Building experimental data together for cell AI

What changes when one gene in a cell is suppressed? Will the same intervention produce the same result in a different cell?

Research institutions, US government agencies and AI companies are jointly investing in the data needed to make some of these responses predictable before an experiment is run. This is the aim of Biohub’s Virtual Biology Initiative. Programme overview ↗

What is needed is more than a large collection of cell measurements. The records must show which responses occur when something is added to or reduced in a particular cell. Without experiments that change the conditions, it is difficult to distinguish observed characteristics from changes caused by an intervention.

The US Department of Energy is contributing measurement instruments and automated laboratory facilities, as well as computing resources. Investment in AI is going into the experiments that produce training data, alongside model development itself. Department of Energy announcement ↗

But it is impossible to measure every experimental condition. The question is whether knowledge gained under one set of conditions can be used under conditions that have not yet been measured.

This year’s Virtual Cell Challenge at Arc Institute tests that ability. Participants receive gene-expression data from cells before an intervention and information about the gene to be suppressed. The changes actually observed after the intervention in the target cells are used to evaluate their predictions. Challenge description ↗

If a model can learn the relationships between a cell’s state, a genetic intervention and the changes that follow, and then predict responses under conditions that have not been measured, knowledge from cells that are easy to study could become useful in a wider range of research on scarce samples or disease states that are difficult to reproduce in culture. That requires testing how well the predictions hold under conditions different from those of the cells used for training.

To assess this initiative, counting records alone will not be enough. Which cells and interventions have been added, and how far can the models now handle conditions they previously could not predict? That is what we need to examine to judge the scientific progress behind the scale of the funding.