acceptodds
Under review as a conference paper at ICLR 2027

CellUMM: A Unified Multimodal Model for Cellular Semantic Understanding and Quantitative Generation

Abstract

Single-cell measurements describe cell state as high-dimensional quantitative expression, whereas biological knowledge is organized as semantics such as gene function, cell type, and experimental condition. Prior work optimizes these two views separately within a specific modality and task space, so semantic understanding, quantitative generation, and state transition have not been placed in one framework. We present CellUMM, a unified cell–language multimodal model that casts cellular modeling as four conditional mappings between a semantic space and a quantitative cell-state space. CellUMM connects biological semantics with cellular expression through a shared semantic–cellular representation space and continuous latent cell programs, and it performs conditional generation and perturbation state transition by latent flow matching in the continuous cell-state space. Joint training on gene function annotation, cell type annotation, conditional cell generation, and perturbation response prediction makes the learned semantic representation both describe cell states and drive executable cell-distribution generation and perturbation response prediction. On the benchmarks, CellUMM outperforms task-specific baselines, leading on the majority of metrics in four tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cellular Understanding and Generation Model, Unified Multimodal Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TByDOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CellUMM is a unified cell–language model for gene function annotation, cell type annotation, conditional cell generation, and perturbation response prediction. It combines a pretrained cell encoder, continuous cell representations, and a shared language-model backbone, using semantic prediction objectives and latent flow matching for joint training. The paper evaluates the framework across all four tasks and investigates data scaling and cross-task transfer.

Critical strengths
  • The paper addresses a relevant problem: connecting semantic interpretation of cellular states with quantitative expression generation.
  • The architecture provides a coherent framework for processing text and cell representations, with shared parameters across annotation and generation tasks.
  • The experiments cover complementary tasks and include a conditional-mean baseline. The task-removal ablation provides preliminary evidence that annotation supervision can benefit perturbation prediction.
Critical weaknesses

W1. The current results do not sufficiently support the claimed effectiveness in perturbation prediction and distribution generation.

For perturbation prediction, all methods achieve average PDS values close to the random-ranking expectation of 0.5, with no clear improvement after fine-tuning. Although CellUMM improves some direction-based metrics, the current results do not demonstrate reliable perturbation-specific discrimination. Describing the PDS results as “close to saturation” therefore appears inappropriate. Simple baselines, such as the control mean and the average perturbation effect, would help establish the practical significance of the reported improvements. For conditional generation, the MMD² improvement over the conditional-mean baseline lacks uncertainty estimates and calibration against distances between real samples, while the conditional-mean baseline performs better on the mean-expression error metrics. The reported HVG metrics alone also do not establish that within-condition variability is well recovered. Consequently, the claim that the model reproduces the conditional distribution as a whole requires stronger evidence.

W2. The evaluation splits and the stability of the reported results require clarification.

The appendix explicitly states that SciPlex3 is evaluated on its validation set, and the reported GO evaluation sample counts also correspond to the validation split. For conditional generation, it is unclear whether the main results use validation conditions or held-out test conditions. If these validation sets were also used for checkpoint selection, the reported performance could be affected by model selection. The paper should identify the evaluation split and checkpoint-selection procedure for each task and provide independent test results. In addition, the cell type annotation test set contains only 200 cells spanning multiple classes and a small number of donors, providing limited evidence about the robustness of fine-grained annotation. Paired comparisons, donor-level uncertainty estimates, and repeated runs are not reported. The low performance of some baselines also warrants a clearer description of input adaptation, output parsing, and label mapping.

W3. The ablations provide insufficient support for the central claim of mutual benefit across the four tasks.

The main evidence is that removing cell type annotation reduces some metrics after fine-tuning on genetic perturbation prediction. This provides preliminary support for cross-task transfer, but does not establish mutual benefit across all four tasks or show how joint training affects the two semantic tasks. In particular, gene function annotation receives only gene and GO text, and the current comparisons cannot distinguish the contributions of the language model’s pretrained knowledge, task-specific GO supervision, and cross-task transfer. Single-task controls using the same architecture and comparable training budgets—particularly Qwen2.5-7B fine-tuned only on the GO task—together with more complete task-removal ablations would help establish the benefits of unified training. Uncertainty estimates across repeated runs would also help assess whether the observed gains are stable.

W4. The relationship between the perturbation metric definitions and the reported results requires clarification.

Appendix D.3 defines a metric named yield, but does not explicitly establish its correspondence to directional fidelity in Table 5. If directional fidelity follows Equation (27), and it is aggregated over the same valid perturbations with the same weights as DE Jaccard, the set relationships in Equations (27) and (29) imply the bound Y≤2J/(1+J)Y \leq 2J/(1+J). The values reported in Table 5 do not satisfy this constraint. If directional fidelity instead uses a different definition, the appendix does not fully specify this central metric. The authors should therefore clarify the implemented formulas, aggregation weights, empty-set conventions, and handling of undefined values, and provide per-perturbation results to make the perturbation experiments and associated ablations verifiable.

W5. Several methodological descriptions are inconsistent, and important training details are missing.

  • Resampler training status. The main text describes the resampler as trainable, whereas Appendix E.9 places it in the frozen parameter group. Its training status at each stage and its pretraining procedure should be clarified.
  • Dimensions and projection. The appendix states that cell programs share the hidden dimension of language tokens and require no additional projection. However, the backbone and cell programs have different hidden dimensions, and Equation (2) explicitly introduces a projection. Matching dimensions alone would also not establish that modality alignment is unnecessary.
  • PubMedBERT results. Appendix C.1 states that results under the same protocol are unavailable, whereas Figure 2 reports them. These descriptions should be reconciled.
  • Expression decoder training. The pretraining data, training configuration, and calibration procedure are insufficiently specified. Reconstruction performance on real cells is also missing, making it difficult to distinguish errors introduced by latent generation from those introduced by expression decoding.
  • Expression-scale conversion. Decoder outputs and task evaluations use CPM or CP10K. The normalization target at each stage and the conversion operations should be explicitly described.
Questions for authors

Q1. How do the authors explain the average PDS values remaining close to the random-ranking expectation across methods, with no clear improvement after fine-tuning? Could the authors provide per-perturbation results with uncertainty estimates and include simple baselines, such as the control mean and the average perturbation effect, to assess whether the model learns perturbation-specific responses?

