A biological-response compound representation allows chemical perturbation prediction across cell lines

Accurately predicting how cells respond to drugs is crucial, as it could reduce the cost and speed up drug discovery. Present computational methods use chemical structures to represent compounds, but they provide limited insight into their activity within biological systems. Researchers have now developed a new framework called BioPert, which leverages gene expression responses measured in a reference cell line as transferable representations of chemical perturbations.

BioPert utilizes a small neural network to predict transcriptional delta responses in different types of cell lines.

When evaluated on the Tahoe-100M and LINCS L1000 datasets, BioPert outperforms both molecular fingerprints and embeddings derived from chemical structure. These traditional methods offer only minimal improvements over random controls. BioPert's prediction correlations reach 0.79 on the Tahoe-100M dataset, marking a significant improvement of 0.34 compared to the next-best representation.

On the LINCS dataset, performance can vary depending on the experimental reproducibility of the test conditions, emphasizing the influence of batch effects.

One key advantage of BioPert is that it surpasses predictions obtained by simply copying the reference response, indicating that it captures context-dependent effects. The researchers tested BioPert's performance using the C32-cobimetinib case, which revealed pathway-level accuracy of the predictions, even when the reference and target responses differed. These findings open up new possibilities for predicting chemical perturbations, potentially reducing the burden associated with phenotypic screening.