acceptodds
Under review as a conference paper at ICLR 2027

CoState: Interpretable Donor Representation Learning from Single-Cell Transcriptomes

16 views

Abstract

Single-cell RNA sequencing offers new opportunities for individual-level disease classification. Disease-associated differences between donors can reflect changes in cell-type composition, within-type expression states, or both. However, existing approaches that average cell features to represent donors do not distinguish between these two sources of variation. We introduce CoState (Composition and State), which jointly learns cell and donor representations by combining observed cell-type proportions with learned within-type state distributions into a joint probability distribution. Using Hellinger geometry, we map this distribution to a donor vector, enabling interpretable comparisons that quantify each cell type's contribution to differences between donors. On a benchmark comprising 1,007 donors from 23 Inflammation Atlas studies spanning 19 diseases and healthy controls, CoState outperforms the evaluated baselines in cross-study disease retrieval, achieving an approximately 16% relative improvement in retrieval macro-F1 over the strongest baseline in independent-study evaluation. Compared with baseline methods in the training-cell scaling experiment, CoState benefits more from increasing the number of training cells, achieving greater performance gains under a fixed training schedule. Code is available at https://anonymous.4open.science/r/costate-0D40.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single Cell, Contrastive Learning, Representation Learning, AI for Biology, AI for Science

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_2MkJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The others present a method that jointly learns cell and patient representations and offers insights into learning patient representations through downstream applications and interpretable analysis. However, I believe the method does not draw a clear conclusion, and the interpretability claims seem overstated.

Critical strengths

The others present a method which can jointly learning the representations of cells and patients, and offer insights in learning the patient representation through downstream applications and interpretable analysis.

Critical weaknesses

The others present a method which can jointly learning the representations of cells and patients, and offer insights in learning the patient representation through downstream applications and interpretable analysis. However, I believe the method does not present a clear conclusion and the interpretability seems overstated. Here are my questions:

  1. I think the learning of cells and patients is not a novel idea as it is very natural to learn the cell information before the patient level, The authors should discuss more in the introduction section to justify their novelty.
  2. The presentation method is also limited. If the data do not have cell type labels, we cannot use this method based on the visualization. Many singles of data do not have precise cell type labels The authors should consider this issue.
  3. The benchmark study is also limited and the authors ignore many related base lines such as MultiMIL and scDiffusion in the patient level. The authors should include them in the benchmark, and justify that they make a fair comparison by tuning other methods to their best performance.
  4. Did the authors split training/testing data through study or batch? If they split through batch, there could be data leakage.
  5. I cannot see the contributions of interpretability, at least the authors need to show how they can use it to interpret the prediction results.
  6. The testing of recent virtual cell foundation models, such as X-atlas and STATE, should considered.
Questions for authors

Please see my comments above.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 2MkJ (1/3)

Thank you for the careful review. We have revised the manuscript in response to your comments. In the revised version, we clarify how CoState uses distance decomposition to explain disease predictions, report results for the requested baselines, and include a worked example showing why a donor was matched to one disease reference.


1 — Novelty

Prior methods such as MrVI, MultiMIL, ProtoCell4P, and PaSCient already connect cell-level information to donor-level analyses. CoState makes this connection explicit by representing each donor as a joint distribution over cell types and learned cell states. This representation supports disease prediction and lets us explain the distances underlying kNN and prototype predictions (§1, lines 59–70, and §2, lines 128–135).

Interpretable donor comparisons. Proposition 1 decomposes the squared Hellinger distance exactly into a term for differences in cell-type proportions and one term per cell type for differences in its state distribution. Squared Hellinger is symmetric and bounded, remains finite when a cell type is absent from one donor, and equals half the squared Euclidean distance between the square-root probability vectors. Across the three cohorts, within-type state differences accounted for 74–78% of the distance between donors with different diseases from different studies (§3.3, lines 195–214, and Appendix G.4, lines 1989–1995 and Figure 9).

We kept the learned donor distributions fixed and varied the distance used for kNN disease prediction (Appendix H.3, Table 16).

Distance on the same donor distributionsTest kNN macro-F1Val. kNN macro-F1Ext. kNN macro-F1
Squared Hellinger0.8910.8600.840
Total variation0.8710.8550.786
Euclidean on the joint probability vector0.8330.8550.746
Cosine on the joint probability vector0.8400.8480.750

Cell-level representations. CoState scored highest on the five cell-type structure metrics shown below. The comparison includes joint-label scANVI and ProtoCell4P, both of which use disease and cell-type labels. Removing CoState’s donor-level losses lowered the mean score on all five metrics across three seeds (§4.3, lines 380–383 and Table 2, and §4.5, lines 434–435 and Table 15).

