acceptodds
Under review as a conference paper at ICLR 2027

Learning Multiome Representations Across Individuals

Abstract

Paired single-cell multiome assays measure gene expression and chromatin accessibility in the same cell. Contrastive learning on the two measurements has therefore become a standard way to learn cell representations. In population studies, however, such representations are useful only if they transfer across donors. We show that the contrastive loss alone does not ensure this. When negatives are sampled from all donors, the contrastive loss rewards features that identify the donor. When negatives are drawn within donors, each donor's representation can be rotated separately without changing the loss. Removing donor information altogether resolves neither issue without also discarding phenotype. We propose Minuet, which combines within-donor contrastive learning with a prediction task shared across donors and a donor penalty applied within groups of donors that share biological covariates. We show that these objectives determine the representation up to one shift for each group of donors connected by shared covariates. Applied to six paired cohorts, Minuet achieves the best average rank among nine methods in integration, cross-donor and cross-dataset transfer, and phenotype prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell multiomics, multimodal representation learning, contrastive learning, conditional regularization, RNA-ATAC integration

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_TEYJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents Minuet, a donor-aware integration method for single-cell multiomics data (scRNA-Seq and scATAC-Seq). The authors provide the theoretical foundations of the limitations of simple contrastive loss for the task, motivating the design of Minuet. Experiments are provided over a synthetic setting designed to challenge all identified weak spots as well as six different published datasets. The paper also provides thorough ablations of the method.

Critical strengths
  1. The paper targets an important task — obtaining a shared donor multiome representation that is designed to transfer across donors

  2. The overall framing is appealing — stemming from three concrete identified limitations of a standard contrastive approach, as motivation for concrete building blocks that are elegant and appealing.

Critical weaknesses
  1. Clarity and writing style.

The text is very dense in prose and leading claims rather than concrete descriptions and layout of intent. Many sentences and sections are very hard to follow. For example, section 5.4 “Ablations”, opens with “DC-PCA uses the target SS itself as the representation and ranks below MINUET” , the reader is expected to infer that this sentence, in the context of ablation, is there to justify that using only SS is not enough. Similarly, Section 3, which is meant to ground the theoretical motivation, uses narrative phrasing and subsection headers rather than directly enumerating and explaining the core leading theoretical limitation of standard contrastive learning.

  1. Presentation of the results.

Although Minuet demonstrates strong empirical performance overall, several evaluations and presentation choices could be better justified. While the authors follow the “scMultiBench benchmark”, the paper presents evaluation over a subset of benchmark metrics, and the main text focuses on ranking, rather than numerical results. The lack of the latter hides similarities in performance across compared methods over many metrics (details can be recovered through tables 6-16) .

Questions for authors
  1. Can the authors provide a concise numerical presentation of the main results (including average and variations) and discuss them in detail.

  2. Can the author explain the choice of the subset of scMultiBench metrics ?

  3. Are the authors willing to commit to rewriting the paper, ensuring that they present the main concepts in a straightforward manner? (Without this, it is unlikely this paper meets the threshold for acceptance)

  4. It will be interesting if the authors would comment on the broader implications of the work, including potential applications of the approach to other modalities or domains.

6 comments

Thank you for the careful reading and the direct advice.

W1 / Q3. Clarity, and the commitment to rewrite.

"The text is very dense in prose and leading claims rather than concrete descriptions and layout of intent."

Yes. We had started rewriting the paper after submission for the same reason you give, and your review pushed us further. In the main text of the revised PDF, red marks text we changed after submission, before the reviews arrived, and blue marks text we changed in response to the reviews.

Section 3 now starts by listing the three problems and the part of Minuet that fixes each one:

"When cells come from many donors, contrastive learning on paired cells has three problems. (i) Negatives from all donors reward encoding the donor (Eq. 2). (ii) Negatives from the same donor leave each donor free to rotate and shift its representation (Proposition 3.1). (iii) Removing all donor information also removes biology that differs between donors (Eq. 7). Minuet draws negatives within each donor to avoid (i), adds a shared prediction head that removes the rotations in (ii), and penalizes donor information only within covariate groups, which aligns the shifts while avoiding (iii)."

Section 4.4 now states what each ablation tests before presenting the result. The sentence you quote is gone. For example, the paragraph on the head now reads:

"Within-donor negatives alone do not transfer well. Adding the head and the penalty raises F1 on every cohort, and Table 1b shows that the head gives most of this gain. To see whether the head only reads out what the encoder already has, we freeze each encoder and fit the best linear head on top. On query donors, this head explains 34% of the target when the encoder was trained with the shared head and 20% when it was not. By Proposition A.2, a linear head cannot change what a linear readout sees, so the gain comes from the encoder."

