PerturbLDM predicts how drugs alter gene activity in cells under experimental conditions that have not yet been measured

PerturbLDM predicts how drugs alter gene activity in cells under experimental conditions that have not yet been measured

A preprint describing PerturbLDM appeared on bioRxiv on August 12. The model was tested on drug, dose, and cell line combinations withheld during training, although each drug, dose, and cell line had appeared separately in the training data.

The same drug can alter gene activity differently across cell lines, which are cell populations grown and studied separately. There are so many combinations of drugs, doses, and cell lines that laboratories can measure only a fraction of them. The model is designed to predict which genes will become more or less active under conditions that have not yet been measured.

In the preprint, PerturbLDM receives the drug name, dose, and a profile of cells from the same cell line in a DMSO control experiment. DMSO is the solvent added in place of the drug. The model first compresses information about the activity of thousands of genes into a compact representation. It then constructs the response step by step using the drug, dose, and control profile, before converting the result back into gene activity values.

The authors used 46 471 experimental conditions from Tahoe-100M, a large dataset of single-cell measurements. They retained 32 529 conditions for training and withheld 13 942 conditions. Both sets included all 379 drugs, 47 cell lines, and three dose levels. Only specific combinations of the three variables were withheld. Each drug, dose, and cell line in a withheld combination had already appeared separately in the training data.

In an evaluation of virtual cell models, a metric that compares the activity of all genes at once could give a high score to an average prediction even when the changes caused by the intervention were lost within the overall profile. The PerturbLDM authors therefore subtracted the matched DMSO control from both the predicted and measured profiles. The resulting change in gene activity represents the effect of the drug in that specific cell line.

PerturbLDM was compared with AdditiveMean, a simple baseline that adds the average cell line profile to the average effect of the drug-dose pair, then adjusts the result using the overall mean profile. This comparison tests whether the more complex model captures features specific to a particular combination of drug, dose, and cell line. In 95.23% of the withheld conditions, the predicted change in gene activity matched the measured change more closely than the AdditiveMean prediction did.

This type of model could be used before direct experiments to prioritize unmeasured combinations of previously studied drugs, doses, and cell lines, then select conditions for laboratory testing.