acceptodds
Under review as a conference paper at ICLR 2027

Beyond Cell Realism: Recovering Disease Signals with Bulk-to-Single-Cell Generation

Abstract

Bulk RNA-seq provides transcriptomic data for large patient cohorts, but averaging across cell populations can obscure disease-associated expression changes within specific cell types. Single-cell profiling can resolve these changes, yet its cost limits cohort-scale measurements, and many existing bulk cohorts lack matched single-cell data. Generating cell-level expression profiles from bulk could extend such analyses to these cohorts. The central question is whether the generated profiles preserve the differences between patient groups that motivate disease analysis. We evaluate this through the recovery of differentially expressed genes (DEGs): genes showing significant expression differences between patient groups within a specified cell type. For example, our breast cancer evaluation compares cancer epithelial expression between triple-negative and estrogen-receptor-positive tumors. Because each cohort poses a different disease question, the relevant patient groups, cell types, and gene-level evidence must be established for each evaluation. We therefore propose an agentic benchmarking workflow that translates a cohort-specific disease question into a traceable DEG-recovery evaluation. It audits patient pairings and disease contrasts, retrieves supporting evidence from publications and supplements, and assembles a reference of disease-relevant DEGs detected in measured cell-type profiles but missed by matched bulk analysis. This design allows the evaluation to be tailored to different cohorts and biological questions while retaining a common patient-level scoring procedure. Furthermore, to reflect the practical use of existing bulk cohorts, models learn from external reference cells and receive only measured bulk from new patients, without target-cell training or calibration. Under this protocol, the tested generative baselines score zero on important DEG recovery in breast cancer and T-cell leukemia. We address this recovery gap with a patient-conditioned generator that combines bulk-derived context with a source-fitted expression shift. Across three paired cohorts comprising 104 participants, our method recovers 25 of 91 important cohort–gene references, compared with nine for a direct adaptation of scSemiProfiler. It recovers 10 of 26 references in breast cancer and six of 11 in T-cell leukemia, while tying that comparator at nine of 54 in sepsis. Together, the benchmark and generator connect bulk-to-single-cell generation to the disease signals needed for patient-cohort analysis.

Primary area:
generative models
Keywords:
bulk-to-single-cell generation, conditional generative models, patient-level evaluation, disease signal recovery, agentic benchmarking

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

This paper investigates whether single-cell profiles generated from patient bulk RNA-seq can recover cell-type-specific disease-associated differential expression signals detected in measured single-cell data but missed by bulk analysis. The authors propose two main components: (1) an agentic benchmarking workflow for cohort auditing, literature-based disease-gene evidence retrieval, and patient-level DEG recovery evaluation; and (2) a patient-conditioned single-cell generation framework that uses a Patient Bulk Encoder (PBE) to condition a CFGen-based flow-matching generator with a count decoder, followed by a Bulk-Conditioned Shift (BCS) module to adjust patient- and cell-type-specific gene expression.

Experiments on three paired bulk/scRNA-seq disease cohorts show improved recovery of important bulk-missed DEGs compared with the evaluated generative models and mean-prediction baselines in some settings. Ablation studies indicate that BCS contributes substantially to the observed improvements. The enhanced agentic workflow retrieves more disease-associated gene evidence than the original retrieval procedure.

Critical strengths
  1. Patient-conditioned single-cell generation. The paper extends a CFGen-based generator with a bulk-derived Patient Bulk Encoder (PBE) and a Bulk-Conditioned Shift (BCS), enabling single-cell generation for unseen patients using only their bulk RNA-seq data. The design addresses a relevant inference setting.

  2. Biologically relevant evaluation. The paper evaluates whether generated single-cell profiles can recover cell-type-specific disease signals detected in measured single-cell data but missed by paired bulk RNA-seq. The use of paired bulk/scRNA-seq data and literature-supported DEG references provides a biologically meaningful framework for assessing the utility of generated single-cell profiles in disease-related analyses.

Critical weaknesses
  1. Insufficient evaluation of the model design and generated results. The generative framework builds on CFGen with the addition of PBE and BCS, but the effectiveness of these components is not thoroughly evaluated. The contribution of PBE is not adequately isolated through matched ablation studies. Although BCS improves DEG recovery, it remains unclear whether it recovers missing patient-specific signals or potentially amplifies those already captured by the generator. More importantly, the evaluation focuses predominantly on cell-type-level DEG recovery, with limited assessment of generated single-cell expression distributions, cell-to-cell heterogeneity, and cell-state distributions. Furthermore, since DEG significance depends on both between-group expression differences and within-group variability, the accuracy of generated patient-level expression variance also requires further investigation to rule out potentially inflated significance. Finally, comparisons with established cell-type-specific expression deconvolution methods are missing.

  2. Limited reproducibility and experimental transparency. The incomplete documentation of source datasets and participant-level data splits makes it difficult to verify data independence and reproduce the experiments. In addition, the historical agentic benchmarking experiments lack complete records of prompts, model/tool versions, retry budgets, search costs and human interventions, preventing full reproduction of the retrieval workflow. Several implementation details of the generative framework, including the flow-matching architecture and inference settings, are also insufficiently specified, and no source code is provided to clarify these details.

Questions for authors
  1. Since the primary objective of this work is to generate single-cell profiles from bulk RNA-seq, evaluating the quality of the generated cells is essential. However, the current evaluation focuses predominantly on cell-type-level DEG recovery. Can the authors provide quantitative comparisons between generated and measured single-cell data, particularly in terms of gene expression distributions, cell-to-cell heterogeneity, and cell-state distributions?

  2. Can the authors provide matched ablation studies to isolate the contribution of PBE, including the effects of patient-bulk conditioning and the auxiliary bulk reconstruction loss?

