Biology sets the pace: five R&D takeaways from World Summit AI 2026
October 29, 2026
By Steve Korver
The 10th edition found a field humbled by biology, more practical in the clinic, and newly alert to public trust.
“You cannot make a human being exhibit disease – or have the intervention reverse the disease – faster by throwing more GPUs at a human,” said Daphne Kolleropens in new tab/window, founder and CEO of insitroopens in new tab/window, during the fireside chat ‘From Data to Targets: Causal AI for Better Medicines’. She added: “You cannot accelerate biology. It has its own pace. But you can accelerate certain things around it.”
At its 10th edition on 7–8 October, World Summit AIopens in new tab/window stretched from founders chasing the next trillion-dollar model to an AI that listens for chainsaws in the rainforestopens in new tab/window. This time, the promise felt more subdued and grounded in the lab, and the peril had shifted from science fiction to opinion polls. Here are five takeaways for those working in research and health.
1. AI in R&D is moving from correlation to causation
AI thrives where two things exist, according to Daphne: internet-scale data and fast trial-and-error. “Neither of the two things that make AI so effective exist in the world of biology,” she said.
Her answer is to generate causal data at industrial scale. The company’s “cell factory” has produced over 30 billion genetically distinct cells, each deliberately perturbed and measured, and combines them with detailed health data from millions of people. “One of the most pernicious failures that we have as a species is not being able to understand that correlation and causation are not the same thing,” she said.
The economics are brutal. Because 90% of drugs fail in clinical developmentopens in new tab/window, every success “has to carry on its back the cost of all the failures,” pushing the costs, by some estimates, past $2.5 billion per drugopens in new tab/window. Daphne is out to raise the success rate by picking targets with human genetic evidence that they cause the disease. Such targets are over twice as likely to make it to approvalopens in new tab/window.
2. When data is scarce, teach AI the rules
Why did AlphaFoldopens in new tab/window succeed where AI in broader biology hasn’t? Daphne gave three reasons: decades of structural data, protein folding that works the same way in bacteria as in humans, and a problem that is “a lot closer to physics than it is to biology”. Human biology offers little of that. “The human being is not an engineered system,” she said. “It is the result of a stochastic evolution process that is a smorgasbord of different pathways doing different things in different cells.”
In her keynote ‘AI-Guided Discovery: Rethinking the Scientific Method’, Bonnie Kruftopens in new tab/window, Managing Director of Microsoft Research AI for Scienceopens in new tab/window, pointed to science’s “silver lining” when data is scarce: the laws of physics, which “describe the universe with really unique and precise accuracy.” So her team trains AI on physics simulations until it can predict how molecules and materials behave “many orders of magnitude faster” than the simulations themselves.
AlphaFold’s heirs think it can go further. In his keynote ‘Isomorphic Labs: AI Models to Solve All Disease’, Michael Schaarschmidtopens in new tab/window, VP and co-head of AI at Isomorphic Labsopens in new tab/window, the Google DeepMind spin-out, said the real challenge is generalization: systems that work for any target and any disease. No single model can do that, so Isomorphic orchestrates many, from predicting how proteins bind drugs to designing new ones. A day after the presentation, it was reported to be raising money at a valuation of up to $50 billionopens in new tab/window, though none of its drugs has yet reached a patient.
Where Isomorphic bundles models, Carina Kernopens in new tab/window, CEO of LinkGevityopens in new tab/window, bundles diseases. In ‘AI Is Transforming the World. How Close Is It to Cracking Aging?’, she argued that drug discovery’s one-target-one-disease approach can’t keep up with aging, which drives many diseases at once. Her target is necrosis, the uncontrolled cell death that contributes to tissue degeneration across organs and was long considered “one of medicine’s undruggable targets.” Aging data are “small, inconsistent and incomplete”, but “we can’t wait, because people are dying of debilitating diseases today,” so her team builds biological rules into its AI. She compared this to teaching it the rules of chess rather than just showing it the board. The company says its lead compound achieved 90% inhibition of necrosisopens in new tab/window in lab and preclinical studies.
3. From the lab to the clinic, at biology’s pace
Daphne’s case study was ALS, “a three-to-five-year death sentence” with “effectively nothing to offer” patients today. Some 97% of ALS patients show TDP-43 pathologyopens in new tab/window, in which the protein stops doing its normal job. AI analysis of 6 billion images mapped how patient-derived motor neurons sicken, and genome-wide screens found targets that reversed it in the dish – producing “an image that still gives me goosebumps.” With Bristol Myers Squibbopens in new tab/window, insitro is close to selecting a lead antisense oligonucleotide (ASO), an RNA-targeting drug, which will need roughly a year of IND-enabling studies before human trials.
