The Awkward Reality of the Virtual Cell Competition: The Winner Relied on Statistical Denoising, and Pure Deep Learning Did Not Win

First, watch the video. An 11-slide deck makes one thing clear: the human body has about 37 trillion cells, almost all carrying the same DNA; the difference is which genes are being read at this moment.

The video cover says 'Every cell carries the same DNA.' This claim is not new. What is truly hard is measurement. To measure a cell's transcriptome, you have to destroy the cell. Measurement itself kills the cell; you cannot watch the same cell before and after perturbation, only compare cell populations. This is where the noise comes from.

Perturbation and Combinatorial Explosion

Perturbation means interfering with a cell and then seeing what happens. A common tool is CRISPRi, which clamps onto a gene so the transcription machinery cannot read it. Genes are connected into networks. Turn off X, and Y may be upregulated to compensate, Z may go out of control, and dozens of genes move with them. Wait for the cascade to stabilize, and the measured transcriptomic snapshot is the perturbation response.

Experiments are expensive. Single-gene knockout has about 20,000 possibilities, two-gene combinations about 200 million, and three-gene combinations about 1.3 trillion. Doing all of them is unrealistic. The goal of virtual cells is to train models to predict experiments that have not been done, answering combinatorial perturbation questions in milliseconds without killing cells. Experiments will not be replaced.

The Arc Challenge Puts the Problem on the Table

Arc Institute's Virtual Cell Challenge 2025 used H1 human embryonic stem cells, about 300,000 cells, and 300 CRISPRi perturbations. More than 5,000 participants from 114 countries and over 1,200 teams took part. The total prize pool was $250,000. Scoring used PDS, DES, and MAE.

MAE sounds reasonable but is actually very misleading. Of 20,000 genes, maybe only 50 to 200 truly change after perturbation. If a model uniformly predicts no change, MAE is still very low; it scores by ignoring biology.

The winner was BM_xTVC from BioMap Research, with the model xTrimoSCPerturb. Its approach was a hybrid approach: transformer architecture, protein language model embeddings, classic statistical features, plus pseudobulk aggregation for denoising. Second was Sichuan University's XLearning Lab, optimizing for the metrics and also using pseudobulk representations and protein embeddings. Third was Team Outlier's TransPert, a purely statistical framework with no deep learning. The General Award went to Altos Labs' go-with-the-flow, which uses flow matching to predict the distribution of cell responses.

All top teams used pseudobulk aggregation. First average cells with the same perturbation, then predict. Classic statistics are helping neural networks guard against noise. Pure end-to-end deep learning did not win. One commenter replied, 'so far the models are just good denoisers.' Blunt, but it matches the winning approach. gabrycina replied, 'Let's see what this year's leaderboard brings.'

The Metric Crisis

The author gabrycina calls the metric crisis the field's core intellectual drama. In a 2025 Nature Methods benchmark, seven foundation models did not beat linear regression, or even beat predicting the training mean. In October 2025, a bioRxiv rebuttal appeared: after calibrating the metrics, deep learning can exceed the baseline. Systema pointed out the root cause in Nature Biotechnology 2025: perturbation experiments mix two types of signal. One is the true biological effect of knocking out a gene. The other is batch effects, selection bias, and technical confounding, which make all perturbed cells look different from controls regardless of which gene was knocked out. If a model looks good, it may be learning the second type of signal.

The old narrative was that deep learning is useless for perturbation prediction. The new narrative is that deep learning may be useful, but the measurement tools are broken. The metric debate turns a vague failure into a concrete, fixable problem.

What Makes It Hard Compared with AlphaFold

Everyone uses AlphaFold as a reference. Protein folding has a single determined structure, and the training signal is clean. Virtual cells face stochastic distributions, 20,000 interacting genes, and responses that change with cell type, neighboring signals, and cell cycle. The same gene can do different things in stem cells and cancer cells. Measuring single cells also requires killing the cells. AlphaFold delivered an answer in 2020. Virtual cells still have not solved it by 2026.

Why It Matters

Drug discovery could screen targets before synthesizing molecules. Disease understanding could simulate states in which gene regulation goes wrong. Personalized medicine could build models from patient biopsies, try 50 treatments, and pick the one effective for that tumor. Animal experiments could also be reduced.

The current state is not pretty. The tools are immature, the metrics are controversial, and simple baselines trade blows with billion-parameter models. But data infrastructure is expanding, competition frameworks are running, and 2025 saw multiple Nature papers. The 2026 leaderboard will give new answers. What is certain now is that models first have to learn not to be fooled by noise.

Original video and article: https://x.com/gabrycina/status/2114399011449008537

Interactive article: https://virtual-cell-deploy.vercel.app