Pharma · · 8 min read
How to choose an AI drug discovery partner
Acumen Editorial · Updated November 21, 2026
Pharma is buying AI for discovery at a pace that makes diligence hard. A November 9, 2026 sourcing analysis by Beroe, published in Clinical Leader, counted 50 publicly disclosed AI deals from January 1 to October 14, 2026, across 18 of the top 30 pharma companies. About 64 percent focused on discovery and target identification.
The deals keep coming. On November 19, STAT reported a new Merck collaboration with Isomorphic Labs that pairs Isomorphic's design engine with Merck's existing wet-lab data for an undisclosed oncology target. A day earlier, an AstraZeneca filing showed it will pay Tempus at least $200 million from 2027 to 2031 for de-identified data, platform access and services.
If your team is next, the question is not whether to partner. It is how to tell which partner will actually move your programs.
Why is choosing an AI drug discovery partner harder now?
Three reasons.
The market is concentrated. Beroe found that a few suppliers account for most deal activity. Chai Discovery works with five top pharma companies. Isomorphic Labs, Owkin, Evinova, Iambic Therapeutics and Tempus each work with three. The authors flag lock-in risk and possible affordability problems for small and midsize R&D organizations.
Headline values hide the real price. Eli Lilly's deal with Insilico Medicine, as reported by Fierce Biotech, is worth up to $2.75 billion, but only $115 million is upfront. Beroe puts it plainly: "Most headline deal values depend on future milestones being met." A big number tells you what the partner hopes to earn, not what the platform will do for you.
Public benchmarks do not settle it. DrugDiscoveryBench, released by Scale Labs and Phylo in June 2026, set 82 agentic tasks written by pharmaceutical scientists, from target identification to patent mining to structure-activity analysis. The best system solved 51.6 percent, and the top three sat within six points of each other. When agents were given the expert's method as a hint, they recovered many of the tasks they had missed. The gap is scientific judgment, and it varies by task. A leaderboard cannot tell you how a platform handles your target class.
What kinds of AI drug discovery partners are there?
Beroe recommends treating general-purpose models and domain-specific discovery platforms as separate sourcing categories, each with its own dependencies, pricing and leverage. We add a third category: AI offered inside CRO and data deals.
| Partner type | What you buy | Typical structure | Main risk | What to test first |
|---|---|---|---|---|
| Discovery platform (design engine, structure prediction, generative chemistry) | Candidate molecules or hypotheses for your targets | Upfront plus milestones and royalties | Paying milestones for output you could not have judged in advance | Hit quality and ranking on your closed programs |
| General-purpose model or enterprise LLM alliance | Reasoning, literature and coding help across R&D | Enterprise license or usage-based | Broad claims, thin evidence on your science | Task-level accuracy on your workflows, graded by your scientists |
| CRO or data partner with AI services | AI-assisted trials, real-world data or analytics | Multiyear service and data commitments | Annual floors and lock-in to one data source | Data fit, provenance and exit terms |
Many deals mix types. Merck's Isomorphic collaboration combines a partner's engine with Merck's own wet-lab data. That makes the question of who owns what comes out of the work central to the contract.
How do you test an AI drug discovery partner before you sign?
Run a paid, time-boxed evaluation on problems the partner has never seen. Here is the checklist we use:
- Pick closed programs. Use finished or stopped series with real assay results and recorded decisions. Keep them unpublished.
- Freeze the evidence. Give the platform only what your team knew at each decision point.
- Write reference answers first. Your scientists record the compound or target advanced, the rationale and the measured outcome before the partner runs.
- Include the hard cases. Activity cliffs, series killed for toxicity or exposure, targets where the obvious answer was wrong.
- Score each task on its own. Target prioritization, hit ranking, property prediction and route feasibility fail differently. Do not accept one blended score.
- Grade blind. At least two of your scientists per reasoning item, with disagreements adjudicated and agreement reported.
- Audit the work. For agentic tools, check that every reported number appears in the logs and that the code reproduces it.