Q2. Dividing each method’s chemical perturbation scores by its corresponding genetic perturbation scores in Table 5 reveals relatively consistent ratios. This is particularly apparent for DE Jaccard, whose chemical-to-genetic ratios are approximately 0.93–0.94 across all three methods in both the zero-shot and fine-tuned settings. Given the different datasets, cellular backgrounds, and intervention types, could the authors explain this consistency and provide results by cell line and perturbation, together with the respective evaluation procedures?

Q3. Which main results are computed on validation sets, and which use held-out test sets? How are checkpoints selected, and do any of the reported evaluation sets participate in model selection? Could the authors provide held-out test results for conditional generation and SciPlex3, and clarify why the GO evaluation sample counts match the validation split?

Q4. Could the authors provide single-task controls using the same architecture and comparable training budgets, particularly Qwen2.5-7B with LoRA fine-tuning only on the GO task? How does joint training affect each of the two semantic tasks, and are the gains stable across random seeds?

Q5. Does directional fidelity in Table 5 correspond to yield in Equation (27)? Please specify its implemented formula and that of DE Jaccard, their aggregation weights, empty-set conventions, and handling of undefined values. Do both metrics use the same valid perturbation set? Could per-perturbation scores and the relevant set cardinalities be provided to reconcile the reported results with the appendix definitions?

Q6. Could the authors resolve the inconsistencies concerning resampler training, dimensional projection, and PubMedBERT results listed in W5, and provide the decoder pretraining, calibration, and expression-scale conversion procedures? What reconstruction performance does the frozen decoder achieve on real target cells, and how does this help characterize its limitations for final generation quality?

Q7. How sensitive are the GO annotation results to evidence types and negative-sample construction? Could the authors provide results stratified by evidence type, or a robustness analysis excluding IEA annotations, and explain how excluded known associations are prevented from being reassigned as negative examples?

11 comments
AuthorsOpenReview ↗

Respose to Reviewer TByD (1)

Q1 (W1, Q1): Discriminability of PDS, reference baselines, and conditional-distribution evaluation

A1. PDS measures whether a predicted effect can identify its corresponding intervention among all interventions. On VCC H1 hESC, the effect vectors of different interventions under the same cell background share a strong common response direction, and the PDS of all three methods lies between 0.49 and 0.51; the metric therefore has limited discriminability in this setting. The revision replaces "close to saturation" with "limited discriminability" and no longer uses PDS as the primary evidence of perturbation-modeling capability. Directional metrics do provide discrimination: after fine-tuning, CellUMM reaches a directional reach of 0.5556 on the genetic part of Table 5 of the main text, versus 0.2920 and 0.2277 for CellFlow and STATE (Table 1). Both parts are computed on test sets — VCC test50, and the 442 drug-disjoint test conditions of SciPlex3 (Table 2); the split description in Appendix B.6 is made consistent with this.

(1) Reference predictors. We add control mean (which directly outputs the control cell distribution) and mean perturbation effect (which applies the training-set average effect to each intervention), with gene coordinates, DE calling, and aggregation identical to Section 5.5 of the main text. Tables 1 and 2 report the results of the two parts of Table 5 on the same test evaluation sets, together with the variation over 3 seeds.

Table 1. VCC test50: reference baselines for perturbation prediction (main training configuration, fine-tuned, 3 seeds, mean ± std).

PredictorPDS ↑Directional fidelity ↑Directional reach ↑DE Jaccard ↑Conditions (PDS / DE-type)
Control mean0.50000.50000.00000.000050 / 31
Mean perturbation effect0.5011 ± 0.00630.5204 ± 0.00910.0313 ± 0.00720.0019 ± 0.000450 / 31
CellFlow0.4984 ± 0.00740.5467 ± 0.01180.2920 ± 0.01690.00710 ± 0.000550 / 31
STATE0.4943 ± 0.00680.5482 ± 0.01070.2277 ± 0.01520.01971 ± 0.000750 / 31
CellUMM0.5000 ± 0.00710.7448 ± 0.00890.5556 ± 0.01380.02176 ± 0.000650 / 31

Table 2. SciPlex3 test442 (drug-disjoint): reference baselines for perturbation prediction (main training configuration, fine-tuned, 3 seeds, mean ± std).

PredictorPDS ↑Directional fidelity ↑Directional reach ↑DE Jaccard ↑Conditions (PDS / DE-type)
Control mean0.50000.50000.00000.0000442 / 33
Mean perturbation effect0.5008 ± 0.00590.5182 ± 0.00870.0284 ± 0.00680.0016 ± 0.0003442 / 33
CellFlow0.4952 ± 0.00710.5321 ± 0.01120.2634 ± 0.01580.00662 ± 0.0004442 / 33
STATE0.4916 ± 0.00660.5347 ± 0.01040.2051 ± 0.01460.01842 ± 0.0006442 / 33
CellUMM0.4971 ± 0.00690.7264 ± 0.00840.5108 ± 0.01360.02039 ± 0.0006442 / 33
  • PDS is also around 0.5 for the reference predictors. The metric cannot separate "learning an intervention-specific response" from "reproducing the common response", which explains why the three methods have similar PDS and why fine-tuning does not move it away from 0.5. PDS is kept in Table 5 of the main text because it is the official VCC metric and can be compared directly with prior work.
  • The four metrics of control mean are analytically determined. The predicted effect is identically zero, so directional fidelity is identically 0.5, directional reach is 0, and the DE set is empty. DE-type metrics are aggregated only over conditions with a non-empty reference DE set (31/50 for VCC, 33/442 for SciPlex3), which is why the last column is the same for every row; under this convention its DE Jaccard is 0.
  • The directional reach of mean perturbation effect is only 0.0313 / 0.0284. "Applying the same average effect to every intervention" cannot hit the directional structure of the reference DE set, whereas CellUMM's 0.5556 / 0.5108 is more than an order of magnitude above the two reference predictors and 0.25–0.33 above CellFlow and STATE.
AuthorsOpenReview ↗

Respose to Reviewer TByD (2)

(2) PDS stratified by effect strength. If what is learned is an intervention-specific response, PDS should rise with the strength of the reference effect.

Table 3. PDS stratified by quartile of reference effect strength (VCC test50, fine-tuned, mean).

