Finding. A latent-diffusion model trained on Tahoe-100M can invent plausible single-cell responses to drug–dose–line combinations it has not seen, and it beats a simple add-the-margins baseline in 95% of 13,942 held-out cases. In PBMCs it keeps interferon programs and, specifically, an interferon-associated FAO-OXPHOS program better than scGen. Mitochondria appear as a transcriptional module the generator is not allowed to blur.
Why this paper matters
Perturbation space is too big to measure. A conditional generator that preserves a metabolic-mitochondrial program under interferon is more useful to this desk than another generic UMAP smoother. It is still a model.
What they actually measured
Held-out combination metrics, compound-neighbor ranking, two smaller biological case studies including FAO-OXPHOS.
How to read the score
Around 60. Score 62.
What to do with it
If you impute perturbation responses, compare PerturbLDM on your mitochondrial modules. Do not skip the wet lab. The directional implication is that conditional diffusion can carry FAO-OXPHOS programs through unseen interferon-like contexts.