AI Drug Discovery: Where the Real Moat Is

There is a pair of numbers that should shape how any founder thinks about AI drug discovery, and most coverage of the field reports only the first one.

A 2024 analysis in Drug Discovery Today, by Madura Jayatunga and colleagues at Boston Consulting Group, examined the clinical pipelines of AI-native biotech companies. In Phase I, AI-discovered molecules succeeded 80 to 90 percent of the time, which the authors call "substantially higher than historic industry averages." Then Phase II, where the success rate was about 40 percent, "albeit on a limited sample size, comparable to historic industry averages."

Read those together and the strategy falls out. AI has become genuinely good at designing a molecule that behaves like a drug inside a human body. It has not yet moved the number that decides whether you have a company, which is whether the drug treats the disease.

That gap is not a temporary modeling problem waiting on the next architecture. It is a data problem, and understanding its shape tells you where the defensible asset in this market actually sits.

What AI drug discovery actually means

AI drug discovery is the use of machine learning to choose a biological target, propose molecules that might act on it, and optimize those molecules before any of them reach a human. Four jobs sit under the term:

• Target identification: deciding which protein or pathway to go after in the first place. Models trained on genomic, proteomic and literature data rank which biological targets are plausibly causal in a disease.
• Hit generation: searching chemical space for molecules that plausibly bind that target. This is where virtual screening (scoring large compound libraries computationally instead of testing them physically), molecular docking (simulating how a molecule fits into a binding site), and generative chemistry (proposing novel structures rather than filtering existing ones) do their work.
• Lead optimization: making a promising molecule more potent, more selective, and less toxic, while keeping the properties that let a drug survive in the body.
• Structure prediction and elucidation: working out the three-dimensional shape of a protein or a small molecule, which is what makes the first three tractable. AlphaFold2 is the landmark here, and cryo-electron microscopy is the experimental counterpart.

What the term does not cover, and this is where founders arriving from software get the economics wrong, is the trial. A candidate that clears every computational filter still has to be manufactured, dosed into animals, then usually into healthy volunteers, then into patients who actually have the disease. That sequence takes years, and it is where nearly all of the cost lives.

Where AI drug discovery sits in the drug pipeline

The whole term covers the first two rows of the table below. Everything after that is unchanged by AI so far, and the durations and attrition rates come from the FDA's own description of the clinical research process.

Stage	What it asks	Typical duration	What AI touches today
Target identification	Which biology should we attack?	Months to years	Heavily, and with the least data behind it
Discovery and lead optimization	Which molecule should we take forward?	Months to years	Heavily, and this is where the gains are real
Preclinical	Is it safe enough to put in a human?	Months to years	Partially, mostly toxicity prediction
Phase 1 (20 to 100 people)	Safety and dosage	Several months	Barely
Phase 2 (up to several hundred)	Efficacy and side effects	Several months to 2 years	Barely
Phase 3 (300 to 3,000 people)	Efficacy and adverse reactions	1 to 4 years	Barely
Regulatory review	Does the evidence hold up?	Months to years	Barely

The FDA puts the attrition plainly: roughly 70 percent of drugs move past Phase 1, about 33 percent past Phase 2, and 25 to 30 percent past Phase 3. Read down the last column of that table and the strategic problem is visible in one glance. The technology is concentrated in the rows before the first human dose, and the killing happens in the rows after it.

Demis Hassabis, who shared half of the 2024 Nobel Prize in Chemistry with John Jumper for protein structure prediction (the other half went to David Baker for computational protein design), is careful about that boundary even while making the strongest public case for the field. Speaking with Sequoia Capital, he describes a protein's 3D structure as an incredibly important thing to know if you want to design medicines, then immediately adds that "it's only one part of the drug discovery process."

That qualifier is the load-bearing part. Structure prediction sits upstream of the decision that actually kills programs.

The number that should shape your AI drug discovery strategy

AI drug discovery succeeds where the question is chemical and stalls where the question is biological. Split the pipeline by the question each stage asks and the pattern is exact.

Phase I mostly asks: is this molecule safe and tolerable in a human? That is close to a chemistry question. Does it dissolve, absorb, and clear the way a drug should? Does it bind things it should not bind and cause toxicity? Those properties are encoded in the molecule's structure, and structure is precisely what fifty years of accumulated chemical and structural data describes. A model trained on that corpus can learn it.

Phase II asks a different question: does this drug make sick people better? That depends above all on whether the target you picked in step one is causally involved in the disease, along with dose, patient selection, and trial design. No amount of chemical data answers the target question, because the answer lives in human biology that has never been written down.