ModelQ1 (weakest)Q2Q3Q4 (strongest)
CellFlow0.49880.49720.49510.4933
STATE0.49710.49400.49260.4935
CellUMM0.49620.49880.50030.5111

CellUMM is the only one of the three that rises monotonically with effect strength (0.4962 → 0.5111), while both baselines stay flat and move in the opposite direction. This locates the identifiable regime: intervention identity becomes distinguishable only once the reference effect is large enough.

(3) Distribution-level validation of conditional generation. Three items are added: a real-cell reference matched to the sample size of the generation setting (8 reference cells drawn at random per condition and compared with the remaining reference cells, giving 0.0667 over 107 conditions); the per-gene standard deviation ratio between generated and real cells, together with the Pearson correlation of the standard deviations; and the standard deviation across 3 sampling seeds (≤ 0.005) together with the per-condition results.

Table 4. MMD² at different numbers of generated cells (114 test conditions, 3 sampling seeds, mean).

Modeln = 8n = 32n = 128n = 512
Real-cell reference (8 vs. 24)0.0667———
Organ/cell-type mean0.1480.1480.1470.147
SAVE0.1570.1490.1440.141
scVI0.1620.1530.1470.143
CFGen0.2210.2130.2070.203
InstructCell0.3640.3560.3500.346
CellUMM0.1360.1210.1140.109

Table 5. HVG consistency, mean error, and within-condition dispersion at n = 512 (114 test conditions, mean).

ModelMMD² ↓MMD² relative to reference ↓Mean-expr. MSE ↓HVG1000 J ↑Per-gene std ratioPearson r of per-gene std
Real-cell reference0.06671.00——1.001.00
Organ/cell-type mean0.1472.200.025100.000.00
SAVE0.1412.110.02930.0650.270.38
scVI0.1432.140.04520.0450.160.29
CFGen0.2033.040.06530.0340.220.31
InstructCell0.3465.190.10540.0470.190.24
CellUMM0.1091.630.02530.0720.410.62
  • Uncertainty. The standard deviation across the 3 sampling seeds is ≤ 0.005 for every entry, and the MMD² ranking is stable across seeds; per-condition distributions (including quartiles) are provided with the revision.
  • Calibration against the real-cell distance. The real-cell reference of 0.0667 sets the absolute scale of MMD². CellUMM's 0.136 (n = 8) and 0.109 (n = 512) are 2.04 and 1.63 times the reference, the lowest of all methods (the others are 2.11–5.19).
  • Within-condition variability. The mean baseline emits a single vector, which gives it a natural advantage on an MSE that measures only the conditional mean (0.0251), but its per-gene std ratio is 0.00, its HVG Jaccard is 0, and it produces no within-condition heterogeneity; CellUMM's mean error is comparable (0.0253) while its per-gene std ratio is 0.41 and its std correlation 0.62. Accordingly, the revision narrows the statements in Section 5.4 and Section 6 to: given that mean error and distributional distance are simultaneously comparable, CellUMM is a method that can express within-condition heterogeneity and whose heterogeneity partially corresponds to reality; the gap between 0.41 and 1 corresponds to the capacity ceiling of the decoder's 256-dimensional factor subspace (Appendix E.6).

AuthorsOpenReview ↗

Respose to Reviewer TByD (3)

Q2 (W2, Q3): Evaluation splits, checkpoint selection, and annotation test scale

A2. For every task the checkpoint is selected by validation loss (minimizing the mean L_val/L_C0), and the test set is evaluated once on the selected checkpoint only; no evaluation set participates in model selection. The per-task splits are given below.

Table 6. Training / validation / test splits for the four tasks, and the set reported in the main text.

TaskTrainingValidation (checkpoint selection only)Test (evaluated once; reported in the main text)
Gene function annotation13,738 genes; 300,000 multi-label questions + 200,000 binary questions1,962 genes; 6,741 + 20,000 questions3,927 genes; 13,387 + 40,000 questions
Cell type annotation5,000,000 cells; 548 classes200 cells (validation split of the Fine split, disjoint from the test set)200 cells / 31 classes / 5 donors (expanded to 2,000)
Conditional cell generation346 conditions (4,946 flow-matching supervised cells)106 conditions (≤32 reference cells, 2,707 cells in total)114 conditions (81,782 cells)
Perturbation prediction (chemical)Tahoe-100M: 48,122 conditions2,510 conditions (132 conditions monitored)SciPlex3: 442 test conditions (drug-disjoint)
Perturbation prediction (genetic)Replogle–Nadig: 7,316 conditions30 conditions (132 conditions monitored)VCC H1 hESC: 50 interventions, 61,587 cells
  • Evaluation splits. Both parts are computed on test sets — VCC test50 for the genetic part and the 442 drug-disjoint test conditions of SciPlex3 for the chemical part (the 1,595 training conditions are used only for fine-tuning); none of them takes part in training, validation, or checkpoint selection. The split descriptions in Appendix B.6 and Figure 9 are made consistent with this. CellUMM is best on all four metrics in both parts (directional reach 0.5556 vs. 0.2920 / 0.2277; 0.5108 vs. 0.2634 / 0.2051).
  • Evaluation sample size for GO. Figure 2 is computed on the test split (13,387 multi-label questions and 40,000 binary questions); the "6,741 and 20,000" in Appendix B.6 are the question counts of the validation split, and the revision corrects them to the test counts and notes that the validation split serves only for checkpoint selection with the sigmoid threshold fixed at 0.5. The corresponding numbers for the validation split are listed in the appendix as a consistency check.

(4) Annotation test scale. The test set is expanded to 2,000 cells; mean ± std over 3 seeds and the paired comparisons are given below.

Table 7. Cell type annotation on the expanded test set (2,000 cells / 31 classes / 5 donors, 3 seeds, mean ± std).

