After the Decision: The Hard Work of Making AI Work in Biology

“AI in biology only works when it understands context, learns iteratively, and is built hand-in-hand with the scientists who use it.”

Building on conversations from our previous events, where the focus centered on the cost of hesitation and why a "wait and see" approach may be one of biotech's most expensive strategies, our AI for Biology Thought Leadership Dinner in Zurich took the discussion one layer deeper. There is now little debate about whether AI has become essential for drug discovery. Instead, the focus shifted to how to make it work within the realities of biology, discovery workflows, and clinical translation.

A striking shift was evident around the table. Leaders from pharma, biotech, data science, and research have moved on from AI adoption in abstract terms and the main agenda became understanding the real texture of implementation: what works, what remains unproven, and what needs to be built before AI can meaningfully change the probability of success in drug development.

Three themes stood out.

Can AI discover new biology?

AI has already produced visible “aha” moments in structural biology, most famously around protein structure prediction. But translating AI into broader clinical success is a different challenge. Biology is more contextual, more dynamic, and less deterministic. The near-term opportunity may therefore be less about promising immediate breakthrough cures, and more about improving efficiency, narrowing uncertainty, and increasing the probability that better scientific decisions are made earlier. The next frontier is parallel science: many hypotheses tested at once, with the strongest moving fastest toward experiment.

How do we speed up lab validation?

Large-scale computation and large-scale experimentation are no longer the only constraints. The harder question is: which experiment should be run, in which biological context, and with which data, so that the result is more likely to transfer into clinical relevance? In other words, the value of AI will depend well beyond the ability to process information; it will lie in helping scientists ask better questions.

What’s the best way to organise ever-expanding data?

Data readiness is strategic infrastructure. Structured, organized, well-annotated data remains one of the most important foundations for AI in biology. Ontology decisions made early can either accelerate future learning or create downstream friction. Just as importantly, the people generating the data and the people using the data need to be part of the same loop. AI tools cannot be built in isolation from the scientists, clinicians, and teams whose workflows they are meant to support.

‍

The next test for AI in R&D is not whether the models are impressive. It is whether scientists use them, trust them, and make better decisions because of them.

Built outside the workflow, AI risks becoming another demo. Built with users, on structured data and clear ontologies, it can start to change how discovery gets done.

The industry is still waiting for its broader “aha” moment. Until then, the metrics that matter are practical: faster teams, more hypotheses tested in-silico, better experiments, less duplication, stronger confidence before costly decisions, and clearer signals at key phase transitions.

Thank you to Jonas Béal, Nadia, Mayssam, and Andrea for representing Owkin, and to Thomas Di Maio, Head of Diagnostics EU-CAN at AstraZeneca, and Nikhil Podduturi, Head of Data Science, Digital Lab at Boehringer Ingelheim, for sharing their insights during our fireside chat. We also extend our thanks to AWS, and in particular Romain Curunet, Senior Account Manager, Digital Natives, Healthcare, Life Sciences & Retail, for partnering with us on our AI Thought Leaders Dinner Series.

More insights to come as we build on our learnings through these dinner series.