Cell embedding on Test donorsLeiden NMILeiden ARIk-means NMIk-means ARIkNN macro-F1
CoState0.7960.7620.7030.5200.846
CoState without donor-level losses0.7510.5980.6510.4190.833
scANVI0.6230.3680.5670.3150.832
scANVI, joint disease × cell type0.5580.2320.4890.2650.828
ProtoCell4P0.6230.5260.4210.2570.584
MrVI (transductive)0.7620.6990.6900.5130.816

Donor-level performance. Among the configurations shown, CoState had the highest point estimate on 10 of 12 readouts across disease prediction (Table 1) and cross-study retrieval (Figure 2 and Tables 8 and 12). Both report disease macro-F1. In cross-study retrieval, each query is matched to its nearest donor from a different study (§4.2, lines 326–331).

MethodTest (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)Cross-study retrieval (Test / Val. / Ext.)
CoState0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.8230.969 / 0.984 / 0.639
MultiMIL0.923 / 0.860 / 0.8440.676 / 0.704 / 0.6860.760 / 0.744 / 0.6770.447 / 0.902 / 0.500
scANVI, joint disease × cell type0.735 / 0.741 / 0.6560.690 / 0.636 / 0.5700.866 / 0.771 / 0.7620.591 / 0.531 / 0.485
ProtoCell4P (4 prototypes)0.829 / 0.862 / 0.6600.715 / 0.689 / 0.6490.788 / 0.758 / 0.6290.706 / 0.500 / 0.595
PaSCient0.825 / 0.855 / 0.6620.706 / 0.737 / 0.5630.762 / 0.739 / 0.5740.368 / 0.796 / 0.421
MrVI (transductive)0.316 / 0.257 / 0.1810.240 / 0.184 / 0.1390.488 / 0.399 / 0.3010.283 / 0.166 / 0.336

We clarified how the donor representation connects prediction and attribution in the revised introduction (§1, lines 59–66) and related work (§2, lines 128–135).

AuthorsOpenReview ↗

Response to Reviewer 2MkJ (2/3)

2 — Dependence on cell-type labels.

CoState can construct and visualize donor profiles for new samples that lack manual cell-type annotations. We infer cell types with a 15-nearest-neighbour classifier on CoState’s frozen cell embeddings, using annotated cells from training donors as references. The inferred labels define the cell-type groups in the visualization and are used to build the donor profile for disease prediction. On the Test split, the inferred cell types matched the manual annotations with 91.8% accuracy. We describe this option in §3.1, lines 160–161 and the label-transfer procedure in Appendix E.1, lines 1160–1165.

To assess the effect on donor-level disease prediction, we rebuilt the donor profiles with inferred cell types using the same trained CoState checkpoints and repeated the evaluations in Table 1. The table reports mean disease prediction macro-F1 across three seeds (Appendix E.1, Table 4).

Cell-type labels used to build donor profilesTest (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)
Atlas labels (manual for Test, scANVI label transfer for Val. and Ext.)0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
Labels inferred from CoState embeddings0.884 / 0.910 / 0.8420.857 / 0.884 / 0.8640.832 / 0.891 / 0.825

Using inferred cell types changed mean donor-level disease prediction macro-F1 by at most about 0.010 across all nine comparisons.


3 — MultiMIL, scDiffusion and fair tuning.

We added MultiMIL and scDiffusion to the donor-level disease prediction benchmark (Table 1). Bold indicates the best score for each readout among the methods shown.

MethodTest (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)
CoState0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
MultiMIL0.923 / 0.860 / 0.8440.676 / 0.704 / 0.6860.760 / 0.744 / 0.677
scDiffusion0.695 / 0.709 / 0.6170.652 / 0.505 / 0.4960.838 / 0.769 / 0.666

We used the official MultiMIL recipe with scVI latents and selected its checkpoint by validation accuracy. For scDiffusion, we used the upstream autoencoder defaults and mean-pooled cell latents to obtain donor embeddings. For ProtoCell4P, we chose the number of prototypes on a 15% holdout of training donors using its validation criterion, donor AUC plus one tenth of cell-type accuracy. We describe these protocols in Appendix E.4, lines 1439–1450.


