acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Weak Supervision for Multilevel Representation Learning with scPACE

Abstract

Scientific and clinical datasets often consist of groups of exchangeable instances, such as thousands of cells sampled from a patient or donor. Yet information is distributed unevenly across these levels: integer-valued molecular counts are measured for individual cells, whereas phenotypes such as age, diagnosis, or disease severity are typically measured only for the donor. Cell-level phenotypes and cell-type labels may be sparse or missing. Although modern generative models can represent single-cell count data, they do not typically model how donor-level biological variation manifests through cellular states, cell-type composition, and gene expression. We introduce scPACE, a hierarchical generative model that represents each donor by an underlying biological state and each cell by its position along a related biological continuum, such as developmental pseudotime, cellular maturity, or disease-associated progression. This structure explicitly captures how donor-level variation is expressed heterogeneously across individual cells. Under explicit assumptions, the latent state hierarchy is identifiable up to a sign convention. We apply scPACE to large-scale real-world single-cell studies spanning mouse cortical development, and Alzheimer’s and Huntington’s disease. Our framework provides a principled foundation for generative modeling of phenotype-linked biological variation across donors and cells, enabling the study of developmental and disease progression across biological scales.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
hierarchical generative models, variational autoencoders, weakly supervised representation learning, semi-supervised learning, multiple-instance learning, identifiability, single-cell transcriptomics, disease progression modeling, computational biology

Reviews

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

The paper defines a multi-level generative model which includes cell state, and phenotype and batch information, and evaluates how well the representations learned by it perform in downstream analysis.

Critical strengths

S1: Presentation

The paper clarifies it's method, assumptions and it's model clearly. The presentation is good.

S2: Quantitative evaluation of the method

Critical weaknesses

The paper is an interesting read, but needs further work. I encourage the authors to make improvements towards (1) clarifying the role of the identification results and (2) motivating their experimental evaluation, and would consider raising my score if both of these concerns are addressed.

  • W1: Irrelevancy of the identification results:

The relevancy of the identification results remains unclear.

  • W2: Motivation of the experimental evaluation:

The paper could improve the motivation for the experimental evaluation for downstream tasks.

Questions for authors
  • See W1: Do the assumptions made hold in practice?
  • See W2: What is the biological relevancy of recovering donor- and cell-level phenotype variation in unseen donors.
  • How does the method compare to other hierarchical representation learning frameworks?
  • Why does the modeling require a multi-level hierarchy?
2 comments
AuthorsOpenReview ↗

Detailed response to reviewer TQuM

We thank the reviewer and note the offer to raise the score if W1 and W2 are addressed. We answer both directly here: the full technical treatment of the identifiability scope is in our response to Reviewer 7i31 (Q5) and in the revised Appendix A.6.

W1. Relevance of the identifiability result; do the assumptions hold in practice?

Why it matters. The quantities we interpret biologically are two scalars, tdt_d per donor and sns_n per cell, and a coordinate of a nonlinear latent-variable model has no meaning unless the data determine it: two equally good fits could order the same cells differently, with the ordering an artifact of initialization (Locatello et al., 2019). Theorem 1 says the model’s structure removes this freedom up to a single joint sign flip, with three concrete consequences:

  1. It justifies the biological interpretations in Sections 4.2–4.4 (any two equally good fits give the same type ordering and the same within-donor ordering, so those orderings are properties of the data).
  2. It fixes the evaluation protocol (we report ∣r∣|r| and calibrate only sign and scale, the only unidentified components).
  3. It predicts what fails when an assumption is dropped (without a cell channel the intercepts mkm_k are unidentified, so type positions are arbitrary while within-type ordering is not, exactly the Setting 1 pattern: pooled 0.08, within-type median 0.45).

Do the assumptions hold? The donor-level assumptions hold in all four settings: the donor phenotype is observed for every donor and measured independently of the transcriptome. The cell-level assumptions hold in full only in Setting 4 (CAG is an independent PacBio measurement, cell types observed). Settings 1 and 3 have no cell channel, so the donor-level conclusions hold and the cell-level ones reduce to within-type ordering. Setting 2’s maturity score is a function of the counts, so it is not an independent proxy and we do not claim the cell-level conclusions there. We have stated these three roles in the revised Appendix A.6 (scope-by-setting paragraphs). The full derivations are in our 7i31 Q5 response.

