CFM-GP

Measuring how every cell type in a tissue responds to a perturbation is experimentally punishing. Each additional cell type means another round of sample preparation, batch handling, and sequencing, so the space of perturbation-by-context combinations stays sparsely sampled. Computational surrogates help, but most established ones carry a structural cost: methods in the scGen, CVAE, and trVAE lineage learn a latent shift from control to perturbed state and generally need refitting for each cell type, reproducing the scaling problem they were meant to solve.

CFM-GP takes a different route by making cell identity an explicit conditioning variable inside a single continuous-time model. Instead of learning a discrete latent offset, it learns a neural vector field that transports the unperturbed expression distribution to the perturbed one along a trajectory parameterised by time, with the cell type supplied as an embedding at every integration step. One trained model therefore covers all cell types in a dataset, sharing the perturbation dynamics that are common across populations while retaining the deviations specific to each.

The model was developed by Abrar Rahman Abir at the Bangladesh University of Engineering and Technology with Sajib Acharjee Dip and Liqing Zhang at Virginia Tech, and is available as a preprint. It occupies the same niche as CPA and GEARS — predicting transcriptional responses to perturbations that were not directly measured — but replaces their autoencoder and graph-based machinery with a flow matching objective borrowed from modern generative modelling.

#Key Features

• Cell-type conditioning inside one model: A learnable cell-type embedding enters the vector field at every step, so a single checkpoint serves every population in a dataset instead of requiring one model per cell type.
• Continuous transport rather than a latent jump: Perturbation is modelled as an ODE integrated from the control state at t=0 to the perturbed state at t=1, which captures intermediate dynamics that discrete latent arithmetic collapses.
• Flow matching training objective: Regressing a velocity field against known interpolation paths avoids the adversarial instability of GAN-based predictors and the sampling cost of diffusion.
• Cross-species transfer: In an out-of-distribution test the model is trained on LPS-stimulated phagocytes from one mammalian species and evaluated on a held-out species, a genuinely unseen context rather than a held-out slice of the same population.
• Released checkpoints: The repository ships trained weights for all five benchmark datasets alongside the preprocessing, training, and evaluation scripts.

#Technical Details

The velocity network is a compact multilayer perceptron: three linear layers with 256 hidden units and ReLU activations, taking the current interpolated expression state, the original control profile, a 16-dimensional cell-type embedding, and a 16-dimensional time embedding, and emitting a velocity vector in gene-expression space. Training uses Adam at a learning rate of 1e-4 with mean squared error loss, batch size 32, over 50 epochs, with each experiment repeated five times; inference integrates the ODE in ten steps.

Evaluation spans five public scRNA-seq datasets with paired control and perturbed profiles: SARS-CoV-2-infected airway and immune cells (GSE145926, six cell types), IFN-β-stimulated PBMCs (GSE96583, seven immune cell types), panobinostat-treated glioblastoma across six patient samples (GSE148842), IFN-β-stimulated lupus PBMCs (eight cell types), and cytokine-driven hematopoietic progenitors (Statefate, GSE140802). Against CoupleVAE, scGen, CVAE, trVAE, scPreGAN, and scDist, CFM-GP reports the highest R² and Spearman correlation in every cell type tested — R² from 0.865 to 0.995 across COVID cell types on all genes, and above 0.99 on the top 100 differentially expressed genes — with the margin widening on that harder DEG subset. In the cross-species setting it beats CoupleVAE in 9 of 12 source-target transfers, the largest gain being rat to mouse at +8.1% R². Pathway enrichment analysis confirms that predicted profiles recover the perturbed pathways rather than merely matching summary statistics.

#Applications

The practical payoff is in experimental triage. A lab that has profiled a stimulus in a few abundant cell types can use CFM-GP to predict responses in rarer populations from control data alone, prioritising which conditions are worth sequencing. The cross-species results extend the same logic to translational work, where a perturbation characterised in mouse can be projected onto another species given only its unperturbed profiles. Because the model operates on gene-expression vectors plus a cell-type label, it fits standard AnnData-based analysis without special featurisation.

#Impact

CFM-GP shows that flow matching, applied mostly to image and molecular generation, transfers cleanly to perturbation response prediction, and that conditioning on cell identity removes the per-cell-type retraining burden constraining older methods. Code and trained weights are public under GPL-3.0, a copyleft license that complicates use inside proprietary pipelines. Two limits deserve emphasis: within each dataset the model is trained and tested on the same set of cell types, so the headline numbers measure generalisation to held-out cells rather than to novel cell identities, and it predicts responses for perturbations it has seen rather than extrapolating to unseen ones, the problem GEARS addresses.