If any part is still hard to follow, please tell us, and we will revise it.

W2 / Q1 / Q2. Numbers instead of ranks, and the choice of metrics.

"The main text focuses on ranking, rather than numerical results … Can the author explain the choice of the subset of scMultiBench metrics?"

We agree that the rank bubbles were hard to read. A rank tells you which method is better, but not by how much. The main figure now shows the scores themselves as bars, and Appendix C shows the same plots for each cohort, with error bars over three seeds.

Regarding the metrics, scMultiBench [1] includes many, several of which measure the same thing. For each question, we kept one or two in the main text and moved the others to the appendix.

QuestionMain textAppendix onlyWhy
Are cell types kept?ARI, NMIcLISI, graph connectivity; silhouette, which we do not rankEvery method scores above 0.92 on cLISI in every cohort, so it does not separate the methods. Graph connectivity measures the same thing as ARI and NMI. Silhouette has been shown to mislead about both batch removal and biology [2].
Are donors well mixed?CiLISI, which measures mixing within each cell typeiLISI, donor ASWiLISI and donor ASW increase as donors become more similar, even when the difference is biological, such as disease status. They are not good metrics for donor-level integration.
Are the two assays of each cell matched?FOSCTTMMRR, Recall@1All three ask the same question. FOSCTTM uses the whole ranking and needs no cutoff.
Can cell types and phenotypes be predicted for new donors?macro-F1, plus macro-AUROC for phenotypeWe use the same classifier for every method, so the scores compare the representations.

The donor-mixing choice matters most when some donors are sick, and others are healthy. ROSMAP is a good example:

ROSMAPDonor iLISI ↑Alzheimer macro-F1 ↑
Apollo0.360.30
MultiVI0.310.28
Minuet0.260.71

Apollo and MultiVI look the best mixed, but they predict Alzheimer status worse than random guessing. We think they look well mixed mostly because the disease signal is gone.

Minuet is not the best everywhere. IndiSeek, for example, transfers cell types to new donors better, but it hardly matches the RNA and ATAC profiles of the same cell, so its FOSCTTM is close to random. Section 4.3 now says this. Does this choice of metrics seem reasonable?

[1] Liu C, Ding S, Kim HJ, et al. Multitask benchmarking of single-cell multimodal omics integration methods. Nature Methods, 2025.

[2] Rautenstrauch P, Ohler U. Shortcomings of silhouette in single-cell integration benchmarking. Nature Biotechnology, 2026.

(continued in the next comment)

(continued)

Q4. Broader implications.

The problem is not unique to multiome data. It arises whenever paired views are grouped by both technical and biological factors: CITE-seq or spatial data across donors, histology and expression across hospitals, or images and text across sites. In each case, contrastive learning allows each group to retain its own coordinates. The solution is the same: use a shared target to establish a common frame, and a penalty that preserves the group differences that matter.

The question we find most exciting comes next. Minuet uses recorded covariates to decide which donor differences to preserve, but for an individual, we often know much more—including their genome. Genetic variants are fixed at conception, so a cell-state difference associated with a variant reflects biology, not batch; we can estimate its effect on disease using Mendelian randomization. The rotated program in our toy represents such a variant. We plan to use genotypes to determine which donor differences the penalty should preserve, then ask which cell states a variant changes and whether those changes cause disease. Population-scale single-cell cohorts such as OneK1K make this feasible.

W1 / Q3. Clarity, and the commitment to rewrite. Thank you for making the changes; Section 3 indeed reads much better now! If possible, it will be great to use the additional page allowed in revisions to (1) extend the presentation / explanation of the results; and (2) write the discussion more directly in general.

W2 / Q1 / Q2. Numbers instead of ranks, and the choice of metrics. Thank you for making the changes in text and clarifying some of the choices here. The choice seems reasonable; however, the response above is a bit cryptically written and it is not very clear what the authors are trying to explain. (e.g. the sudden "jump" to "The donor-mixing choice matters most when some donors are sick..[]" ).

Related to the evaluations, the synthetic experiment now considers a different setting, focusing on cross-donor alignment. It no longer tests all three failure modes or compares against all baselines (it only includes contrastive controls and PCA, which already performs quite well). The experiment is designed to favor the PCA-based target, and PCA indeed performs quite well. Could the authors explain why the published baselines were not included in the new experiment setting?

