acceptodds
Under review as a conference paper at ICLR 2027

FRESCO: FLOW MATCHING FOR RENDERING EXPRESSION-SPECIFIC CELLULAR MORPHOLOGY

Abstract

Predicting how cells respond to unseen perturbations is central to phenotypic drug discovery and functional genomics. High-throughput screens now measure both cell morphology and single-cell gene expression, but typically in different cells, paired only by treatment. Generating morphologies consistent with a treatment's transcriptional response therefore requires learning a distribution over cells from this group-level pairing. We introduce FRESCO, a conditional rectified-flow model that generates multi-channel single-cell images from gene expression resampled across a treatment's cells. FRESCO represents expression as transcriptional program tokens, conditions on a batch-matched control cell to account for acquisition effects, and factorizes attention across spatial positions and channels. Because standard image metrics overlook perturbation-specific responses, we also introduce metrics of response direction and retrieval. On two transcriptome–morphology screens (28 compounds; 1,000 gene knockouts), FRESCO reaches a median response cosine of 0.57 and 25.9% top-1 retrieval on 56 held-out knockouts, against 0.28 and 11.1% for the strongest baseline, and roughly halves FID relative to CellFlux on three morphology-only benchmarks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, conditional generative models, perturbation modeling, diffusion transformer, cross-modal generation, single-cell biology, cell painting, high-content imaging, gene expression, AI for scientific discovery

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uoZoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FRESCO presents a conditional flow matching model that generates multi-channel single-cell images from a treatment's gene expression and a same-batch control image. Images and expression come from different cells and are paired only by treatment. Expression is given as an average over a random subset of the treatment's sequenced cells and encoded as transcriptional-program tokens. The flow model transforms noise into a simulated treated cell image, conditioned on program tokens and on the control "anchor" image. It uses a pretrained variational autoencoder for image encoding and a channel-factorized transformer (MM-DiT) backbone. The authors also introduce perturbation-specific metrics: response cosine, medcos and top-1 retrieval. The paper reports results on five datasets. Three of them are image-only (JUMP, RxRx1 and BBBC021), where the model is conditioned on perturbation embeddings instead of expression: Morgan fingerprints, Gene2Vec embeddings and HyenaDNA representations. On these, the method outperforms CellFlux and IMPA in FID and KID. The main results are on the two datasets with sequencing data, scGeneScope and OPS Explorer. The model shows better results than MorphDiff, CellFlux and MultiVCDiff, including on 56 held-out (out-of-distribution) genes: top-1 25.9% vs 11.1% for the strongest baseline.

Critical strengths

Group-level conditioning makes sense and is well motivated. It avoids artificial cell-to-cell pairing between imaging and sequencing.

On the three image-only benchmarks, FRESCO achieves better FID and KID than CellFlux. On BBBC021 it also improves mode-of-action (MoA) accuracy, for both seen and held-out compounds.

Code has been released (anonymously).

Critical weaknesses

The main concern is the held-out generalization. It is the key result, but it rests on one dataset, 56 genes and a single seed, with no confidence intervals. Held-out genes could have transcriptionally similar genes in training, yet there is no simple kNN baseline. The authors also do not discuss what drives the generalization, or the batch structure of the sequencing data.

CellFluxMM: Multimodal Generative Modeling of Perturbation-Induced Cellular Responses (ML4H 2025) is not cited. It is a CellFlux-based flow matching model on treatment-matched expression and Cell Painting data. Given this work and MorphDiff, the claim “we formulate expression-conditioned morphology generation for treatment-matched datasets” might overstate the novelty.

Cross-channel attention seems to give the largest gain, but it is evaluated only with FID. The model size behind each table is not stated, so the FID gains over CellFlux may partly come from a larger model.

Questions for authors

Overall, I find the paper interesting, particularly the group-level conditioning and the results on held-out genes are promising. Still, several points need clarification before I can support acceptance:

  • Can the authors show that the held-out results are driven by expression and not by batch information from sequencing path? How does FRESCO compare with a simple kNN baseline that uses the most transcriptionally similar seen gene?
  • How does FRESCO relate to CellFluxMM, and how does it differ?
  • Which model size was used for each table? Can the authors report the cross-channel ablation with the conditioning metrics (top-1, medcos)?
  • Can the authors comment on why one-hot conditioning performs significantly better than the expression embedding on seen genes (Table 3: top-1 52.0% vs 30.0%)?
  • Did the authors try dropping out one control image during training?
4 comments
AuthorsOpenReview ↗

Held-out generalisation: uncertainty and expression controls (W1, Q1a)

We thank the reviewer for the questions about generalisation and conditioning. The new analyses support higher held-out top-1 and medcos than the evaluated MorphDiff run, and the expression controls argue against a batch-only explanation. Compared with nearest-neighbour image transfer, FRESCO has similar scores on the original split and higher response-direction agreement when whole clusters are held out. We also clarify the relation to CellFluxMM, report model sizes and cross-channel conditioning metrics, and address the one-hot comparison and control-image dropout.

W1. Held-out uncertainty and training seeds.

We resampled the 54 eligible genes with replacement 10,000 times, using the same sampled genes for both methods and keeping the retrieval candidates and references fixed. The table reports 95% percentile intervals:

MetricFRESCOMorphDiffPaired difference
top-125.9% [14.8, 38.9]11.1% [3.7, 20.4]+14.8 pts [+1.9, +27.8] (≤0 in 1.3% of resamples)
medcos0.573 [0.453, 0.679]0.277 [0.174, 0.357]+0.30 [+0.19, +0.42]

Across three independent FRESCO training runs, held-out top-1 is 25.9%, 27.8% and 20.4% (mean ± s.d.: 24.7 ± 3.9%), each above the evaluated MorphDiff run (11.1%). All runs use the same evaluation samples and guidance weight, selected on seed 0. The gene-bootstrap intervals quantify variation across query genes for fixed outputs; the three runs separately assess FRESCO’s training variability.

