Loading…
Single-cell perturbation response prediction by conditional latent diffusion, trained on the Tahoe-100M atlas of 100 million drug-treated cells.
No providers recorded yet. Browse all providers
Take the mean expression profile of a cancer cell line, add the mean effect of a compound at a given dose across every other cell line, subtract the global mean, and the arithmetic reproduces 97.6% of the variance in the measured post-treatment profile. That baseline knows nothing about how a particular drug behaves in a particular cellular context, yet it is nearly indistinguishable from the truth on absolute-expression metrics, because most of a transcriptome is cell identity rather than response. What a perturbation model must supply is the residual: the interaction between compound, dose and cell line that separable marginals cannot reach.
PerturbLDM targets that residual directly. It pairs a variational autoencoder, which compresses normalised expression into a compact latent space, with a conditional diffusion model that denoises latent cell states under an explicit conditioning vector. In the drug-response setting that vector carries a learned drug embedding, dose as a scalar, and the mean expression profile of DMSO-treated cells from the target cell line. Because the matched control enters as an input rather than something the network must reconstruct, sampling is anchored to the requested background and the denoiser's job reduces to learning how the intervention moves a cell away from it.
The framework comes from Yulin Dai, Xiaoqian Jiang and Zhongming Zhao's groups at UTHealth Houston, with collaborators at MD Anderson Cancer Center and Vanderbilt University Medical Center. It joins CPA, scDiffusion, PerturbDiff, PerturbGen and STATE in a crowded field, distinguished less by its components than by its evaluation design: fit once at atlas scale, then reused without refitting on a separate assay.
The autoencoder is a multilayer-perceptron VAE; the denoiser is a residual MLP that fuses a condition descriptor with a sinusoidal timestep embedding. Diffusion uses a DDPM scheduler over 1,000 steps with a linear variance schedule and v-prediction. The Tahoe configuration has a 1,024-dimensional latent space, hidden dimension 2,048 and a 13,784-gene feature space; the fetal-colon and PBMC fits use 64- and 128-dimensional latents. Within the 100-million-cell Tahoe-100M atlas, 46,471 measured drug-dose-cell-line conditions span 379 drugs, 47 cell lines and three doses; 32,529 were used for fitting and 13,942 withheld. Against an additive marginal baseline, PerturbLDM raised mean R² from 0.9761 to 0.9853 and cut mean absolute error by 20.7%; on matched-control-relative effects it achieved higher gene-effect Pearson correlation in 95.2% of held-out conditions and higher Hallmark programme correlation in 94.4%, and it produced lower MMD and optimal-transport distances than CPA and chemCPA in over 99.9% of conditions. In smaller settings it recovered a withheld mid-gestational fetal-colon state with 67% lower mean per-gene error than Squidiff, and six of seven interferon and antiviral GO programmes in each held-out PBMC lineage against scGen.
The intended use is experimental triage: given a compound, dose and cell line no screen has assayed, generate the expected transcriptional response and decide whether the experiment is worth running. The PANACEA analysis extends this to compound characterisation — ranking drugs by the similarity of their predicted Hallmark effects placed Afatinib first for Dacomitinib and Nilotinib first for Bafetinib, recovering EGFR/ERBB and BCR-ABL inhibitor relationships from generated profiles alone. The developmental and immune fits show the conditioning interface also accepts categorical labels such as gestational stage or stimulation status, for interpolating unobserved cell states within a study.
PerturbLDM's contribution is methodological discipline as much as architecture: benchmarking against a marginal baseline that most comparisons omit makes explicit how little of this field's apparent accuracy comes from modelling perturbation at all. The authors state its limits plainly. The Tahoe evaluation withholds combinations of drugs, doses and cell lines that all appear elsewhere in training, so it demonstrates interpolation within an observed factor space rather than generalisation to unseen compounds, and the drug embedding is a learned lookup tied to the fitted dataset's compound vocabulary. The PANACEA mechanism-of-action concordance of 51.9% exceeded the 39.6% expected by chance but did not reach significance across 27 evaluable drugs. Source code is MIT-licensed, but no trained checkpoints or per-cell predictions are distributed, so reproducing the atlas-scale model requires staging Tahoe-100M and refitting. The work is a preprint awaiting peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.