- Compare at least two partners on the same set. A single result has nothing to measure against.
- Rerun after updates. A new model version is a new product. Write a rerun right into the contract.
This also prepares you for regulators. The FDA and EMA guiding principles of good AI practice in drug development, released January 14, 2026, call for risk-based performance assessment and for data quality and governance across the AI life cycle. A written evaluation plan with expert graders and recorded results is the start of that file.
What should you ask an AI drug discovery partner?
Ask for evidence and terms, in that order:
- Which results come from public data, and which from unpublished programs like ours?
- How do you rule out training-data exposure for the problems you report?
- Will you run on our private evaluation set before signature, and share task-level results?
- Who on your side reviews scientific output, and what are their roles and experience?
- Who owns models, molecules and data generated from our data? Can you use our data to improve your platform for others?
- What happens to our data and models if we exit? Is there an annual floor or minimum commitment?
- Which payments depend on milestones we can verify independently?
- How will you show that a new model version performs at least as well on our set?
Pricing is also changing. Beroe reports a shift from per-seat and milestone-driven SaaS toward outcome-based and usage-based pricing, while noting outcome-based terms are harder to apply in discovery. If you want outcome-based terms, your private evaluation set is how you define the outcome.
Who should run the evaluation?
Not the partner. A vendor grading its own pilot has every reason to choose easy problems. Your team has the programs and the assay data but rarely the spare scientist hours to write reference answers and grade blind at volume. That is the gap independent expert evaluation fills: medicinal chemists, biologists, DMPK and computational scientists who build and grade the set against your data, with no stake in which partner wins.
How Acumen helps
We build the evaluation data behind scientific AI. Our life sciences model evaluation service supplies private evaluation sets written and graded by verified scientists: frozen-evidence problems from your closed programs, reference answers with rationale, task-level rubrics, agent trace audits and reruns on every partner update.
Use it to compare partners before you sign, to verify milestone claims, or to set the outcome measures in an outcome-based deal. For how to test the chemistry itself, see our guide on how to evaluate AI for medicinal chemistry.
FAQ
How do you choose an AI drug discovery partner?
Test before you sign. Build a private set from your own closed programs, have your scientists write reference answers, and run two or more partners on the same problems. Score each task separately and grade blind. Then compare data rights, exit terms and which milestones you can verify, not the headline deal value.
How many AI drug discovery deals has pharma signed in 2026?
A Beroe analysis in Clinical Leader counted 50 publicly disclosed deals from January 1 to October 14, 2026, across 18 of the top 30 pharma companies. About 64 percent focused on discovery and target identification, 24 percent on clinical development and 12 percent on enterprise-wide platforms. The authors note it is not a complete census.
Are public benchmarks enough to compare AI drug discovery platforms?
No. DrugDiscoveryBench, from Scale Labs and Phylo, found the best system solved 51.6 percent of 82 expert-written tasks, with the top three within six points. Public problems may also sit in training data. Only problems built from your own unpublished programs show how a platform handles your targets.
What contract terms matter most in an AI drug discovery deal?
Ownership of models, molecules and data generated from your data; whether the partner can reuse your data for others; exit rights; annual floors or minimum commitments; and which milestones you can verify independently. Beroe notes that most headline deal values depend on future milestones, so the upfront amount is often small.
What is vendor lock-in risk in pharma AI?
Lock-in happens when your data, models or workflows depend on one supplier you cannot easily leave. Beroe found a few suppliers hold most pharma AI deals, with Chai Discovery working with five top pharma companies. Reduce the risk with data portability terms, a private evaluation set you own, and a second partner benchmarked on it.
Do FDA and EMA expect AI partners to be evaluated?
The FDA and EMA guiding principles of good AI practice in drug development, released January 14, 2026, call for risk-based performance assessment and for data quality and governance across the AI life cycle. They apply when AI generates or analyzes evidence. A documented, expert-graded evaluation of a partner's platform supports that expectation.