Why Do Linear Baselines Still Matter for Virtual Cell Models?

Fabian Theis says linear baselines still matter because a field rushing toward a virtual cell sometimes skipped them. On the plain task of naming a cell type from a new transcriptome, some models close to the state of the art are logistic regression. CellTypist does a good job. A paper from his group about five or six years ago said, in a plain title, that deep learning is not needed for cell type prediction. The claim holds inside one organ. Across organs the task gets subtler, because cell types show up again in different organs and immune cells change with the tissue.

Early perturbation analyses used classical linear models. Around 2019 his lab introduced linear shifts inside a latent space and still compared them with simpler linear shifts. The ARC Institute ran a competition it called a virtual cell challenge. The fixed metrics were not very specific to single-cell data and could be exploited. The model crowned the winner is a flow-matching model, which he takes as a sign that some nonlinearities are needed once the exploitation is set aside. On simple setups with a few cell lines and gene knockouts, especially when the effects are subtle, fancier models sometimes overfit, so the linear baseline stays the check. For drugs he tells students that beating the mean, short of even a linear model, is the baseline to clear, and the measurements are very noisy.