OPS remains the only transcriptome–morphology dataset here that tests held-out perturbations. scGeneScope holds out acquisition replicates; BBBC021 tests held-out compounds with the identity-conditioned model. We will make these different evaluation settings clear in the revision.

Q1a / W1. Expression or sequencing batch?

OPS: each perturbation’s cells are spread across the 32 sequencing libraries in nearly the same proportions. Library composition alone retrieves the perturbation near chance (2.6% versus 1.8%). Giving FRESCO the wrong gene’s expression reduces top-1 to 1.9–5.6% (chance 1.9%). The wrong gene has a similar library mix to the correct one. Together, these checks argue against library composition alone explaining retrieval, but do not exclude all technical variation in the expression input.

scGeneScope: round-1 test sequencing batches differ from training, whereas round 2 shares batches. Expression still distinguishes compounds within a round-2 batch (54.8% versus 12.9% chance). We also tested four combinations: expression from the correct or a shuffled compound, each taken from the same or the other sequencing round. The two rounds have separate batches and different compound groupings:

Expression inputRound 1: 28 compounds (top-1 / medcos / cos(e,d))Round 2: 27 compounds (top-1 / medcos / cos(e,d))
Correct compound, same round (paper)32.1% / 0.48 / 0.6437.0% / 0.55 / 0.62
Correct compound, other round’s sequencing39.3% / 0.39 / 0.5929.6% / 0.41 / 0.61
Shuffled compound, same round0.0% / −0.05 / −0.033.7% / −0.09 / −0.01
Shuffled compound, other round7.1% / −0.04 / −0.030.0% / 0.03 / −0.01

Expression from the correct compound still gives positive response alignment when it comes from the other round. Shuffling the compound reduces alignment to near zero (chance top-1: 3.6% / 3.7%). This supports a compound-related signal that transfers across rounds.

AuthorsOpenReview ↗

Nearest-neighbour baseline and cluster-held-out evaluation (Q1b)

We agree that a nearest-neighbour baseline is needed to assess how much the held-out results depend on similar training genes. For each held-out gene, the baseline selects the training gene with the most similar average expression using cosine similarity and returns its real training images. We score these against the same references as FRESCO.

Original split. FRESCO and the lookup achieve similar point estimates. The paired intervals include zero for all three metrics, establishing neither an advantage nor equivalent performance.

OPS held-out, 54 genestop-1medcoscos(e,d)
FRESCO25.9%0.5730.682
kNN-1 lookup, pseudobulk cosine25.9%0.5820.636
FRESCO − lookup [95% CI]0.0 pts [−12.7, +12.7]−0.009 [−0.118, +0.089]+0.046 [−0.130, +0.209]

Cluster-held-out split. To test transfer beyond shared transcriptional clusters, we retrained FRESCO with six entire clusters withheld (123 genes; 122 eligible for evaluation). No test gene shares a cluster with a training gene, although similar genes may remain in other clusters. Expression preprocessing includes held-out genes’ expression; their treated images are excluded from generator training.

On this split, FRESCO achieves higher medcos and cos(e,d) than the lookup, with no detected top-1 advantage. We also retrained MorphDiff and Jones et al. (2026) on this split. FRESCO has the highest point estimates on all three metrics, tying Jones et al. on top-1.

OPS cluster-held-out, 122 genestop-1medcoscos(e,d)
FRESCO4.1%0.3710.403
kNN-1 lookup, pseudobulk cosine3.3%0.2370.251
FRESCO − lookup [95% CI]+0.8 pts [−2.5, +4.1]+0.135 [+0.009, +0.266]+0.151 [+0.012, +0.286]
MorphDiff, retrained on this split1.6%0.1260.227
Jones et al. (2026), retrained on this split4.1%0.2550.306

Intervals are from 10,000 paired gene-bootstrap samples. We will include both splits and the kNN baseline in the revision.

AuthorsOpenReview ↗

Prior work, model sizes and cross-channel attention (Q2–Q3)

Q2 / W2. Relation to CellFluxMM.

We agree that our formulation claim should acknowledge this prior work more clearly. We will cite CellFluxMM alongside MorphDiff and MultiVCDiff and narrow the claim accordingly. CellFluxMM jointly generates perturbed images and population-level L1000 expression. It starts from controls and processes the concatenated image and expression latents. FRESCO generates multi-channel single-cell images conditioned on resampled scRNA-seq summaries. It uses the control image as context, starts the flow from noise, and has explicit spatial and cross-channel attention.

We will replace the sentence the reviewer quotes with: “We study expression-conditioned single-cell morphology generation under treatment-level pairing, building on MorphDiff, MultiVCDiff and CellFluxMM.”

Q3a / W3. Model size.

The OPS and scGeneScope results in Table 1 use FRESCO’s 8-layer S model. The corresponding model sizes are shown below. Counts refer to trainable generative-network parameters; frozen pretrained autoencoders are excluded, and trained autoencoders are listed separately.

ModelModel size (OPS / scGeneScope)Trained autoencoder size (OPS / scGeneScope)
FRESCO56.8M / 56.8Mnone
CellFlux56.7M / 56.7Mnone
MultiVCDiff61.7M / 61.8Mnone
MorphDiff378.5M / 378.5M55.3M / 55.3M

Q3b / W3. Cross-channel ablation with conditioning metrics.