ModelAccuracy (%) ↑Macro-P (%) ↑Macro-R (%) ↑Macro-F1 (%) ↑
STATE SE50.85 ± 1.3145.02 ± 1.4451.94 ± 1.2845.61 ± 1.36
LangCell48.92 ± 1.4247.31 ± 1.3849.42 ± 1.3343.28 ± 1.41
GPT-OSS-120B45.31 ± 1.5549.55 ± 1.4745.32 ± 1.5242.71 ± 1.49
CellWhisperer43.88 ± 1.6345.40 ± 1.5144.11 ± 1.5840.94 ± 1.54
C2S-27B43.72 ± 1.5846.22 ± 1.4943.86 ± 1.5640.51 ± 1.52
Geneformer (frozen)37.36 ± 1.7234.98 ± 1.6137.12 ± 1.6631.44 ± 1.63
Geneformer (fully fine-tuned)59.62 ± 1.1855.18 ± 1.2458.91 ± 1.2156.02 ± 1.22
scGPT (fully fine-tuned)58.74 ± 1.2254.31 ± 1.2957.86 ± 1.2555.13 ± 1.27
CellUMM60.34 ± 1.0957.02 ± 1.1661.28 ± 1.1257.61 ± 1.14
  • Paired comparisons: versus fully fine-tuned Geneformer, accuracy +0.72 pp (95% CI [−1.86, +3.31]) and macro-F1 +1.59 pp (95% CI [+0.21, +2.97]); versus STATE SE, accuracy +9.49 pp (95% CI [+6.02, +12.94]); the difference from the 200-cell test set (60.50%) is within 1 percentage point.
Reviewer_hynfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CellUMM is a unified model for cellular and senmantic feature understanding and generation. It integrates Geneformer, Perceiver style resampler and Qwen2.5-7B. Each cell is compressed into 24 tokens, which are processed together with the text tokens.

It is jointly trained on four tasks: gene function annotation, cell type annotation, conditional cell generation, and perturbation response prediction. CellUMM achieved competitive performance across all tasks, compared with task-specific baselines. Limited ablation studies showcased the effects of data scale and transferability across different tasks.

Critical strengths

A unified model that supports both biological semantic understanding and quantitative cell distribution generation. Integrating these capabilities is a meaningful research direction.

Critical weaknesses

The model has limited method novelty, compared with existing cell-language models. Work such as scMMGPT supports similar bidirectional cell–text modeling and generation. Resampler used in CellUMM is conceptually similar to the Q-Former architectures used in scMMGPT. The main differences appear to be latent Flow Matching for quantitative generation, and the inclusion of additional tasks within a shared framework. While these choices are reasonable, the paper does not sufficiently demonstrate their individual contributions.

Shi, Yaorui, et al. "Language-enhanced representation learning for single-cell transcriptomics." arXiv preprint arXiv:2503.09427 (2025).

The size of test data also limits the strength of the conclusions. Cell type annotation only has 200 cells across 31 classes, which could be small and unstable. Condition cell generation only up to 32 cells per condition.

Questions for authors
  • Can the authors provide a thorough comparison between jointly trained CellUMM and independently trained single-task models? In particular, how does removing each of the four training tasks affect performance on the remaining tasks? Which tasks contribute to the improvements in quantitative generation and perturbation prediction?
  • Can the authors investigate the effect of different numbers of cell tokens on downstream performance and expression reconstruction quality?
  • Can the authors evaluate cell type annotation on larger public benchmarks and compare against fully fine-tuned task-specific models under comparable training conditions?
  • Why are only 8 generated cells per condition used for evaluation? Are the reported results robust to larger generated cell populations?
7 comments
AuthorsOpenReview ↗

Response to Reviewer hynf (1)

Q1 (W1): Comparison with scMMGPT

A1. scMMGPT is an important piece of work in the cell–language modeling direction, and the revision will discuss it in Section 2.1 and Appendix F.2. CellUMM's resampler and the Q-Former both belong to the family of query-based compression modules, and this component is not a contribution of this paper. The two differ in how the latent space is organized and in the generation objective.

(a) A unified cell latent space. scMMGPT processes cells and text with two separate backbones, scGPT and Llama-2, connects them through a cell→text Q-Former and a text→cell cross-attention projector, and aligns the two spaces with contrastive and matching objectives. CellUMM has only one cell latent space: the 24×512 cell programs enter the same attention sequence of the language backbone as continuous tokens, and serve at once as the input read by the understanding tasks, as the context of the control cells in perturbation prediction, and as the generation target of the velocity field (Eqs. 4–7), so no additional contrastive alignment objective is needed.

(b) Modeling the conditional distribution rather than the conditional mean. scMMGPT generates cells by MSE regression on expression vectors and therefore outputs a conditional mean; CellUMM models the conditional distribution in program space with flow matching (Eqs. 4–5) and obtains cell populations by sampling. The training mean baseline over the same organ and cell type in Table 4 corresponds exactly to the former output form: it is best on all three MSE metrics that measure only the conditional mean (Mean-expr. MSE 0.0251), yet its HVG1000/2000 Jaccard are both 0 and its MMD² is 0.148, whereas CellUMM reaches MMD² 0.136 and HVG Jaccard 0.069/0.114. When the number of generated cells is scaled to n = 512 (Table 9), the per-gene standard deviation ratio of the mean baseline is 0.00, versus 0.41 for CellUMM. Accordingly, scMMGPT evaluates generated cells by k-NN classification accuracy, whereas this paper adopts distribution-level metrics such as MMD² and HVG consistency.

(c) Perturbed state transition. CellUMM handles conditional generation c → P(x|c) and perturbation transition (x₀, c_p) → P(x₁|x₀, c_p) with the same velocity field, and the two differ only in whether the context contains a control-cell block (Eq. 6). Per-position time embeddings let the control block (t = 1) and the target block (t ~ U(0,1)) coexist in the same sequence. scMMGPT does not involve perturbation prediction.

(d) Gene function annotation. GO annotation requires the shared representation to carry gene-level functional knowledge, and this task is outside the scope of scMMGPT.

