2026-07-30
Two Weizmann scientists argue in Cell that scaling won't yield a universal model of biology, as life is emergent, not reducible to its parts; they call for process-based 'world models'.
AI has been winning real battles in biology. AlphaFold turned protein-structure prediction into a routine tool. Evo 2 pushed genome modeling and design across all domains of life. AlphaGenome predicts the effects of variants in regulatory regions. There is a whole class of transformers trained on single-cell transcriptomes, and the Human Cell Atlas team wants to unify them into a foundation model. Following that line, the field has started dreaming bigger: build ever-larger models on ever-more data until you get a universal model of biology, even a virtual cell that simulates a whole cell.
This Cell perspective pours cold water on that. The authors are Yonatan Stelzer (molecular cell biology) and Amos Tanay (computer science and applied mathematics, also molecular cell biology) at the Weizmann Institute. Their counter-thesis is blunt: bigger models and more data will not produce a universal model of biology that actually reads life, because life is not the sum of its parts.
That is a real problem, not a philosopher's hobby. It decides where the bio-AI compute and data spent every year should actually go.
This is a perspective, with no experiments and no benchmarks. The full text sits behind Cell's paywall; what is public is the abstract and the reference list. The argument structure below is reconstructed from those, plus the keyword list.
The case rests on two classic pieces of thinking. One is the physicist Anderson's 1972 essay "More is different": the behavior of a complex system cannot be derived from the laws of its components, because new regularities emerge at larger scales. The other is the evolutionary biologist Mayr's stance that biological systems are shaped by contingent evolutionary history, so there is no set of universal physical laws you can compute them from.
Put together, the authors' point is that AlphaFold solves a problem at the "parts" level (what shape a single protein folds into), the kind of reductionist task where big models excel. But once the problem moves up to cells, tissues, and developmental processes, behavior emerges from interactions and time; feeding an even bigger model the full inventory of parts and sequences does not give you the whole. They give the current line a critical name: explanatory reductionism, the expectation that lower-scale mechanisms will explain higher-scale phenomena.
What they want instead, from the keywords and references, is to anchor AI in canonical biological processes and build data-driven "world models" with explicit mechanistic links across molecules, cells, and their dynamics in space and time, instead of end-to-end black boxes. Their own spatiotemporal modeling of gastrulation, cited in the references, is the template.
A perspective has no quantifiable results. The paper runs no controlled experiments and reports no numbers on any biological task.
Strictly, what it delivers is a judgment and an agenda, not an experimental conclusion: the universal-model route will not work, and bio-AI resources should shift toward process-based, mechanistic, multi-scale world models. It does not run a head-to-head numerical comparison against AlphaFold or the virtual-cell program. That should be stated plainly so readers do not mistake this for an empirical comparison.
For AI practitioners, especially in AI4Science, this argument deserves attention. It does not deny the AlphaFold-class achievements. It denies the infinite extrapolation of that success line. The prevailing bio-AI narrative is "more data, bigger model," structurally identical to the scaling laws of language models. This paper says in a top biology venue that for biology, on the scales where emergence and evolution dominate, that paradigm hits a wall.
The practical implication is route selection. If you work at the sequence level on a well-defined target (structure, regulatory effects, variant prediction), large models keep helping. If you work on dynamic systems like cell behavior, development, or disease processes, the claim here is that more compute is not enough; you have to build the mechanistic structure into the model. It nudges the bio-AI agenda away from "build a bigger universal model" toward "custom mechanistic models for specific biological processes."
The biggest limitation is the form of the article itself. It is a perspective, not a proof. It does not build a world model along its own prescription and then show experimentally that it beats a general large model. So the claim that "the world-model route is better than the universal-model route" is still a claim, not a verified conclusion.
Second, the full text was not retrieved (Cell paywall; no open version in Europe PMC or PubMed Central). Everything above about the positive proposal (canonical processes, multi-scale mechanisms, world models) comes from the public abstract, keywords, and reference list; the finer argument and cases in the body were not seen. Anyone who wants to follow the prescription has to go back to the original.
Third, one doubt from a practitioner's read: the critique targets explanatory reductionism, but the public material is thin on how the proposed world model is actually built and how it avoids becoming just another black box. A program against black-box scaling whose substitute is not concrete enough in engineering risks being read as another call to arms, with no actionable route behind it.