We retrained FRESCO without cross-channel blocks on both transcriptomic datasets, keeping the total block count fixed. Guidance was chosen as for FRESCO, and evaluation used the same samples:

OPS held-out (FID / top-1 / medcos)OPS seenscGS round 1scGS round 2
FRESCO3.28 / 25.9% / 0.572.79 / 30.0% / 0.6311.76 / 32.1% / 0.489.40 / 37.0% / 0.55
no cross-channel12.65 / 16.7% / 0.3912.25 / 18.0% / 0.5286.53 / 17.9% / 0.3777.24 / 40.7% / 0.50

The point estimates favour cross-channel attention on all three metrics for OPS and scGeneScope round 1, and on FID and medcos for round 2. Round-2 top-1 is higher without it (40.7% versus 37.0%, a difference of one compound out of 27). The ablation therefore measures more than FID. However, image quality also changes substantially, so it does not separate better conditioning from better image quality. We will include these results in the revision.

AuthorsOpenReview ↗

One-hot conditioning and control-image dropout (Q4–Q5)

Q4. Why one-hot beats expression on seen genes (52.0% vs 30.0%).

A one-hot label gives a fixed, unambiguous identity for a perturbation already seen during training, allowing the model to learn its associated morphology directly. An expression summary varies with the cells sampled and may overlap with those of other perturbations. The model must therefore infer perturbation-specific morphology from a variable input, which may make exact-identity retrieval harder. The comparison also uses different backbones, so it does not isolate the effect of conditioning alone.

Q5. Dropping out the control image during training.

We did not use control-image dropout in the reported runs. We trained a separate model with no control image at all. This shows whether the control helps, but not whether occasional dropout would make the model more robust when the control is missing or uninformative.

That model shows the control helps. Removing it worsens FID and retrieval in both scGeneScope rounds (round 1: FID 15.35 versus 11.76, top-1 28.6% versus 32.1%; round 2: 12.29 versus 9.40, 22.2% versus 37.0%). On held-out OPS genes, top-1 drops from 25.9% to 20.4% and medcos is similar (0.56 versus 0.57), while FID is slightly lower without the control (3.10 versus 3.28). In the intended use, a control image from the target plate is always available, so we did not optimise for a missing control.

We thank the reviewer for these questions and would be glad to discuss any of the results further.

Reviewer_skRPOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

the authors proposed a conditional fm based cellular morphology generation model (FRESCO), to learn mapping from transcriptomic resposne to multi-channel cellular morphology distribution under treatment-level correspondence without exact cell-level pairing. it trained a MM-DiT module in frozen vae latent space with program pooling, control anchoring and channel factorization. it uses two RNA-morphology screen dataset from scGeneScope and OPS explorer, as well as three other morphology-only dataset. FRESCO beats the best baseline on top-1 retrieval and median response cosine on OPS explorer data.

Critical strengths
  1. solves group-level multimodal generation problem with certain level of unseen perturbation generalization.
  2. control anchoring and cross-channel modeling introduces reasonable inductive bias supported by experiment result. e.g., separating gaussian nosie and control-image conditioning helps model utilize both acquisition context and stochastic generation, which seems to improve cross-replicate response matching.
Critical weaknesses
  1. baseline coverage and evaluation evidence is not sufficient to support the superior performance claim. CellFlux and IMPA are unfortunately outdated, meanwhile, the method proposed in Jones et al. (2026), though cited in the manuscript, was not compared.
  2. evaluation protocols across different methods are not fully disclosed.
  3. FID only evaluates the RGB channels but leave pre-defined morphology features untested, thus cannot fully evaluate biological meaningful morphology distributions across all channels.
  4. lack ablation on dense expression vs pragram-token vs joint condition, as well as fixed treatment-mean vs resampled expression, under same backbone, guidance and compute resource.
  5. missing paired uncertainty analysis on held-out gene top-1 retrieval. i'm open to raise my score significantly if all concerns can be addressed.
Questions for authors
  1. can u use cellprofiler to validate major claims regarding the performance gain on morphological level?
  2. can u reveal the per gene results, bootstrap condifence interval, or multi-seed variabiltiy in OPS explorer held-out gene retrieval task?
  3. please address other comments in weakness section
3 comments
AuthorsOpenReview ↗

Response to Reviewer skRP (1/3)

We thank the reviewer for the specific requests and for recognising the treatment-level formulation, control anchoring and cross-channel modelling. We answer each point below. Unless stated otherwise, OPS held-out results use the 54 eligible held-out genes (chance top-1 1.9%), and all methods are scored on the same fixed evaluation samples.

In brief.

  • W1: we add Jones et al. (2026) to all four transcriptome-conditioned settings. FRESCO has lower FID and higher top-1 and medcos in all four.
  • W2: one protocol table now lists expression input, guidance, control image, image space and model size for every method.
  • W3 / Q1: re-scoring with CellProfiler features on all channels preserves the conclusions of Table 1.
  • W4: we compare programme-token, dense and joint conditioning, and fixed-mean versus resampled training.
  • W5 / Q2: paired bootstrap intervals, per-gene results and three training seeds on held-out OPS genes.

