AI Drug Discovery Is Getting Faster. Choosing the Right Biology Is Still the Hard Part

Around 850 drugs developed with the help of AI are currently moving through the pipeline. As of early 2026, not one of them has been approved in the United States.

That gap is worth sitting with. There is more money going into AI drug discovery than ever, and less and less time before it has to show results. Since 2022, over 100 billion dollars has gone into applying AI across the life sciences. The investment is there, the pipeline is there, but the approvals are not.

The Money Is Ahead of the Evidence

It would be easy to assume the bubble is deflating, but the opposite is happening. Venture funding for computational biotech went from around 599 million dollars in 2023 to 2.2 billion in 2025, and one of the field's most active investors, Dimension Capital, recently closed an 800 million dollar fund built specifically around this thesis. Money that used to head for software is increasingly heading for biology.

So this isn't a story about investors losing faith. It's something more interesting. The money is arriving well before the field has produced its clearest proof: an approved drug. The bet has been placed, and the result simply isn't back yet.

What Speed Solved, and What It Didn't

It's worth being precise about what AI has changed, because it has genuinely changed something.

The early stretch of drug discovery really is faster now. AI-enabled workflows have compressed early discovery timelines by an estimated 30 to 40 percent, and brought the preclinical candidate stage down from the usual three or four years to somewhere between 13 and 18 months. This is where AI is strongest: once you have a target, designing and refining a molecule against it is exactly the kind of well-defined problem it handles well.

And the clinical data backs this up, in a very specific way. A peer-reviewed analysis of AI-discovered molecules found they clear Phase 1 trials at a rate of 80 to 90 percent, well above the historical industry average of around 52 percent. When the job is to produce a molecule that is safe and well-behaved in the body, AI is genuinely good at it.

Then comes Phase 2, and the picture changes completely. There, the success rate for AI-discovered molecules sits at roughly 40 percent, essentially the same as the historical rate of about 37 percent for everything else. The advantage doesn't shrink. It disappears.

To be fair, these are early numbers on a small sample, and the Phase 1 figures especially deserve some caution. But the shape of the gap is telling, and it has a name: the translational gap, the distance between predicting what a molecule is and predicting whether it will actually work in a person. AI has largely closed the first. The second is still wide open.

The Hard Part Is Understanding the Disease

The reason is worth being clear about, because it isn't really a limitation of the models. It's about where the real difficulty in drug discovery lives.

To treat a disease, you first have to understand what actually causes it. And the diseases we still can't treat, many cancers, autoimmune and neuropsychiatric conditions, tend to be exactly the ones whose causes we don't yet fully understand. They are complex, driven by many things at once rather than a single broken gene or pathway.

This is the part speed alone can't fix. AI can design and optimise a molecule against a target faster than any team could by hand. But if that target was chosen based on an incomplete understanding of the disease, faster design just gets us to the wrong place sooner. The bottleneck was never really the molecule. It was understanding the disease well enough to know what to aim at in the first place.

AI Can Help Here Too, Just Not Alone

None of this means AI can only help with the molecule. It may matter just as much, perhaps more, for the harder question that comes before it: not just what to build, but what to aim at.

Working across decades of prior results and linking findings through knowledge graphs, AI can surface connections a researcher reading one paper at a time would never see, associations that were always present in the data but never assembled in one place. Pointing toward a possible cause, a patient subgroup, or a mechanism that no single study made obvious is real, and it matters.

But it is a starting point, not an answer. A connection AI surfaces still has to be tested at the bench, weighed against what a researcher knows from experience, and confirmed in the messy reality of biology. AI can narrow the search and speed up the reasoning. It cannot close the loop on its own.

The Regulators Are Asking the Same Question

It's a sign of where things stand that the people who approve drugs are converging on the same point.

In January 2026, the FDA and the EMA jointly issued a set of ten Guiding Principles of Good AI Practice in Drug Development. They are not a checklist for approval. They read more like a shared expectation: that AI should support human decisions rather than replace them, and that a model's workings should be documented well enough for a reviewer to evaluate its behaviour independently.

Underneath the regulatory language sits a familiar question. What did the model do, when can it be trusted, and how was its output checked before anyone relied on it? It's the same question a careful scientist would ask of any finding. What's new is that it is no longer just good scientific practice; it is becoming the standard that approval itself depends on.

Building for Direction, Not Just Speed

At R2H, this shapes the way the work gets done. AI is genuinely good at going fast, but speed was never the only reason to use it. It also helps make sense of the biology, and that matters more. Still, speed and understanding only take you so far. Deciding which biology is sound and which target is worth pursuing is where a human has to stay in the loop.

The same thinking sits behind NeoRevive, a drug rescue project currently in development. A drug that failed isn't necessarily a drug that doesn't work; often it failed because the biology behind it was only partly understood the first time. That is exactly where the work starts. A failed drug is run through Clarisyn to rebuild the picture of its biology: what it actually does, which mechanism it acts on, which patients were ever likely to respond. From that understanding, a team led by clinicians, working alongside toxicologists and biologists, looks for where it genuinely fits: a better-defined patient group, a different dose, or another disease driven by the same biology. That mix matters. It keeps the reasoning broad, and it brings in the kind of hands-on clinical and experimental judgement no model has. Clarisyn rebuilds the picture and lays out the evidence; the judgement about where the biology actually leads stays with the people who understand it.

Which is really the shape of the year. 2026 may not turn out to be the year AI discovers its first approved drug. It may instead be the year the field gets clearer about what actually counts as evidence that AI is helping, and honest that going faster was never the hard part.