  3. Since the conditional generator already incorporates patient-specific information, how do the authors determine whether BCS recovers missing expression signals or amplifies those already captured by the generator? Can the authors provide a more detailed analysis of patient-specific expression estimates before and after BCS correction, particularly at the single-cell level?

  4. Since the primary DEG evaluation is performed at the patient and cell-type levels, could the authors include comparisons with established deconvolution methods that infer sample-specific, cell-type-specific gene expression from bulk RNA-seq, such as bMIND, BayesPrism, and CIBERSORTx, under the same data constraints?

  5. DEG significance depends on both between-group expression differences and within-group variability across patients. How well do the generated cell-type-specific pseudobulk profiles reproduce the patient-level expression variance observed in matched measured single-cell data? Could the authors investigate whether the generated profiles underestimate within-group variance, potentially leading to inflated DEG significance?

  6. Can the authors provide the complete source dataset manifests, clarify donor-level data separation, and release the implementation details or source code? Additionally, can the agentic benchmarking workflow be reproduced with complete execution records, including prompts, model/tool versions, search budgets, and human interventions?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Vaau (Part 1 of 2)

Thank you for such a thoughtful review. We were struck by how precisely your questions identified what our emphasis on disease-related DEGs had left open. They prompted us to compare expression within individual patients as well. Our model detected the most measured cell-type DEGs among generators evaluated on the full cohort in each disease, and ranked first in T-ALL and second in sepsis for patient mean expression error (Table 1).

Weakness 1 / Q1 · How well do cells match measured distributions, heterogeneity and cell states? · Tables 1, 4

Answer. Our model better matched the measured cell distributions than scSemiProfiler in T-ALL and sepsis. The three external baselines were closer in breast.

We compared each gene’s expression distribution between generated and measured cells from the same patient using Wasserstein distance. Across 2,000 genes, our paired distance minus scSemiProfiler’s was +0.051 in breast (14/14 patients), −0.067 in T-ALL (26/37) and −0.028 in sepsis (53/53). Our cell-to-cell variance error was lower than that of all three baselines in sepsis and higher than that of all three in breast. After removing each gene’s mean, Ours also had a smaller distribution distance than scSemiProfiler in T-ALL and sepsis (Table 4). These comparisons measure heterogeneity one gene at a time. A frequency for a state defined by several genes requires their joint expression in the same cells.

Weakness 1 / Q2 · What do bulk conditioning and reconstruction each contribute? · Table 1 and a separate matched retraining experiment

Answer. Giving the generator each patient’s bulk RNA through PBE was enough to detect reference DEGs in all three cohorts, even without the auxiliary reconstruction loss. BCS then adjusted the size of group differences already present in the generated cells, bringing more genes above the detection threshold.

Table 1 shows the two configurations reported in the paper. On the same PBE output, applying BCS raised the count from 3 to 10 of 26 reference genes in breast cancer, from 5 to 6 of 11 in T-ALL, and from 5 to 9 of 54 in sepsis. The new control trained without PBE detected none in any cohort, with or without BCS.

In a separate experiment outside Table 1, we retrained under matched conditions to isolate the auxiliary loss. Before BCS, patient conditioning alone yielded 2, 3 and 4 DEGs (breast, T-ALL and sepsis). With BCS, those counts were 7, 3 and 5. Adding the auxiliary loss during training gave 6, 6 and 9 after the same correction. Thus the loss added detections in T-ALL and sepsis, while breast cancer had one fewer. Patient conditioning already supplied a detectable disease signal before either reconstruction or correction.

Weakness 1 / Q3 · Does BCS restore missing signals or amplify existing ones? · Tables 1, 4, 5, 6

Answer. In the PBE model, BCS amplified existing group differences enough for additional genes to pass the DEG criteria.

We kept the generated cells fixed and used each patient’s bulk measurement to adjust their expression within each cell type. All 12 newly detected cohort–gene pairs already had the measured direction before correction (Table 1), and mean squared error in the disease-group effects fell in every cohort (Table 4). In breast cancer, BCS raised the median absolute-effect ratio from 0.255 to 0.336 and the SE ratio from 0.501 to 0.560, while detections rose from 3 to 10 of 26 (Table 5). For individual cells, the average-centered distribution distance fell after BCS, while the distance on the original expression scale rose (Tables 1 and 4). Corrected profiles also more often matched their own patient than another patient in the same clinical group, although scSemiProfiler led on that measure (Table 6).

AuthorsOpenReview ↗

Response to Reviewer Vaau (Part 2 of 2)

W1/Q4 · How does patient-specific expression compare with deconvolution? · Tables 2–3, 9, 13

Answer. Across all three cohorts, our model recovered more DEGs and better matched expression within each patient. BayesPrism better ranked patients for each gene.

Stage 1 estimates cell-type proportions. Cells sampled separately from the same patient can differ in composition from the bulk specimen. Table 9 uses pseudobulk summed from each patient’s annotated cells, so the exact fractions are known.

On measured bulk from the same patients, our model detected more reference DEGs than BayesPrism in each cohort. In T-ALL, for example, it detected 6 of 11 versus 0 of 11 (Table 13).

In our run of the public CIBERSORTx High-Resolution web tool, it returned no gene expression estimates for breast cancer epithelial cells. In T-ALL, it estimated expression for 326 genes, but each gene had the same predicted value in all 37 patients. Only sepsis supported a comparison of patient differences. On the same 283 genes with varying CIBERSORTx estimates, CIBERSORTx tracked patient rankings less closely than BayesPrism (Spearman ρ of 0.51 vs. 0.60) and also reproduced disease-group differences less accurately (Pearson r of 0.43 vs. 0.62). We report this partial comparison here because Table 13 compares methods across all three cohorts. We did not have time to complete the bMIND comparison during the rebuttal period. For the camera-ready version, we will complete that comparison and rerun CIBERSORTx after checking its reference matrix, batch settings and processing code. We will report both analyses and their limitations in the supplement.