Q4. Broader implications. Thank you for the clarification. It will be nice to add this to the discussion.

======== Q4. Broader implications.

The problem is not unique to multiome data. It arises whenever paired views are grouped by both technical and biological factors: CITE-seq or spatial data across donors, histology and expression across hospitals, or images and text across sites. In each case, contrastive learning allows each group to retain its own coordinates. The solution is the same: use a shared target to establish a common frame, and a penalty that preserves the group differences that matter.

The question we find most exciting comes next. Minuet uses recorded covariates to decide which donor differences to preserve, but for an individual, we often know much more—including their genome. Genetic variants are fixed at conception, so a cell-state difference associated with a variant reflects biology, not batch; we can estimate its effect on disease using Mendelian randomization. The rotated program in our toy represents such a variant. We plan to use genotypes to determine which donor differences the penalty should preserve, then ask which cell states a variant changes and whether those changes cause disease. Population-scale single-cell cohorts such as OneK1K make this feasible.

Reviewer_J5E9OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to learn representations from paired single-cell RNA and ATAC measurements that transfer across individuals. It identifies three challenges: pooled negative sampling can reward donor-specific shortcuts; within-donor contrastive learning leaves donor-specific transformations unconstrained; and unconditional removal of donor information can discard biologically meaningful variation.

The proposed method, MINUET, combines within-donor contrastive learning, a shared prediction head trained on donor-centered PCA targets, and donor mean and covariance alignment within biological covariate groups. Frozen encoders subsequently embed query cells without adaptation. The paper provides theoretical analyses of the contrastive objectives and a conditional identification bound relating the learned representation to the shared prediction target.

Experiments cover six multiome cohorts comprising 636,509 cells from 110 donors. Evaluation includes observed-cell integration, held-out-donor transfer, leave-one-dataset-out transfer, and phenotype prediction. MINUET achieves the best reported average rank across all four protocols, although individual metrics favor different methods.

Critical strengths
  1. Clear problem formulation with well-motivated methodological choices. The paper distinguishes matching modalities within cells from learning representations that are comparable across donors. Its analyses of donor shortcuts, donor-specific transformations, and the conflict between donor invariance and phenotype preservation provide a coherent explanation of why modality matching alone is insufficient. The proposed objectives address these distinct issues, and the synthetic experiments help examine their mechanisms.
  2. Broad evaluation with meaningful attention to donor-level generalization. The experiments extend beyond integration of training cells to unseen donors, unseen studies, and donor phenotype prediction. Query donors are excluded from preprocessing fitting, feature selection, and representation training, while predicted phenotypes are withheld from the conditioning covariates in the phenotype-prediction protocol. Treating donors as the statistical units and including the PCA prediction target itself as a baseline strengthen the evaluation. Together, these experiments provide substantial evidence that the method learns useful representations beyond the reference cells.
Critical weaknesses
  1. The theoretical results provide a conditional characterization of the representation, rather than a guarantee of biological recovery or cross-donor generalization. The identification bound controls distance to a transformed PCA-based target, conditional on a full-column-rank prediction head, with constants depending on its singular values. It does not establish that this target captures the relevant biological variation or that the learned head is sufficiently well conditioned. Furthermore, the analysis uses squared prediction errors, whereas training uses SmoothL1, so the quantitative connection to the implemented objective requires clarification. The unseen-donor result depends on prediction and alignment errors measured on the query donor; it does not establish that these errors remain small under distribution shift. The synthetic experiment addressing donor-specific rotations also evaluates held-out cells from reference donors, rather than unseen donors with independently drawn transformations. These points do not negate the empirical transfer results, but they limit how strongly the theory explains or guarantees them.
  2. The aggregate rankings obscure task-specific trade-offs and do not use an identical metric set for every method. Although MINUET has the best average rank, it does not consistently lead on individual transfer outcomes. For example, Table 6 reports held-out-donor cell-type transfer macro-F1 of 0.640 for MINUET versus 0.737 for IndiSeek, and LODO macro-F1 of 0.514 versus 0.540. The phenotype-prediction ranking also includes donor mixing and paired-cell matching, rather than only phenotype prediction. In addition, FOSCTTM is excluded from the averages for Apollo and DC-PCA but included for other methods, making the aggregate ranks less directly comparable. The detailed numerical results are useful, but a common-metric ranking and task-specific comparisons with donor-level uncertainty are needed to establish the robustness and practical meaning of the overall superiority claim.
Questions for authors
  1. Applicability of the theoretical bound. What are the minimum singular values and condition numbers of the learned prediction heads across cohorts and seeds? How does the squared-error analysis relate quantitatively to the SmoothL1 objective used in training? For covariates that vary within donors, how large is the target discrepancy term (B_S)?
  2. Diagnostic value for unseen donors. Do the query prediction error and donor-mean discrepancy in Corollary 3.1 predict actual failures in cell-type transfer or biological conservation? Demonstrating this relationship would help establish the practical value of the theoretical characterization beyond proximity to the PCA target.
  3. Role of the shared target. How sensitive are the results to the PCA target’s dimensionality and construction? Is there evidence that the target preserves the biological signals needed for transfer, particularly when those signals explain little variance?
  4. Robustness of the comparative conclusions. Does MINUET retain its advantage when rankings are restricted to metrics available for all methods? For phenotype prediction, what are the conclusions based solely on macro-AUROC and macro-F1, with paired donor-level confidence intervals? Please also discuss the settings in which stronger paired-cell alignment does not translate into better cell-type transfer.
4 comments

