How Far Generative AI Has Come in Drug Discovery, and How Far You Can Trust It
I published a report with BRIC View on generative AI in biotech and drug discovery. This post is the plain-language version, written so you can follow it without a background in the field. One question runs through the whole report: generative AI is said to be changing biology, but how far can you trust what comes out of it?
The answer depends on where you look. Protein design has reached the point where results are confirmed in the lab. Small molecules are standing at the threshold of the clinic. Genome-scale design is still inside the computer. The rest of this post walks through that gap.
What it takes to produce one drug
Bringing a new drug to market means screening tens of thousands of candidate compounds down to one, then testing it in animals and then in people. That takes well over a decade and costs on the order of a billion dollars. Even among candidates that reach first-in-human dosing, roughly one in ten becomes an approved drug.
Hence an old question: can we pick the candidates faster, and pick them better?
Where generative AI fits
The same class of model that writes sentences in ChatGPT also generates proteins and compounds. Earlier AI was mostly asked to predict: will this molecule be toxic? Generative AI answers a different request: draw me something new with these properties. The work moved from searching to designing.
That takes material to learn from. UniProt holds over 227 million protein sequences. The AlphaFold structure database holds over 214 million predicted structures. ChEMBL holds over 20.3 million bioactivity measurements. Public resources like these are the training material.
Protein design has reached experimental confirmation
The clearest results have come from protein design. RFdiffusion designs, from scratch, proteins that bind to a chosen target. These designs were not left on paper. They were expressed, purified, and imaged, and the designed shape matched the measured structure to within 0.63 Å of backbone RMSD. A single atom is roughly 1 Å across, so the error was smaller than one atom.
The success rate, though, was about 19%. Design five and one works. That is a large improvement over earlier methods, and the 2024 Nobel Prize in Chemistry went to protein structure prediction and computational protein design.
The funnel narrows toward the back
Generating candidates inside a computer is cheap and plentiful. Passing through the lab and then through clinical trials, the number that survives drops sharply while the evidence demanded of each one gets stronger.
The numbers trace that shape. Among AI-discovered drugs, 21 of 24 advanced out of Phase I, which mainly tests safety. That is about 88%, well above the industry average of 40 to 65%. In Phase II, where the question becomes whether the drug actually works, only 4 of 10 advanced. At 40%, that does not beat the industry average.
The most advanced case sits in the same place. Rentosertib, a TNIK inhibitor aimed at idiopathic pulmonary fibrosis, met its primary safety endpoint in a Phase 2a trial of 71 patients and entered Phase 3 in July 2026. Efficacy showed a favorable trend, but only as a secondary endpoint, in arms of roughly 18 patients each. Confirmation is still ahead.
How a scorecard gets inflated
Sometimes the scorecard itself deserves suspicion. If the practice problems show up on the exam, the score looks good, and the same thing happens in data. Material used for training leaks into the evaluation set.
This has been documented in at least 294 papers across 17 fields and sorted into eight distinct types. With leakage, benchmark scores come out high and then fail to reproduce when someone measures again elsewhere. I ran into the same problem a few days ago when I measured it myself on ADMET models, written up here in Korean.
Risk grows with capability
As generative capability grows, so does the room for misuse. When researchers inverted the objective of a drug discovery model, telling it to seek toxicity instead of avoiding it, roughly 40,000 toxic molecules were designed in under six hours. Nothing was synthesized or verified; these are computational results.
There are moves in the other direction. Evo 2, a genome generative model, deliberately excluded viral genomes that infect eukaryotes from its training data, blocking dangerous sequences at the design stage rather than after the fact.
Where the rules stand
Regulation is catching up. The US FDA issued a draft framework for judging whether AI-generated data is credible enough to support regulatory decisions. The EU set obligations for high-risk AI under the AI Act and pushed the main compliance dates to December 2027 and August 2028. The USPTO concluded that AI cannot be named an inventor and that a significant human contribution is required. In Korea, the MFDS issued a guideline for LLM-based digital medical devices in June 2026, listing seven hazard categories along with validation metrics.
What is still missing is a separate standard for how evidence produced by generative models should be weighed in drug review.
Where this started
Last spring I wrote a BRIC View report on how bio big data is being pooled and expanded. Once that was done, the next question was sitting right there: what is actually being built on top of that data?
Starting the survey, I found no shortage of material listing technologies and cases. What was missing was a separation of how strong the evidence is at each stage. So instead of widening coverage, I followed one question through.
Read it
The full report is available at BRIC View, in Korean.
Every number in this post comes from the report, and the primary sources are in its reference list.
Next
The part of the report that took longest was separating which claims rest on what grade of evidence. Wanting to do that with my own hands, I measured how the choice of data split changes reported performance in ADMET models, and wrote that up separately. Next I want to widen the task and model set and work out why some of my results diverged from prior work.
The sentence the report closes on is this. In the era of generative AI, competitiveness in biotech will rest less on how fast you can generate something, and more on how credibly you can demonstrate that what you generated holds up.