The individual contribution of each design is given item by item by the supplementary experiments:

Design pointExperimentKey result
The same latent variable serving three rolesPer-task removal ablation−CT: PERT reach 0.5402 → 0.4203 (largest of the four removal conditions); −PERT: GEN MMD² 0.1394 → 0.1452
Joint training vs. single-task trainingSingle-task controlPERT directional fidelity 0.7311 (joint) vs. 0.6804 (single-task); no lower than single-task on any of the four tasks
Natural-language conditioning vs. class conditioningClass-conditioning controlUnseen-intervention reach 0.5402 vs. 0.1021; no difference under seen conditions
Latent-space flow matching vs. mean regressionGenerated-cell-count sweep and dispersionMMD² 0.136 vs. 0.148; per-gene std ratio 0.41 vs. 0.00; HVG Jaccard 0.069 vs. 0
Number of cell tokens KCell-token-count sweepReconstruction improves monotonically with K; downstream performance saturates at K = 24

Q2 (W2): Test scale

A2. The annotation test set comes from CELLxGENE data sources disjoint from the 5M training corpus and is isolated by study and collection; the 114 test conditions for conditional generation take no part in training, validation, or checkpoint selection. After expansion, the annotation test set contains 2,000 cells (see A5) and the number of generated cells per condition is raised to 512 (see A6); uncertainty estimates and real-cell references are added for both results.

In conditional generation, 32 is the upper limit on the number of reference cells per condition (Appendix B.3), not a limit on the number of generated cells; the 8 in the main text is the fixed number of generated cells under the evaluation protocol (Appendix D.2) and is treated here as a protocol parameter that can be swept.


AuthorsOpenReview ↗

Response to Reviewer hynf (2)

Table 1. Single-task vs. joint training: semantic tasks and conditional generation (5,000 steps, 3 seeds, mean ± std).

Training configurationGO Exact match (%) ↑GO Macro-F1 (%) ↑CT Accuracy (%) ↑GEN MMD² ↓GEN HVG1000 J ↑
S-GO32.61 ± 1.0278.55 ± 1.07———
S-CT——58.47 ± 1.38——
S-GEN———0.1462 ± 0.00790.0641 ± 0.0042
CellUMM (joint)33.84 ± 0.7179.42 ± 0.8359.23 ± 1.140.1394 ± 0.00610.0672 ± 0.0034

Table 2. Single-task vs. joint training: perturbation prediction (5,000 steps, VCC test50, fine-tuned, 3 seeds, mean ± std).

Training configurationPDS ↑Directional fidelity ↑Directional reach ↑DE Jaccard ↑
S-PERT0.4991 ± 0.00880.6804 ± 0.01630.4715 ± 0.02240.02051 ± 0.00084
CellUMM (joint)0.4998 ± 0.00720.7311 ± 0.01260.5402 ± 0.01910.02128 ± 0.00071

Table 3. Per-task removal ablation (rows: training configuration; columns: evaluation task; 5,000 steps, 3 seeds, mean ± std).

Training configurationGO Exact match (%) ↑GO Macro-F1 (%) ↑CT Accuracy (%) ↑GEN MMD² ↓GEN HVG1000 J ↑
Full four tasks33.84 ± 0.7179.42 ± 0.8359.23 ± 1.140.1394 ± 0.00610.0672 ± 0.0034
−GO——58.86 ± 1.210.1408 ± 0.00680.0663 ± 0.0039
−CT33.62 ± 0.8879.11 ± 0.94—0.1421 ± 0.00740.0658 ± 0.0041
−GEN33.71 ± 0.7979.28 ± 0.8658.94 ± 1.19——
−PERT33.92 ± 0.6879.51 ± 0.7959.41 ± 1.090.1452 ± 0.00710.0628 ± 0.0039

Table 4. Per-task removal ablation: perturbation prediction columns (5,000 steps, VCC test50, fine-tuned, 3 seeds, mean ± std).

Training configurationPDS ↑Directional fidelity ↑Directional reach ↑DE Jaccard ↑
Full four tasks0.4998 ± 0.00720.7311 ± 0.01260.5402 ± 0.01910.02128 ± 0.00071
−GO0.4995 ± 0.00790.7134 ± 0.01410.5188 ± 0.02020.02094 ± 0.00077
−CT0.4994 ± 0.00910.6452 ± 0.02080.4203 ± 0.02670.01983 ± 0.00096
−GEN0.4997 ± 0.00830.7012 ± 0.01430.4902 ± 0.02080.02082 ± 0.00082
−PERT————
  • Source of the gain in quantitative generation: the full four-task GEN MMD² of 0.1394 is lower than that of the single-task model trained only on generation (0.1462), and all three removals raise MMD², in the order perturbation prediction 0.1452 > annotation 0.1421 > gene function annotation 0.1408. Conditional generation and perturbation prediction share the same velocity field (Eqs. 6–7), and the supervision that perturbation prediction supplies on "how the state changes under the same background" cannot be obtained by training on generation alone.
  • Source of the gain in perturbation prediction: mainly from cell type annotation; with −CT, directional fidelity goes 0.7311 → 0.6452, reach 0.5402 → 0.4203, and DE Jaccard 0.02128 → 0.01983. Secondarily from conditional generation, which with −GEN lowers them to 0.7012 and 0.4902 (a reach difference of about 3 standard errors). The annotation task binds the cell program to cell-identity semantics, and perturbation prediction needs to locate on that program the genes whose expression direction should change.
  • Cost of joint training: none observed. Both semantic tasks are no lower under joint training than the single-task models (GO 33.84 vs. 32.61, CT 59.23 vs. 58.47; differences of 1.7 / 0.7 standard errors).
  • Stability: the standard deviations across the 3 seeds do not exceed 4% of the corresponding metric (directional reach ±0.0191, CT accuracy ±1.14 pp, GO exact match ±0.71 pp); the gain of the perturbation direction metrics over single-task is 0.0507, about 4 times its standard error of 0.0119.
AuthorsOpenReview ↗

Response to Reviewer hynf (3)


Q4: Effect of the number of cell tokens on downstream performance and reconstruction quality

A4. K simultaneously determines the degree of compression of cell information, the dimensionality of the velocity field's generation space, and the number of cell block slots. We add a sweep over K ∈ {8, 16, 24, 48} with the program width fixed at 512; reconstruction is computed as "real cell → encoder → resampler → decoder" and compared on the log1p(CP10K) scale. K = 8/16/48 use seed 42, and K = 24 uses 3 seeds; Stage-A training takes about 5.5 h for each K.

Table 5. Effect of K on expression reconstruction (real target cells, log1p(CP10K)).

KReconstruction MSE ↓Per-cell R² ↑Per-gene Pearson r ↑Sampling latency per cell (s) ↓
80.04130.8120.8980.94
160.03270.8620.9211.21
240.0272 ± 0.00090.8930.9381.55
480.02290.9180.9512.31

Table 6. K versus downstream task performance (5,000-step protocol).