Thank you for the review. All numbers are for the revised Minuet, which computes the target from normalized counts (see the general response).

W1 / Q1. The theory is conditional: head conditioning, SmoothL1, BSB_S, new donors.

"The identification bound … [is] conditional on a full-column-rank prediction head, with constants depending on its singular values … the analysis uses squared prediction errors, whereas training uses SmoothL1."

We agree that the theory is conditional, and the paper now says so. If the head predicts the target well, the representation is one linear map of the target, up to one shift per group. The theorem does not say that the target holds all the biology. The Discussion now reads: "First, it aligns only what the target captures. The head places donors in common coordinates by predicting the target, so a signal absent from the target is not aligned across donors."

We measured each quantity you asked about:

QuantityValue
Condition number of the trained head500 to 1800
Query residuals in the quadratic range of SmoothL1 (below β = 1)80% to 84%
BSB_S as a share of target variance1.1% (Wang), 3.7% (Mannens)
Toy, 4 new donors: angle between frames/readout R²Minuet 6° / 0.78; within-donor contrastive 32° / 0.32

Head conditioning. The heads are poorly conditioned, so a bound with 1/σmin⁡21/\sigma_{\min}^2 would be loose. We therefore also state the result in the target coordinates, where it is an identity with no singular values (Appendix A.4): the distance of PzPz to the target equals the prediction error plus the between-donor variance within groups. The singular values enter only when we measure the error in the zz coordinates, and only in directions in which zz varies. Does this form of the result address your concern about conditioning?

SmoothL1. SmoothL1 equals half the squared error in its quadratic range, and most query residuals fall there. Also, the stability result holds for any zz and PP, regardless of how they were trained, and we measure its error term after training.

BSB_S. BSB_S enters only where a donor spans several groups, which happens in Wang and Mannens (Table 3). It is small next to the prediction error.

New donors in the toy. You are right that the old toy tested only held-out cells of the training donors, while our claim is about new donors. We therefore rebuilt it. In the new toy, donors differ only by the angle of one gene program and a technical factor, so any loss in the readout can come only from donors using different frames. We train on 6 donors and test on 4 new donors with angles not seen in training (Section 4.2). Minuet places them in nearly the same frame as the reference donors; within-donor contrastive learning does not.

Q2. Do the two terms of the query-donor bound predict failures?

We computed both terms in Corollary A.1 for each query donor (72 donor–seed pairs across six cohorts). Across cohorts, both terms are associated with lower mapping F1 (Spearman correlations of −0.33 and −0.34). Within a cohort, however, they rank donors only weakly; with about four query donors per cohort, these rankings have little power. Nor do we expect a tight relationship: the corollary bounds how far a new donor lies from the target, while cell-type F1 also depends on how well the target separates cell types in that cohort. We therefore view the two terms as a label-free check for flagging a donor or cohort that is far from the reference, rather than as predictors of F1.

Q3. How much does the shared target matter?

On Wang and Crohn, mapping F1 changes little with the rank of the PCA target, and more components help a little:

Target rankHalfDefaultDouble
Mapping F10.4970.5170.542

The theory needs only that the target is the same function of the cell state in every donor (A2); donor-centered PCA is the simplest such choice. The target fixes the frame, and the contrastive loss keeps what RNA and ATAC share, whatever its variance. So low-variance biology is not lost: phenotype is predicted much better from Minuet than from PCA, the target itself (Figure 7).

