Why Simple Maths Keeps Beating Billion-Parameter AI Cell Models

The uncomfortable result the field cannot ignore#

Picture a laboratory that spends months training a vast artificial intelligence model on the genetic profiles of tens of millions of individual cells, only to find that a statistics method from a first-year textbook does the same job just as well. That is roughly where a growing part of computational biology now finds itself, and the discomfort is spreading through conference halls, preprint servers and journal clubs across the world.

Over the past year, a run of careful benchmarking studies has landed on a blunt conclusion. The single-cell foundation models that were meant to become the AlphaFold of cell biology often fail to beat plain, old-fashioned linear methods on the tasks that matter most to working scientists. One analysis in Nature Methods put it plainly in its title: deep-learning-based gene perturbation prediction does not yet outperform simple linear baselines (Nature Methods, 2025). That single sentence has become a rallying point for a wider debate about hype, evidence and where the field should spend its money next.

This is not a story about AI failing. It is a story about scientists doing their job properly: testing a bold claim hard, and reporting what they find even when the answer is awkward.

Background: what a single-cell foundation model actually is#

To follow the argument, it helps to understand three ideas.

The first is single-cell RNA sequencing, often shortened to scRNA-seq. Every cell in your body carries the same DNA, yet a neuron behaves nothing like a liver cell. The difference comes from which genes are switched on. Single-cell sequencing measures the activity of thousands of genes inside one individual cell at a time, producing a long list of numbers that describes that cell's state. Do this for millions of cells and you get an enormous table: cells down one side, genes across the top, activity levels in between.

The second idea is the foundation model. This is the same recipe behind chatbots. You take a huge neural network, feed it staggering amounts of unlabelled data, and let it learn general patterns without being told what to look for. Afterwards you adapt, or "fine-tune", that pre-trained model for specific jobs. The bet in biology was simple and seductive. If a model reads the gene-activity profiles of tens of millions of cells, it might learn the underlying grammar of the cell, in the way a language model learns the grammar of English.

The third idea is the analogy that launched the whole enterprise. Researchers describe treating "each cell like a sentence and genes like words", so that the model learns the language of cells (Experimental & Molecular Medicine, 2025). Two models became famous. Geneformer was trained on tens of millions of cells and published in Nature (Nature, 2023). scGPT followed with a generative approach and appeared in Nature Methods (Nature Methods, 2024). Both promised a general-purpose engine for cell biology: annotate cell types, integrate messy datasets, and even predict what happens when you switch a gene off.

Against these giants sit the "baselines". A baseline is a deliberately simple method used as a yardstick. Principal component analysis, which compresses the big table into a handful of summary numbers, is one. Ordinary linear regression is another. The simplest of all is the "no change" model, which predicts that nothing happens. If a model with hundreds of millions of parameters cannot beat "assume nothing happens", something is wrong with the story.

The benchmark that set the tone#

The Nature Methods paper by Constantin Ahlmann-Eltze, Simon Anders and Wolfgang Huber has become the reference point for the whole discussion. The team compared five foundation models and two other deep-learning methods against those deliberately simple baselines, testing how well each could predict the changes in gene activity after one or two genes were perturbed (Nature Methods, 2025).

The result was stark. None of the sophisticated models beat the baselines. For double perturbations, an "additive" model that simply added the effects of two single changes held its own. For the harder job of predicting the effect of a perturbation never seen in training, none of the models did better than assuming no change at all (Nature Methods, 2025). The same group had already shown that a lean linear technique they call LEMUR handles multi-condition single-cell data efficiently, reinforcing the point that much of the useful signal is linear and accessible (Nature Genetics, 2025).

The authors were careful about their framing. The word "yet" in the title matters. Their argument is not that deep learning is doomed, but that critical benchmarking must guide method development rather than trail behind the marketing.

Five reasons the simple methods keep winning#

1. The biological signal is more linear than we assumed#

The most surprising lesson is about biology, not computing. A large share of the structure in single-cell data can be captured by linear combinations of gene activity. If the relationships that matter are mostly additive, then a linear model is not a crude approximation. It is close to the right tool. That is why principal component analysis and regression keep matching networks that are thousands of times larger.