“There are people out there saying that we soon won’t need clinical trials,” she said. “I wish we had that level of predictiveness. We’re so far away from being able to say what’s going to happen in a human being.”
AI is already working in the clinic. On the panel ‘AI in Healthcare: How Safe Is Safe Enough?’, Helia Mohammadiopens in new tab/window, Chief Research Technology Officer at City of Hopeopens in new tab/window, said AI now summarizes the hundreds of pages of records cancer patients arrive with, citing the source for each point, and matches patients to clinical trials at the point of care. As for the panel’s question, she argued that the right benchmark is today’s care, with its human error and fatigue, not some ideal: “It’s not right to compare AI with perfection. Our health system right now is not perfect.” But she also named the catch: a clinician’s mistake affects one patient, while a model’s mistake repeats wherever that model is deployed.
And the reports AI reads are messy. Dustin Raneyopens in new tab/window of SuperTruthopens in new tab/window cited a US studyopens in new tab/window in which one in five patients who read their doctors’ notes found an error, and 40% of them judged it serious. Hence his caution: “When it comes to care, I don’t believe that we’re ready for inference.” He wants AI to apply clear clinical rules at the bedside and keep probabilistic guesswork for discovery. BenevolentAIopens in new tab/window founder Ken Mulvanyopens in new tab/window offered the rule drug development has always lived by: evidence must be “commensurate with risk.”
4. Should AI be regulated like a medicine?
A recurring theme came straight from the life sciences: regulate AI like a drug – an idea that might make anyone who has sat through a Phase III wince. In ‘A Conversation with AI Visionary Gary Marcus‘, the NYU professor emeritus Gary Marcusopens in new tab/window called for “FDA-style review, and then auditing after.” He’s not worried about scientific tools like AlphaFold, which don’t leak private data or hack other systems, but about agents that go rogue on the open internet because they can’t reliably follow rules. Those, he said, should be recalled like “cars with exploding gas tanks.”
Gaia Marcusopens in new tab/window (no relation to Gary) is Director of the Ada Lovelace Instituteopens in new tab/window, an independent research institute which has been researching AI regulation since 2020. In the fireside chat ‘What We Know About What the Public Think About AI – and Why You Should Care’, which I moderated, she argued along similar lines: “I want to know that if something goes wrong, you can actually act to take it off the market.” She suggested food-safety-style issue tracking, and perhaps even professional licensing for the people who build these systems.
5. Public trust is now a force in R&D
Ada’s pollingopens in new tab/window found that 89% of UK adults think AI products shouldn’t be rolled out until they are proven safe, even if doing so slows things down. Americans are warier still: half now say they are more concerned than excited about AI, up from 37% in 2021, and only one in ten say the reverseopens in new tab/window. Gaia said people in Ada’s deliberations were “frankly astounded” that AI assistants reach the public with only the guardrails their makers provide.
As she put it: “Public opinion eats politics eats policy for breakfast.” In other words, policymakers who don’t proactively address risks upstream will be left making repeated ad hoc interventions in response to public outcry. For R&D, that public license is not abstract: it decides who gets access to patient data.
Nor is the public as uninformed as the industry likes to claim: “People are experts in their own lives. They may not be able to tell you how LLMs work… they do understand impacts.”
On the panel ‘Who Shapes AI’s Destiny? Power, Control, & Responsibility’, Audrey Tangopens in new tab/window, Taiwan’s Cyber Ambassador and former digital minister, described “alignment assemblies” that text 200,000 random people to recruit citizens’ panels to help draft codes of conduct for AI. One result: after Taiwan required ads on platforms such as Meta and Google to carry a verified digital signature, making the platforms jointly liable for unsigned scam ads, scam ads fell by more than 94%opens in new tab/window.
Gaia warned of the stakes: “If we’re not to have a generation of angry young men and women, which I personally am really worried about, I think we need to be clear that where we govern we’re governing in the public interest.”
The long game
Daphne ended on a sobering note. “As a medical profession, we have cured very few things outside of infectious diseases and a few childhood cancers. We can manage disease, but we don’t really cure disease.” Her hope for the next five to ten years is a repeatable path from data to disease-modifying drugs, and AI diagnostics that catch disease before it starts. Biology will set the pace. But as Audrey reminded the summit: “We, the people, are truly the superintelligence, and we shape AI.”
Related Articles:
Adopting AI for systematic literature reviews — quickly, but cautiously
The cost of waiting: why delaying AI adoption is a strategic risk for R&D
Contributor
SK