W2. Biological relevance of recovering donor- and cell-level variation in unseen donors; comparison to other hierarchical frameworks; why a multi-level hierarchy.

Unseen donors are both the test and the use. The inferred donor state td^\hat{t_d} estimates from expression how far a donor has progressed along a process (development, pathology, disease burden). When a label exists, td^\hat{t_d} can differ from it usefully (a donor molecularly ahead of or behind its chronological age or neuropathological stage). When none exists (a new cohort without quantitative neuropathology, a model system with no clinical stage), td^\hat{t_d} gives a stage from expression alone. At the cell level, where per-cell phenotypes are rarely measured, sn^\hat{s_n} places every cell on the same axis as the donor phenotype, so one can ask which types move first as donors progress and which cells within a donor are ahead.

Why two levels. Information is measured at two levels, and a single-level model discards one. Copying the donor label to every cell is pseudoreplication: hundreds of thousands of cells treated as independent observations of a quantity that varies across tens of donors, with every cell of a donor pushed toward the same value despite real within-donor differences. Summarizing cells into one donor prediction (as scVIP does) leaves no per-cell state, so within-donor biology cannot be inferred. HD makes this concrete: the cause is one number per donor (inherited CAG, summarized with age in CAP), but the damage is cell by cell (somatic CAG expands at different rates in different types, and the spiny projection neurons that expand most are the ones that die, Handsaker et al., 2025). Within one donor an SPN and an oligodendrocyte are at very different disease stages despite sharing a CAP score, and CAP per donor and somatic CAG per cell are distinct measured quantities, so a one-level model must drop one.

Comparison to other frameworks. We benchmark scVIP because it is the hierarchical single-cell model that predicts donor phenotypes, just one of the task in question. The general group-level VAEs in Section 2 (Neural Statistician, ML-VAE, FHVAE, Ladder VAE, NVAE, VDVAE) were built for images or speech and have no count likelihood, phenotype channels, or cell-type model. The single-cell multi-group models (multiGroupVI, scPoli, MrVI) learn per-sample embeddings but predict neither a phenotype nor a scalar cell state. To our knowledge no existing method combines all of a group latent, an instance latent, a generative group-to-instance edge, donor and cell phenotype channels, a count likelihood, semi-supervised cell types, and an identifiability result, which is the gap scPACE fills.

I thank the authors for their thoughtful response. They have addressed my concerns for W1 and W2.

Reviewer_7i31OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes scPACE, a hierarchical generative model for learning donor- and cell-level representations from single-cell RNA sequencing data. The method combines (1) a one-dimensional latent hierarchy linking donor and cell states through a linear relationship, and (2) cell-type-stratified attention-based MIL that aggregates cellular states and cell-type proportions to infer donor-level phenotypes.

Critical strengths

Meaningful problem formulation and interpretable methodology. The paper addresses an important challenge in connecting donor-level phenotypes with heterogeneous cellular states. The hierarchical structure provides an interpretable inductive bias, supported by a conditional identifiability analysis.