4 — Data split.

We held out donors for Test and Validation, and entire studies for External evaluation. Test and Validation donors did not enter CoState training and came from studies represented in training. All seven External studies were held out from CoState training. We split the main cohort by donor and sequencing-library component, keeping each donor’s cells in a single split. We held out 10% of cells within the training donors for checkpoint selection under a fixed training budget. We describe the splits in Appendix E.1, lines 1147–1154 and checkpoint selection in Appendix E.4, lines 1564–1565. Cross-study retrieval excludes all candidates from the query donor’s own study (§4.2, lines 326–331).

AuthorsOpenReview ↗

Response to Reviewer 2MkJ (3/3)

5 — Interpretability.

We added a practical example in §4.5, lines 471–477 to show how CoState explains a prediction. We compared one held-out donor from COMBAT [1] with two reference profiles, built by averaging the profiles of 24 other COVID-19 donors and eight donors with non-COVID sepsis from the same study. Each profile describes how common each cell type is and which gene-expression patterns its cells show. The donor’s profile was closer to the COVID-19 reference (squared Hellinger distance 0.08325 vs. 0.26367 for the sepsis reference). Within-type monocyte states accounted for 75.2% of the distance margin favoring the COVID-19 reference (Appendix H.4, lines 2335–2340).

Of the 250 cells sampled from this donor, 67 were monocytes (26.8%). Monocytes made up 31.6% of the COVID-19 reference and 38.8% of the sepsis reference. CoState assigns each monocyte weights across 16 learned expression patterns. One pattern accounted for 85.1% of the donor’s average monocyte assignment weight, compared with 82.5% in the COVID-19 reference and 10.0% in the sepsis reference (Appendix H.4, lines 2341–2344).

CoState’s monocyte state analysis highlighted IFI27 in the comparison between COVID-19 and sepsis (q = 0.023 after correction across 2,000 genes, see §4.5, lines 473–475 and Appendix H.4, lines 2353–2358). Saichi et al. independently reported elevated IFI27 RNA in CD14⁺ monocytes from patients with moderate COVID-19 [2]. This example shows how CoState can trace a donor match to a cell state and highlight genes that help interpret that state.

Reference

  1. COMBAT Consortium. A blood atlas of COVID-19 defines hallmarks of disease severity and specificity. Cell, 2022. DOI.
  2. Saichi M et al. Single-cell RNA sequencing of blood antigen-presenting cells in severe COVID-19 reveals multi-process defects in antiviral immunity. Nature Cell Biology, 2021. DOI.

6 — X-Atlas and STATE

We agree that STATE is a relevant addition to our benchmark alongside UCE, scGPT, Geneformer, and Nicheformer. We averaged frozen STATE (SE-600M) cell embeddings within each donor view and evaluated the resulting donor embeddings on disease prediction (Table 1) and cross-study retrieval (Table 12). The revised manuscript describes the protocol in Appendix E.4, lines 1546–1548.

MethodTest (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)Cross-study retrieval (Test / Val. / Ext.)
CoState0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.8230.969 / 0.984 / 0.639
STATE (SE-600M)0.725 / 0.703 / 0.7160.659 / 0.561 / 0.5390.807 / 0.783 / 0.7560.234 / 0.310 / 0.368

We also evaluated the individual STATE cell embeddings on held-out Test donors and added these results to Table 2.

MethodLeiden NMILeiden ARIk-means NMIk-means ARIkNN macro-F1Type firstTransfer F1
CoState0.7960.7620.7030.5200.8460.8830.901
STATE (SE-600M)0.4900.2390.4440.2380.7310.2540.723

X-Atlas/Orion is a Perturb-seq dataset for studying gene perturbations in HCT116 and HEK293T cell lines. Its labels describe perturbations, while our benchmark requires disease labels for individual donors. X-Atlas/Orion dataset card

Reviewer_M3ZfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CoState learns donor-level representations from single-cell transcriptomes by combining observed cell-type proportions with learned within-type state distributions. A square-root transformation gives a Hellinger geometry in which donor distances decompose exactly into composition and cell-type-specific state contributions. The model combines cell-level reconstruction and cell-type objectives with disease-supervised donor learning and same-donor consistency. Experiments on the Inflammation Atlas compare the method with a broad set of baselines for disease prediction, cross-study retrieval, and cell-level representation quality.