W1/Q5 · Does low within-group variance inflate DEG significance? · Tables 1, 4, 5, 14

Answer. Generators must capture group means and within-group patient variation. Ours had smaller absolute errors in between-patient variance than scSemiProfiler for 19/26 genes in triple-negative and 17/26 in ER-positive tumors. Case/control counts were 36/54 and 38/54 in sepsis, and 4/11 and 7/11 in T-ALL.

Breast comparator (patients)Effect magnitude and Welch SE both closerSigned effect closerDEGs: Ours vs comparator
Official CFGen (14)23/2626/2610 vs 0
scSemiProfiler-direct (14)16/2620/2610 vs 0
Bulk2Space (11)18/2425/267 vs 1

Joint counts use absolute deviations from 1 in predicted/measured ratios (log scale: 15/26 vs scSemiProfiler). Two Bulk2Space genes had undefined tests.

SE was below measured for 24/25 hits. With effects fixed, we reran BH-adjusted Welch tests using our gene-wise group variances. scSemiProfiler went from 0/0/9 to 6/4/9 hits (breast/T-ALL/sepsis). For 4/6 new breast hits its triple-negative variance exceeded 5× measured (TRPS1: 0.450 vs 0.050 measured, Ours 0.025). For RASGRP1, one of 4 new T-ALL hits, its control variance was 0.658 vs 0.282 measured (Ours 0.165). Substitution lowered adjusted p from 0.557 to 0.0097. CFGen stayed at 0/0/0. Bulk2Space stayed at zero in T-ALL and sepsis (1 to 0 in 11 matched breast patients). At equal variances Ours led in breast and T-ALL and tied in sepsis (10/6/9). Signed effects favored Ours over CFGen for all 26 breast genes.

Table 14. Between-patient variance within each clinical group

Median gene-wise ratio of generated to measured between-patient variance within cases/controls (1 = measured).

MethodBreastT-ALLSepsis
Ours (PBE + BCS)0.511/0.2350.317/0.5310.364/0.213
Official CFGen0.196/0.0690.023/0.0300.052/0.316
Bulk2SpaceNA (11/14 profiles)0.007/0.0290.002/0.006
scSemiProfiler-direct0.719/0.0330.635/1.3990.835/1.572

Weakness 2 / Q6 · Can others reproduce the splits, model and agent workflow? · Tables 10–12 · Supplement Tables 4–5

Answer. Yes, the literature search can be rerun under the documented settings. We ran each strategy three times per cohort with Sonnet 5.5 high. Under the same evidence criteria, our search found more measured DEGs than the adapted SCTA search in eight of nine comparisons (Tables 11–12). The reproducibility package includes the search instructions, execution records and scoring code.

Supplement Tables 5 and 4 list focus-type reference studies and BCS settings. Training/validation used 307/76, 28/9 and 91/25 specimens, and BCS refits used 307/27/89 source IDs (breast IDs are specimen-level). The recorded source and evaluation patient IDs were disjoint in all three cohorts. Appendix A specifies PBE widths and decoder Eq. 5. Inference used adaptive Dopri5 with 20 trajectory points. Both original literature workflows used Sonnet 5 high. Table 10 compares model versions and effort, Table 11 gives three runs per method and Table 12 gives unique hits and list overlap. Separate Sonnet 5.5 runs allowed 24 searches, 60 documents and 45 minutes without human intervention during scored runs. We will release code upon acceptance.

Reviewer_T5pjOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper aims to recover patient-level disease signals from bulk RNA-seq transcriptomic data. First, it contributes an agentic workflow-based evaluation protocol for three use cases (breast cancer, T-cell leukaemia and sepsis). Second, it proposes a generation framework that uses external reference cells during training and individual measured bulk data during inference for patient-conditioning. For validation, paired single-cell and bulk measurements identify differentially expressed genes missed by bulk analysis under fixed thresholds. The agentic workflow retrieves literature evidence and recorded priors to identify the disease-relevant subset of those genes. The paper then compares the proposed generative framework with three generative baselines under the specified protocol and find relatively higher recovery across the full cohort in two of the three studied use cases.

Critical strengths

The paper proposes a participant-level evaluation protocol that assess whether generated cell-type profiles recover disease relevant signals between patient groups. This goes beyond measuring correlation with observed single-cell expression and provides a more direct test of their utility for hypothesis generation in disease analysis. However, a short note on the importance of recovering disease relevant signals is needed for the readers not familiar with the domain.

Critical weaknesses

As the following weaknesses concern the core contributions of the paper, I recommend rejection in its current form:

  1. The paper does not position its contribution of agentic workflow against existing work in the Related Work section. It would be beneficial to distinguish the framework from existing single-cell agentic frameworks such as SCTA [1], for methodological similarities although they address different tasks. The paper also does not mention what AI model (and version) was used for the agentic benchmark construction. There are also some concerns regarding the evaluation of the agentic workflow: (1) although W0 and W1 are described as legacy and enhanced retrieval workflows, which of these are agentic AI-based and which retrieved evidence informs the results in Section 4.2, are not fully clear; and (2) the accuracy of the cohort-audit, retrieval, and evidence-audit skills is not independently evaluated. These concerns make the assessment of the agentic benchmark contribution difficult.
  2. Before assessing whether the generated profiles recover disease-related differences, the paper should show how closely they match measured expression for those genes at the patient and cell-type level. Expression error for matched patient–cell-type profiles and per-gene correlation across patients, reported by disease group and alongside the other generators, would make the comparison more convincing. This would show whether the recovered group differences are backed by accurate patient-level predictions rather than appearing despite large errors. Missing generated profiles should be reported separately.
  3. The paper focuses on quantitative recovery metrics, but would benefit from a biological interpretation of the results. Which disease-relevant genes were selected, and why is this subset biologically meaningful? Which genes did the generators recover, and what is the biological significance of those they failed to recover? Addressing these questions would help contextualise the numerical results.
  4. The methodology section is difficult to follow because some equation symbols are not explicitly defined (e.g. in Eq 2 and 3). In addition, important details, such as PBE training and BCS fitting, are deferred to the appendix. Please define the symbols when first introduced and bring the essential methodological details into the main text.