Critical weaknesses
  1. Restrictive cross-level coupling and insufficiently established practical applicability. The model assumes that the expected scalar cell state shifts with donor state at the same unit slope across all cell types, allowing only type-specific intercepts and variances. Although the downstream nonlinear model permits heterogeneous expression responses, this shared scalar coupling requires biological justification. When cell-level supervision is used, its compatibility with the donor phenotype depends on the chosen target, yet the paper does not establish how robust the method is to alternative, weakly related, or conflicting cell-level targets. This matters in practice because an appropriate cell phenotype may be unavailable or must be constructed by the user from marker genes. In the mouse experiment, adding maturity-score supervision improves query cell correlation from 0.08 to 0.97, but donor correlation decreases from 0.91 to 0.86 and donor RMSE increases from 0.40 to 0.50. These single-run results suggest a possible trade-off rather than a demonstrated general effect, and the cell-channel weight analysis further indicates sensitivity to the supervision configuration. Without clearer target-selection guidance and robustness analyses, it remains uncertain when the coupled model provides a reliable practical advantage.
  2. The evidence does not clearly isolate the benefit of cross-level supervision, and the theoretical guarantee has limited coverage of the applications. In the mouse donor-only setting, removing donor supervision changes median within-type cell correlation only from 0.45 to 0.42, while demeaned correlation increases from 0.22 to 0.34. With direct cell supervision, removing both generative donor-to-cell connections leaves pooled cell correlation at approximately 0.97, although donor prediction declines. Thus, the ablations support some benefits of coupling but do not clearly establish that donor-phenotype supervision materially improves cell-state inference. A corresponding donor-supervision ablation is missing for Alzheimer's disease. Furthermore, most principal scalar baselines do not receive the cell-level targets supplied to scPACE in Settings 2 and 4. The supervised scVI-plus-MLP comparison partially addresses this issue for mouse maturation, but an analogous comparison is missing for Huntington's disease. Finally, the identifiability result requires fully observed phenotype channels and cell types, known observation-noise variances, and independent phenotype measurements conditional on the latent states. These assumptions do not directly cover the expression-derived mouse target, the absence of cell supervision in AD, or the partially measured HD target. Consequently, neither the current comparisons nor the theorem fully establishes the added value and biological interpretability of the coupled states across the reported applications.
Questions for authors
  1. Choice of cell-level target and validity of the coupling. What criteria should practitioners use to select a cell-level phenotype compatible with the donor phenotype? How sensitive is performance to alternative targets, particularly when their association with the donor phenotype is weak or differs across cell types? Would allowing type-specific slopes in the donor-to-cell bridge improve robustness?
  2. Trade-off between donor and cell prediction. In the mouse experiment, adding cell supervision improves cell prediction but reduces query donor correlation from 0.91 to 0.86. Is this pattern reproducible across training seeds and donor splits? How should the cell-supervision weight be selected when performance at both levels matters?
  3. Contribution of donor supervision. Removing donor supervision has little effect on median within-type cell correlation in the mouse experiment. Does removing it in the AD experiment materially reduce the association between the inferred cell state and the DAM score? This would help distinguish the contribution of donor-phenotype supervision from that of expression structure and the hierarchical architecture.
  4. Matched supervision in HD. How does scPACE compare with a predictor trained on scVI embeddings using the same measured CAG targets, analogous to the mouse comparison in Appendix G? Does its advantage persist when evaluated within cell types and individual query donors?
  5. Scope of the identifiability result. Given the theorem’s full-observation and independent-proxy assumptions, which identifiability conclusions, if any, remain valid in the donor-only, expression-derived-target, and partially observed settings used in the experiments?
4 comments
AuthorsOpenReview ↗

Detailed response to reviewer 7i31 - Q1 and Q2

Q1. Choice of cell-level target, robustness to alternative/weak/conflicting targets, and type-specific slopes.

A selection criterion follows from the model. From yn∣sn∼N(δC+aCsn,τC2)y_n | s_n \sim N(\delta_C + a_C s_n, τ_C^2) and sn∣td,cn=k∼N(td+mk,σk2)s_n | t_d, c_n=k \sim N(t_d + m_k, σ_k^2) (Eq. 1), the expected cell target in type k of donor d is E[yn∣td,k]=δC+aCmk+aCtdE[y_n | t_d, k] = \delta_C + a_C m_k + a_C t_d: linear in the donor state, with one slope aCa_C shared across types and type-specific intercepts aCa_C mkm_k. Since E[yd∣td]=δD+aDtdE[y_d | t_d] = \delta_D + a_D t_d, a compatible cell target is one whose per-type donor means move with the donor phenotype across donors at a common slope aC/aDa_C/a_D, with type-specific offsets. This is checkable before committing to a target: regress per-type donor-mean cell target on the donor phenotype and look for a shared slope with type-specific offsets. We will add this as a practical criterion.

Weak or conflicting targets. The behavior is predictable from the two gradients acting on sns_n: the prior pulls sns_n toward td+mkt_d + m_k, and the cell channel pulls it toward the affine readout of yny_n. When the two targets disagree, the gradients compete, and raising λcell\lambda_{cell} moves sns_n toward the cell target and away from the donor-predictive part of its variation. This is exactly the trade visible in the mouse λcell\lambda_{cell} sweep (Appendix Table 10, Figure 8).

Type-specific slopes. The unit slope is a modeling choice, and the model already absorbs heterogeneous coupling strength through σ_k²: a type whose target responds weakly to the donor state takes a larger σk2\sigma_k^2 and is coupled more loosely, without changing the common scale of sns_n. That common scale is what makes states comparable across types, which the cross-type readings in Figure 2c and Figure 3c (SPNs highest among HD classes) depend on. Type-specific slopes would make one unit of sns_n mean a different amount in each type and remove that comparability; they would also break the identifiability argument, which uses the shared unit slope to rule out reflecting S without T (Appendix A.5, Step 4).