Critical strengths
  1. The distinction between changes in cell-type composition and changes within a cell type is biologically meaningful. The exact, non-negative decomposition provides a clear way to attribute donor differences to these two sources.
  2. The evaluation covers a broad range of baselines. CoState consistently outperforms the disease-supervised comparators, with a particularly substantial advantage on the validation cohort.
  3. I want to record that the internal-test exposure during gene selection, the absence of guaranteed biological disentanglement, and the label-assisted nature of the clustering evaluation were all stated by the authors themselves. This is better practice than is typical, and the criticisms below rest in part on those disclosures. My concern is that several of them are load-bearing for the paper's claims and are not addressed beyond disclosure.
Critical weaknesses
  1. The experiments do not adequately isolate the proposed contribution. CoState uses disease labels during representation learning, whereas most baselines do not. These comparisons demonstrate predictive performance but do not establish how much improvement comes from the proposed geometry rather than supervision. The existing "without donor losses" ablation also replaces the donor representation with mean cell embeddings, changing both the objective and aggregation. Supervision-matched comparisons and ablations that preserve the representation while changing one objective at a time are needed. A simpler cell encoder with the same donor head would also help establish the value of the architectural complexity. The external results make this concern concrete: Table 1 reports 0.840 kNN macro-F1 for CoState against 0.834 for frozen UCE and 0.834 for pseudobulk + PCA, neither of which uses disease labels, and the former of which does not use the gene panel at all. Given the disease–study confounding in this cohort, this narrow margin does not establish that the architecture or supervision is ineffective, but it limits the strength of the performance advantage demonstrated in this setting.

  2. The evaluation protocol limits the generalization claims. Appendix E.1 states that the disease-informed gene panel was selected before donor splitting and included internal test donors. Although transparently disclosed, this prevents the internal test results from constituting a fully held-out evaluation. It also distributes the exposure asymmetrically across the comparison, since panel-based methods share this panel while the frozen foundation models use full-transcriptome inputs. In the external cohort, six of seven studies contain only one disease label, making disease prediction difficult to separate from study effects. Excluding same-study retrieval candidates is useful, but does not by itself eliminate all study-related confounding; the resulting external evaluation contains only 19 donors and two classes. Broader cross-study, multi-disease generalization therefore remains insufficiently established.

  3. The headline result needs clearer context and uncertainty reporting. The approximately 16% relative retrieval improvement highlighted in the abstract corresponds to the 19-donor healthy-versus-SLE evaluation, rather than the full 19-disease benchmark. The sample size and class count should accompany this claim. The paper should also report the actual scored classes and per-class donor counts for each readout. Tables 1 and 2 report point estimates only, and uncertainty estimates for the main comparisons would be particularly relevant where performance differences are small.

  4. The biological interpretation of the learned states is not sufficiently validated. The distance decomposition is mathematically interpretable, but this does not establish that its learned state coordinates correspond to distinct biological states. The state head and fusion layer are directly optimized only by the disease-contrastive and same-donor objectives, although their upstream features are also shaped by reconstruction and cell-type learning. Consequently, the state codes are encouraged to capture disease-discriminative variation, but are not explicitly constrained to represent distinct biological states. The interferon-associated annotations are biologically plausible, but do not independently validate the learned states, given the use of disease supervision and expression-based characterization. The claims should distinguish attribution of disease-discriminative variation from identification of biological states. Sensitivity to the number of state codes, cell-type annotation quality, and limited cell counts in rare populations would further strengthen this interpretation.

Minor. The paper does not provide a consolidated training configuration. Please report the loss weights, temperature, donor batch size, optimizer, learning rate, training steps, parameter count, and computational cost.

Questions for authors
  1. Can the authors provide supervision-matched baselines and remove disease supervision while retaining the same Hellinger donor representation? How does a simple MLP encoder perform with the proposed donor head?
  2. As an initial sensitivity check, how much does a gene panel selected using training donors only overlap with the current panel? This would help characterize the effect of test-donor exposure, although overlap alone would not establish that the reported performance is unaffected.
  3. Given that six of seven external studies carry a single disease label, can the authors scope the cross-study claims to what this cohort supports, and report the external results alongside an explicit statement of the disease–study confound?
  4. Can the authors report scored class counts, uncertainty estimates, and sensitivity to the number of state codes and per-type cell availability?
5 comments
AuthorsOpenReview ↗

Response to Reviewer M3Zf (1/3)

Thank you for the careful review. We have revised the manuscript in response to your comments. In the revised version, we compare CoState with baselines that also use both disease and cell-type labels during training. We add ablations that retain the Hellinger donor head, report results for a gene panel selected solely on training donors, and clarify the scope of the external evaluation. We also report the scored donor counts and confidence intervals for key comparisons.

