Single-cell perturbation prediction drops the pretraining crutch: GFM wires knowledge graphs into flow matching and trains up to 30x faster

GFM moves prediction of single-cell transcriptomes after unseen gene perturbations away from large-scale pretraining and toward a generative route conditioned on biological knowledge graphs. Earlier models such as scGPT and CellFlow rely on pretraining over observational single-cell data or a protein language model, and GEARS and TxPert add graph neural networks and biological priors but do not place those priors inside an effective generative framework. On four CRISPRi Perturb-seq datasets, K562, RPE1, HepG2 and Jurkat, split 75% training, 10% validation and 15% test and repeated over three splits, GFM beat CellFlow, GEARS, scGPT and a linear baseline on mean metrics and 2-Wasserstein distance over the top 20 differentially expressed genes, while that linear baseline had itself outperformed the deep models in earlier reports.

GFM transports Gaussian noise along a path set by the perturbation condition inside a latent space, then decodes the endpoint back into gene expression. Three parts do the work: a variational autoencoder compresses log-normalized expression profiles into a latent space and decodes them back; a graph attention network propagates a trainable perturbation embedding for each gene over three concatenated knowledge graphs; and a neural ordinary differential equation module learns a velocity field conditioned on that embedding. Before training, optimal-transport couplings between Gaussian noise and target latent states are computed separately for each perturbation condition, which keeps paths close to straight and avoids the repeated minibatch optimal-transport computation CellFlow needs during training; inference uses classifier-free guidance with a scale of 1.5 and integrates with the dopri5 solver.

Removing the flow-matching objective or scrambling graph connectivity both degrade GFM's predictions, so each part contributes. With the model structure held fixed, replacing optimal-transport velocity matching with direct regression from the noise latent state to the target latent state, or randomly permuting graph source nodes while keeping the edge count, both lower performance. On the same four datasets, dropping flow matching mainly hurts distributional metrics, while randomizing graph connectivity significantly hurts both mean and distributional metrics; scrambling the perturbation similarity graph alone hurts markedly more than scrambling the GO or PPI graph alone, and permuting any two graphs together lowers performance further.

The demonstrated scope is single-gene perturbation prediction on four CRISPRi Perturb-seq datasets, K562, RPE1, HepG2 and Jurkat, plus a two-gene task of 131 two-gene and 152 one-gene CRISPRa perturbations in K562; evaluation focuses on the top 10 to 100 differentially expressed genes and generates 100 cells per perturbation. Trained on the K562 genome-wide CRISPRi data with a 90% training and 10% validation split and no held-out test set, the model generated predictions for every single-gene perturbation and ranked similar perturbations against HSP90AA1 knockdown as the query.

GFM and the other tested models showed similar or higher energy distance than randomly sampled perturbed cells, so none adequately captured the overall dispersion of the perturbed population. In the two-gene task, when both genes were seen in training, the additive baseline performed best on the top 20 differentially expressed genes, consistent with earlier reports that most two-gene perturbations in this dataset are additive. Prediction accuracy tracks how similar a test perturbation's transcriptional effect is to its training-set graph neighbors: the most similar training neighbor of PSMB2 knockdown was PSMB5 knockdown at cosine similarity 0.965, with an accurate prediction, whereas the most similar neighbor of HIST1H2BN knockdown was MON1A knockdown at 0.172, with poor prediction. The authors also note that each new biological prior will need evaluation for specific use cases rather than assuming any prior improves prediction.