Q2. Is the donor/cell trade-off reproducible across seeds and splits? How should λ_cell be chosen?

Direction is expected from the objective. One scalar sns_n per cell serves both the cell readout and the donor prior, and within a type the maturity score is not an exact function of age, so increasing λcell\lambda_{cell} necessarily trades donor-predictive variance for target fit.

New seeds (Appendix Table 7). We trained Settings 1 and 2 at two further seeds (seeds 1 and 2), same data, split, and hyperparameters as the paper (seed 0). Over the three seeds (mean±SDmean \pm SD, min–max), query donor ∣r∣|r| is 0.948±0.0380.948 \pm 0.038 (0.9050.905–0.9740.974) for Setting 1 vs 0.843±0.0230.843 \pm 0.023 (0.8200.820–0.8650.865) for Setting 2, and donor RMSE is 0.305±0.0850.305 \pm 0.085 (0.2500.250–0.4030.403) vs 0.543±0.0560.543 \pm 0.056 (0.5030.503–0.6070.607). Query cell ∣r∣|r| is 0.087±0.0520.087 \pm 0.052 (0.0410.041–0.1440.144) for Setting 1 vs 0.972±0.0010.972 \pm 0.001 (0.9710.971–0.9720.972) for Setting 2. Setting 2 has lower donor ∣r∣|r| than Setting 1 at every seed and the two ranges do not overlap, so the trade-off is reproducible across seeds. (The donor split is held fixed across runs by design, since changing it changes the query set; the seed varies initialization and sampling.) We note the seed-0 run reported in Table 1 (donor ∣r∣=0.905|r|=0.905 for Setting 1) sits at the bottom of the three-seed range, so the paper’s headline donor number is conservative.

Choosing λcell\lambda_{cell}. We do not tune λ\lambda's on the real cohorts. λdonor\lambda_{donor} and λcell\lambda_{cell} are selected once on the simulation by recovery of the known states (Appendix C), landing on the plateau where cell recovery saturates (λcell=500\lambda_{cell}=500) and donor recovery is near its peak (λdonor=10\lambda_{donor}=10). When donor-level performance matters more, the sweep is the guidance (Appendix Table 10, Figure 8): λcell=10\lambda_{cell}=10 keeps donor ∣r∣|r| at 0.970.97 while raising pooled cell ∣r∣|r| from 0.080.08 to 0.910.91, i.e. a small cell weight anchors the type positions at almost no donor cost.

AuthorsOpenReview ↗

Detailed response to reviewer 7i31 - Q3 and Q4

Q3. Does removing donor supervision reduce the cell-state/DAM association in AD, as it barely does in mouse?

New ablation (Table 1b, Appendix G). We removed the donor phenotype head (λdonor=0\lambda_{donor}=0) and retrained AD Setting 3 unchanged (seed 0, same split). The association between the inferred cell state and the DAM score on the 6,417 query microglia drops from ∣r∣=0.548|r|=0.548 to 0.1100.110 (nRMSE 0.8410.841 vs 0.9980.998, where 1 is no better than the query mean), below CNA (0.250.25) and Sceptic (0.190.19). In AD, therefore, the cell-state/DAM association comes from donor-phenotype supervision passed through the hierarchy. This differs from mouse, where expression structure alone already orders cells within each type (median within-type ∣r∣|r| 0.450.45 with vs 0.420.42 without donor supervision; Appendix Table 8), and the difference reflects the biology. Mouse visual-cortex development is an easier task: cells mature from progenitors into neurons and glia with large, coordinated expression changes across many genes, and the donors are inbred laboratory mice in a controlled environment, so most donor-to-donor variation is age. Maturation is thus a strong signal in the counts, and the count likelihood alone orders cells within type. AD is more complex: human donors differ in age, sex, genetic risk (e.g. APOE), co-occurring pathologies, and postmortem quality, and the DAM response (Keren-Shaul et al., 2017) is a graded within-type change that is only one of many sources of microglial variation. The model needs the donor phenotype to pick the disease axis out of the rest, which is why removing it collapses the association to 0.11.

Q4. Matched supervision in HD (predictor on scVI embeddings with the measured CAG target), within types and per donor.