(continued in the next comment)

(continued)

W2 / Q4. Rankings on common metrics and task-specific comparisons.

"FOSCTTM is excluded from the averages for Apollo and DC-PCA but included for other methods … a common-metric ranking and task-specific comparisons with donor-level uncertainty are needed."

Common metrics. FOSCTTM cannot be applied to Apollo and DC-PCA. Without this metric, Minuet still has the best overall rank (2.34, against 3.89 for scPairing). It is first in integration, mapping to new donors and phenotype, and second in mapping to a new study, just behind scGLUE.

Phenotype. The phenotype rank now uses only macro-AUROC and macro-F1. Donor mixing and paired-cell matching answer other questions, and donor mixing would reward a method for removing the phenotype itself. Donors are the units here, so we compared methods with a paired donor-level bootstrap: 2000 resamples of donors within each phenotype class, with the same resample for both methods. Macro-F1 difference, mean over the three phenotype cohorts:

Minuet minusDifference [95% interval]
Strongest baseline in each cohort−0.01 [−0.17, 0.16]
MultiVI+0.22 [0.08, 0.37]

With 16 to 28 donors per cohort, the intervals are wide, and we do not claim that Minuet beats the strongest baseline on phenotype. What we take from this is that Minuet keeps the phenotype while it removes donor information: it is on par with the best baselines and clearly above MultiVI, which mixes donors best (Figure 7).

IndiSeek. We rechecked IndiSeek, and the numbers are correct. It has the highest mapping F1, but its FOSCTTM is close to random pairing. The two metrics ask different questions: mapping F1 asks whether cell types can be read from the embedding, and FOSCTTM asks whether the RNA and ATAC profiles of the same cell lie close together. A method can keep cell types in each assay without matching the two, and we think IndiSeek does this. Minuet is built to do both, and Section 4.3 now states this trade-off.

Dear Reviewer J5E9,

Thank you again for your review. As the discussion period closes on November 18, we wanted to ask whether our response addressed your concerns. As you suggested, we have completely updated the toy model. It now tests donors using rotations not seen during training.

Best regards, The authors

Thank you for the detailed responses and substantial additional experiments. Overall, I am satisfied with the revision and maintain my original rating.

Reviewer_T7kPOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies why contrastive learning on paired single-cell RNA and ATAC measurements may fail to transfer across donors. It identifies donor shortcuts under pooled negatives, donor-specific coordinate ambiguities under within-donor negatives, and the loss of biological phenotype under unconditional donor suppression. Minuet combines within-donor contrastive learning, a shared prediction head anchored to donor-centered PCA targets, and donor penalties conditioned on biological covariates. Theoretical results characterize identification relative to this reference frame, while experiments evaluate integration, donor and dataset transfer, and phenotype prediction across six cohorts.

Critical strengths
  1. The problem diagnosis is clear and technically substantive. The donor-shortcut analysis, invariance result for donor-specific transformations, and information-theoretic account of phenotype loss explain related limitations of contrastive objectives. The synthetic experiments test these mechanisms directly, including shared-versus-donor-specific readouts, over twenty seeds.
  2. The evaluation protocol is carefully designed. Donors are the unit of evaluation, and held-out donors are excluded from preprocessing and feature selection as well as model fitting. This makes the transfer results more informative than evaluations based only on random cell splits.
  3. The controls are well chosen. Architecture-matched contrastive controls help separate objective design from encoder choice, and DC-PCA directly tests the value added beyond Minuet's own prediction target. A common configuration across cohorts and protocols further supports the empirical comparison.
  4. The method addresses a meaningful biological trade-off. Conditional penalties distinguish nuisance donor variation from differences associated with recorded biological covariates. Minuet achieves the best average rank under all four protocols, and the paper explicitly discusses how the choice of covariates limits what can be preserved or removed.
Critical weaknesses
  1. A shared prediction constraint does not by itself establish an improvement in the backbone representation. Appendix A requires a full-column-rank head and uses target and representation dimensions of 64. Exact centered prediction then constrains within-donor variation to a shared linear transform of the PCA target, with component-specific shifts retained. This rules out the stated donor-specific coordinate freedom under the theorem’s assumptions, but does not establish that such misalignment occurs in real training or that the encoder must learn better biological features. If its existing features already admit a common linear predictor of the target, adjusting the head can satisfy prediction without a substantial encoder change. The theorem therefore identifies representations relative to a chosen reference; it does not alone explain improvement over DC-PCA. Learned geometry, retained offsets and nonzero prediction error may still affect downstream performance. A frozen-backbone/head-only control and a simple learned linear-encoder control, alongside DC-PCA, would help distinguish an effective readout from a useful change in representation. Prediction error, head conditioning and transfer quality should be assessed together.