W1. Baseline coverage (Jones et al. 2026 not compared; CellFlux / IMPA outdated)

Table 1 already includes MorphDiff (2025) and MultiVCDiff (2026). We have now added Jones et al. (2026) to all four transcriptome-conditioned settings. We retrained it with the authors' code, keeping their DiT architecture (scaled to 61.7M parameters to match FRESCO's 56.8M), rectified-flow objective, EMA, optimiser and gradient clipping, with the same training compute as FRESCO. Its published one-hot labels cannot represent genes absent from training, so the only component we add is the conditioning adaptor: an MLP over the same cNMF programme vectors that FRESCO uses. On OPS, FRESCO also receives a dense expression vector, so this compares the two methods as configured rather than isolating the architecture.

SettingMethodFID ↓cos(e,d) ↑top-1 ↑medcos ↑
scGeneScope, round 1Jones et al.49.250.64321.4%0.37
FRESCO11.760.64032.1%0.48
scGeneScope, round 2Jones et al.45.860.61325.9%0.22
FRESCO9.400.61637.0%0.55
OPS, seen genesJones et al.5.860.4957.0%0.25
FRESCO2.790.77830.0%0.63
OPS, held-out genesJones et al.6.380.5815.6%0.37
FRESCO3.280.68225.9%0.57

FRESCO is better on 15 of 16 entries: lower FID and higher top-1 and medcos in all four settings. cos(e,d) is similar on scGeneScope (within seed noise, ≈0.03) and higher for FRESCO on OPS. On held-out OPS genes, top-1 is 25.9% vs. 5.6% and FID is 3.28 vs. 6.38.

Guidance. The table samples Jones et al. without guidance amplification (weight 1). With its own guidance weight fitted on training data (2.24 on OPS, 1.46 on scGeneScope), its held-out OPS top-1 rises to 9.3% and medcos to 0.48, still well below FRESCO (25.9%, 0.57). FRESCO keeps lower FID and higher top-1 and medcos in all four settings. We will add Jones et al. to Table 1 and describe its adaptation and training in App. D.3.

Table 2 (BBBC021, RxRx1, JUMP) is a secondary, image-only benchmark; the main task of the paper is transcriptome-conditioned generation (Table 1). On the image-only benchmarks, Jones et al. use a much larger model (~700M parameters vs. 56.8M for FRESCO) together with a pretrained encoder (MolGPS, a 3B-parameter pretrained molecular encoder) that is not publicly available, so we cannot reproduce their setup and their published numbers are not comparable at our model size. We therefore compare against Jones et al. where it matters most: on the main task, retrained at matched model size and compute on both transcriptome–morphology datasets (table above).

AuthorsOpenReview ↗

Response to Reviewer skRP (2/3)

Response to Reviewer skRP (2/3)

W2. Evaluation protocols

We summarise the protocol of every method in one table, which we will add to the appendix:

MethodExpression inputGuidanceControl imageImage spaceSize (OPS / scGS)
FRESCOresampled scRNA-seq cells, pooled into cNMF programme tokens (plus a dense adaLN vector on OPS)CFG, weight chosen once (training split for OPS, validation split for scGS)matched control as contextfrozen SD-VAE latents56.8M / 56.8M
CellFluxtreatment expression profile over highly variable genesCFG, published weightcontrol image as flow sourcepixels56.7M / 56.7M
MorphDiffone fixed z-scored pseudobulk per treatmentnonenone (sampled from noise)latents of its own autoencoder378.5M / 378.5M (+55.3M autoencoder)
MultiVCDiffnone: perturbation identity (ECFP4 on scGS, Gene2vec on OPS); expression generated jointlynonenonefrozen SD-VAE latents61.7M / 61.8M
Jones et al.treatment-mean cNMF programme vectorweight 1 (fitted weight also reported, W1)none (sampled from noise)pixels61.7M / 61.8M

Sizes count trainable generative-network parameters; frozen pretrained autoencoders are excluded and trained ones are listed separately.

The remaining protocol is already in the paper:

  • Baselines (architecture, descriptor, guidance and sampling mode): App. D.3.
  • FRESCO configuration and training schedule: App. D.4 and App. C.
  • Guidance: FRESCO's weight is chosen once (on the training split for OPS, the validation split for scGeneScope) and then fixed for all evaluations; baselines use their published settings.
  • Seen vs. held-out genes: the seen-gene reference pools crops across splits, including training crops, so held-out genes are the primary test of generalisation on OPS (App. D.2).

New in the revision: the model sizes above, and an explicit statement that all methods are scored on the same fixed evaluation samples and real-image references.

W3 / Q1. FID uses RGB channels only; CellProfiler validation

We agree that FID alone does not assess all channels. Our conditioning metrics (cos(e,d), top-1 and medcos) already use every channel: 95 features for the five-channel and 138 for the seven-channel images, covering intensity, texture, granularity, shape and pairwise colocalisation (App. F.2).

CellProfiler validation (new). As requested, we re-scored all generated and real images with standard CellProfiler features measured on every channel, and recomputed the three conditioning metrics:

SettingMethodcos(e,d) ↑top-1 ↑medcos ↑
scGeneScope, round 1CellFlux0.66132.1%0.384
Jones et al.0.68225.0%0.360
FRESCO0.68539.3%0.452
scGeneScope, round 2CellFlux0.80140.7%0.771
Jones et al.0.61333.3%0.216
FRESCO0.66440.7%0.558
OPS, seen genesCellFlux0.4314.0%0.056
Jones et al.0.5464.0%0.225
FRESCO0.78125.0%0.511
OPS, held-out genesCellFlux0.2667.4%0.116
Jones et al.0.6087.4%0.181
FRESCO0.74424.1%0.459

The conclusions of Table 1 hold on CellProfiler features. FRESCO is best on 9 of the 12 metric–setting pairs and tied on one more, and it is ahead of Jones et al. on all 12. On OPS, FRESCO is best on every metric, for both seen and held-out genes, with top-1 of 25.0% and 24.1% against at most 4.0% and 7.4% for the baselines. On scGeneScope, FRESCO leads round 1 and CellFlux leads round 2, as in Table 1.

