Lab Newsletter — August 15, 2026: A Turing Test for the Cell

AlphaFold made a protein’s shape something you could look up. The field’s next moonshot is bolder: a virtual cell you can query. The dream, stated plainly, is this — give a model a starting cell state and a perturbation (knock down this gene, add this drug), and have it predict the gene-expression response you would otherwise have to run an experiment to see. A cell you can ask “what if?” Two things happened in 2025 that make this the right moment to look hard at that ambition: the field built itself a scoreboard, and the scoreboard delivered a humbling first verdict.

🎯 The ambition: a cell you can query

The clearest expression of the goal comes from Arc Institute, which has turned the dream into a public contest. “Virtual Cell Challenge: Toward a Turing Test for the Virtual Cell” (Roohani et al., Cell, 2025) is explicitly modeled on CASP — the protein-structure competition whose annual, honest benchmarking helped summon AlphaFold. The 2025 edition set a deliberately hard task: predict the effect of single-gene CRISPRi perturbations in a held-out cell type — H1 human embryonic stem cells, a distribution shift from the K562/A375 lines models usually train on — scored on a purpose-built dataset of ~300,000 single-cell profiles across 300 perturbations. The response says a lot about where the field’s head is at: over 5,000 people from 114 countries and more than 1,200 teams entered. Alongside it, Arc released State, a first-generation virtual cell model that “captures the cell-type-specific effects of genetic, chemical, and cytokine perturbations,” trained on 167 million observational cells and over 100 million perturbational cells across 70 human cell contexts. Why it matters for the lab: this is our Human Cell Simulator ambition, made concrete — and handed a shared, honest yardstick.

🧬 The tools: single-cell foundation models

The instruments people bring to this contest are single-cell foundation models — the transcriptomic cousins of the language models that reshaped NLP. Geneformer (Theodoris et al., Nature, 2023) is an attention-based transformer pretrained on ~30 million single-cell transcriptomes (its successor, 95 million) that, in the authors’ words, “gained a fundamental understanding of network dynamics, encoding network hierarchy in the attention weights … in a completely self-supervised manner.” scGPT (Cui et al., Nature Methods, 2024) is a generative pretrained transformer built over 33 million cells, offering one backbone for cell-type classification, multi-omic integration, gene-network inference, and perturbation modeling. The bet behind all of them is the recipe that worked for text and, arguably, for proteins: pretrain on the ocean of unlabeled data, then fine-tune to predict what a cell will do. Why it matters for the lab: these are the open, buildable models — the BioImage Model Zoo / BioEngine ethos — that a lab can fine-tune to its own biology rather than treat as a black box.

🧭 The honest reckoning — and why it’s the right story for us

Here is where the year earned its keep. A Nature Methods benchmark by Constantin Ahlmann-Eltze, Wolfgang Huber and Simon Anders (EMBL / Heidelberg), titled with unusual candor — “Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines” — pitted five foundation models and two other deep-learning methods against deliberately simple controls, and none of them won. For combinations of two genes whose individual effects were known, the deep models did no better than a simple additive model; for genes never seen in training, they did no better than predicting the average of the training perturbations. A separate benchmark found even the mean of the training examples beat scGPT and scFoundation. This is not a reason to look away — it is exactly why a lab building a virtual cell should build the way this one does: with validation and benchmarking front and center, treating a predicted perturbation as a hypothesis, and measuring against CASP-style challenges rather than self-reported wins. As the authors put it, the result “highlights the importance of critical benchmarking in directing and evaluating method development.” And the pieces connect: a predictive cell model is precisely the instrument an autonomous research agent would need to decide which experiment to run, and spatial context is a tissue layer these dissociated-cell models still lack. The honest frontier is stark — perturbation data is scarce, “beating the mean” for a truly novel gene is unsolved, and a transformer’s confidence is not a measurement. That’s the same prove-it discipline we keep returning to, and here it cuts closest to home. The AlphaFold moment for the cell hasn’t arrived. But the honest scoreboard that could one day summon it just went up — and that, for a field prone to hype, is its own kind of progress.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits; a Grok-based replacement is wired and awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.