The moat decides whether you can play. The asset decides whether you get paid.
What a data advantage in drug discovery is actually worth
Schrödinger and Nimbus Therapeutics found a psoriasis drug together. Nimbus took it through Phase 2b and sold the subsidiary holding it to Takeda for $4 billion in cash, with up to $2 billion in sales milestones behind it. Schrödinger had held equity in Nimbus since the company was founded in 2009. By the end of 2022 that stake was around 4%, and it paid out about $147 million.
Schrödinger’s contribution was priced in 2009, before anyone knew whether the molecule worked, and in a currency that dilutes. Equity is not the usual arrangement, and the usual arrangement is not any better.
Drugs are tested in people in three numbered phases. Phase 1 checks safety in small numbers, Phase 2 is the first real look at whether the drug does anything, and Phase 3 is the large trial that supports approval. Each phase costs several times the one before, which is why the phase you sell in decides most of the price.
A target is the protein or gene a drug is built to act on, and the firms that find them are usually paid in milestones and royalties. They only pay if the drug arrives, and the vast majority does not. So the discoverer holds either a claim that shrinks with every round funding the clinic, or one that never triggers. The company that ran the trials holds the drug. Schrödinger had the better version of that trade, equity from the founding and cash at the exit, and still took only 4%.
Defensibility and value capture are separate questions. A data moat tells you nobody can copy what you do. It says nothing about who pays you for doing it. That is the frame I want to put on the archetype I said I would back first: companies generating proprietary human biology to choose better drug targets. In this business a single drug programme is called an asset, and the question that decides your return is whether you own one or hold a share of someone else’s.
The one layer models cannot eat
About nine in ten drugs entering clinical trials fail, and the largest single category of failure is lack of efficacy. Arrowsmith and Miller put it behind 59% of Phase 2 failures and 52% of Phase 3 failures for 2011 and 2012. Harrison’s 2013 to 2015 cut landed near half for both.
A failed efficacy readout does not prove the target was wrong. Dose, endpoint and population produce the same result. But the target sits underneath all three and cannot be corrected afterwards. Get the chemistry wrong and you make another molecule. Get the target wrong and you find out eight years and several hundred million dollars later.
That is also what separates this layer from the one next to it. Molecule design is being commoditised. As models are increasingly good, doing it well no longer distinguishes anyone, and a company built on it fails the test I set in the first piece: would this be worth more or less if the best models were free. Target choice passes that test, because the data that fixes it is the part nobody can retrieve. A model reasons over what has been published. The step from an association to a target is mostly unpublished, because it has mostly not been done.
That is the moat. The rest of this piece is about why it does not pay, and what to do about it.
A moat that does not convert
Selling targets has a credibility problem. If your target were genuinely good, why are you selling it instead of developing it? Every business development team asks that, and the honest answer is usually that you could not fund the development. Keeping your own programmes is what answers it. When a company licenses what it chose not to keep and develops the rest, the licensed targets carry the signal that the company puts its own capital behind the others.
It does not mean it always works. Verge Genomics is the cautionary version. Human brain and spinal tissue, genetically supported targets, validation in human neuron models, two pharma partners in Lilly and Alexion, everything the pitch asks for. In December 2025 its lead ALS candidate missed a pre-specified efficacy analysis at Phase 1b and the company cut about 90% of its staff. It is now trying to sell target data from a position where everyone can see it has nothing left in the clinic. The moat was real, but it was never the business.
Credibility as a target seller requires keeping programmes. Keeping programmes requires clinical capital and data advantage does not supply it. As you need about $1.4 billion in cash per approved drug on DiMasi’s 2016 estimate (and $2.6 billion once you add the cost of capital over the decade it takes), the question is how to get it.
Partnerships are the obvious answer, but they do not close it either. Across 516 licensing transactions in 2025, upfront cash ran somewhere between 7% and 15% of the announced total. The rest is milestone payments, owed only if the drug clears hurdles that most drugs never clear. Split the difference and a $500 million partnership puts $35 million to $75 million in the door, spread over a multi-year term, part of it equity, usually against partner-selected targets. Verge’s Lilly deal was up to $25 million near term against a $694 million headline. The Alexion deal was up to $42 million against $840 million. Both were strong validation, and neither was clinical capital.
Partnership money buys time without selling equity, which is worth more at seed prices than it looks, and good for early investors like us because we get diluted less. Eventually you still need equity to fund the asset.
Sell the asset to keep the engine running
The asset market is merit-based. Pharma buys quality assets from whoever has them, which is why five of the ten largest R&D licensing partnerships between January and October 2025 involved companies based in China. Nobody needs to know your name as long as your data package is good.
Selling the company outright at Phase 1 can bring real money. Sanofi paid $1.15 billion upfront for Vicebio, whose lead had only just entered Phase 1. That sells the machine that finds the drugs along with the drug itself, and a vaccine platform is an easier thing to buy whole than a target-finding engine.
Keeping the engine while extracting real cash for one asset is a Phase 2 move, and the deal that opened this piece is proof it can be done. It is also proof of what it costs. Trackers put Nimbus’s private funding above $423 million before the sale. Nimbus sold after positive Phase 2b results and kept everything else, including an early-stage cancer programme and the computational engine that produced both.
This structure is the way out of the loop. Takeda did not buy Nimbus, but its subsidiary holding one programme. Put each asset in its own vehicle with its own syndicate, keep the engine at the parent, sell one at a time. You still need a lot of capital, and you do not need to sell the company.
Why I would back target discovery anyway
It is hard and does not necessarily pay well, as the Schrödinger case would attest.
The reason to back it anyway is that target choice is the input deciding what a programme is worth, so the archetype that chooses better converts each dollar of clinical capital into a better asset than the company next to it. That advantage shows up in the pipeline they keep, not in anything they sell. A target-first company is worth backing when it is on its way to becoming an asset owner.
So the questions are about structure more than science. Enough of your own pipeline that the licensing side is credible. A shape where one asset can be sold without selling the engine. Indications that answer inside a decade, which turns out to matter more than the technology and is a separate piece.
If you are building this, I would like to talk.