[1] https://arxiv.org/abs/2607.23821v1

Questions for authors
  1. How many patients were used to train each component of the proposed generator? The BCS counts reported in the appendix do not answer this for the whole generator. Because the external references are (rarely available) patient-level data, it would be beneficial to show how sensitive the results are to source-cohort size.
  2. The ablation study highlights the contribution of the BCS, but an ablation for PBE is missing.
  3. For the methods with assigned zero, can the paper report output coverage separately, and show NA when the analysis was not completed? This helps distinguish output coverage and analysis failures from failure to recover the disease signal.
  4. The authors should consider explaining the terms like measured bulk, source-derived pseudo bulk, donor identities before introducing them for a general audience not specialised in this domain.
  5. Why is an absolute mean difference of 0.5 in natural-log units selected? Is this fixed before assessing the generator results? Are the recovery comparisons in Table 2 robust to different cut-offs?
  6. How is the value of library size L selected at inference?
3 comments
AuthorsOpenReview ↗

Response to Reviewer T5pj (Part 1 of 3)

Thank you for reading the paper from the perspective of an AI researcher new to single-cell biology. We agree with your concern. In fitting the experiments into the page limit, we gave the results more space than the explanations needed to interpret them. We are sorry that readers had to supply that background themselves. In this response, we explain what each measurement captures and why it matters for the biological question. Before the camera-ready deadline, we will revise the main text and supplement with the same reader in mind.

Your question about individual patients illustrates the point. In T-ALL, our model estimated each patient’s average gene expression more accurately than scSemiProfiler. When we then compared the two patient groups, our model detected six of the eleven disease-related genes whose expression differences were visible in measured cells but missed in bulk RNA. scSemiProfiler detected none (Tables 1–2).

Weakness 2 / Q3 · How accurate are patient profiles, and which zeros indicate missing output? · Tables 1–3, 5, 7, 13

Answer. Our model estimated patient-level expression more accurately than scSemiProfiler in T-ALL and sepsis by mean absolute error. For individual T-ALL genes, scSemiProfiler more accurately ranked patients within each clinical group (Tables 1–3). We also distinguish an incomplete evaluation (NA) from a completed test with zero DEGs. Bulk2Space lacked breast cancer epithelial profiles for 3 of 14 patients, whereas CFGen produced all required profiles and detected zero prioritized DEGs in all three cohorts (Table 1).

Table 5 shows why good patient ordering can coexist with zero detected DEGs. In T-ALL, scSemiProfiler missed the direction for one reference gene, fell below the effect cutoff for five and missed the adjusted significance cutoff for the other five. On T-ALL reference genes, mean absolute error in cases and controls was 0.543 and 0.548 for our model, compared with 0.735 and 0.923 for scSemiProfiler. Median per-gene correlations across patients in those same groups were 0.362 and 0.462 for our model and 0.744 and 0.632 for scSemiProfiler. Across all patients in each cohort, BayesPrism likewise ranked patients more accurately per gene, while our model better matched expression across genes within each patient (Table 13).

For Bulk2Space, the paper’s end-to-end rule assigned zero for incomplete coverage. Only 11/14 breast patients had profiles, so the full-cohort DEG count is unavailable. The matched 11/14 rows of Table 5 show 7/26 original reference genes detected by Ours and 1/26 by Bulk2Space. Requiring each gene to remain a measured cell-type DEG missed by bulk in those 11 patients gave 6 and 0, respectively (Table 7).

Weakness 3 · Why these genes, and what do hits and misses mean biologically? · Table 1

Answer. We wanted to test what generating single cells adds to a bulk RNA analysis. When RNA from several cell types is measured together, a disease-related change in one cell type can fall below the DEG threshold. We therefore asked whether generated cells could reveal biologically important differences that, in these same patients, were detectable only when measured cells were analyzed by type.

Our candidates combined genes linked to disease in prior studies identified by our literature search with genes nominated by our bioinformatician. We selected them before inspecting model outputs and retained only those that met the DEG criteria in the measured target cell type but fell below them in matched bulk RNA. In breast tumor epithelial cells, our model recovered the lower expression of XBP1 in triple-negative versus ER-positive tumors. One biologically informative miss was FBP1, whose loss has been linked to increased glycolysis in basal-like breast cancer (Dong et al., 2013). Our model predicted the observed decrease but underestimated its magnitude (−0.47 versus −1.46 measured), and FBP1 did not meet our DEG criteria. Outside the priority lists, our model also recovered 3 T-ALL and 7 sepsis bulk-missed genes (Table 1).

AuthorsOpenReview ↗

Response to Reviewer T5pj (Part 2 of 3)

Weakness 1 · How do W0, W1 and SCTA differ, and what independent checks support the workflow? · Tables 10–12

Answer. Thank you for pointing us to SCTA, a relevant comparison we had overlooked. Your suggestion helped us test what our literature workflow contributes beyond the original search (W0). When we gave our enhanced search (W1) and the adapted SCTA workflow the same cohort question and applied the same criteria for literature support, W1 identified more genes that were differentially expressed in measured cells in eight of nine paired runs (Tables 11–12). W1 also found more of these genes than W0 in all six Sonnet 5.5 settings (Table 10).

