Akram Baharlouei examines the engineering challenges of applying foundation models to single-cell biology. She describes cellular rejuvenation and virtual-cell systems as research goals, then connects those ambitions to a lengthy drug-development process. The talk focuses on what can actually be measured and modeled rather than presenting a clinical intervention or a completed virtual human.
RNA sequencing provides scalable snapshots of gene activity, but cells vary naturally and laboratories introduce additional batch effects. Baharlouei compares RNA with protein, microscopy and spatial measurements, explaining why a very large collection of snapshots still does not reveal the continuous behavior of the same cell. She questions whether treating genes as tokens and compressing their measurements into latent representations preserves the information needed for downstream tasks.
Baharlouei stresses benchmark comparisons against simpler methods, including cases where linear or fit-to-data approaches match or outperform more elaborate foundation models. She then discusses flow matching as a way to predict distributions of cellular responses instead of only an average outcome. Her conclusion is that better measurements, multimodal data and task-specific evaluation must accompany scale; the examples do not establish that one architecture solves biological prediction generally.
Watch on YouTube