  2. Repairing an ambiguity introduced by the sampling design does not establish the value of the overall design. Restricting negatives to the same donor addresses the pooled-negative shortcut but leaves independent donor coordinate systems unconstrained; the shared prediction head is then introduced to repair this consequence. The theoretical diagnosis and remedy are informative, but their success within this design chain does not establish that the chain is preferable to retaining some cross-donor comparisons. These components should therefore be assessed jointly, rather than treating the coordinate repair as independent evidence for the necessity of the full method. The central missing comparison is a sweep over intermediate cross-donor negative fractions, holding the remaining objectives and sampling budget fixed, both with and without shared prediction. This would test whether a simpler mixed-negative alternative achieves comparable transfer while controlling donor shortcuts. Neither the sign of the shortcut incentive nor the existence of a loss symmetry establishes the magnitude of the practical problem. Cross-donor comparisons also need not remove every ambiguity; their benefit requires empirical evaluation.

  3. The mechanism evidence on real cohorts needs a more direct interpretation. The appendix already reports within-group donor dispersion and donor-prediction results, so real-data diagnostics are not absent. However, residual within-group donor predictability remains substantial: Table 12 reports 0.755 for Minuet on Wang and 0.911 on ROSMAP, compared with 0.663 and 0.582 for MultiVI. These single-seed values do not prove failure of the penalties: aligning means and covariances is weaker than removing all donor information, and stronger suppression can also remove biology. Nevertheless, the residuals deserve discussion alongside biological retention and architecture-matched ablations. The real-data ablation figure covers only two cohorts; numerical results with uncertainty and a cross-validated shared-versus-donor-specific readout diagnostic would better connect the synthetic mechanism to practical transfer. Reporting cohort-specific graph connectivity and spectral gaps would also clarify where that part of the theory is informative.

  4. The headline rank should be accompanied by task-specific trade-offs. Minuet's best average rank does not imply the best cell-type transfer or donor mixing on every cohort. Unconditional donor mixing is also imperfectly matched to a method designed to retain between-covariate differences. A version stratified jointly by cell type and covariate group would assess the intended objective more directly. Since paired-cell matching is unavailable for some methods, ranks over a common metric subset and effect sizes against the closest comparators would make the summary easier to interpret. The query-donor diagnostic also requires donor-level target statistics, whereas embedding a new cell does not; these different inference requirements should be stated explicitly.

Questions for authors
  1. How do a frozen-backbone/head-only control and a simple learned linear encoder compare with Minuet and DC-PCA in target prediction and donor transfer? What evidence attributes the benefit of shared prediction to an improved backbone representation rather than a readout of existing features?
  2. Does the full within-donor-plus-prediction design outperform a simpler mixed-negative alternative? Please compare intermediate cross-donor negative fractions, with the remaining objectives and sampling budget fixed, both with and without shared prediction, and report donor transfer alongside shortcut and registration diagnostics.
  3. Can the real-data ablations be tabulated with uncertainty, and can residual within-group donor predictability be related to biological retention, registration diagnostics, and the donor-graph quantities?
  4. Can the summary include conditional donor-mixing metrics, common-subset ranks, and explicit distinctions between per-cell embedding and donor-level diagnostic requirements?
5 comments

Thank you for the careful review. Your questions led to most of the new experiments in the revision: ablations on all six cohorts with three seeds, a sweep over the fraction of cross-donor negatives with and without the head, a head-only control, and the donor graph for each cohort. All numbers below are for the revised Minuet, which computes the target from normalized counts (see the general response).

W1 / Q1. Does the shared head improve the encoder, or only the readout?

"A frozen-backbone/head-only control and a simple learned linear-encoder control, alongside DC-PCA, would help distinguish an effective readout from a useful change in representation."

We agree that this is the key question, and we ran both controls.

Head-only control. We freeze the encoder and fit the best linear head to the target. If the shared head merely read out features already present in the encoder, this frozen head would perform equally well whether or not the encoder had been trained with the shared head. It does not: on query donors, it explains much more of the target when the encoder was trained with the head, and transfer is also better in Table A. Since a linear head cannot change what a linear readout of the encoder can use (new Proposition A.2), we interpret this gap as evidence of a change in the encoder itself. Section 4.2 shows the same toy example.