2. Benchmarks were flattering the big models#

Several 2026 studies argue that early evaluations were not testing the right things. A benchmark of single-cell foundation models for real-world RNA-seq data integration found that performance drops sharply once you move from clean, curated datasets to the messy data labs actually generate (bioRxiv, 2026, preprint, not peer reviewed). A separate multi-dimensional benchmark, VCBench, reached similar conclusions about the gap between headline scores and practical use (bioRxiv, 2026, preprint, not peer reviewed).

3. Zero-shot claims do not always survive contact with data#

A big selling point of foundation models is "zero-shot" use, meaning you apply them straight out of the box without task-specific training. Yet a benchmark of zero-shot embeddings for reconstructing how cells change over time found that traditional approaches based on highly variable genes remained more robust, because the large models tended to over-compress subtle temporal signals and flatten branched biological trajectories into artificial straight lines (bioRxiv, 2026, preprint, not peer reviewed).

4. Attention did not add the magic ingredient#

Foundation models rely on a mechanism called attention, which is supposed to learn which genes influence which. When researchers examined the regulatory relationships pulled from that attention in models like Geneformer and scGPT, they found trivial gene-level baselines could match them, with attention contributing little extra signal for perturbation prediction. If the clever internal machinery is not carrying the load, the case for the added complexity weakens.

5. Parameter-free methods are catching the models flat-footed#

Perhaps the sharpest challenge comes from work showing that parameter-free representations, which involve almost no training at all, can outperform single-cell foundation models on downstream benchmarks (arXiv, 2026, preprint, not peer reviewed). When a method with essentially no learned parameters beats one with hundreds of millions, it forces a hard question about what the extra scale is buying.

The other side of the argument#

Good science resists a tidy narrative, and the foundation-model camp has a real reply. A study in Genome Biology argued that the debate has been distorted by poor choice of metrics. When you evaluate the whole integrated representation rather than one narrow prediction task, the space learned by single-cell foundation models preserves biologically meaningful relationships between cell types better than simple baselines, a strength earlier work had overlooked (Genome Biology, 2025).

A 2026 evaluation in Advanced Science struck a middle note, finding that the models can perform many tasks well while still struggling with rare cell types and data quality (Advanced Science, 2026). The honest summary is that foundation models are neither the revolution their loudest promoters promised nor the failure their critics imply. They are early tools that shine in some settings and stumble in others, and the field is still working out which is which.

There is also the matter of maturity. Protein structure prediction looked shaky for years before AlphaFold, and today's cell models are far younger than the language models they borrow from. "Not yet", as the Nature Methods authors wrote, leaves the door open.

Why this matters far beyond one subfield#

The stakes are not academic. Single-cell foundation models sit at the centre of the "virtual cell" vision, the goal of simulating a living cell in software well enough to predict how it responds to a drug or a mutation. Huge public and private investment is flowing toward that dream. If the current generation of models cannot reliably beat linear regression on perturbation prediction, funders and companies need to know before they commit another round of capital.

For the global scientific community, the episode is a case study in research integrity. It shows the value of shared, adversarial benchmarks, of publishing negative results in top journals, and of baselines chosen to be genuinely hard to beat rather than easy to flatter. Those habits protect every data-heavy discipline now reaching for foundation models, from climate science to materials discovery. The single-cell field is, in effect, running an early stress test of how AI hype meets scientific scrutiny, and the rest of science should watch closely.

Key takeaways#

The most important insights from this debate are these. First, across several careful studies, current single-cell foundation models often do not beat simple linear baselines on core tasks such as perturbation prediction. Second, a large part of the useful signal in single-cell data appears to be linear, which is a finding about biology as much as about algorithms. Third, headline benchmark scores can flatter big models, and performance frequently drops on the messy, real-world data that laboratories actually produce. Fourth, the picture is genuinely mixed, and peer-reviewed work shows foundation models can preserve biological structure better than baselines when judged with the right metrics. Fifth, the wider lesson is about method, not machinery: strong baselines, honest benchmarks and published negative results are how a field keeps itself honest.

Frequently asked questions#

Are single-cell foundation models useless? No. The evidence says they do not consistently beat simple methods on certain benchmark tasks today, especially perturbation prediction. On other jobs, and with better metrics, they can capture biological relationships that baselines miss (Genome Biology, 2025).