We reproduced the Appendix G protocol on HD: ridge and an MLP on the frozen 10-dimensional scVI embedding, trained to predict CAG length on the measured training nuclei, same HD donor split and calibration, evaluated on the 11,047 CAG-measured query nuclei (Appendix Table 12).

MethodPooled |r|nRMSEDemeaned |r|Median |r|
scPACE0.2500.9690.1180.047
MLP on scVI0.2790.9620.1580.044
Ridge on scVI0.2310.9740.0390.041

On HD the MLP is competitive with scPACE, because the CAG signal is concentrated in the spiny projection neurons and is weak in the smaller types. We do not claim a within-type advantage for scPACE on this cohort. We note the comparison rests on three query donors with small per-type counts, so these point estimates are noisy. However, on the mouse comparison, with 20 query donors, is where the within-type structure can be read with more confidence, and there scPACE’s representation is better organized within type than the frozen scVI embedding (median rkNNr_{kNN} 0.7990.799 vs 0.7150.715, Appendix Figures 9–10; median within-type ∣r∣|r| 0.8180.818 vs 0.7340.734 for the MLP, Appendix Table 11).

The reason we report HD at all despite this is that it answers a different question from "which method predicts CAG best." The MLP is trained for one task, the prediction of CAG length. scPACE is designed to solve four tasks with one model: donor-state inference, cell-state inference, representation learning, and cell-type imputation. In scPACE the counts, the cell types, the CAP score, and the measured CAG lengths enter one objective, and all four sources of information shape the same cell representation znz_n and the same cell state sns_n. With that one representation, scPACE also predicts the CAP score of query donors with lower error than scVIP (RMSE 1.03 vs 1.29), imputes cell types at accuracy 0.99, and mixes donors within cell types more than scVI (0.66 vs 0.33, Table 1c). A predictor trained on top of a frozen scVI embedding gives only the CAG prediction, and it can only read out the structure the unsupervised embedding already contains. However, it cannot feed the cell-level signal back to the donor channel, the composition model, or the representation. That gradients from every channel reshape one shared latent space is the same design principle behind semi-supervised joint models such as scANVI (Xu et al., 2021), where propagating label supervision into the representation, rather than classifying a fixed embedding, is what lets annotation transfer across batches and tolerate missing labels. The matched predictor is a fair baseline for the CAG task specifically, and on that narrow task it performs comparably, but what it does not provide is the jointly structured representation that the other three tasks draw on.

AuthorsOpenReview ↗

Detailed response to reviewer 7i31 - Q5 and ethics flag

Q5. Scope of the identifiability result across the donor-only, expression-derived, and partially observed settings.

We have added this analysis to Appendix A.6. The theorem factors into a donor-level part (Steps 0–1) and a cell-level part (Steps 2–3) with different requirements, and the three regimes keep different parts.

Donor-only (Settings 1, 3). Each cell contributes only its type and counts, Ob={Cn,Xn}O_b=\{C_n, X_n\}. Steps 0–1 and the composition half of Step 2 hold, so three things are identified up to the common sign: the donor state tdt_d, the composition parameters α\alpha, β\beta, and the law of the counts given (T,C,b)(T, C, b). What is not identified is the split of that count law into bridge and kernel: a per-type shift of the cell states (with mkm_k and the kernel moved to compensate) and a transfer of Gaussian variance from bridge to kernel both leave the count law unchanged.

Under these, within-type, within-donor ordering is invariant while the relative positions of types are free. This is why Settings 1 and 3 are evaluated within type (Eqs. 60–61) and why the Setting 1 pooled correlation is not an identified quantity. In Setting 1 scPACE has the highest median within-type ∣r∣|r| (0.450.45 vs 0.310.31–0.400.40 for the six scalar baselines). In Setting 3 the cell metrics use only microglia, a single type, so the Table 1b pooled ∣r∣|r| is already a within-type ∣r∣|r|, and scPACE reaches 0.55 vs at most 0.25.

Expression-derived target (Setting 2). The maturity score is a function of the counts, so the noise is computed from the counts and YC⊥X∣SY^C \perp X | S fails. Lemma 1 does not apply to the cell observational model, and we do not claim the cell-level conclusions of Theorem 1 for Setting 2. The donor-level conclusions are unaffected, and the within-type ordering of the donor-only regime carries over. Here the cell observational model fixes the relative type positions and the scale of sns_n to agree with the score. This is a reporting convention that makes the pooled metric and the score-trained-baseline comparison meaningful, not identification of an independent quantity, which is why Setting 2 is labeled auxiliary supervision.