Linear encoder and PCA. A linear encoder trained with the full objective transfers about as well as the MLP, so the gain does not come from the MLP. PCA (DC-PCA in the previous version) is the target itself, with no learning, and Minuet outperforms it on every task in Figure 6.

Table A. Query donors; mean over six cohorts and three seeds unless noted.

ControlMinuetComparison
Frozen-head R² of the target0.340.20 (encoder trained without head)
Mapping F10.670.59 (without head)
Mapping F10.670.56 (PCA)
Phenotype F10.620.52 (PCA)
Toy readout R²0.850.39 (head on frozen within-donor encoder)
Mapping F1, Wang and Crohn0.52 (MLP)0.53 (linear encoder)

Do these controls answer your question, or would another comparison be more direct?

W2 / Q2. Is the full design better than mixing in some cross-donor negatives?

"The central missing comparison is a sweep over intermediate cross-donor negative fractions, holding the remaining objectives and sampling budget fixed, both with and without shared prediction."

We ran the sweep. The fraction pp of negatives from other donors takes the values 0, 0.25, 0.5, and 1, with and without the shared head; the batch, the sampling budget, and all other terms stay the same. Details are in Table B and Table 5 in the paper.

Table B. Mean over six cohorts and three seeds. p=0p = 0 is the default Minuet; the other values use plain InfoNCE.

pp00.250.51
Donor accuracy within groups, with head ↓0.540.570.620.76
Mapping F1, with head ↑0.670.650.670.67
Mapping F1, without head ↑0.590.580.590.60

We see two results, and both hold in every cohort. First, the donor becomes easier to predict as pp grows, as Eq. (2) predicts; Proposition A.1 now shows that the reward for a donor feature grows in proportion to pp. Second, cross-donor negatives do not replace the head: without the head, transfer stays low at every pp. With the head, transfer changes little with pp, so we chose p=0p = 0, which encodes the donor least. We agree that the components should be judged together, and the sweep supports the full design: within-donor negatives keep the donor shortcut small, and the head is needed at every pp.

(continued in the next comment)

(continued)

W3 / Q3. Mechanism evidence on real cohorts.

"Residual within-group donor predictability remains substantial … numerical results with uncertainty and a cross-validated shared-versus-donor-specific readout diagnostic would better connect the synthetic mechanism to practical transfer."

Uncertainty and cohorts. The ablations now cover all six cohorts, with mean and s.d. over three seeds in Tables 1 and 5.

Residual donor information. We agree that the residual deserves discussion next to biological retention. We now plot phenotype macro-F1 against excess donor information within phenotype groups in Figure 7 (detailed panels in the appendix). On Crohn, Minuet retains the most phenotype of all methods while having less excess donor information than every method except MultiVI; on ROSMAP, it retains clearly more phenotype than PCA. We view the residual as donor variation within a group that the penalty does not target. The penalty aligns group means and leaves the rest, and removing all of it would also remove phenotype (Eq. 7).

Shared versus donor-specific readout. We fit, with cross-validation, one shared linear readout of the target and one readout per donor. On query donors, the donor-specific readout is only slightly better, and the gap is smaller with the head (R² gap 0.033) than without it (0.048). So on real data, too, the head brings donors closer to one frame.

Donor graph. Table 3 of the paper now lists the components and the constant γ for each cohort:

CohortGroups fromDonors in several groupsGroups → componentsγ
ROSMAP, Crohn, Alopecia, RAdonor-level covariatesnoneeach group is a component1
Wangstage × region9 of 2210 → 70.52
Mannenstissue5 of 67 → 10.50

So the general bound of Appendix A.7 matters only for Wang and Mannens, where the factor 1/γ is about 2.

W4 / Q4. Trade-offs behind the headline rank.

"A version stratified jointly by cell type and covariate group would assess the intended objective more directly … ranks over a common metric subset … would make the summary easier to interpret."

Common metrics. FOSCTTM is the only metric that some methods lack. Without it, Minuet still has the best overall rank (2.34, against 3.89 for scPairing). It remains first in integration, mapping to new donors and phenotypes, and second in mapping to a new study, just behind scGLUE.

Trade-offs. Section 4.3 now says where Minuet does not lead. IndiSeek transfers cell types best, but its FOSCTTM is close to random pairing. In our view, it keeps cell types without matching the two assays of a cell.

