2026-09-28
MultiFlow couples RNA and ATAC flows with bidirectional cross-attention, lifting held-out top-100 gene iLISI from 0.716 to 0.867 on Perturb-multiome.
The same genetic perturbation can rewrite a cell in completely different ways depending on cell type. Most single-cell perturbation models still predict the transcriptome alone and drop the chromatin-accessibility layer. The combinatorial space of perturbations times cellular contexts is also larger than any screen can cover.
Perturb-multiome already reads RNA and ATAC in the same cell. What was missing is a generative model that treats the two modalities as a coupled system and transfers perturbation effects to cell types never seen under that perturbation at training time.
MultiFlow first compresses paired scRNA and scATAC profiles into 128-dimensional latent spaces, then learns two interacting flow-matching trajectories. A shared Gaussian source is interpolated in a straight line toward the observed paired latents, and the network regresses the velocity. The coupling piece is a bidirectional Cross Flow Module: at each time, RNA features attend to ATAC keys and values, and ATAC attends back, so the two velocity fields share one dynamical system.
Cell-state generation turns the perturbation channel off and samples unperturbed paired profiles from a cell-type embedding. Perturbation prediction uses the mean latent state of control cells as context; each perturbation is a learned embedding, with the control vector fixed at zero. Training is leave-one-cell-type-out: perturbed cells of the held-out type never enter training, and only that type's control cells build the inference-time context. After sampling, a training-only latent mean shift is applied before decoding.
Generation is trained for 600 epochs on OpenProblems (69,249 cells, 22 types) with 100 midpoint ODE steps. Perturbation prediction is trained for 1,600 epochs on GSE274113 (133,022 cells, 10 types, 19 transcription-factor perturbations) with 50 ODE steps.
On OpenProblems generation, RNA iLISI is 0.898, 0.052 above scDiffusion-X. ATAC iLISI is 0.917 and MMD is 0.011, both best in the comparison. Cell-type-averaged Pearson correlation for CD4+ T cells reaches 0.9960 (RNA) and 0.9897 (ATAC). On independent PBMC10k, ATAC iLISI is 0.650, about 18.83% above scDiffusion-X. Mean-profile correlations are high for every strong method; distribution metrics are what separate them.
On held-out cell types, top-100 responsive genes improve from 0.716 to 0.867 iLISI (relative +21.1%) and from 0.093 to 0.034 MMD. ATAC composite absolute score is 0.839 versus 0.660 for the best baseline, with 0.672 correlation and 0.934 directional accuracy on significant peaks. Peak-gene neighborhood concordance is 0.31 versus 0.15; cell-level RNA-ATAC neighborhood coupling is 0.765 versus 0.288. Dropping the ATAC flow consistently hurts RNA prediction; adding RNA to ATAC helps less.
| Setting | Metric | MultiFlow | Best baseline |
| OpenProblems RNA gen. | iLISI | 0.898 | scDiffusion-X 0.846 |
| OpenProblems ATAC gen. | iLISI / MMD | 0.917 / 0.011 | worse |
| Held-out RNA top-100 | iLISI / MMD | 0.867 / 0.034 | 0.716 / 0.093 |
| ATAC perturbation | absolute score | 0.839 | 0.660 |
For virtual-cell and perturbation-screen work, one coupled flow both generates paired multiome profiles and moves perturbation effects into unmeasured cell types. Cross-attention buys cross-modal coordination, not just two pretty marginals. Code and a tutorial are public.
This is still a generative improvement, not a new biological mechanism. The data diet is also narrow.
Perturbation transfer is validated mainly on Perturb-multiome; cross-platform generalization is untested, as the paper states. The architecture is locked to two paired modalities, with no proteomics or missing-modality path. Cross-attention recovers association, not causal regulation. The PDF and the GitHub tutorial disagree on epoch counts for generation versus perturbation; the methods section in the paper should be treated as canonical.