The Virtual Cell
Everything you need to understand AI virtual cell simulation — from first principles to state of the art — with no biology background required.
What Is a Cell Actually Doing?
Your body has ~37 trillion cells. Every single one contains the same DNA — the same complete instruction manual for building a human. A liver cell and a neuron have identical DNA. What makes them different is which instructions are currently being read.
This reading process is called gene expression. At any moment, a cell is reading some genes loudly, some quietly, and most not at all. The result — the complete snapshot of which genes are active and how much — is called the cell's transcriptome.
Think of DNA as a massive cookbook with 20,000 recipes. The transcriptome is the list of which recipes are being cooked right now, and how many portions of each. The cookbook never changes. The cooking does — constantly.
Measuring this snapshot is now possible with a technology called scRNA-seq (single-cell RNA sequencing). It reads the transcriptome of a single cell and gives you a vector of ~20,000 numbers — one per gene — representing activity levels.
To measure a cell's transcriptome, you have to destroy the cell. The measurement kills it. You can never watch the same cell before and after something happens — you can only compare populations of cells. This is a core source of noise in everything that follows.
Three Distinct Concepts
| Term | What It Is | Stability |
|---|---|---|
| Genome | Full DNA — all 20,000 genes. The complete instruction manual. | Fixed. Same in every cell of your body. |
| Transcriptome | Which genes are being read right now, at what level. A snapshot vector of ~20,000 numbers. | Dynamic. Changes with cell type, environment, signals, time. |
| Proteome | Which proteins are currently built from those transcripts. | Dynamic. Lags behind transcriptome by hours. |
What Is a Perturbation?
A perturbation is simply: you interfere with the cell and see what happens.
The most common type used in this field is CRISPRi — a molecular tool that physically clamps onto a specific gene in the DNA, blocking the cell's reading machinery from transcribing it. You're saying "gene X, shut up" at the source.
CRISPRi acts at the DNA level — it prevents the gene from ever being read. No RNA gets produced. No protein gets built. The effect is upstream of everything. (A different technique, RNAi, cuts RNA after it's produced — but CRISPRi is the dominant tool in this field.)
Why does silencing one gene matter? Because genes don't operate in isolation. They form networks. Each gene's protein product can regulate other genes. Silencing gene X might:
Gene Y goes up
Another gene activates to compensate for the loss of X's function. The cell tries to maintain homeostasis.
Gene Z explodes
Gene X was suppressing Z. Remove X, and Z goes unbraked — potentially overactivating an entire pathway.
50 genes shift
Z's protein activates A and B, which suppress C, which releases D... the ripple propagates across the network.
The transcriptome snapshot
After all the cascades settle, you measure the new transcriptome. That's the "perturbation response" — the end state of the whole network reaction.
The perturbation response reveals the hidden wiring of the cell. By watching what cascades when you knock out gene X, you learn what X was connected to, what it controlled, and how the network compensates. This is one of the primary ways we understand gene function.
Why Predict It?
Right now, running a perturbation experiment means: silencing a gene, growing cells for days, killing them to measure the transcriptome, analyzing the data. Per experiment: days of work, thousands of dollars, cells destroyed.
The scale problem is staggering:
| Experiment Type | Number of Possibilities | Status |
|---|---|---|
| Single gene knockouts | ~20,000 | Partially covered by Replogle dataset |
| Two-gene combinations | ~200,000,000 | Tiny fraction measured |
| Three-gene combinations | ~1,300,000,000,000 | Essentially unmeasured |
| Single gene × cell type × condition | Astronomical | Unmeasured |
Train a model on a fraction of these experiments. Have it accurately predict the rest. A virtual cell you can query like a database — "what happens if I knock out genes X and Y together in a stem cell under oxidative stress?" — answered in milliseconds, for free, without destroying a single real cell.
This is the complete value proposition. Not replacing experiments, but making each experiment do the work of thousands — multiplying scientific reach by orders of magnitude.
Gene → RNA → Protein → Drug
This is the chain that connects all the concepts. Understanding it is key to understanding why the transcriptome is the right thing to measure and predict.
