A landmark Cell viewpoint from Weizmann Institute researchers Yonatan Stelzer and Amos Tanay asks a question most of AI biotech doesn't want to hear: what if scaling alone can't make machines understand life?
The paper, "Why machines don't speak biology: Toward native biological language models," doesn't argue that AI is failing in biology. It argues that the field is optimizing for the wrong thing — and that the real bottleneck isn't data, GPUs, or model size. It's that biology doesn't behave like a language model's training corpus.
Why AlphaFold worked — and why it can't be replicated
AlphaFold's success looks like proof that "just scale it" works. But the authors break down exactly why it succeeded: protein folding space is not random. Natural proteins are shaped by evolution, which has produced modular, repeated, conserved structure families. Multiple sequence alignments gave AlphaFold evolutionary constraints, and decades of experimentally solved structures gave it reliable ground truth. In other words, AlphaFold didn't learn biology from scratch — it learned the statistical residue of evolution.
That success also defines its boundary. AlphaFold is superb at interpolating within the evolutionary-explored structure space, but it can't reliably say which mutations cause misfolding, and it doesn't magically solve de novo design of large complex proteins. The general lesson: AI generalizes broadly only when a problem has rich, stable prior structure and the training data covers it well.
Cells and tissues are not just bigger proteins
Moving from proteins to genomes, cells, and tissues isn't a matter of adding a few orders of magnitude of data. A protein is a few hundred amino acids; a genome has tens of thousands of genes and over a million regulatory elements; a cell holds on the order of 10⁶–10¹⁰ protein molecules; a tissue involves billions of interacting cells. The dimension, organization, and dynamical complexity all change as you scale.
The deeper problem: life isn't a machine assembled from static parts. Cells and tissues are processes unfolding in time and space, with feedback, regulation, and interaction across levels. That cross-scale organization is the core challenge a general biology AI has to confront — and it's precisely what scaling alone doesn't address.
The game-engine test: generation is not understanding
The paper's sharpest analogy is a physics engine. A model that truly understands a car doesn't just generate a plausible video of a car accelerating through snow — it computes the interaction of engine output, gravity, friction, and the environment. By that test, today's biology foundation models mostly generate plausible videos. They interpolate within the distribution of data evolution left behind; they cannot relive the stochastic history that produced it.
This is why "emergence at scale" is a hope, not a strategy. Within a well-defined problem scale — predicting expression from DNA, predicting how an enhancer variant shifts expression, predicting drug effects on a pathway — AI is already hugely valuable. But the authors are explicit: deep learning is strong at interpolation and weak at extrapolation beyond its training distribution. Feeding bigger models more gene, protein, and cell data does not guarantee they will "emerge" an understanding of life.
The proposed fix: from parts to canonical processes
Stelzer and Tanay propose shifting the fundamental unit of modeling from isolated components to canonical biological processes — processes that repeat observably across experiments, with relatively stable spatial, temporal, and mechanistic structure. Canonical processes give AI a coordinate system: DNA replication, transcription, translation; the cell cycle; embryonic development; even viral infection and whole-brain connectomes.
Process-aligned modeling has concrete advantages: you get data about the process itself, not static snapshots; variability between individuals becomes quantifiable; observations from different experiments can be mapped onto a common axis; and deviations caused by mutations, drugs, or environmental perturbations gain real interpretable meaning.
That reframes the data problem. The real scarcity isn't more single-cell data — it's process-aligned, spatiotemporally resolved, in-situ, perturbation-aware multi-scale data. Data should be recorded in native environments (regulatory elements in real epigenomic context, cells in native tissue), mapped to explicit spatiotemporal coordinates, and prioritized for early, direct responses to perturbation rather than terminal states after multi-level feedback.
Toward a causal biological world model
The end goal is a process-based biological world model that reconstructs mechanisms and causal relationships rather than reproducing observations. To get from today's single-scale foundation models to such a multi-scale world model, the authors offer three principles:
- Distillation — strip away redundant correlations from existing models, keeping causally meaningful mechanisms (e.g., separating true regulatory drivers from incidentally co-evolving mutations).
- Standardization — build reusable modules (a standardized protein structure/function model as a component for a cell world model; a standardized cell stimulus-response model as a component for a tissue world model).
- Reconnection — place distilled, standardized models back into the spatiotemporal coordinate system of canonical processes so protein, cell, and tissue models actually connect mechanistically. This, the authors say, is where the real frontier lies.
What this means for the industry
For virtual cell efforts, AI drug discovery, and the "unified biology model" race, this paper is a warning shot. It suggests that companies and labs competing purely on model size, data volume, and GPU count may be scaling the wrong axis. The durable moat will be process-aligned experimental infrastructure: in-situ multi-omics, time-resolved perturbation data, and standardized process ontologies that connect molecular, cellular, and tissue scales.
It also reframes evaluation. If your model can only reproduce what it was trained on, you don't yet have a biology model — you have a very expensive autoencoder. The test of progress is causal, out-of-distribution prediction: does the model correctly predict the early consequences of a perturbation it has never seen?
Actionable takeaways
- If you're building a bio foundation model, invest in perturbation and time-course data collection, not just scale — early, direct responses beat terminal states.
- If you're evaluating one, ask for out-of-distribution causal prediction results, not reconstruction benchmarks.
- If you're choosing a research direction, bet on process-aligned, in-situ, multi-scale data infrastructure — that's where the field's real scarcity lies.
Reference: Stelzer, Y. & Tanay, A. "Why machines don't speak biology: Toward native biological language models." Cell (2026). doi:10.1016/j.cell.2026.07.003
Frequently Asked Questions
Doesn't AlphaFold prove that scaling works?
AlphaFold succeeded because protein folding space has structure: evolution produced modular, repeated, conserved families; multiple sequence alignments gave evolutionary constraints; decades of solved structures gave ground truth. It learned the statistical residue of evolution, not biology from scratch.
Why can't AI speak biology yet?
Because the bottleneck is process, not data or compute. Cells and tissues are processes unfolding in space and time, and their cross-scale organization cannot be reached by scaling alone. The paper proposes moving from parts to canonical processes, and ultimately to causal biological world models.
What does this Cell viewpoint mean for AI drug discovery and virtual cells?
It is a warning shot: teams competing purely on model size, data and GPUs may be scaling the wrong axis. The durable moat is process-aligned experimental infrastructure (in-situ multi-omics, time-resolved perturbation data), and the test of progress shifts from reconstructing training data to out-of-distribution causal prediction.
