Live·Open questions in longevity research
All news
Scientific ComputingScience Research

The competition’s top computer models predicted nearly identical responses to the same intervention in cells of different origins

21 November 2026· 261121001

The competition’s top computer models predicted nearly identical responses to the same intervention in cells of different origins

On November 19, the research organization Arc Institute announced the results of the Virtual Cell Challenge 2026. Participants predicted the effects of gene knockdown in three cell cultures, consisting of cells of different origins grown in the laboratory. Illumina’s BioInsight AI Lab won among 805 teams that submitted final predictions.

A biologist knocks down one gene and measures how the expression of other genes changes in response. A virtual cell is a computer model that predicts these changes from the cells’ initial state and a specified intervention. Researchers could use such a model to select experiments for cells that are difficult to study in the laboratory. To do this, the model must account for differences in how cell cultures respond to the same intervention.

In the 2025 competition, participants predicted the effects of new interventions within a single cell culture. This time, Arc provided only data on cells before the experiment and a list of 300 genes to knock down individually. The institute measured the actual responses itself and withheld them from participants. For the comparison, it selected experiments in which the median reduction in expression of the targeted gene was at least 80%, so predictions would be assessed at comparable knockdown strengths.

Almost all of these genes had already been knocked down in other cell cultures. All three prize-winning teams transferred the changes measured in those cultures, adjusted them for the initial state of the new culture, and calculated gene expression in individual cells. Two independent researchers on the FlowState team took second place using only publicly available data. The other two prize-winning teams also had their own measurement datasets.

According to Illumina’s description, its CellularAI algorithm gives greater weight to measurements with less random variation and to data from cultures similar to the new one. It applies the calculated changes to gene expression data from individual cells before the intervention. This preserves the natural variation in gene expression between cells in the prediction.

The winner scored 0,44 on a scale combining six metrics across the three final cell cultures. Zero corresponds to a prediction that gives the same response to every knockdown: the average response across all interventions in that culture. The reference for a score of one comes from an actual experiment. The measured cells are divided into two nonoverlapping halves, their responses are compared, and the results are averaged over several such splits. The final score averages how far the model progresses between these reference points across several capabilities: distinguishing interventions, predicting the magnitude and direction of changes, and identifying the affected genes.

A model can distinguish the effects of knocking down different genes while still making nearly identical predictions for different cell cultures.

Arc separately compared the ten best solutions on three validation cultures used before the final round. Predictions for the same knockdown in different cultures were much more similar to one another than the actual responses were: the models mainly reproduced the shared response. When assessed with the same statistical test, the models also overestimated the number of genes whose expression changed. The median in the measurements was 22–44 genes per knockdown, compared with about 500 in the winner’s predictions. To transfer findings from one culture to another, a model needs to capture the features specific to that culture’s response.

“For now, data from previous experiments and the care taken to transfer them to a new context appear to matter at least as much as model architecture,” Arc Institute stated.
Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#virtual-cells#gene-knockdown#gene-expression#cell-cultures#cellularai