Partially observed target (Setting 4). The theorem treats YCY^C as observed for every cell; CAG is measured in 29,030 of 302,191 nuclei. Steps 2–3 run on the measured nuclei provided the measurement indicator is independent of (S,YC,X)(S, Y^C, X) given type and batch (missing at random given the observed design). Under that assumption every conclusion holds for all seven classes, each of which has measured nuclei (Appendix Figure 5b), with cell types and the donor channel fully observed (unmeasured nuclei still contribute through the count likelihood). Setting 4 is thus the one regime where the full theorem applies, and CAG length is the one cell-level target in the paper that satisfies the independent-proxy assumption.

Ethics flag

The items listed under "ethics concerns" are a copy of the technical questions above and, respectfully, they are not ethics issues. We have addressed them as scientific questions.

Thank you for the detailed responses and additional analyses. Most of my concerns have been addressed. However, I remain concerned about donor-level generalisation, which is highlighted as one of the key advantages of the framework (as stated in the final sentence of the Introduction). While scPACE performs well on mouse data, its performance is considerably weaker in AD (Table 1). Given the greater heterogeneity of human patient cohorts, this raises questions about its generalisability to complex disease settings. I therefore consider this an important unresolved limitation and would prefer to maintain my original rating.

Reviewer_AkcGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To address the challenge of multi‑level incomplete supervision in grouped single‑cell data, this paper proposes scPACE, a hierarchical VAE. Using a type‑shifted Gaussian bridge, its scalar donor state generates donor phenotype, cell‑type composition and per‑cell scalar states; cell states, types and an auxiliary variable produce embeddings and NB counts. Phenotypes act only as Gaussian observation channels instead of inference inputs for prediction under missing targets, and its scalar hierarchy is identifiable up to a sign given full observation and independent phenotype proxies. Evaluated via held‑out‑donor protocol on over 1.3 million cells from mouse cortical development, Alzheimer’s and Huntington’s cohorts, scPACE is compared against scVI, scANVI, scVIP and six scalar trajectory methods.

Critical strengths

1.Well-motivated problem. Joint modeling of donor- and cell-level supervision with missingness addresses real pseudoreplication issues.

2.It carries out thorough ablations covering hierarchy edges aggregator weights and schedules. It provides donor‑cluster bootstrap intervals and runs a simulation study for weight selection.

Critical weaknesses
  1. The strong cell‑level results come from supervision signals instead of model design. Under donor‑only supervision, scPACE shows poor pooled‑cell correlation and falls behind all baseline methods. Even a basic MLP built on fixed scVI embeddings can achieve comparable performance when given the same cell‑level supervision.

  2. Donor‑level results are not stable across datasets. scPACE performs much worse than scVIP on the Alzheimer’s disease dataset. The Huntington’s disease test only has three query donors and leaves out donor correlation metrics due to instability. This weakens its main argument about good donor‑state inference.

  3. There exists circularity between supervision and evaluation. For two out of three study cohorts, cell targets and donor phenotype scores are calculated from the transcriptomic data being analyzed. This goes against the idea of using independent weak supervision to constrain latent states.

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Detailed response to reviewer AkcG - weakness 1

W1. "The strong cell-level results come from supervision signals instead of model design. Under donor-only supervision, scPACE shows poor pooled-cell correlation and falls behind all baseline methods. Even a basic MLP built on fixed scVI embeddings can achieve comparable performance when given the same cell-level supervision."

We separate two claims: that the pooled Setting 1 correlation is low, and that the model adds nothing over a matched predictor. The first is a property of the metric. The second is contradicted by our ablations and by the matched comparison once it is evaluated within cell type.

Why the pooled Setting 1 correlation is low. Cell type cnc_n alone explains 86.3%86.3\% of the maturity-score variance among annotated query cells, because progenitor types sit at low maturity and mature neuronal types at high maturity. The pooled correlation is therefore dominated by the relative positions of types along the state axis. Those positions are set by the intercepts mkm_k, which the counts alone do not identify: the conditional prior pθ(zn∣un,cn,sn)p_\theta (z_n | u_n, c_n, s_n) receives onehot(cn)onehot(c_n) and can absorb any type-specific shift in sns_n without changing the count likelihood (the exact symmetry in Appendix A.6). Without a cell-level target to anchor the mkm_k, the pooled number collapses even when within-type ordering is good. Within each type, scPACE in Setting 1 attains the highest median correlation of any donor-only method (median ∣r∣=0.45|r|=0.45 vs 0.400.40 for PhenoPath and 0.310.31–0.400.40 for the other five scalar baselines, Appendix Table 9).

