10/26/2026  |  Ján Búza  |  The Recursive

https://www.therecursive.com/a-drug-has-to-answer-5-questions-ai-made-only-one-cheap-who-pays-for-the-rest/ 

A monthly column on biotech for founders and builders in Central and Eastern Europe. The first asks why cheaper design hasn't led to cheaper drugs.

AI has made it cheap to design a drug candidate, and yet drugs are getting slower and more expensive to develop. Most people who come to biotech from software assume that once a model has designed the drug, the company has won. The hard part is proving that the drug works in a person, and that proof keeps costing more.

In August, Anthropic reported that its model, Claude, designed proteins that bind 14 of the 15 targets it was given. Of 1,320 designs, 354 worked. Two outside labs, Adaptyv Bio and Twist Bioscience, had to make and measure every protein before anyone knew that. A protein that sticks to its target in a dish is only the first of five steps a drug needs to reach patients.

This column follows a drug through those five steps and asks who pays for each. It draws on what our fund learned from backing 30 companies along that path.

The bill keeps growing
In 2012, Jack Scannell and colleagues showed that the number of drugs approved per billion dollars spent on research, adjusted for inflation, had roughly halved every nine years since 1950. The newest figures point the same way. Deloitte follows the 20 drug companies that spent most on research in 2020 and put their average cost of taking one drug from discovery to launch at $2.67 billion in 2025, up from $2.23 billion a year earlier.

Clinical trials The drugs also take longer. Researchers at Tufts University found that the clinical phase lasted nearly seven months longer for drugs approved in 2014 to 2018 than for those approved in 2008 to 2013, and IQVIA reported in March that clinical timelines grew again in 2025. Fewer drugs get through, too. Citeline followed the drugs that entered Phase 1 trials between 2014 and 2023, and 6.7% of them reached approval, the lowest share in any of its analyses.

The five questions
A drug has to answer five questions before anyone pays for it.

1. Can we make a molecule that hits the target?
2. Can the medicine reach the right tissue?
3. Does it work for a person?
4. What does that proof cost in this disease?
5. Can we make itit be made at a price anyone will pay?

Each answer costs money and time, and a company must raise the funds for the next answer before the last round runs out. AI has cut the cost of the first one.

"The third is the expensive one, because it needs patients and years, and nothing so far has made that part of the process cheaper."

So, a cheaper first answer mostly lengthens the queue for the third, and that queue is already long. Citeline counted 5,995 drugs in development in 2001 and 22,940 at the start of 2026. Each of them needs trials before anyone knows whether it works.

What 30 bets taught us
ZAKA is an early-stage fund investing in health and industrial technology. As Michael Porter wrote in his 1996 essay What Is Strategy?, “the essence of strategy is choosing what not to do.” Since 2022, we have backed 30 companies working on these five questions. We watched some do well and others stall, and from that we wrote a few rules for how we invest in biotech.

Today we turn down four kinds of companies:

• The first designs molecules with a model, has no other edge, and will not own the drugs it finds. 
• The second builds a carrier that takes medicine into an organ and has no drug of its own to put in it.
• The third sells a service whose revenue grows only as fast as it hires people.
• The fourth owns a drug but cannot say who will pay for testing it in people. A seed round pays for lab and animal work. The first human trial needs more: a larger round, a pharma partner, a grant, or investors who have paid for a trial like it before. A company that cannot name one of these can run out of money before its drug reaches a single patient.

Why this matters here
Antonín Holý, a chemist at IOCB Prague, made the compounds that Gilead turned into medicines for HIV and hepatitis B. IOCB says they earned the institute more than €1 billion in license fees between 2001 and 2025. The institute was paid for 25 years because it held rights in the drugs themselves.

Companies from the region work on these questions too. Sensible Biotechnologies, which ZAKA first backed in 2022, works on the fifth. It makes mRNA inside living cells instead of in a reaction tube, to cut what a dose costs to make. Its Slovak founders, Miroslav Gašpárek and Marián Kupčulák, run labs in Bratislava and Oxford, and Gašpárek is also an investment partner at ZAKA. In October, Sensible raised a €42 million Series A led by Oxford Science Enterprises, as The Recursive reported.

"Most founders in the region pitch me only the science. The investors who fund them ask first what the next answer costs and who will pay for it."

What this means for founders
If your company owns a drug, price the next answer before you raise. Name who will pay for the step after this round and the cheapest result that would persuade them, then check that the round buys that result. The disease you pick sets that price. 

FinalDose, a startup founded in Oxford that we backed this year, goes after diseases where the proof is cheap, and bladder cancer that has not grown into the muscle is on its list. Merck's Keytruda and Ferring's Adstiladrin were both approved for that cancer once it stops responding to the standard BCG treatment, Keytruda on a single-arm trial with 96 evaluable patients and Adstiladrin on one with 98.

If you sell software or services to drug developers, the same chain decides how much you can earn. Schrödinger sells drug-design software and has worked with Nimbus Therapeutics on several programmes. When Takeda paid Nimbus $4 billion upfront for its TYK2 inhibitor, Schrödinger's 3.8% stake brought it about $147 million. Nimbus and its other shareholders kept the rest. Build for a buyer who pays you again and again, or for a way to own part of the drugs your product helps find.

"Pharma buyers are slow, so raise enough to wait for them."

AI will keep making the first answer cheaper, which helps any founder who designs molecules. It does little for the trials, which still need patients and years. Before each round, know which answer the money buys and who will pay for the one after it.