The Central Dogma
Information in a cell flows in one direction:
How Drugs Connect to Gene Expression
Most drugs work by binding to a protein — not a gene. Ibuprofen blocks COX proteins. Statins block cholesterol-synthesis enzymes. Cancer drugs block proteins that drive cell division.
But proteins regulate genes. Many proteins are transcription factors — their job is to bind DNA and switch genes on or off. So when a drug blocks protein X:
Drug blocks protein X → X can no longer activate genes A, B, C (those go quiet) → X can no longer suppress genes D, E, F (those become active) → D and E's proteins now regulate more genes → cascade propagates...
The transcriptome after the drug is the full footprint of everything that happened. Predicting it means predicting the drug's complete downstream effect — including intended effects and side effects.
Why Signals From Other Cells Matter
Your liver cells are constantly receiving molecular messages from neighboring cells, the bloodstream, and distant organs — hormones, cytokines, growth factors. Each of these lands on a receptor, triggers a transcription factor, and reshapes your transcriptome.
This means the transcriptome is not just internally determined — it's a continuous conversation between the cell and its entire environment. This makes prediction harder: you need to know the cell's internal state and the context it's sitting in.
Cell Identity, Attractors & Why Cells Don't Randomly Drift
If the transcriptome is constantly fluctuating, why doesn't your liver cell randomly become a neuron? This question leads to one of the most important concepts in cell biology.
The Epigenetic Landscape (Waddington, 1957)
Think of all possible cell states as a landscape with hills and valleys. A liver cell sits in the "liver valley." Fluctuations happen constantly, but the valley's shape pulls it back. Getting to the "neuron valley" would require climbing over a huge hill.
What Creates the Valleys — Gene Regulatory Networks
Liver cells stay liver cells because of self-reinforcing gene circuits. Master regulator genes (like FOXA2 for liver) activate each other and simultaneously suppress the master regulators of other cell types. It's a self-locking system.
Three Layers of Protection Against Drift
Gene Regulatory Networks
Master regulators activate each other and suppress competing identity programs. Small fluctuations get absorbed and corrected.
Epigenetics
Neuron genes in a liver cell are physically wrapped tight around protein spools (histones). The transcription machinery literally can't reach them. This constrains the transcriptome's range of fluctuation.
Neighbor Signals
Surrounding liver cells constantly send "stay liver" signals. Remove a cell from its tissue context and identity starts to drift — this is why cells in a dish behave differently than in an organ.
Cancer
Cancer is not a cell becoming another normal cell type — it's a cell whose regulatory network is corrupted enough to lock into a pathological attractor. One with no brakes on division.
Shinya Yamanaka proved cell identity is not permanent. By forcing expression of just 4 genes (Oct4, Sox2, Klf4, c-Myc) in a skin cell, he kicked it over the valley wall into a pluripotent stem cell state — capable of becoming almost any cell type. This was shocking: it proved cell identity is a stable attractor, not a locked destiny.
Why This Matters for the Virtual Cell
The attractor landscape picture is exactly what models are trying to learn. Not just "what gene expression level does gene X land at after a perturbation" — but the shape of the landscape itself. Where are the valleys? How deep? What perturbations tip a cell from one to another?
This is why perturbation prediction is not just a regression problem. It's learning a dynamical system refined by billions of years of evolution.
Key Datasets — The "ImageNet" of This Field
Just as computer vision built on ImageNet, cell biology perturbation prediction has a few canonical datasets that everyone benchmarks on. These datasets are the shared currency of the field.
| Dataset | Technique | Scale | Significance |
|---|---|---|---|
| Adamson et al. 2016 | CRISPRi knockdown of ~87 genes during ER stress response | ~65K cells | Small but widely used early baseline. One of the first at this scale. |
| Norman et al. 2019 | CRISPRa activation of 105 genes; 131 two-gene combinations | ~100K cells | First large-scale combinatorial perturbation dataset. Opened the question of multi-gene interaction prediction. |
| Replogle et al. 2022 | Genome-wide CRISPRi — nearly 10,000 genes, 2.5M cells | ~2.5M cells | The current gold standard. Large enough to reveal that many "impressive" models were memorizing patterns. Changed what was possible to claim. |
Before Replogle, models trained on small datasets could appear to work well — but were actually fitting to statistical noise. At genome-wide scale, these artifacts disappear and only genuine signal survives. Replogle is why the field knows deep learning doesn't yet decisively beat simple baselines: there's enough data to see through overfitting.