The hierarchy, not just the supervision, organizes cell states (design). Our ablation removing both generative donor-to-cell edges (no bridge, β=0\beta=0) reduces the median within-type correlation from 0.450.45 to 0.180.18 while leaving donor correlation unchanged at 0.910.91 (Appendix Table 8, Figure 7). This isolates a structural contribution of the generative hierarchy to cell-state organization that is separate from donor-level prediction. At the level of the learned representation, we have added a per-type comparison of scVI and scPACE (Appendix Figures 9–10). For each query cell we take its 15 nearest neighbours in a UMAP of the method’s 10-dimensional latent space, average their maturity scores, and correlate each cell’s own score with that neighbour average, within each type (rkNNr_{kNN}: 1 means neighbours share the same maturity, 0 means the representation does not organize maturation within the type). scPACE’s representation is ahead in 10 of the 12 types with at least 100 query cells, with median rkNNr_{kNN} of 0.7990.799 vs 0.7150.715 for scVI, a difference of 0.0840.084 (95%95\% donor-bootstrap interval [0.052,0.162][0.052, 0.162]). The margin is about 0.010.01 in the two large neuronal classes where scVI is already strong (IT Glut, 46,513 cells, 0.7850.785 vs 0.7730.773; nonIT Glut, 20,842 cells, 0.7380.738 vs 0.7320.732) and 0.030.03 to 0.190.19 in the eight smaller classes (e.g. Astro-Epen 0.8020.802 vs 0.6260.626, Vascular 0.7610.761 vs 0.5680.568, Immune 0.8120.812 vs 0.6560.656). An MLP can map a frozen scVI embedding to the score, but the embedding itself orders maturation less well within a type, and the embedding, not the score, is what downstream analysis uses.

The supervision is used well (new AD ablation). In AD Setting 3 no cell-level phenotype enters the model, yet the inferred cell state is associated with the held-out DAM score at ∣r∣=0.548|r|=0.548 on the 6,417 query microglia, versus at most 0.250.25 for the six scalar baselines that receive the same donor labels (Table 1b). When we remove the donor phenotype head (λdonor=0\lambda_{donor}=0) and retrain with everything else unchanged, the association drops to ∣r∣=0.110|r|=0.110 (Appendix G). The donor supervision is therefore load-bearing, and the hierarchy is what transmits it from the donor state tdt_d to the cell states sns_n to produce an association no baseline with the same labels reaches.

AuthorsOpenReview ↗

Detailed response to reviewer AkcG - weakness 2 and 3

W2. "Donor-level results are not stable across datasets. scPACE performs much worse than scVIP on the Alzheimer’s disease dataset. The Huntington’s disease test only has three query donors and leaves out donor correlation metrics."

HD. We agree the donor-level comparison rests on three query donors, which is why donor ∣r∣|r| is omitted and donor RMSE carries the comparison (Table 1c, footnote). HD is included for the one feature no other cohort offers: an independently measured, non-expression-derived cell-level target (somatic HTT CAG repeat length by PacBio), which is the setting our identifiability result actually covers (Appendix A.6). We have made this scope explicit where HD is introduced.

AD. scVIP does predict the donor phenotype better in this cohort (query ∣r∣=0.58|r|=0.58 vs 0.130.13), and this follows from what each model is. scVIP is a discriminative multiple-instance predictor whose only output is the donor phenotype, pooling the full 10-dimensional cell embedding per type. scPACE predicts the donor phenotype through a single scalar tdt_d that must simultaneously generate cell-type composition, the per-cell states, and the counts, and it additionally produces cell states, cell-type imputation, and a mixed representation that scVIP does not. On the same cohort scPACE gives the strongest query cell-state association (0.5480.548), the lowest cell nRMSE, and the highest donor mixing (Table 1b), none of which scVIP produces. We think the trade is worth stating plainly: on the single task scVIP is built for, sometimes a specialized discriminative predictor can beat a unified generative model, and that is the cost of the unified design, with the gain of a structure latent space that carries information across tasks.