These metrics measure agreement in the treatment response; they do not show that the full morphology distribution is reproduced. We will add this analysis to the appendix and state this distinction next to the FID results.

AuthorsOpenReview ↗

Response to Reviewer skRP (3/3)

W4. Dense vs. programme vs. joint conditioning; fixed mean vs. resampling

Conditioning input (new). On scGeneScope we trained two variants alongside FRESCO: programme tokens plus a dense expression vector (same MM-DiT-S backbone as FRESCO), and the dense vector alone (ChannelSiT-S backbone). The dense vector enters through adaptive layer norm. Each variant's guidance weight is chosen on the validation split.

ConditioningBackboneRound 1: FID / top-1 / medcos / cos(e,d)Round 2: FID / top-1 / medcos / cos(e,d)
Programme tokens (FRESCO)MM-DiT-S11.76 / 32.1% / 0.478 / 0.6409.40 / 37.0% / 0.546 / 0.616
Programme tokens + dense vectorMM-DiT-S12.91 / 21.4% / 0.365 / 0.5768.68 / 40.7% / 0.678 / 0.660
Dense vector onlyChannelSiT-S14.34 / 32.1% / 0.426 / 0.6069.02 / 40.7% / 0.656 / 0.629

Programme tokens alone give the best round-1 results (top-1 tied with the dense vector); adding the dense vector gives the best round-2 results (top-1 again tied). Programme tokens are therefore an effective conditioning input, but they do not consistently beat a dense vector, and the dense-only variant uses a different backbone.

Fixed mean vs. resampled expression (new, OPS). We tested the effect of expression pool size using NMF conditioning computed from draws of 20, 100, 200 or 500 cells, with guidance fitted separately for each setting to match the observed effect size. Pooling 500 cells gives a conditioning vector close to the treatment mean, although all settings still use resampled inputs.

N pooled per drawcos(e,d)
2060.2%
10065.9%
20068.2%
50063.2%

Pooling more cells improves alignment between N=20 and N=200 and reduces the guidance needed to match the observed effect size. The apparent decline at N=500 is driven by three perturbations; excluding these, alignment increases monotonically through N=500. This supports using sufficiently large cell pools, but does not directly establish a benefit over fixed-mean training. We train with resampling because it reflects the realistic setting: at deployment, expression always comes from a finite K-cell sample, never from an exact population mean.

W5 / Q2. Paired uncertainty, per-gene results and training seeds (OPS held-out genes)

Paired bootstrap (new). We resampled the 54 eligible held-out genes with replacement 10,000 times, using the same genes for both methods and keeping the retrieval candidates and references fixed. 95% percentile intervals:

FRESCOMorphDiffPaired difference
top-125.9% [14.8, 38.9]11.1% [3.7, 20.4]+14.8 pts [+1.9, +27.8] (≤ 0 in 1.3% of resamples)
medcos0.573 [0.453, 0.679]0.277 [0.174, 0.357]+0.30 [+0.19, +0.42]

Per-gene results (new). FRESCO ranks the correct gene higher than MorphDiff for 33 of 54 genes, MorphDiff for 16, with 5 ties (Wilcoxon signed-rank p = 0.0017; sign test p = 0.021). The median rank of the correct gene among 54 candidates is 5 for FRESCO and 11.5 for MorphDiff. We will release a per-gene CSV (rank, cosine, cell counts, eligibility) and add a plot of the paired ranks to App. E.

Training seeds (new). We trained FRESCO three times independently and evaluated all runs on the same samples, with the paper's guidance weight (λ = 2.24, fitted on seed 0):

SeedFID ↓top-1 ↑medcos ↑cos(e,d) ↑
0 (paper)3.2825.9%0.5730.682
13.3927.8%0.4930.679
23.4420.4%0.5560.656
mean ± s.d.3.37 ± 0.0824.7 ± 3.9%0.541 ± 0.0420.672 ± 0.014

All three runs have higher top-1 and medcos than MorphDiff (11.1%, 0.277) and than Jones et al. with or without guidance (at most 9.3%, 0.48). The bootstrap varies the query genes for fixed outputs, while the seeds measure FRESCO's training variability; MorphDiff and Jones et al. have one run each.

We thank the reviewer for motivating these analyses. We hope they address the concerns, and if so we would be grateful if the reviewer would consider updating their assessment. We would be glad to discuss any remaining point.

Reviewer_PjL2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors present FRESCO, a latent flow-matching model with a channel-factorized diffusion transformer that generates single-cell fluorescence images conditioned on the image of a control cell from the same acquisition batch and an "expression profile" estimated from a random subset of the treatment's sequenced cells. The authors mix matrix factorization with deep model embeddings to yield tokens capturing transcriptomics (they call this "program pooling"), and utilize the control image as context to update a velocity field in the latent space (they call this "control anchoring"). The authors use two datasets, from Liu et al 2026 and Dapello et al 2025, to demonstrate this. Overall, the work is interesting and potentially a powerful new model architecture relevant to how biology is collected in these datasets. However, I found several of the claims about what the trained model learned to be unsubstantiated, and the baselines to be somewhat weak, which point to methodological issues in how the data is sampled and its ultimate utility in practice.