The Virtual Cell Challenge 2025
Organized by the Arc Institute (affiliated with Chan Zuckerberg Initiative and Stanford), announced at NeurIPS 2025. This is the field's flagship competition — the closest thing to a shared Turing test for virtual cells.
5,000+ people
Across 114 countries. 1,200+ teams submitted results. 300+ final submissions.
$250K total
$100K first place, $50K second, $25K third, $100K generalist prize.
~300K cells
H1 human embryonic stem cells. 300 CRISPRi perturbations. Purpose-built for the challenge.
The Task — Why It's Hard
The challenge used H1 human embryonic stem cells — a cell type no model had ever seen during training. Competitors received the list of 300 genes to perturb and had to predict the resulting transcriptome for each.
A model trained on cancer cell lines (K562, etc.) must transfer its "understanding of biology" to stem cells — where the same gene plays a completely different role, embedded in a completely different regulatory network. It's like asking someone who learned chess to play Go. The pieces move differently. Does your model understand biology, or has it memorized cancer cell patterns?
Scoring Metrics
The challenge used three metrics to capture different aspects of prediction quality:
| Metric | What It Measures | Why It Matters |
|---|---|---|
| PDS — Perturbation Discrimination Score | Can you distinguish which gene was knocked out from the predicted expression profile? Are predictions specific enough to be differentiable? | Tests whether the model is actually capturing perturbation-specific biology vs. generic "cells look different when perturbed." |
| DES — Differential Expression Score | Do you identify the correct set of upregulated and downregulated genes? | Tests whether the model gets the direction of change right for the genes that matter. |
| MAE — Mean Absolute Error | Raw numeric accuracy across all 20,000 genes. | Sounds like the right metric — but is deeply misleading (see Metric Crisis section). |
Of 20,000 genes, maybe 50–200 actually change after a perturbation. The other 19,800+ stay roughly the same. If your model predicts "nothing changes" for everything, your MAE is still very low — you're right about 19,800 genes. The model gets rewarded for ignoring the biology. This is why most models underperformed baselines on MAE, and why the Generalist Prize used 7 metrics instead.
Winners & What Worked
Hybrid deep learning + classical statistics. Combined an improved scFoundation transformer architecture with protein language model embeddings (so the model chemically "knows" what each protein does, not just its expression statistics), plus classical statistical features as explicit biological hints. Used pseudo-bulk aggregation to denoise signal before prediction.
Metric-driven strategy. Pseudo-bulk representation with protein embeddings. Heavily optimized for the specific scoring metrics rather than general biological accuracy.
Statistical framework using only summary-level data with similarity-aware aggregation. No deep learning at all — pure statistics.
Flow matching generative model. Instead of predicting a single "answer," it predicts the full distribution of cell responses — capturing that different cells respond differently to the same perturbation. Pretrained on ~7M single cells from public datasets plus internal perturbation screens. Won the prize for most robust performance across all 7 evaluation metrics.
All top teams used pseudo-bulk aggregation — essentially averaging over cells before prediction as a denoising step. This is classical statistics saving neural networks from noise. The winning approach combined billion-parameter transformers with statistical cheat sheets. Pure end-to-end deep learning did not win.
The Four Dominant Technical Approaches
Foundation Models (the GPT analogy)
Treat genes like words. A cell's expression profile is a "sentence." Pretrain a transformer on millions of cells, then fine-tune for perturbation prediction.
Hybrid: Statistics + Deep Learning
Neural networks augmented with biological knowledge — protein embeddings, pseudo-bulk denoising, explicit gene feature engineering. The winning approach.
Generative / Flow Matching
Predicts the full distribution of cell states, not just the mean. Models cellular heterogeneity — different cells respond differently to the same knock-out.
Mechanistic / White-Box
Encodes known biological networks (KEGG pathways, protein interactions) as structure. Uses LLM reasoning to trace cascades step by step. Interpretable.