W3. "Circularity between supervision and evaluation: for two of three cohorts, cell targets and donor phenotype scores are calculated from the transcriptomic data."

We would like to correct a factual point first, because it is central to the rating. No donor phenotype in any of the three cohorts is derived from the transcriptome. Mouse donor supervision is chronological age, a calendar measurement. AD donor supervision is the continuous pseudoprogression score of Gabitto et al. (2024), which orders donors by quantitative neuropathology, not by RNA. HD donor supervision is the CAP score, a function of inherited CAG length and age. The only expression-derived quantities in the paper are the two cell-level evaluation scores (mouse maturity, AD DAM), and they enter a model only in mouse Setting 2.

Taking the settings in turn: mouse Setting 1 and AD Setting 3 never see the cell-level score during fitting. It is held out and used only for evaluation, so there is no circularity. HD Setting 4 uses an independently measured cell target. That leaves mouse Setting 2 as the only place an expression-derived score is used for supervision, and the manuscript flags exactly this (Section 4, lines 252–256): the scores "are not independent measurements of the underlying biology" and are "used as evaluation targets"; the expression-derived score is excluded from the identifiability claim (Appendix A.6); and Setting 2 is labeled an auxiliary-supervision setting.

Even in Setting 2, evaluating sns_n against the score is informative because the model is never given gene identities or a marker list. It sees 21,847 genes (mouse) with no indication of which mark maturation: to align sns_n with the score it must separate the maturation axis from cell cycle, library size, dataset, and batch. The paper also contains a validation that uses no expression score at all (Figure 2): at fixed chronological age (P28), dark rearing lowers the inferred L2/3 state from 0.770.77 to 0.480.48 and light re-exposure restores it to 0.730.73.

New control. To remove any residual doubt for Setting 2, we dropped the 22 target-defining genes (10 mature, 12 progenitor markers) from the model input and retrained Setting 2 unchanged (seed 0, 21,825 input genes; Appendix G). Query cell ∣r∣|r| changes only from 0.9720.972 to 0.9630.963 and cell nRMSE from 0.2330.233 to 0.2680.268. scPACE therefore predicts the maturity score from the rest of the transcriptome and does not rely on reading back the 22 marker genes that define the target.

Thank you for the detailed response.

It has addressed some of my concerns (W3). However I find the additional evidence for W1 and W2 still not fully convincing and some clarifications are still needed. I will maintain my score.

In addition the revised version exceeds the 10 page limit and does not comply with ICLR regulations (lines 540-543). Please verify this yourself.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common response to all reviewers

We thank all three reviewers for the detailed and constructive reviews, we paid attention to all their feedback, this has certainly made the manuscript better. During the discussion period we have revised the manuscript (changes marked red in the PDF) to add every experiment requested across the reviews, and we point to the specific new tables and figures in the per-reviewer replies below. The additions are:

  1. Three-seed runs for the mouse cohort, showing the donor/cell trade-off is reproducible across training seeds (Appendix Table 7).
  2. A marker-gene dropout control for Setting 2: removing the 22 target-defining genes from the input changes query cell ∣r∣|r| only from 0.972 to 0.963 (Appendix G), so the recovered state does not depend on reading back the genes that define the target.
  3. A donor-supervision ablation on the Alzheimer’s cohort (λdonor\lambda_{donor}=0): the cell-state/DAM association drops from 0.548 to 0.110 (Table 1b, Appendix G), isolating the contribution of donor supervision.
  4. Matched scVI-embedding baselines (ridge and MLP) for Huntington’s disease on the independently measured CAG target (Appendix Table 12).
  5. A per-type comparison of how each representation organizes maturity (Appendix Figures 9–10), with scPACE ahead in 10 of 12 types (median rkNNr_{kNN} 0.799 vs 0.715; bootstrap interval [0.052, 0.162]).
  6. An expanded scope analysis of the identifiability result by setting (Appendix A.6).

One factual correction we would like to make up front, as it recurs across the reviews: no donor phenotype in any of the three cohorts is derived from the transcriptome (chronological age, the SEAAD pseudoprogression score from quantitative neuropathology, the CAP score from CAG length and age). Only the two cell-level evaluation scores are expression-derived, and an expression-derived score is used for training in exactly one of the four settings (mouse Setting 2), where the manuscript already labels it auxiliary supervision and excludes it from the identifiability claim.

All reported figures are the current manuscript values.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.