So the Jayatunga finding is not a curiosity. It is the field telling you exactly which half of the problem it has solved. AI is doing well on the half that was already the cheaper, faster half, and it is performing at the historical baseline on the half that consumes the budget.

For a founder, three things follow:

1. Do not build a company whose entire thesis is speed to development candidate. Compressing a two-year discovery stage into six months is real value, but it is value on the short leg of a ten-year journey. It is a feature that a pharma partner will pay for once, not a franchise.
2. Target selection is the high-leverage bet, and it is the one with the least data behind it. If you have an edge, it should be here.
3. Assume your Phase II odds are the industry's odds until you have your own evidence otherwise. Building a plan on the Phase I number is how you run out of money in year six.

AI drug discovery is best at exactly the part that was already cheapest

BenevolentAI ran the aggregate statistic as a single public experiment, and the result was a molecule that passed safety and failed efficacy.

In April 2023, BenevolentAI reported top-line Phase IIa results for BEN-2293, a topical pan-Trk inhibitor for mild-to-moderate atopic dermatitis, and the lead clinical asset from a company built on AI-driven target identification. The drug hit its primary endpoint: it was safe and well tolerated. It missed on efficacy, failing to show a statistically significant effect across the intention-to-treat population on either secondary endpoint, the Eczema Area and Severity Index and the pruritus numerical rating scale. Fierce Biotech covered the fallout, which included a collapsed out-licensing plan and, a month later, roughly 180 layoffs.

That is the aggregate statistic playing out in a single program. The molecule was fine. The molecule was, in the narrow sense the models are trained to optimize, a success. The bet on what to point it at did not pay.

The pattern repeats at the top of the market. Recursion merged with Exscientia in November 2024 to create the largest pure-play in the category. Six months later, on May 5, 2025, it halted four programs and paused a fifth, discontinuing REC-994 after a long-term extension study showed no promising trends, and REC-2282 after limited tumor shrinkage.

As of this writing, no molecule discovered or designed by AI has completed the path to regulatory approval anywhere. The furthest along is Insilico Medicine's rentosertib, which reported Phase 2a results in Nature Medicine and only entered Phase III in 2026. AI-repurposed drugs are a different category and worth separating out: BenevolentAI's knowledge graph identified the existing arthritis drug baricitinib as a COVID-19 treatment, and that indication won full FDA approval in 2022. Finding a new use for a known, already-approved molecule is a genuinely different problem from designing one.

None of this means the field is fake, and the baseline it is being measured against is genuinely terrible. Max Hodak, co-founder of Neuralink and founder of the neural engineering company Science, put the industry's track record bluntly in a Y Combinator interview: "Humanity just isn't very good at drug discovery." Matching a bad baseline is not an achievement, but it does mean the field is at the stage where its strongest claims are still forward-looking, and a founder should price them that way.

Why AI models cannot close the gap with existing data

The gap persists because the models have already ingested the available data, and the missing answers were never recorded in it. The clearest account of that comes from a researcher with nothing to sell.

In a YC Root Access conversation on molecular discovery, recorded at a NeurIPS afterparty, Ellen Zhong, assistant professor of computer science at Princeton, walked through what is and is not solved in her field. Her lab works on cryo-electron microscopy: imaging proteins and reconstructing their 3D structures from extremely noisy two-dimensional data. Before her MIT PhD she worked on molecular dynamics algorithms for drug discovery at D. E. Shaw Research, and she interned on the AlphaFold team at DeepMind during the release of AlphaFold2.

She grants the headline result readily enough. That specific problem, she says, sequence to relatively static protein structure prediction, may well have been solved. And then she describes what that leaves:

"everything is jiggling everything is moving in order to actually perform functions that lead to life"

Proteins are not the still images the folding models predict. They are machines in motion, and their motion is how they do their jobs. The static structure is a photograph of something that only matters because it moves. That is the first layer of what is missing.

The second layer is that the models have run out of road on existing data. As Zhong puts it, "right now the machine learning and like AI models have ingested all the data that we have," and while you can distill useful things from what is already in there, the remaining questions are not in the corpus. Then, mid-answer, she poses the question herself and does not hedge:

"will machine learning alone be able to do that? No."

Later in the same conversation she names the gap that matters most for a founder, and it is the Phase II gap in scientific dress:

"So I think to really bridge that gap, there's going to need to be new experimental technologies."

The gap she is bridging is the one between molecular biology and human health. That is the same gap the clinical statistics describe, stated in scientific rather than financial terms.

One more line from that conversation matters more than the rest. Talking about which problems in her field are clean and which are messy, Zhong notes that structure reconstruction can be posed as a single objective function you optimize against, but protein design is different, because "it's unclear how you validate."

