- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
PerturbRIFT predicts single-cell expression distributions for genetic perturbations whose response cells are unavailable during training. It first retrieves 16 related training perturbations using biological resources, including STRING, CollecTRI, and pathway memberships. A learned weighting network aggregates their measured gene-wise means, zero probabilities, and conditional nonzero variances. These statistics define a factorized zero-inflated normal prior.
Samples from this prior are encoded using a pretrained, frozen autoencoder. A conditional flow model then transports the encoded populations toward the latent distributions of observed responses, using flow matching and an endpoint MMD objective. The decoder generates expression profiles while calibrating their expected population mean to the retrieved prior mean. Finally, Cube sampling couples output-mask decisions across cells to reduce finite-population imbalance. These components are trained in separate stages.
The evaluation covers unseen perturbations within four datasets—K562, RPE1, HepG2, and Jurkat—with five shared perturbation-level splits per dataset. The reported results show consistent improvements in Top100 DEG mean-response metrics, perturbation discrimination, and several distributional metrics. K562 ablations attribute most mean-prediction gains to biological anchor selection and learned weighting, while flow improves PCA-space distributional agreement and recovery of correlation changes.
Critical strengths
On the perturbations that do have the necessary information, PerturbRIFT performs well. The Top100 results improve consistently over strong linear baselines across all four datasets. I think the experiments and ablations are well-done and relatively comprehensive overall as well. There are four datasets and five shared perturbation splits, and the ablations help explain where the gains come from. In particular, KG-based anchor selection substantially improves the prior mean prediction, while the flow improves correlation-change recovery and PCA-space distributional agreement. These are useful findings.
Critical weaknesses
It seems like a complicated method, with three phases of training, but I am not yet convinced that the experiments establish the need for the proposed source distribution and generative refinement. The retrieved response statistics enter the source, the conditioning, and the decoder’s mean calibration. Table 3 keeps the source fixed, so it does not show whether this source is better than control cells or noise when the same statistics are available through conditioning. There is also no comparison against simply sampling cells from the same retrieved perturbations and rescaling them to the same predicted mean. That would preserve measured gene dependencies and help show what the flow contributes beyond retrieval and a good mean predictor.
Evaluation is restricted to perturbations covered by the required resources, and it is unclear how much coverage is lost or what happens when useful related perturbations are unavailable. Also, some distributional wins are very small, e.g. Jurkat MMD is 4.228 versus 4.256 for SLIM in Table 2, and the main tables provide no uncertainty estimates. There are five data splits, but repeated training seeds on fixed splits are not documented.
Questions for authors
-
Could the authors compare against latent flow matching with MMD and the same OT pairing, but starting from control cells or a generic noise source? This would directly test whether putting the response information into the source distribution is useful.
-
Could the authors report uncertainty across the existing splits and run additional training seeds for the key comparisons?
-
How many perturbations are excluded from each dataset by the resource-coverage filtering? How would the method handle a query with missing annotations or few related training perturbations?
-
Also, compute permitting, it would be great to see performance on slightly larger datasets like PBMC
-
Could the authors provide the missing citations for the baseline methods?
-
Why is endpoint MMD computed against the OT-resampled target batch, which can contain duplicate cells, instead of the original observed batch?
-
What values are used for the prior-loss coefficients, the probability and variance floors, and the minimum nonzero-observation count?
-
What are the training and inference costs of the three-stage pipeline compared with the strongest simple baselines?