is real, and it is indifferent to how the molecule was found. A compound discovered by a generative model and one found by a chemist's intuition face exactly the same statistics once they reach human beings.

Watch the Phase III readouts, not the discovery announcements. The discovery claims have been arriving for years. The verdict has not.

Frequently Asked Questions

Has AI actually created a new drug yet?

AI has produced candidates that reached human trials, which is a genuine milestone. No AI-derived drug has yet completed the full approval path and become standard treatment. The candidates are real, the final verdict is pending, and Phase III results over the next stretch will settle it.

How is AI drug discovery different from normal computer modelling?

Traditional computational chemistry simulates physics directly, which is accurate and extremely slow. Machine learning models learn patterns from known chemistry and predict outcomes without simulating every interaction. That trade gives up some theoretical rigour and buys enormous speed, which is what allows millions of compounds to be assessed instead of dozens.

Will AI make medicines cheaper for patients?

Not directly, and not soon. Discovery is a small share of what a drug costs to bring to market compared with clinical trials, manufacturing and marketing. Cutting discovery time and improving which candidates enter trials reduces waste, but drug pricing is driven far more by patents, negotiation and market structure than by research cost, as anyone comparing what GLP-1 drugs cost each month will recognise.

Which companies lead in AI drug discovery?

Eli Lilly has moved most visibly, with its own AI supercomputer and the Insilico agreement. Insilico Medicine is among the better-known discovery platforms. Most large pharmaceutical companies now run internal programmes or partner with a specialist rather than building everything themselves.

Can AI predict drug side effects too?

Partly. Models are reasonably good at flagging known problem patterns, such as molecules likely to be toxic to the liver or to interfere with heart rhythm. They are far weaker at predicting effects nobody has seen before, because there is no data to learn from. That is a large part of why human trials cannot be skipped.

Judge it on the readouts

AI drug discovery has moved past the stage where it is only a pitch. Millions of compounds are being assessed in hours, a major pharmaceutical company has built dedicated hardware for it, and billions are committed to molecules the models picked.

What has not happened is the part that matters to a patient. Finding a promising molecule has never been the hard part of making medicine. Proving it works in people is, and that proof arrives on the same slow schedule it always has. The honest position is that the tools have clearly improved and the outcome is still being tested.