Both literature searches are agent-driven. The original runs used Claude Sonnet 5 high. W0 searched canonical studies and signatures. W1 also checked supplements, gene identity, direction and duplicate sources. Section 4.4 scores these search lists against all measured DEGs in the tested gene set; §4.2 scores generators on the separately curated bulk-missed reference. Native SCTA starts with single-cell measurements to nominate therapeutic targets. We adapted its candidate and filtering stages to the same question-only task. With the same published-evidence filter, SCTA sometimes had higher precision (breast run 1: 75.3% vs 61.1%). The search advantage also varied by model: Sonnet 5 W0 led in breast at both efforts. In a blinded automated source check, 39/45 sampled enhanced-search entries had a verifiable direction (Table 11). This is a check of individual evidence entries. The repeated runs test the complete workflow. We will clarify the task difference in Related Work.

Q1/Q2 · How many training donors, how sensitive to sample size, and what does PBE add? · Table 1 · Supplement Tables 4–5

Answer. In the submitted supplement, Table 4 reports the BCS fitting and validation splits, Table 5 lists the external reference studies for each cohort’s focal cell type, and Appendix A describes the Stage 1 training mixtures. For the generator itself, we report the training cells and specimens below. At these training sizes, removing PBE reduced reference DEG detection to zero in every cohort, with or without BCS (rebuttal Table 1).

The generator used 153,548 external cells from 307 training specimens in breast cancer, 11,731 from 28 in T-ALL, and 48,582 from 91 in sepsis. The corresponding validation sets contained 76, 9 and 25 specimens. For BCS, combining the “Inner fit” and “Inner select” columns in Supplement Table 4 gives the final fitting pools of 307, 27 and 89 source IDs. For Stage 1, Appendix A specifies 800 training and 200 validation mixtures, each made by pooling 1,000 cells. The ablation in rebuttal Table 1 holds the training data fixed and evaluates PBE trained with its reconstruction objective. It does not vary the number of training donors independently.

Q5 · Why 0.5, when was it chosen, and do rankings change? · Tables 4, 8

Answer. Our bioinformatician chose the 0.5 cutoff before seeing generator outputs. scSemiProfiler led on the fixed sepsis list only at the two stricter cutoffs.

To choose the cutoff, our bioinformatician used agent-retrieved literature and measured single-cell DEGs to retain biologically expected genes. With adjusted p ≤ 0.05 and the 26/11/54 references fixed, lowering the absolute natural-log cutoff from 0.5 to 0.25 increased our recoveries from 10/6/9 to 12/10/28. In sepsis, raising it to 0.75 changed a 9/54 tie to 3/54 for ours and 7/54 for scSemiProfiler. At the strictest cutoff (1.0), even the measured cells retained only 20 of 26, 2 of 11 and 3 of 54 original genes. scSemiProfiler detected none of the original breast or T-ALL genes at any cutoff. CFGen detected none in any cohort, and Bulk2Space detected none in its two complete cohorts.

AuthorsOpenReview ↗

Response to Reviewer T5pj (Part 3 of 3)

Weakness 4 / Q4/Q6 · How are the terms, Eqs. 2–3, PBE/BCS training and library size L defined? · Eqs. 2–3 · Appendix A

Answer. Figure 1 fixes the number of cells generated per bulk specimen at 1,000. The library size L in Eq. 3 has a different unit: it is the expected total count across the 2,000 modeled genes in one cell. We will distinguish these quantities when we introduce the equation.

Measured bulk is RNA from mixed tissue; source pseudobulk sums measured reference cells. Donor IDs link repeated specimens from one person. Eq. 2 counts recovery of gene effects in the measured direction. Eq. 3 learns a flow conditional on cell type, log library size and encoded bulk. PBE learns its context jointly with bulk reconstruction; BCS uses external reference data to learn how much each cell type’s expression departs from its average for a given bulk measurement.

Our follow-up analysis illustrates how to estimate L from external training cells. Within each cell type, we sum the counts across the modeled genes for each reference cell and estimate the mean and standard deviation of the log totals. For a generated cell of that type, we draw L from the corresponding log-normal distribution. We will add a worked example to Appendix A showing how a patient’s bulk RNA determines the allocation of the 1,000 cells and how L sets the expected counts for one of those cells. In the Introduction, we will use the breast cancer example to explain why preserving a difference between patient groups matters for biomarker discovery. In Methods, we will define each biological term and symbol at first use and show what the bulk encoder and the correction step each learn.

Reviewer_NpgoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a patient-conditioned method for generating single-cell gene expression profiles from bulk RNA-seq data, with the goal of recovering disease-associated expression signals that are missed by bulk analysis. It tests whether generating single-cell RNA-seq profiles from bulk data can actually recover cell-type-specific differentially expressed genes (DEGs) between patient groups that were missed in the bulk data. The authors introduce an agentic workflow that pulls literature evidence and pairs cohort data to build a grounded list of disease-relevant reference genes, and a patient-conditioned generator. The generator adds a Patient Bulk Encoder (PBE) and a linear post-generation adjustment called Bulk-Conditioned Shift (BCS) on top of a CFGen flow-matching backbone, trained strictly on external single-cell references and their pseudobulks. The paper provides experiments across three paired cohorts (Breast cancer, T-ALL, and Sepsis; 104 patients total), and the proposed model recovers 25 of 91 reference DEGs, beating adapted generative baselines like scSemiProfiler-direct (9/91) and CFGen (0/91).

Critical strengths
  • The authors include simple non-generative mean-prediction controls and openly show that their post-hoc shift (BCS) improves thresholded DEG recall without uniformly lowering overall MSE.
  • The proposed method recovers disease-associated signals from bulk RNA-seq data when single-cell measurements are unavailable for new patients.
  • The authors evaluate the method across three distinct disease cohorts and includes a matched ablation study showing that the BCS module improves the reported gene-recovery counts.
  • In general, the evaluation setup is realistic and addresses a real blind spot in single-cell generation. Instead of just measuring cell realism or UMAP overlap, the authors evaluate whether generated cells recover participant-level disease DEGs missed by bulk, under a strict zero-shot target regime.
Critical weaknesses
  • The evaluation protocol makes the baseline comparison difficult to interpret because a method receives a zero score when any required participant–cell-type profile is missing. This combines output-generation failures with disease-gene recovery failures.