Stratified donor mixing. We computed donor CiLISI within each stratum of cell type × covariate group, for mapping to new donors (Mannens left out, since each tissue has only one query donor). Stratification raises the scores of all methods but barely changes the order:

MethodMean rank, by cell typeMean rank, by cell type × groupROSMAP phenotype F1
MultiVI1.41.80.28
Apollo1.81.80.30
Minuet3.84.00.71
scCLIP4.44.60.52

The two methods that mix donors best also lose most of the phenotype signal (below random guessing, 0.5). We take this as the trade-off you describe: more mixing is not always better when donors differ in biology. Is this the stratified metric you had in mind?

Inference. Embedding a new cell needs only the frozen encoders. The bound for a new donor (Corollary A.1) also needs the donor's group and its donor-level means, and Appendix A.5 now states this.

Dear Reviewer T7kP,

Thank you again for your review. As the discussion period closes on November 18, we wanted to ask whether our response addressed your concerns.

Best regards, The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers for their helpful comments. The reviewers found the problem well motivated (T7kP, J5E9, TEYJ) and the donor-level evaluation careful (T7kP, J5E9). The main concerns were that the paper is hard to follow (TEYJ) and that the main text gives ranks instead of numbers (T7kP, J5E9, TEYJ).

We revised the paper substantially. The main changes are:

  • Section 3 now starts with the three problems of contrastive learning across donors and gives one component of Minuet for each. We merged the theory and method sections and moved the secondary results to the appendix.
  • Main figure. We changed it from rank bubbles to bar plots of the scores so that close results are visible; the last column retains the overall rank, as scIB and scMultiBench do. The appendix shows the same plots for each cohort, with the s.d. across three seeds. Appendix C now explains why we use each metric and why we exclude some scMultiBench metrics. For example, donor mixing without conditioning rewards removing biology that differs between donors, a point that Reviewer T7kP also makes.
  • Names. Some of our terms are unusual. We now use the scMultiBench task names (Integration, Mapping, Phenotype). We call negatives drawn from all donors "global negatives,” distinguishing them from Minuet’s within-donor negatives, which are an example of conditional contrastive learning [1]. We renamed DC-PCA to PCA; it remains PCA after centering each donor.
  • Synthetic experiments. All of them now use one simple generator, and the query donors have rotations not seen in training.
  • Prediction target. After submission, we found that the target should be computed from library-size-normalized, log-transformed counts, because on raw counts the leading components mostly follow sequencing depth. All Minuet results in the revised PDF use this target. It mainly helps cell-type transfer to held-out donors:
Mapping macro-F1Raw countsNormalized
Mean over six cohorts0.640.67
Crohn0.600.68
Rheumatoid arthritis0.870.94
  • Appendix A is rewritten. The proofs are shorter and use the notation of the main text, and each subsection restates the main-text result it proves. The main changes:
    1. The identification result now states its assumptions (A1)–(A3). It gives the exact result at the objective minimum (Theorem A.1) and then a stability bound (Theorem A.2). In the target coordinates, this bound is an identity with no singular values.
    2. A new proposition shows that a linear head on a frozen encoder cannot change what a linear readout can use (Proposition A.2).
    3. Appendix A.7 treats covariates that vary within donors. It defines the constant γ directly and computes it for each cohort in Table 3.
    4. The description of the target now includes its normalization.
  • Appendix layout. The per-cohort results are now bar charts with the mean and s.d. over three seeds for every metric, cohort, and method, including donor-prediction F1. We removed the cohort-mean table because the main figure shows the same means. The per-cohort ablations, including the choice of normalization layer, are now one table that matches the ablation table of the main text.
  • New experiments. In the paper: the ablations now cover all six cohorts with three seeds, including a sweep over the fraction of cross-donor negatives with and without the shared head and a head-only control on a frozen encoder (T7kP); the donor graph of each cohort (T7kP); and the query-donor diagnostic (J5E9). In our replies: a linear-encoder control and the shared versus donor-specific readout (T7kP); the condition of the head, the share of SmoothL1 residuals in the quadratic range and the size of BSB_S (J5E9); ranks on the metrics that all methods share (T7kP, J5E9); donor mixing within strata of cell type × covariate group (T7kP); and paired donor-level intervals for phenotype (J5E9).

In the main text of the revised paper, red marks text changed after submission, before the reviews arrived, and blue marks text changed in response to the reviews. Green marks text changed in the second round. The appendix is not colored because we reorganized much of it.

[1] Tsai YHH, Li T, Ma MQ, Zhao H, Zhang K, Morency LP, Salakhutdinov R. Conditional contrastive learning with kernel. ICLR, 2022.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.