Critical strengths
  • The problem is well explained, and relevant modeling and dataset literature is cited.
  • The response metrics are well designed, with permutation and bootstrap checks.
  • The architecture is potentially interesting.
Critical weaknesses
  • Baselines are weak: the introduction states that models based on point-estimates might collapse -- is this seen in practice? Baseline sizes and tuning methods are unreported. Does MorphDiff use classifier-free guidance or any other schemes? Why "program pooling" -- the section states that programs provide an alternative to using a single dense vector, but this is never evaluated/ablated in the paper. Similarly, for the "control anchoring" claim, there is no ablation to show that using the control image as context is better than using it as the source. Given these are the two primary innovations motivating the architectural design, this seems to be the largest gap.
  • When forming and utilizing transcriptional programs, I do not understand how you select random subsets to include during inference. Were the programs/clusters formed using only the training data? For every gene that is held-out, have you previously seen cells corresponding to the same cluster? While this scheme avoids the transcriptomic out-of-distribution issue, I see this as a methodological issue here that is essentially train/test leakage. The central challenge is that you must generate these clusters independently of the cells being tested. This is not discussed in the paper at all, which is a significant gap.
  • Important but less severe: the following are claimed without evidence (either temper claims or provide evidence) (1) "deterministic model trained to match independently sampled expression-image pairs can collapse toward a treatment-level average rather than generate realistic individual cells" (2) "the model learns the coupling induced by treatment rather than an artificial cell-to-cell correspondence."
Questions for authors
  • Please explain evidence for the claims listed above.
  • Please justify why the sampling scheme is realistic and what the authors hope to learn from the model.
4 comments
AuthorsOpenReview ↗

Baseline configuration and collapse (W1a–W1c)

We thank the reviewer for the careful review. We address the concerns in four comments: baseline configuration, architectural ablations, programme fitting and generalisation, and claim clarification and intended use. In brief:

  • Baselines and ablations (W1): we report model sizes and tuning, and add two ablations. Using the control image as an anchor beats using it as the flow's starting point on OPS; on scGeneScope the result depends on the evaluation round. Programme tokens work well as a conditioning input but do not consistently beat a single dense expression vector.
  • Leakage (W2): we state exactly which data were used to fit the programmes (on OPS, this includes expression from held-out genes), and add an experiment that holds out entire transcriptional clusters.
  • Claims (W3/Q1, Q2): we clarify the theoretical scope of the collapse argument and replace the ambiguous coupling terminology.

W1a. Is collapse seen in practice?

We have not demonstrated collapse in the baselines. We agree that the wording should distinguish the theoretical argument from empirical evidence more precisely. We will replace the final two sentences of the Introduction's group-level pairing paragraph with:

A deterministic image predictor trained under squared-error loss has the conditional mean as its population optimum (Bishop, 1994). Generating a population that reflects variation among cells within a treatment instead calls for modelling the conditional distribution, for example through diffusion or flow matching.

This explains why we use a generative model. The diffusion and flow baselines can also generate varied cells, so this argument does not explain differences in performance between them and FRESCO.

Reference: Christopher M. Bishop. Mixture Density Networks. Technical Report NCRG/94/004, Aston University, 1994.

W1b. Baseline sizes and tuning.

The baseline guidance policy is stated in the Table 1 caption; the table below adds the checkpoint-specific parameter comparison. The table counts trainable parameters of each generative network from its checkpoint. Frozen pretrained autoencoders are not counted; trained ones are listed separately.

ModelGenerative network (OPS / scGeneScope)Trained autoencoder (OPS / scGeneScope)Generates in
FRESCO56.8M / 56.8Mnonefrozen SD-VAE latents
CellFlux56.7M / 56.7Mnonepixels
MultiVCDiff61.7M / 61.8Mnonefrozen SD-VAE latents
MorphDiff378.5M / 378.5M55.3M / 55.3Mlatents of its own autoencoder
Jones et al.61.7M / 61.8Mnonepixels

FRESCO is about the same size as CellFlux, slightly smaller than MultiVCDiff, Jones et al. and about 6.7 times smaller than MorphDiff's denoiser. Baselines use their published configurations and guidance settings. FRESCO's guidance weight is chosen once (on the training split for OPS, the validation split for scGeneScope) and then fixed for all evaluations. We will add this table to the paper, with each model's training epochs, guidance setting, conditioning input and whether it sees a control image.

W1c. Does MorphDiff use classifier-free guidance?

No. As specified in Appendix D.3, we use noise-initialised sampling with 100 DDIM steps and η = 0, without classifier-free guidance.

AuthorsOpenReview ↗

Architectural ablations (W1d–W1e)

W1d. Programme pooling versus dense expression.

Table 3 compares expression with identity conditioning; we now add the requested comparison with dense expression conditioning. For programme conditioning, we trained two more scGeneScope variants: programme tokens plus a dense expression vector (same backbone as FRESCO), and the dense vector alone (ChannelSiT-S backbone). The dense vector enters through adaptive layer norm. Each variant's guidance weight is chosen on the validation split. We report FID, top-1 retrieval, medcos and cos(e,d) below.

Conditioning (scGeneScope)BackboneRound 1: FID / top-1 / medcos / cos(e,d)Round 2: FID / top-1 / medcos / cos(e,d)
Programme tokens (FRESCO)MM-DiT-S11.76 / 32.1% / 0.478 / 0.6409.40 / 37.0% / 0.546 / 0.616
Programme tokens + dense vectorMM-DiT-S12.91 / 21.4% / 0.365 / 0.5768.68 / 40.7% / 0.678 / 0.660
Dense vector onlyChannelSiT-S14.34 / 32.1% / 0.426 / 0.6069.02 / 40.7% / 0.656 / 0.629

Programme tokens alone give the best round-1 results (top-1 tied with the dense vector); adding the dense vector gives the best round-2 results (top-1 again tied). So programme tokens are an effective way to condition the model, but these results do not show a consistent advantage over a dense vector, nor do they isolate the effect of the programme structure. We do not yet have paired confidence intervals. We will report these results and this limitation in the revision.