  • CFGen obtains zero recovery across all three cohorts, but the paper does not adequately establish why it performs poorly under the evaluation setup.

  • The reproducibility of the proposed method is difficult to assess because I could not identify a publicly available implementation. Releasing the code for the model, evaluation pipeline, and agentic benchmarking workflow would help readers reproduce the reported results and verify the comparisons with baselines.

Questions for authors
  • How do the baseline comparisons change when gene recovery is reported both under the current end-to-end scoring rule and among participants for whom all required profiles are available?

  • In Section 4.3, you ablate BCS while holding the parent generator fixed, which shows PBE + CFGen alone gets 3/26, 5/11, and 5/54. What happens if you apply BCS directly to the unconditioned CFGen baseline?

  • Bulk2Space gets an automatic 0/26 in Breast because it generated zero cancer epithelial cells for 3 of 14 patients. What is its recall on the remaining 11 patients if you evaluate on the non-missing subset?

  • Could the method be evaluated on independent cohorts to better assess how well it generalizes beyond the three datasets?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Npgo

Thank you for asking us to compare both methods on the same patients. Among the 11 breast patients both covered, Bulk2Space estimated average expression more accurately, while our model recovered more of the measured differences between tumor subtypes.

Weakness 1 / Q1/Q3 · What does the original zero mean, and how many genes does Bulk2Space recover on the 11 patients it covers? · Tables 1, 5, 7

Answer. Our DEG advantage persisted on the 11 patients for whom both methods generated tumor epithelial cells. Bulk2Space lacked the required cells for three patients, triggering the original 0/26 score. We report coverage (11/14) and mark full-cohort recovery as unavailable (NA).

We compared both methods with measured bulk and single cells from the same 11 patients (2 cases, 9 controls), using 1,785 genes and the original effect cutoff of 0.50.

  1. Which measured differences did each model reproduce? In these 11 patients, 88 genes differed between the two groups in measured cancer epithelial cells. Our model recovered 22/88 in the measured direction and Bulk2Space 2/88.
  2. How many of the paper’s 26 priority genes did each model recover? On the original list, Ours detected 7/26 and Bulk2Space 1/26 in the direction measured in these 11 patients (Table 5, rows marked 11/14).
  3. Which of those hits still met the original selection criteria in the smaller group? Of the 26 genes, 20 still differed between groups in measured epithelial cells, and 17 of those were still missed by bulk. Counting hits only among those 17, while keeping the original denominator of 26, gave 6/26 for our model and 0/26 for Bulk2Space.

On the same 11 patients, Bulk2Space had lower mean expression error across 2,000 genes (0.285 vs 0.528 for ours). The disease-gene comparison includes only two cases, so its precision is limited.

Weakness 2 / Q2 · Why did official CFGen recover zero reference genes? Does adding BCS help? · Tables 1, 4, 5

Answer. CFGen generated all required profiles for the unseen patients, but the disease-group differences it predicted were too small to recover any reference DEGs. BCS raised the median effect ratios from 0.048 to 0.115 in breast, 0.033 to 0.059 in T-ALL and 0.049 to 0.107 in sepsis. The effect correlations improved (Table 4), but no reference DEG passed in any cohort (Table 5).

CFGen can condition on a known donor’s ID, but an unseen patient has no learned ID representation (submitted supplement, Appendix C). Both generators received the same cell counts per patient and cell type. CFGen has no input for the new patient’s bulk RNA to guide expression within those cells. Our patient bulk encoder (PBE) derives that information directly from the new patient’s measured bulk RNA and supplies it during generation, without requiring any single-cell measurements from that patient. With PBE, applying the same BCS correction increased the reference DEG counts from 3 to 10 in breast cancer, 5 to 6 in T-ALL and 5 to 9 in sepsis (Table 1).

Weakness 3 · Will the code be available?

Answer. Upon acceptance, we will release the generation, evaluation and agent-workflow code, model settings and scoring criteria.

Q4 · Has the model been validated beyond the three cohorts? · Table 1 and external validation plan

Answer. We evaluated 104 patients whose recorded IDs did not overlap with training. The three cohorts pose different biological tests. Our model recovered 10/26 prioritized DEGs in breast tumor epithelial cells (14 patients), 6/11 in T-ALL blasts (37), and 9/54 in sepsis monocytes (53). Finding disease signals in all three provides initial evidence that the approach works across distinct diseases and cell types (Table 1). Validation beyond these cohorts is still pending.

Many external single-cell cohorts could support training. Evaluation requires two measurements from the same patient specimen: bulk RNA as input and measured single cells to check the predicted disease signal. Few cohorts provide such pairs, and verifying their provenance takes time. We therefore could not include many candidate cohorts in the paper’s evaluation. We continue to seek pairs in other diseases and will evaluate any cohorts verified before camera-ready and report the results in the paper or supplement.

Your question led to a fairer patient comparison. Across all three cohorts, our generator had the lowest disease-effect error (Table 4) and recovered more measured cell-type DEGs than the original generators (Table 1) and the newly added BayesPrism baseline (Table 13). Another reviewer suggested SCTA for evaluating our literature workflow. Given the same cohort question and evidence criteria, our search identified more measured DEGs than adapted SCTA in eight of nine paired runs (Tables 11–12). We would welcome your feedback on both contributions and would be grateful if you would reconsider the paper in light of these results.