What is a "baseline" and why does it matter? A baseline is a deliberately simple reference method, such as linear regression or "assume no change". It sets the bar a fancier model must clear. If a complex model cannot beat a strong baseline, the extra complexity is hard to justify (Nature Methods, 2025).

Does this mean AI has hit a wall in biology? Not at all. The same period has seen major AI advances in protein and genome analysis. The finding is specific: for some single-cell tasks, today's models have not yet earned their complexity, which is a normal stage in a young field.

Why would simpler methods win? Because much of the structure in single-cell data seems to be linear, meaning simple additive relationships carry a lot of the signal. When that is true, linear methods are close to the correct tool rather than a rough shortcut.

What should I trust: journals or preprints? Peer-reviewed papers carry more weight. Several of the most recent benchmarks are preprints and are labelled here as not peer reviewed. Treat their numbers as promising but provisional until reviewed.

What happens next? Expect better benchmarks, clearer reporting standards, and models designed to beat strong baselines by design rather than by press release. The "yet" in the debate suggests improvement is expected, not ruled out (Nature Methods, 2025).

Glossary#

Single-cell RNA sequencing (scRNA-seq): A technique that measures the activity of thousands of genes inside individual cells, revealing what each cell is doing rather than averaging across a whole tissue.

Foundation model: A large neural network pre-trained on vast unlabelled data to learn general patterns, then adapted to specific tasks. The same idea underlies large language models.

Baseline: A deliberately simple method used as a yardstick. If a complex model cannot beat it, the added complexity is questionable.

Perturbation prediction: Forecasting how a cell's gene activity changes after a gene is switched off or altered, a key test of whether a model understands cellular cause and effect.

Principal component analysis (PCA): A classic statistical method that compresses many measurements into a few summary numbers that capture most of the variation.

Zero-shot: Using a pre-trained model on a new task without any task-specific training, straight out of the box.

Attention: The mechanism inside modern neural networks that weighs which inputs, here which genes, matter most for a prediction.

Virtual cell: The long-term goal of simulating a living cell in software accurately enough to predict its response to drugs, mutations or environmental change.

References#

1. Ahlmann-Eltze C, Anders S, Huber W. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods, 4 August 2025. https://www.nature.com/articles/s41592-025-02772-6
2. Biology-driven insights into the power of single-cell foundation models. Genome Biology, 3 October 2025. https://link.springer.com/article/10.1186/s13059-025-03781-6
3. Theodoris CV, et al. Transfer learning enables predictions in network biology (Geneformer). Nature, 2023. https://www.nature.com/articles/s41586-023-06139-9
4. Cui H, et al. scGPT: toward building a foundation model for single-cell multi-omics. Nature Methods, 2024. https://www.nature.com/articles/s41592-024-02201-0
5. Single-cell foundation models: bringing artificial intelligence into cell biology. Experimental & Molecular Medicine, 2025. https://www.nature.com/articles/s12276-025-01547-5
6. Ahlmann-Eltze C, Huber W. Analysis of multi-condition single-cell data with latent embedding multivariate regression (LEMUR). Nature Genetics, 2025. https://www.nature.com/articles/s41588-024-01996-0
7. Evaluating the utilities of foundation models in single-cell data analysis. Advanced Science, 2026. https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202514490
8. Benchmarking single-cell foundation models for real-world RNA-seq data integration. bioRxiv, 2026. Preprint (not peer reviewed). https://www.biorxiv.org/content/10.64898/2026.04.17.719314v1
9. VCBench: A multi-dimensional benchmark for single-cell foundation models. bioRxiv, 2026. Preprint (not peer reviewed). https://www.biorxiv.org/content/10.64898/2026.06.18.733146v1
10. Benchmarking zero-shot single-cell foundation model embeddings for cellular dynamics reconstruction. bioRxiv, 2026. Preprint (not peer reviewed). https://www.biorxiv.org/content/10.64898/2026.03.10.710748v1
11. Parameter-free representations outperform single-cell foundation models on downstream benchmarks. arXiv, 2026. Preprint (not peer reviewed). https://arxiv.org/pdf/2602.16696