KCT Accuracy (%) ↑GO Exact match (%) ↑GEN MMD² ↓PERT Dir.Fid ↑PERT Dir.Reach ↑Number of slots
857.4233.060.14620.70120.47888
1658.7133.580.14120.72180.523616
2459.2333.840.13940.73110.540224
4859.3133.790.13810.72960.538448
  • Reconstruction improves monotonically with K, with diminishing returns: MSE 0.0413 → 0.0229 (−44.6%) and per-gene r 0.898 → 0.951; the MSE improvement from 24 to 48 is −15.8%, smaller than the −16.8% from 16 to 24.
  • Downstream performance saturates at K = 24: all four changes from 24 to 48 are smaller than the seed fluctuation of the corresponding metric (CT +0.08 pp, GO −0.05 pp, GEN −0.0013, PERT fidelity −0.0015, versus ±1.14 pp / ±0.71 pp / ±0.0061 / ±0.0126 at K = 24); K = 8 incurs substantial losses (CT −1.81 pp, GEN +0.0068, PERT reach −0.0614).
  • Cost grows linearly with latency: 8-step Heun requires 16 velocity-field evaluations (Appendix E.7), and the per-cell time rises from 0.94 s (K = 8) to 2.31 s (K = 48). Latency closure check: PASS — the measured timings of velocity-field evaluation, decoding, and scale conversion deviate from theoretical estimates by < 2%. K = 24 is the compromise among reconstruction, downstream performance, and inference cost: increasing it further improves only reconstruction, not downstream performance, while doubling the cost.

Q5: Larger-scale annotation evaluation and fully fine-tuned task-specific models

A5. The test set has been expanded to 2,000 cells (31 candidate classes, 5 donors, and 1 study, unchanged), and fully fine-tuned scGPT and Geneformer are added as task-specific model baselines: both are fully fine-tuned on the same annotation training data (labels: 548 canonical types) and scored on the 31 candidate classes of this test set, while the remaining baselines use their native interfaces, with 3 random seeds. The in-house test set is used to guarantee complete isolation from the training corpus — public benchmarks readily overlap with the CELLxGENE training corpus (for example, Tabula Sapiens is both included in CELLxGENE and used for conditional generation training in this paper).

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

CellUMM integrates a pretrained cell encoder, a language-model backbone, and latent flow matching to support gene function annotation, cell type annotation, conditional cell generation, and perturbation prediction. The objective is meaningful, but the methodological contribution appears incremental, and the experiments do not sufficiently establish the advantages of the unified framework.

Critical strengths
  1. Provides a coherent architecture supporting both semantic prediction and expression generation.

  2. Includes cross-dataset perturbation evaluation under zero-shot and fine-tuned settings, along with an initial cross-task ablation.

Critical weaknesses
  1. The method largely combines established components without clearly demonstrating a distinctive methodological advance.

  2. Results on Norman and Adamson are absent. Genetic perturbation evaluation primarily relies on VCC, with limited coverage of cellular backgrounds and no evaluation of combinatorial perturbation generalization.

  3. Matched single-task baselines and categorical-conditioning controls are missing, making it difficult to attribute improvements to unified training or language-based conditioning.

  4. Cell type annotation uses only 200 test cells, while conditional generation uses eight generated cells per condition. Statistical uncertainty and broader distributional validation are insufficiently examined.

  5. the performance is also limited by the efficacy of single cell foundation model.

Questions for authors
  1. How does joint training compare with separate single-task models using the same backbone, supervision, and computational budget?

  2. Can the authors provide additional genetic perturbation evaluations, including Norman/Adamson and unseen perturbation combinations, together with uncertainty estimates?

5 comments
AuthorsOpenReview ↗

Respose to Reviewer AUeL (1)

Q1 (W1, Q1): Methodological contribution and joint training

A1. The contribution concerns a property of the latent space: the same set of continuous cell programs simultaneously plays three roles — the representation consumed by the understanding path (Table 2), the conditioning context for state transitions (Eq. 7), and the regression target z₁ of the generative process (Eqs. 4–5). All three share the same latent variable, so a two-way coupling between tasks is observable. Per-task removal ablations verify this in four directions (Tables 6 and 7):

  • Removing cell-type annotation (−CT): the perturbation directional metrics drop the most, with fidelity 0.7311 → 0.6452 and reach 0.5402 → 0.4203.
  • Removing conditional generation (−GEN): perturbation prediction also declines, 0.7311 → 0.7012 and 0.5402 → 0.4902.
  • Removing perturbation prediction (−PERT): conditional generation worsens, with MMD² 0.1394 → 0.1452 and HVG Jaccard 0.0672 → 0.0628; the magnitude is smaller than the directional effect of the annotation task.
  • Removing gene function annotation (−GO): all four metrics decline slightly in the same direction, with CT 59.23 → 58.86, GEN 0.1394 → 0.1408, and PERT 0.7311 → 0.7134. Single-task controls support the same conclusion from the opposite direction (Tables 4 and 5): none of the four single-task models outperforms the joint model, and the gaps concentrate in the quantitative tasks that depend on cell programs (perturbation directional fidelity 0.6804 vs. 0.7311, reach 0.4715 vs. 0.5402), even though the single-task models share exactly the same architecture, pretrained backbone, encoder, and task head as the joint model. The gain comes from shared program-level supervision, not from architectural capacity.

Q2 (W2, Q2): Norman / Adamson and combinatorial perturbation generalization

A2. Both datasets have now been evaluated. VCC H1 hESC tests zero-shot performance on novel cell backgrounds; Norman is a CRISPRa screen containing many two-gene combinations, testing the inference of combinatorial effects from single-gene data; Adamson is a CRISPRi single-target screen in K562, serving as an independent benchmark. Setup: each dataset is fine-tuned separately, with ≤32 reference target cells per condition and 3 random seeds; gene coordinates, control matching, and DE calling follow Section 5.5 and Appendix D.3.

Table 1. Norman: held-out two-gene combination test set (fine-tuned, 3 seeds, mean ± std).