Comments on the paper

14 comments
AuthorsOpenReview ↗

Author response to all reviewers

All reviews helped us see where our paper needed a clearer explanation. We thank the reviewers for asking what the generated cells retain about individual patients. To illustrate why we made this question central to our paper, consider a researcher comparing tumors from two groups of patients. That comparison could reveal a biomarker that distinguishes the groups or a pathway worth testing as a treatment target in one of them. A pathway goes on that shortlist because it behaves differently in the two groups. If a generator gives every patient the same plausible cancer cells, even a million more generated cells will not bring back the missing contrast. A promising lead for one group could disappear from a researcher's shortlist. Our paper contributes a test for whether this contrast survives generation, and a generator designed to preserve it.

Our breast cancer cohort illustrates the problem. A gene can be expressed more strongly in tumor epithelial cells in one subtype than in another. Bulk RNA measures all cell types in a tissue sample together. Other cells contribute expression of the same gene, and their proportions vary between patients. In that mixture an epithelial difference can fall below the detection threshold. For a new patient, the Patient Bulk Encoder (PBE) turns the bulk measurements into a vector that guides cell generation. The Bulk-Conditioned Shift (BCS) then adjusts expression for each cell type using relationships learned from external reference cells. We average the generated epithelial cells for each patient and compare the two subtypes. We then check which genes differ in the same direction in epithelial cells measured from those very patients. We call genes with a statistically supported difference in expression between groups differentially expressed genes (DEGs). Our literature-search agents and team bioinformatician helped identify which genes to prioritize for this test.

In the same 14 breast cancer patients, scSemiProfiler estimated average epithelial expression across 2,000 genes more accurately than our model. Yet it detected none of the 26 prioritized genes that differed between subtypes in the measured epithelial cells and fell below the detection criteria in bulk. Our model detected 10. It also led in T-cell acute lymphoblastic leukemia (T-ALL; 6/11 versus zero for scSemiProfiler) and tied scSemiProfiler in sepsis (9/54 each). A low average error can therefore coexist with the loss of the very difference that motivates comparing patients. When we expanded the test to all genes with a measured cell-type difference, our model detected the most among full-cohort generators in all three diseases (Table 1).

The reviews prompted us to test how well the generated cells match patients at other levels. In T-ALL and sepsis, our model more closely matched both each patient’s mean expression and the distribution across that patient’s cells than scSemiProfiler did (Table 1). On mixtures of known composition, Ours Stage 1 estimated cell-type proportions more accurately on average than CIBERSORTx and BayesPrism in all three cohorts (Table 9). We also gave our model and BayesPrism measured bulk RNA from the same 104 patients and compared the expression they predicted within the target cell type. In every cohort, our model detected a higher fraction of both the prioritized DEGs missed by bulk and all measured cell-type DEGs (Table 13). In the literature search, our workflow found more measured DEGs than the original search in all six Sonnet 5.5 settings and more than adapted SCTA in eight of nine comparisons under the same evidence criteria (Tables 10–12).

We will include these comparisons in the revised paper and supplement. Each shared table identifies the reviewer questions it addresses. Taken together, the results show that cells generated from a patient’s bulk RNA can retain that patient’s expression patterns and reveal disease-group differences observed in the same patients’ measured cells. We believe the generation method and the paired evaluation can help other researchers build and test models for the biological comparisons they want to make. We respectfully ask the reviewers to reconsider the paper’s contribution in light of this evidence.

AuthorsOpenReview ↗

Table 1. Cell-type-specific disease signals and expression fidelity

Cohort order and target cells: breast tumor epithelial cells (5 triple-negative, 9 estrogen receptor–positive), T-ALL blasts (28 cases, 9 controls), and sepsis monocytes (46 cases, 7 controls). For patient-level tests we average each gene’s expression across that patient’s target cells. For cell-distribution tests we keep the individual cell values.

Table 1. Cell-type-specific disease signals and expression fidelity [Vaau Weakness 1/Q1–Q3, Q5 · T5pj Weakness 3/Q2–Q3 · Npgo Weakness 1–Weakness 2/Q1–Q3]

Can a model reproduce differences between clinical groups that are visible in one cell type but fall below the detection threshold when RNA from all cell types is measured together? We scored two sets of differentially expressed genes (DEGs). The primary set comprises literature- or expert-nominated genes that differed between groups in measured cells of the focal type and were missed by matched bulk RNA. The second set contains all measured DEGs in that cell type. A model recovers a gene by reproducing the measured direction with an absolute natural-log expression difference ≥ 0.5 and adjusted p ≤ 0.05.

To compare expression itself, we used all 2,000 genes. Mean absolute error (MAE) compares mean expression across the focal cells of each patient. Wasserstein-1 distance measures how far each gene’s distribution over generated cells lies from its distribution over measured cells. Smaller errors and distances are better. Throughout Tables 1–8, triples give breast cancer, T-ALL and sepsis in that order. ① and ② mark the best and next distinct displayed scores per cohort and metric among full-cohort generators. Ties share ranks. Subset rows and measured bulk are unranked.