If you cannot cheaply tell whether an output is correct, you cannot run the tight iteration loop that makes software companies compound. That single sentence explains most of the difference between building in this market and building a SaaS product.

Is the wet lab a checkpoint or the source of the data?

Two credible camps disagree: one treats the wet lab as a final checkpoint, the other treats it as the only source of new signal. Which one you believe determines what you build.

Hassabis wants to shrink the wet lab. He is running that bet himself, as founder and CEO of Isomorphic Labs, the DeepMind spin-out that exists to turn AlphaFold into drug programs, and he describes it in the same conversation as a later spin-out he now runs. His stated ambition is that "the dream is to do almost all the exploration" computationally, roughly 99 percent of the work and the time done in silico, with the wet lab reserved for a final validation step. If that works, he argues, drug discovery timelines could compress from "an average of 10 years down to months, maybe even weeks and perhaps even days one day." In this view the lab is a checkpoint, and the value accrues to whoever has the best simulation.

Zhong's account points the other way. In her framing the lab is not a checkpoint but the source. New discoveries require new experimental data, because the existing data is exhausted. She describes her collaborations with structural biologists and chemists as working "with experimentalists who have the data, who have the expertise to know like what's a discovery or not," and says that for genuinely new discoveries "I do think for like truly new discoveries it's going to be a combination of new experimental data sources."

These are not the same bet, and the disagreement is not resolvable from the armchair. It is an empirical question about whether the information needed to predict human biological outcomes is latent in data we already have, or whether it has to be generated.

They are also not quite answering the same question. Hassabis is describing compound design against targets we can already model, and Zhong is describing the discovery of biology we cannot yet describe at all. The two views still imply different companies.

Both positions make the experimental apparatus strategically central. In Hassabis's version, the lab is the scarce validation resource that gates everything. In Zhong's, it is the only source of new signal. Either way, a company that has models and no experimental loop owns the part that is commoditizing fastest.

The AI drug discovery moat is the loop, not the model

Y Combinator has been explicit about the shape it wants to fund here. In its request for AI-native discovery engines, it describes the closed loop directly: "Models propose candidate molecules, automated labs synthesize and test them, and the results feed back in to iteratively improve candidates." The conclusion it draws is the strategic one:

"The companies that make meaningful contributions to scientific progress won't just sell research copilot."

A research copilot is a wrapper on the fastest-commoditizing layer of AI drug discovery. Model architectures diffuse in months, the public structural databases are public, and your competitor can fine-tune on the same corpus you did. There is no accumulating advantage in that position.

The loop is different, and the reason is mundane rather than technical. To run design, make, test, analyze as a closed cycle, you need an assay. Building one takes capital, time, and specialist hiring. Every cycle produces data that exists nowhere else because nobody else ran that experiment, on that target, with that readout. Your competitor cannot download it, cannot scrape it, and cannot buy it with a larger funding round, because what it costs is elapsed experimental time.

This is the data-rich versus data-poor trade in its sharpest form. Data-rich problems are easy to start and hard to defend, because everyone trains on the same corpus. Data-poor problems are hard to start and much easier to defend once you solve the supply. Drug discovery is emphatically the second kind, which is why the discipline it demands looks like the discipline physical AI teams have had to develop: most of the effort goes into manufacturing training data rather than tuning models.

The practical test for whether you have a loop or a wrapper: if your model got 20 percent better tomorrow, would you learn anything your competitor could not also learn? If the answer is no, the model is not your moat.

What AI drug discovery does to your timeline and funding math

The defensibility is real and so is the bill. Most of the AI strategy that applies to startups generally assumes you can ship, measure, and correct inside a quarter. Almost none of that holds here, so four things need planning up front.

Your first revenue is probably a partnership, not a drug. Pharma companies buy discovery capability. That funds the loop while the loop is still proving itself, and it is not a compromise, it is the normal financing structure for the category.

Your evidence milestones are set by someone else. In software you decide what counts as validation. Here the FDA does, and the schedule is measured in years. Build the fundraising plan around trial readouts rather than product releases.

Instrument the loop from day one. Every experiment you run is a training example you will want in three years. Teams that treat the assay as an operational cost rather than a data asset end up with results in lab notebooks and spreadsheets that no model can consume. Decide the schema before you have the data, because retrofitting it is close to impossible.

Be honest in your own reporting about which half you improved. If you compressed discovery, say you compressed discovery. Founders who let a Phase I result imply a Phase II result eventually have to explain the gap to the people who funded them on it. The same measurement discipline that keeps an AI product team honest applies with more force when a readout is years away and there is nothing to check against in the meantime.