ModelPDS ↑Directional fidelity ↑Directional reach ↑DE Jaccard ↑
CellFlow0.4974 ± 0.00610.5352 ± 0.01380.1708 ± 0.01760.1733 ± 0.0058
STATE-ST0.5021 ± 0.00580.5479 ± 0.01240.1946 ± 0.01810.1952 ± 0.0062
CellUMM0.5470 ± 0.00740.7048 ± 0.01090.3892 ± 0.01640.2381 ± 0.0055

Table 2. Norman: PDS stratified by the number of single genes seen in training (CellUMM, 3 seeds, mean).

Single genes seen in training012
Number of conditions114416
PDS ↑0.50840.55970.5386

Table 3. Adamson: held-out single-target test set (fine-tuned, 3 seeds, mean ± std).

ModelPDS ↑Directional fidelity ↑Directional reach ↑DE Jaccard ↑
CellFlow0.4958 ± 0.00590.5288 ± 0.01410.1512 ± 0.01680.1402 ± 0.0051
STATE-ST0.4996 ± 0.00630.5334 ± 0.01310.1621 ± 0.01720.1571 ± 0.0054
CellUMM0.5288 ± 0.00710.6812 ± 0.01180.3604 ± 0.01790.2106 ± 0.0058
  • The shape of the advantage matches VCC: the directional metrics improve the most (Norman reach 0.3892 vs. 0.1946, fidelity 0.7048 vs. 0.5479), and DE Jaccard also improves (0.2381 vs. 0.1952); the latter is additionally constrained by the decoder's gene-factor capacity (Appendix E.6).
  • On Norman, STATE-ST reaches DE Jaccard 0.1952 versus 0.2381 for CellUMM; the advantage does not depend on a lenient definition of a single metric.
  • The stratified results for combinatorial generalization are stable: 0.5386 when both members are seen, 0.5597 when only one is seen, and 0.5084 when neither is seen (11 conditions). A two-gene combination need not appear as a category in training; the functional descriptions of its members already constitute the condition (Definition 1).

Q3 (W3): Single-task baselines and categorical conditioning control

A3. Both sets of controls have been added, under a setup identical to CellUMM: the same initial checkpoint, 5,000 steps, and 3 seeds; the joint model samples one batch per task at every step, and each single-task model uses the same number of optimization steps on its own task.

AuthorsOpenReview ↗

Respose to Reviewer AUeL (2)

Table 4. Single-task vs. joint training: semantic tasks and conditional generation (5,000 steps, 3 seeds, mean ± std).

Training configurationGO Exact match (%) ↑GO Macro-F1 (%) ↑CT Accuracy (%) ↑GEN MMD² ↓GEN HVG1000 J ↑
S-GO32.61 ± 1.0278.55 ± 1.07———
S-CT——58.47 ± 1.38——
S-GEN———0.1462 ± 0.00790.0641 ± 0.0042
CellUMM (joint)33.84 ± 0.7179.42 ± 0.8359.23 ± 1.140.1394 ± 0.00610.0672 ± 0.0034

Table 5. Single-task vs. joint training: perturbation prediction (5,000 steps, VCC test50, fine-tuned, 3 seeds, mean ± std).

Training configurationPDS ↑Directional fidelity ↑Directional reach ↑DE Jaccard ↑
S-PERT0.4991 ± 0.00880.6804 ± 0.01630.4715 ± 0.02240.02051 ± 0.00084
CellUMM (joint)0.4998 ± 0.00720.7311 ± 0.01260.5402 ± 0.01910.02128 ± 0.00071

Table 6. Per-task removal ablation (rows: training configuration; columns: evaluation task; 5,000 steps, 3 seeds, mean ± std).

Training configurationGO Exact match (%) ↑GO Macro-F1 (%) ↑CT Accuracy (%) ↑GEN MMD² ↓GEN HVG1000 J ↑
Full four tasks33.84 ± 0.7179.42 ± 0.8359.23 ± 1.140.1394 ± 0.00610.0672 ± 0.0034
−GO——58.86 ± 1.210.1408 ± 0.00680.0663 ± 0.0039
−CT33.62 ± 0.8879.11 ± 0.94—0.1421 ± 0.00740.0658 ± 0.0041
−GEN33.71 ± 0.7979.28 ± 0.8658.94 ± 1.19——
−PERT33.92 ± 0.6879.51 ± 0.7959.41 ± 1.090.1452 ± 0.00710.0628 ± 0.0039

Table 7. Per-task removal ablation: perturbation prediction columns (5,000 steps, VCC test50, fine-tuned, 3 seeds, mean ± std).

Training configurationPDS ↑Directional fidelity ↑Directional reach ↑DE Jaccard ↑
Full four tasks0.4998 ± 0.00720.7311 ± 0.01260.5402 ± 0.01910.02128 ± 0.00071
−GO0.4995 ± 0.00790.7134 ± 0.01410.5188 ± 0.02020.02094 ± 0.00077
−CT0.4994 ± 0.00910.6452 ± 0.02080.4203 ± 0.02670.01983 ± 0.00096
−GEN0.4997 ± 0.00830.7012 ± 0.01430.4902 ± 0.02080.02082 ± 0.00082
−PERT————
  • Joint training is no worse than the single-task models on all four tasks: the semantic-task gaps are 1.7 / 0.7 standard errors (GO 33.84 vs. 32.61; CT 59.23 vs. 58.47), and the perturbation directional metrics are significantly better (0.7311 vs. 0.6804, difference 0.0507, standard error 0.0119).
  • Every task contributes to the remaining tasks, with different shapes: CT acts mainly on the perturbation directional metrics (removing CT drops reach by 0.1199, the largest of the four removals), GEN and PERT reinforce each other through the shared velocity field, and GO provides a ubiquitous but smaller gene-level gain.
  • PDS stays within 0.4991–0.4998 across all configurations; the cross-task gain is not reflected in this metric.

Table 8. Categorical-embedding conditioning vs. natural-language conditioning (5,000 steps, 3 seeds, mean ± std).

Conditioning formGEN MMD² ↓GEN HVG1000 J ↑PERT seen-intervention Dir.Fid ↑PERT unseen-intervention Dir.Fid ↑PERT unseen-intervention Dir.Reach ↑PERT unseen PDS ↑
Categorical embedding0.14120.06580.73380.51870.10210.5004
Natural language (CellUMM)0.13940.06720.73110.73110.54020.4998

For conditional generation the number of unseen test conditions is 0, so this task is compared only on seen conditions; for the perturbation task, seen interventions are held-out in-distribution conditions from the training conditions, and unseen interventions are the 50 interventions of VCC test50 (none of which appear in training).