Approach 1: Foundation Models
Models: scGPT (pretrained on 33M cells, masked gene prediction like BERT), Geneformer (ranks genes by expression level, uses rank order as input to remove technical noise from absolute counts), scFoundation (continuous expression encoding + read-depth recovery task).
The bet: If the model "understands" gene expression deeply enough through pretraining, it can generalize to unseen perturbations and cell types.
The problem: Often can't beat a linear model on perturbation prediction specifically. The "language" analogy breaks down — genes don't have grammar, and the statistical structure of scRNA-seq data is different from text.
Approach 2: Hybrid (the winner)
The key innovation is not replacing deep learning but augmenting it with structured biological knowledge:
- Protein embeddings: Encode what each protein structurally and functionally is (from models like ESM), not just its expression statistics. The model knows gene X encodes a kinase before it even sees any data.
- Pseudo-bulk: Average cells with the same perturbation into a single profile before prediction. Eliminates single-cell technical noise.
- Explicit biological features: Tell the model which genes are typically differentially expressed, as a statistical prior.
Approach 3: Generative / Flow Matching
Traditional models predict: "after knocking out gene X, gene Y will be at level 4.2." But that ignores a biological reality: different cells respond differently. The population of perturbed cells has a distribution, not a single point.
Flow matching learns to transform one distribution into another. Input: the distribution of unperturbed cells. Output: the distribution of perturbed cells. This is related to diffusion models but faster and more stable to train.
Approach 4: Mechanistic (VCWorld)
Instead of learning correlations from data, VCWorld encodes known biology — signaling pathways, protein interaction networks — as explicit structure, then uses an LLM to reason through the cascade step by step. Each prediction comes with a traceable mechanistic explanation: "Gene X was silenced → TF Y lost its activator → Genes A, B, C went down → which caused..." — verifiable against known biology.
Trade-off: More interpretable and generalizes better to out-of-distribution cases, but limited by what's already known biologically. Can't discover truly novel mechanisms.
The Metric Crisis
This is the central intellectual drama of the field right now. The debate has evolved into a precise diagnosis — which is itself progress.
"Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines" — Seven foundation models (scGPT, scFoundation, Geneformer, etc.) benchmarked against simple baselines. None beat linear regression or even just predicting the training mean. Widely circulated. Caused alarm.
Rebuttal: "Deep Learning-Based Genetic Perturbation Models Do Outperform Uninformative Baselines on Well-Calibrated Metrics" — Deep learning does outperform baselines, but only with better metrics. The commonly used metrics are broken:
| Broken Metric | Why It's Broken | What It Rewards |
|---|---|---|
| MSE (Mean Squared Error) | Dominated by the ~19,800 genes that barely change. A model predicting "nothing changes" gets low MSE. | Predicting the mean. Ignoring perturbation signal. |
| Pearson(Δctrl) — correlation of change-from-control | When the control population has systematic bias, predicting "no change from control" still scores well. | Mode collapse — every perturbation predicted identically. |
The rebuttal paper introduces the Dynamic Range Fraction — a calibration test that measures whether a metric can even distinguish a perfect model from a useless one. Conventional metrics often score near zero on this test.
Systema identifies the root cause: perturbation experiments mix two signals that current metrics can't separate:
Perturbation-specific signal
The actual biological effect of knocking out gene X. What we want models to learn and predict.
Systematic variation
Batch effects, selection bias, technical confounders that make all perturbed cells look different from controls — regardless of which gene was perturbed.
Most of what looks like model "performance" is actually learning Signal 2 — systematic variation. The model learned "perturbed cells look different from controls in this dataset," not "this specific gene knock-out causes this specific cascade."
Systema strips out systematic variation before scoring. Available at github.com/mlbio-epfl/systema.
Old narrative: "Deep learning is useless for perturbation prediction."
Current narrative: "Deep learning is probably useful — but we can't tell yet because our measuring tools are broken." The metric debate is turning a vague failure into a specific, fixable problem. That's real progress.