W1e. Control anchor vs. control as source.

We trained a variant in which the flow starts from the control cell instead of Gaussian noise, with no control token (ChannelMMDiT-S on both datasets). It uses the same data, expression input and training budget as FRESCO; guidance is chosen as described in comment 1, W1b. We also include a Gaussian-noise model with no control image.

Each entry below reports FID / top-1 / medcos.

ModelscGeneScope round 1scGeneScope round 2OPS seenOPS held-out
Noise start, control anchor (FRESCO)11.76 / 32.1% / 0.489.40 / 37.0% / 0.552.79 / 30.0% / 0.633.28 / 25.9% / 0.57
Noise start, no control15.35 / 28.6% / 0.3412.29 / 22.2% / 0.212.66 / 26.0% / 0.613.10 / 20.4% / 0.56
Control-as-source, no control anchor9.24 / 21.4% / 0.509.91 / 44.4% / 0.693.68 / 8.0% / 0.494.46 / 7.4% / 0.32

Compared with using the control as the starting point, using it as a control anchor is better on OPS (higher top-1 and medcos, lower FID, for both seen and held-out genes) and on scGeneScope round-1 top-1. Starting from the control gives higher round-2 top-1 and medcos, whereas the control anchor gives lower round-2 FID (9.40 versus 9.91). On round 1, starting from the control gives lower FID and higher medcos.

The copying diagnostics provide additional context:

  • 26.7% (round 2) and 4.6% (OPS) of control-as-source outputs fall below the real-control distance threshold, compared with 0% for FRESCO in both settings.
  • 17.4% (round 2) and 47.8% (OPS) are near-copies of another generated cell, compared with at most 0.6% for FRESCO.

With the expression input removed from both models, the control-as-source model still reaches 18.5% round-2 top-1 (FRESCO 7.4%; chance 3.7%).

Together, these results suggest that the control-as-source model more often produces outputs close to controls or to other generated cells, and that some round-2 retrieval performance persists without expression input. Shared round-2 controls (Appendix D.2) are a plausible contributor to this retrieval performance. In round 1, with no shared controls, the control-as-source model falls behind (21.4% vs. 32.1% top-1). We will report these checks and the distance thresholds with the comparison.

AuthorsOpenReview ↗

Programme fitting and held-out generalisation (W2)

W2. Programme fitting, inference sampling and possible leakage.

Two separate things are involved: the programmes (how expression is represented) and the clusters (how perturbations are split into training and test). We address each.

Programme fitting.

On OPS, the 20-programme cNMF fit used expression from all perturbations, including held-out ones. This corrects the training-only cNMF statement in Appendix D.2.

On scGeneScope, cNMF used only the 115,308 training cells (Appendix D.2).

For the OPS cluster-held-out split, we also refit cNMF using only training perturbations and obtained highly similar programmes (matched cosine 0.991–1.000). cNMF learns programmes from expression alone without image supervision, and held-out images enter neither programme fitting nor generator training. The high similarity after training-only refitting is reassuring about the stability of the programme representation.

Clusters.

Yes, a held-out gene can share a cluster with training genes: the split samples 5% of perturbations from each of 35 transcriptional clusters (Appendix D.2). These clusters are used to construct the split, not to pool the expression input.

Shared cluster membership alone does not establish train/test leakage, but this split tests new perturbations within represented transcriptional neighbourhoods, rather than transfer to unseen clusters. Performance may therefore partly reflect interpolation from similar training genes. The whole-cluster exclusion experiment below tests the stronger generalisation question.

Sampling at inference.

For each generated image, we average a fresh sample drawn with replacement from the requested perturbation, not its cluster (Eq. 5; Appendix A.3): 200 cells for OPS and 20 for scGeneScope (Table 6). The gene panel and programme loadings stay fixed. The model therefore needs measured expression for the requested perturbation. For held-out perturbations, treated images are excluded from generator training and are not inputs at inference.

New experiment: holding out whole clusters.

We thank the reviewer for raising the question of generalisation beyond transcriptional clusters represented in training. To test this, we held out six entire clusters (123 genes), so no held-out gene has a training gene from its cluster, and retrained FRESCO, MorphDiff and MiT on this split. As a simple baseline, we return images of the training perturbation whose average expression is most similar (nearest neighbour). Top-1 and medcos are reported for the 122 held-out genes that meet the eligibility criteria.

OPS, whole clusters held outFID ↓top-1medcoscos(e,d)
FRESCO2.434.1%0.3710.403
Nearest training perturbation (real images)0.81*3.3%0.2370.251
MorphDiff2.491.6%0.1260.227
MiT (Jones et al., 2026)5.184.1%0.2550.306

The nearest-neighbour baseline returns real training images, so its FID measures how close those images are to the test distribution, not generation quality. Bold FID marks the best trained generator.

FRESCO beats the nearest-neighbour baseline on medcos (paired difference +0.135, 95% CI [0.009, 0.266]), with no detectable difference in top-1. FRESCO also has higher medcos and cos(e,d) than MorphDiff and MiT. The revision will state the preprocessing used for each model and split.