Validation is a governance problem, not only a science problem

Provenance is the control that decides whether your results count as evidence or just outputs. In AI-heavy science companies it gets treated as paperwork until the moment it becomes the whole company.

The regulator has already moved on this. In January 2025 the FDA published draft guidance on using AI to support regulatory decision-making for drugs and biological products, its first on the subject, built around a risk-based credibility assessment of the model for a specific context of use. The agency notes it had already seen more than 500 submissions with AI components between 2016 and 2023. Read that as the shape of the question you will eventually be asked: not whether your model is impressive, but whether it is credible for the specific claim you are using it to make.

Four things are worth settling early, in writing, in any AI drug discovery company:

• Provenance for every training example. When a regulator or an acquirer asks why you believed a target was valid, the answer has to trace to a specific experiment, on a specific date, with a specific protocol. A model output with no chain back to a physical measurement is not evidence, and reconstructing that chain retroactively across three years of experiments is a project nobody budgets for.
• Data rights with every partner. A pharma collaboration generates exactly the proprietary data your moat depends on. Who owns it, who can use it to train, and what survives the end of the agreement are terms that decide whether the partnership builds your asset or rents it out. Settle this before the first joint experiment, not at renewal.
• An integrity boundary around the results. If a model selects which experiments to run and also helps interpret the readouts, you have a loop that can quietly confirm itself. Keep the analysis path auditable, keep negative results in the dataset, and make sure nobody can silently drop a run.
• An explicit line on autonomy. Automated labs execute physical protocols. Decide in advance which classes of experiment a model can queue without a human approving it, and treat that boundary as a controlled setting rather than a convention.

None of this slows you down meaningfully at the start. All of it is extremely expensive to add later, which is the usual shape of security debt.

What to do this week

1. Write down which half of the pipeline your product improves, discovery or clinical, and stop describing it as the other one in your deck.
2. Answer the loop test in one sentence: if the best available model improved 20 percent tomorrow, what would you know that a competitor would not? If you cannot answer, you are building a wrapper.
3. Cost out the smallest closed experimental loop in your domain. Not the full platform, the smallest assay that produces data nobody else has and that a model can consume.
4. Pick your target selection edge and name it. Proprietary patient data, an unusual assay, a clinical collaboration, a physical measurement nobody else takes. If you do not have one, you are competing on the commoditizing layer.
5. Set the data schema before the experiments. Decide today how a result gets recorded so that in three years it is a training set rather than an archive.
6. Put data rights on the agenda of your next partnership conversation, ahead of economics. The economics are negotiable later; the data terms are not.

Startups that win in AI drug discovery will not be the ones with the best molecular model. They will be the ones that built the experimental loop nobody else can run, then pointed good models at it. The model is the part everyone can buy.

If you are building an AI company and want a system for making these calls rather than making them one at a time, that is what I teach in AI Operating System for Startups.

Sources

• The Future of AI Molecular Discovery, YC Root Access, the interview with Ellen Zhong this article distills.
• "How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons", Jayatunga, Ayers, Bruens, Jayanth and Meier, Drug Discovery Today, 2024, for the Phase I and Phase II success rates, with the BCG summary of the same work.
• Ellen Zhong's Princeton faculty profile and her E.Z. Lab page, for her role, research focus, and prior positions at D. E. Shaw Research and DeepMind.
• Demis Hassabis: We're Three Quarters of the Way to AGI, Sequoia Capital, for the in-silico exploration argument and the timeline claim.
• AI-Native Discovery Engines, Y Combinator, for the design, make, test, analyze loop and the discovery-engine framing.
• BenevolentAI's Phase IIa top-line release for BEN-2293, with Fierce Biotech on the trial miss and, a month later, the 180 layoffs.
• Recursion halts four pipeline programs, Genetic Engineering & Biotechnology News, May 2025.
• FDA, Step 3: Clinical Research, for the trial phase purposes, participant counts, durations, and advancement rates in the pipeline table.
• FDA on AI in drug development: the January 2025 draft guidance Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products and CDER's Artificial Intelligence for Drug Development page.
• Phase 2a results for rentosertib in Nature Medicine, the most clinically advanced AI-discovered molecule, and a peer-reviewed account of the baricitinib repurposing for the distinction between AI-discovered and AI-repurposed.
• Isomorphic Labs, for Hassabis's role as founder and CEO.
• The Nobel Prize in Chemistry 2024, for the Baker, Hassabis and Jumper citations.
• The Future Of Brain-Computer Interfaces, Y Combinator, where Science founder Max Hodak makes the case that "Humanity just isn't very good at drug discovery."