MethodDisease-relevant single-cell DEGs missed by bulk ↑ /26 · /11 · /54All measured single-cell DEGs ↑ /44 · /60 · /177Patient mean expression (MAE ↓)Cell expression distributions (Wasserstein-1 ↓)
Ours (PBE + BCS)10① · 6① · 9①19① · 23① · 62①0.541 · 0.469① · 0.260②0.2965 · 0.3582② · 0.1630
Ours (PBE only)3② · 5② · 5②9② · 22② · 49②0.561 · 0.480② · 0.257①0.2785 · 0.3474① · 0.1435②
Ours (BCS only)0 · 0 · 00 · 0 · 00.404② · 0.509 · 0.2740.2563 · 0.3912 · 0.1599
Ours (no PBE or BCS)0 · 0 · 00 · 0 · 00.417 · 0.515 · 0.2670.2381① · 0.3800 · 0.1387①
scSemiProfiler-direct0 · 0 · 9①0 · 5 · 480.349① · 0.574 · 0.3620.2456② · 0.4250ᵃ · 0.1914
Official CFGen0 · 0 · 00 · 0 · 00.482 · 0.554 · 0.3680.2482 · 0.3829 · 0.1591
Official CFGen + BCS0 · 0 · 00 · 0 · 00.468 · 0.546 · 0.3790.2664 · 0.3936 · 0.1799
Bulk2Space0/26ᵇ · 0 · 02/88ᵇ · 0 · 00.285ᵇ · 0.658 · 0.2930.2182ᵇ · 0.4715 · 0.2067
Ours (PBE + BCS), matched to Bulk2Spaceᵇ6/26ᵇ · — · —22/88ᵇ · — · —0.528ᵇ · — · —0.2826ᵇ · — · —
Measured bulk (control)0 · 0 · 0ᶜ16 · 21 · 710.563 · 0.449 · 0.433NA (no cells)

Our model recovered 10/26 and 6/11 prioritized genes in breast and T-ALL, where scSemiProfiler-direct recovered none. Both detected 9/54 in sepsis. Removing the literature/expert restriction gives a broader set of cell-type DEGs missed by bulk. Our model recovered 10/28 · 9/39 · 16/106, including 0 · 3 · 7 genes outside the prioritized lists. scSemiProfiler recovered 17/106 in sepsis. Across all measured cell-type DEGs, our model detected 19/44 · 23/60 · 62/177. One breast example is XBP1: our model reproduced its lower expression in triple-negative tumors (estimated effect −0.863, adjusted p=0.022). DEGS2 changed in the measured direction but fell short of the significance cutoff (−0.524, adjusted p=0.065). The measured bulk control itself detected 71/177 in sepsis. The all-DEG ranking above compares generators; the prioritized endpoint asks which cell-type differences bulk missed.

The two PBE rows are the configurations reported in the paper. Both use the same trained generator. Applying the bulk-conditioned shift (BCS) to its output raised prioritized recovery from 3 to 10, 5 to 6 and 5 to 9. Applying BCS to official CFGen yielded zero prioritized hits in all three cohorts. For the two new ablations we trained a generator without PBE and scored it before and after BCS. All 12 newly detected cohort–gene pairs had the measured direction before BCS.

ᵃ T-ALL scSemiProfiler cell distances cover 26/37 patients. ᵇ Both matched breast rows use 11/14 patients for profiles (2 cases, 9 controls, 88 measured DEGs) and 10/14 for cell distances. Bulk2Space full-cohort recovery is unavailable. The marked primary counts apply both reference criteria to the same 11 patients (Table 7). ᶜ Measured bulk misses every primary reference gene by definition.

AuthorsOpenReview ↗

Table 2. Patient-specific accuracy on the 2,000-gene panel

Cohort order throughout: breast cancer, T-ALL and sepsis.

Table 2. Patient-specific accuracy on the 2,000-gene panel [Vaau Weakness 1/Q4 · T5pj Weakness 2]

Do two patients in the same clinical group still differ in the generated data? Across 2,000 genes, patient-wise correlation (r) compares each patient’s generated and measured profiles, then takes the median across patients. Centered MAE measures error after removing each gene’s group mean separately in measured and generated profiles. For each patient pair, difference MAE measures the error in their expression difference. Concordance scores whether the model ranked the pair in the measured order (half credit for a predicted tie).

MethodPatient-wise r ↑Centered MAE ↓Patient-pair difference MAE ↓Patient-pair concordance ↑
Ours (PBE + BCS)0.720 · 0.781① · 0.8850.249 · 0.287 · 0.1820.377 · 0.414 · 0.2590.564① · 0.561② · 0.535①
Ours (PBE only)0.724② · 0.779② · 0.893①0.254 · 0.287 · 0.1810.383 · 0.414 · 0.2580.554② · 0.556 · 0.528②
Ours (BCS only)0.759① · 0.781① · 0.891②0.243② · 0.266 · 0.176②0.365② · 0.376 · 0.250②0.529 · 0.533 · 0.519
Ours (no PBE or BCS)0.759① · 0.772 · 0.8860.250 · 0.271 · 0.1800.375 · 0.383 · 0.2550.508 · 0.503 · 0.495
scSemiProfiler-direct0.634 · 0.626 · 0.7510.221① · 0.257① · 0.1950.325① · 0.364① · 0.2740.548 · 0.586① · 0.525
Official CFGen0.701 · 0.688 · 0.8640.260 · 0.276 · 0.1890.389 · 0.391 · 0.2730.493 · 0.504 · 0.499
Official CFGen + BCS0.714 · 0.698 · 0.8680.254 · 0.270 · 0.1860.380 · 0.383 · 0.2690.516 · 0.532 · 0.519
Bulk2Space0.756ᵇ · 0.442 · 0.8490.197ᵇ · 0.265② · 0.165①0.294ᵇ · 0.371② · 0.228①0.489ᵇ · 0.500 · 0.498
Ours (PBE + BCS), same breast patientsᵇ0.739ᵇ · — · —0.211ᵇ · — · —0.329ᵇ · — · —0.554ᵇ · — · —
Measured bulk (control)0.554 · 0.710 · 0.7590.256 · 0.205 · 0.2170.387 · 0.292 · 0.3120.667 · 0.728 · 0.616

Our model had higher patient-wise correlation than scSemiProfiler in all three cohorts. For differences between patients, scSemiProfiler had lower pairwise error in breast and T-ALL, and our model in sepsis. A close match across genes for one patient does not guarantee that the model ranks patients correctly for any given gene.

ᵇ Both matched breast rows use the same 11 patients (2 cases, 9 controls), with group means and within-group pairs recomputed on those patients.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.