AuthorsOpenReview ↗

Claim clarification and intended use (W3/Q1, Q2)

W3 / Q1. Evidence for “collapse” and “learns the coupling”.

As explained in comment 1, W1a, we have not demonstrated baseline collapse and will limit the conditional-mean argument to deterministic image regression. Appendix A.1 states that no cell-level correspondence is recovered. We agree that “learns the coupling” is ambiguous and will replace that sentence in the paragraph introducing FRESCO with:

Because expression and morphology are measured in different cells, FRESCO models the distribution of cell morphologies conditioned on a treatment's expression summary and acquisition context, without assuming cell-to-cell correspondence between the two assays.

Q2. Why this sampling scheme is realistic, and what we hope to learn.

The intended use is to predict morphology for a perturbation that has been sequenced but not imaged, given a suitable control image from the target acquisition context. Its expression profile is estimated from finitely many cells, so we train on fresh m-cell averages to expose the model to sampling variability, using the same procedure at inference (Appendix A.3).

The OPS explorer dataset of Liu et al. (2026) which forms the multimodal genetic perturbation benchmark in this paper profiles the same gene knockouts through imaging and scRNA-seq, motivating us to ask which morphological phenotypes are predictable from the measured transcriptional response. We hope to learn how the distributions of these phenotypes differ between treatments and identify candidate phenotypes for follow-up imaging of sequenced but unimaged perturbations to identify morphological associated with them. The aim is treatment-level prediction, not recovery of an individual sequenced cell's morphology.

We hope these results address the reviewer's concerns and would be glad to discuss any of them further.

Comments on the paper

2 comments

We thank the reviewers for their constructive feedback. All three found the work informative, and PjL2 rated the presentation clear and easy to follow. Reviewers highlighted the well-motivated treatment-level formulation (uoZo), well-designed response metrics (PjL2), and control anchoring and cross-channel modelling as experimentally supported inductive biases (skRP). uoZo also recognised the image-only benchmark gains and released code.

The new analyses address four shared concerns; the individual replies give the full results.

  1. Stronger baselines and uncertainty (skRP W1, W5/Q2; uoZo W1). Across 54 eligible OPS genes, top-1 is 25.9% versus MorphDiff’s 11.1%, a paired gain of +14.8 pts [95% CI: +1.9, +27.8]; medcos improves by +0.30 [+0.19, +0.42]. All three FRESCO training runs exceed MorphDiff and Jones et al. on both metrics; those baselines have one run each. We added Jones et al. at a similar model size: FRESCO has lower FID and higher top-1 and medcos in all four settings, including with Jones et al.’s fitted guidance.

  2. Nearest-neighbour comparison and harder holdout (PjL2 W2; uoZo Q1b). FRESCO and nearest-neighbour image transfer achieve similar scores on the original split. To test transfer beyond shared transcriptional clusters, we introduced a second, harder split withholding six entire clusters. FRESCO improves medcos by +0.135 [95% CI: +0.009, +0.266] and cos(e,d) by +0.151 [+0.012, +0.286], with no detected top-1 advantage. Against MorphDiff and Jones et al. retrained on this split, FRESCO has higher response-direction scores and ties Jones et al. on top-1.

  3. Expression dependence and all-channel validation (uoZo Q1a; skRP W3/Q1). Wrong-gene expression reduces OPS top-1 to 1.9–5.6% (chance 1.9%). On scGeneScope, correct-compound expression retains positive alignment across sequencing rounds; shuffled compounds lose it. These controls support perturbation-related expression and argue against a batch-only explanation. With CellProfiler features from every channel, held-out OPS top-1 is 24.1% versus 7.4% for both CellFlux and Jones et al. FRESCO leads scGeneScope round 1 and CellFlux leads round 2. MMD-based cos(e,d) assesses distributional response alignment, complementing RGB-based FID.

  4. Architectural ablations (PjL2 W1; skRP W4; uoZo Q3–Q5). On held-out OPS genes, control-as-context achieves 25.9% top-1 versus 7.4% for control-as-source. Removing cross-channel blocks lowers top-1 25.9% → 16.7% and medcos 0.57 → 0.39. These are point estimates; image quality also changes, and scGeneScope results vary by round and metric. Programme tokens provide effective conditioning, with performance varying across settings. The pool-size analysis supports sufficiently large expression pools; a benefit over fixed-mean training remains unestablished.

We will add these results, clarify model protocols and preprocessing, narrow the collapse and coupling claims, and acknowledge CellFluxMM alongside MorphDiff and MultiVCDiff. We thank the reviewers and welcome further discussion.

Revised manuscript. We have uploaded a revised manuscript, with changes in blue. Following the reviewers' requests, it:

  • adds Jones et al. (2026) to Table 1, and a protocol table listing each method's expression input, guidance, control image, image space and model size (Table 6; skRP W1–W2, PjL2 W1b, uoZo Q3a);
  • replaces the collapse and coupling sentences and narrows the contribution statement, citing CellFluxMM (Section 1; PjL2 W1a, W3/Q1, uoZo Q2);
  • corrects the OPS programme-fitting statement and describes how the transcriptional clusters define the held-out split (Appendix D.2; PjL2 W2);
  • reports the new analyses in Appendix E: control as context versus control as flow source (E.3), cross-channel attention on all metrics (E.4), programme tokens versus dense expression (E.5), expression pool size (E.6), paired bootstrap, per-gene ranks and training seeds (E.7), nearest-neighbour lookup and whole-cluster held-out split (E.8), expression versus sequencing batch (E.9), and CellProfiler features (E.10).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.