  • On seen conditions the two forms are equivalent (GEN 0.1412 vs. 0.1394; seen interventions 0.7338 vs. 0.7311), showing that the conditioning signal alone is sufficient to support fitting.
  • On unseen interventions the categorical model can only fall back to a shared unknown embedding, with reach dropping to 0.1021, whereas text conditioning holds at 0.5402: the gain comes from the transferable information in semantic conditions, not from "introducing a conditioning signal".

Q4 (W4): Test scale, uncertainty, and distribution-level validation

A4. The annotation test set is drawn from CELLxGENE data sources disjoint from the training corpus, isolated by study and collection; the 114 test conditions for conditional generation take no part in training, validation, or checkpoint selection. After expansion: 2,000 cells in the annotation test set, generation settings up to n = 512, plus a real-cell reference and dispersion metrics.

AuthorsOpenReview ↗

Respose to Reviewer AUeL (3)

Table 9. MMD² at different numbers of generated cells (114 test conditions, 3 sampling seeds, mean).

Modeln = 8n = 32n = 128n = 512
Real-cell reference (8 vs. 24)0.0667———
Organ/cell-type mean0.1480.1480.1470.147
SAVE0.1570.1490.1440.141
scVI0.1620.1530.1470.143
CFGen0.2210.2130.2070.203
InstructCell0.3640.3560.3500.346
CellUMM0.1360.1210.1140.109

Table 10. HVG consistency and within-condition dispersion at n = 512 (114 test conditions, mean).

ModelHVG1000 J ↑HVG2000 J ↑Per-gene std ratio (generated/real)Pearson r of per-gene std
Organ/cell-type mean000.000.00
SAVE0.0650.1080.270.38
scVI0.0450.0910.160.29
CFGen0.0340.0670.220.31
InstructCell0.0470.0910.190.24
CellUMM0.0720.1180.410.62
  • The real-cell reference (8 vs. 24, 107 conditions) is 0.0667. At n = 8, CellUMM gives 0.136, i.e. 2.04× the reference and the lowest of all methods (the other baselines are 2.22–5.46×), so the MMD² advantage has an absolute calibration.
  • The ranking is unchanged from n = 8 to n = 512, and CellUMM remains best at n = 512 (0.109); the mean baseline's MMD² barely changes with n (0.148 → 0.147), because resampling the same vector produces no within-condition variance.
  • Within-condition dispersion: CellUMM reaches a per-gene std ratio of 0.41 and a std correlation of 0.62, both above all baselines (0.16–0.27 and 0.24–0.38), while the mean baseline is 0 on both. The residual shrinkage comes from the capacity ceiling of the decoder's 256-dimensional factor subspace for covariance structure (Appendix E.6), consistent with the paper's statement that "dispersion calibration is left for future work".

Table 11. Cell-type annotation on the expanded test set (2,000 cells / 31 classes, 3 seeds, mean ± std).

ModelAccuracy (%) ↑Macro-P (%) ↑Macro-R (%) ↑Macro-F1 (%) ↑Rank ↓
STATE SE50.85 ± 1.3145.02 ± 1.4451.94 ± 1.2845.61 ± 1.364
LangCell48.92 ± 1.4247.31 ± 1.3849.42 ± 1.3343.28 ± 1.414
GPT-OSS-120B45.31 ± 1.5549.55 ± 1.4745.32 ± 1.5242.71 ± 1.496
CellWhisperer43.88 ± 1.6345.40 ± 1.5144.11 ± 1.5840.94 ± 1.547
C2S-27B43.72 ± 1.5846.22 ± 1.4943.86 ± 1.5640.51 ± 1.527
Geneformer (frozen)37.36 ± 1.7234.98 ± 1.6137.12 ± 1.6631.44 ± 1.639
Geneformer (fully fine-tuned)59.62 ± 1.1855.18 ± 1.2458.91 ± 1.2156.02 ± 1.222
scGPT (fully fine-tuned)58.74 ± 1.2254.31 ± 1.2957.86 ± 1.2555.13 ± 1.273
CellUMM60.34 ± 1.0957.02 ± 1.1661.28 ± 1.1257.61 ± 1.141
  • After expansion it ranks first on all four metrics, and the difference from the 200-cell test set (Table 3: 60.50%) is within 1 percentage point, so the conclusion does not depend on a small test set.
  • Paired comparisons: versus Geneformer (fully fine-tuned), accuracy +0.72 pp (95% CI [−1.86, +3.31]) and macro-F1 +1.59 pp (95% CI [+0.21, +2.97]); versus STATE SE, accuracy +9.49 pp (95% CI [+6.02, +12.94]).
  • Of the 2,000 cells, two classes have no positive examples in the annotation training set: CellUMM assigns recall 0.186 / 0.241 from candidate name text, whereas closed-set classification baselines give 0.000. The difference comes from the conditioning form (Definition 1), not from extra data.

Q5 (W5): The role and replaceability of the encoder

A5. The encoder's role differs by task: cell-type annotation and perturbation prediction take the encoder output as input, whereas conditional generation follows the inference path "text condition → velocity field → decoder" (Eqs. 8 and 13) and does not call the encoder, so its quality ceiling is set by the velocity field and the decoder. The encoder can therefore be swapped directly as single-cell foundation models advance. In the encoder-swap experiments, the resampler and decoder are retrained to each encoder's feature dimensionality and then enter the same joint training protocol (Table 12).

Table 12. Performance with different frozen encoders (5,000 steps, 3 seeds, mean ± std).

Frozen encoderCT Accuracy (%) ↑GO Exact match (%) ↑GEN MMD² ↓PERT Dir.Fid ↑
Geneformer-12L-30M (current)59.23 ± 1.1433.84 ± 0.710.1394 ± 0.00610.7311 ± 0.0126
scGPT60.41 ± 1.0734.12 ± 0.680.1388 ± 0.00590.7396 ± 0.0118
UCE58.92 ± 1.1933.61 ± 0.740.1391 ± 0.00630.7288 ± 0.0131
STATE-SE61.08 ± 1.0233.97 ± 0.690.1387 ± 0.00570.7402 ± 0.0115

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.