Three Competing Evaluation Frameworks
No consensus exists yet on which is canonical:
- PerturBench — unified evaluation across multiple models and datasets
- Systema — removes systematic variation bias before scoring
- Arc Challenge PDS/DES system — rank-based, discrimination-focused; most influential practically due to prize money and prestige
SOTA Models Landscape 2025–2026
| Model | Origin | Approach | Key Claim |
|---|---|---|---|
| VCWorld | Shanghai Jiao Tong + NeoLife AI, Nov 2025 | Mechanistic white-box + LLM reasoning | SOTA on drug perturbation benchmarks with step-by-step interpretable predictions and verifiable hypotheses |
| C2S-Scale | Google, 2025 | Cell2Sentence scaled to 27B parameters | Treats scRNA-seq as text. Outperforms GPT-4o on single-cell biology QA. Major step toward "virtual cells as model systems." |
| CellForge | 2025 | Agentic — AI agents design and iterate cell models | First attempt at autonomous virtual cell model construction. Very early stage but conceptually significant. |
| Noetik OCTO | Noetik, 2025 | 1.5B parameter multimodal spatial model | Trained on proprietary multimodal spatial dataset. Commercially deployed for precision oncology. Bridges spatial biology and gene expression. |
| Alpha Cell | SciLifeLab Sweden, early 2026 | Foundation model for stem cell behavior | New program directly inspired by AlphaFold. Predicting stem cell behavior with ML. Just launched. |
The AlphaFold Comparison
Everyone in this field invokes AlphaFold. Understanding why requires understanding what AlphaFold actually did — and why the virtual cell problem is harder.
AlphaFold (DeepMind, 2020) solved protein structure prediction: given an amino acid sequence, predict the 3D shape of the folded protein. This problem had been open for 50 years. AlphaFold solved it at near-experimental accuracy in 2020, and the impact was transformative — drug discovery, disease understanding, and basic science all accelerated overnight.
| AlphaFold | Virtual Cell | |
|---|---|---|
| Core question | What 3D shape does this protein fold into? | What does a cell do when you perturb it? |
| Ground truth | Clean — a single deterministic 3D structure | Noisy — a stochastic distribution across a heterogeneous population |
| Variables | One molecule, hundreds of amino acids | 20,000 interacting genes, embedded in tissue context |
| Training signal | Tens of thousands of clean protein structures from PDB | Noisy single-cell measurements that kill the cell to measure it |
| Generalization | Sequence → structure is consistent across organisms | Same gene does different things in different cell types |
| Status | Solved (2020) | Unsolved (2026) |
Protein folding has a single right answer per sequence. A cell's response to a perturbation depends on which cell type, what other genes are active, what signals it's receiving from neighbors, what time in the cell cycle it's at, and dozens of other contextual factors. The virtual cell problem is harder in almost every dimension — more variables, noisier measurements, less consistent ground truth, more context-dependence.
Why Any of This Matters
Screen millions of targets cheaply
Most drugs fail in clinical trials because we didn't understand the biology. If you can simulate "what happens when you inhibit gene X in a cancer cell" without running the experiment, you can screen millions of drug targets computationally before synthesizing a single molecule.
Model what goes wrong
Most diseases involve changes in gene regulation — oncogenes overactive, tumor suppressors silenced, immune genes misfiring. A virtual cell lets you model the diseased state computationally and find where to intervene.
Patient-specific digital twins
Take a biopsy from a patient, build a virtual model of their specific cells, simulate 50 different treatments, pick the one that works for their tumor's specific genetic profile. Not average patients — this specific person.
Reduce preclinical experiments
If cellular responses can be simulated accurately, a large fraction of animal experiments during drug development become redundant. Faster, cheaper, more ethical.
Where the Field Actually Is
The dream is clear. The tools are immature. The metrics are disputed. Simple baselines are embarrassingly competitive with models containing a billion parameters trained on millions of cells.
But: the data infrastructure (Perturb-seq at genome scale), the competition frameworks (Arc Challenge with $250K prizes and 5,000 participants), and the research output (multiple Nature papers in 2025 alone) are all signs of a field that is accelerating fast.
Can we build AlphaFold for cell biology?
The answer so far: "Not yet — but we're getting better at knowing exactly why not." The metric crisis revealed that we were measuring the wrong things. Systema revealed systematic confounds in the data. The challenge revealed that hybrid approaches beat pure neural networks.
In science, turning a vague failure into a specific, diagnosable problem is real progress. The field knows more precisely what it doesn't know — which is how breakthroughs happen.