Test (82 donors) and validation (144 donors) hold out donors from studies represented in training. External (46 donors) holds out seven entire studies. Disease classification scores are macro-F1 for kNN, prototype, and linear readouts, in that order (§4, lines 272–279 and Table 1).


Q1 — Baseline comparisons and donor-loss ablations

CoState scored higher than joint-label scANVI and ProtoCell4P on all three validation readouts. Joint-label scANVI trains on the disease and cell-type combination. ProtoCell4P uses cell-type labels to learn prototypes and disease labels to train its donor classifier. We also added disease-supervised MultiMIL. See Table 1 and Appendix E.4, lines 1434–1447 for the results and training protocols.

MethodSupervisionTestVal.Ext.
CoStateDisease + cell type0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
scANVI, joint disease × cell typeDisease + cell type0.735 / 0.741 / 0.6560.690 / 0.636 / 0.5700.866 / 0.771 / 0.762
ProtoCell4PDisease + cell type0.829 / 0.862 / 0.6600.715 / 0.689 / 0.6490.788 / 0.758 / 0.629
MultiMILDisease0.923 / 0.860 / 0.8440.676 / 0.704 / 0.6860.760 / 0.744 / 0.677
PaSCientDisease0.825 / 0.855 / 0.6620.706 / 0.737 / 0.5630.762 / 0.739 / 0.574
scANVI, disease labelsDisease0.519 / 0.539 / 0.5250.598 / 0.521 / 0.4800.740 / 0.629 / 0.514
scANVICell type0.402 / 0.444 / 0.4210.385 / 0.369 / 0.3070.604 / 0.505 / 0.442

Donor losses and readout. The original Figure 5 ablation removed both donor losses and used mean cell embeddings for the donor readout. We separately removed only the disease contrastive loss, keeping the Hellinger donor head, the same-donor consistency loss, and disease-balanced sampling. Disease kNN fell to the composition-only scores in all three cohorts (see the revised manuscript, §4.5, lines 434–441 and Table 14, and Appendix H.2, lines 2264–2267).

Variant and donor readoutTest (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)
CoState with both donor losses and the Hellinger head0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
Disease contrastive loss removed, Hellinger head retained (new)0.204 / 0.248 / 0.2400.209 / 0.236 / 0.2200.463 / 0.658 / 0.480
Both donor losses removed, mean cell embedding used for readout (Figure 5, w/o donor losses)0.637 / 0.622 / 0.5570.583 / 0.501 / 0.4290.807 / 0.777 / 0.607

For the separate MLP encoder check, we kept the same donor head and training objectives (see the revised manuscript, Appendix H.2, lines 2281–2284).


Q2 — Overlap with the train-only panel

The panel selected on the 735 training donors shares 1,940 of 2,000 genes (97%) with the original. We used the same disease-specific ranking and voting procedure for both panels and applied each resulting panel across cohorts (see the revised manuscript, Appendix E.1, lines 1124–1133 and Table 3).

Gene panelTestVal.Ext.
Original panel0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
Train-only panel, CoState retrained0.941 / 0.915 / 0.8970.840 / 0.875 / 0.8590.893 / 0.800 / 0.733

All three Test scores in the train-only run exceeded the original three-seed means. External kNN increased and the prototype and linear scores decreased. The single retraining run combines the effects of the gene panel and training variability.


Q3 — What the external cohort supports

Six of the seven external studies contain a single disease or control label, so within-cohort classification can exploit study identity (see the revised manuscript, Appendix H.1, lines 2190–2194).

External cross-study retrieval evaluated 19 donors from two classes: healthy controls and SLE. Using the nearest reference from a different study, CoState scored 0.639 macro-F1 versus 0.595 for ProtoCell4P. The confidence interval for the difference includes zero (see the revised manuscript, §4.2, lines 326–331 and Tables 12–13).

AuthorsOpenReview ↗

Response to Reviewer M3Zf (2/3)

Q4 — Scored donors, uncertainty, and sensitivity

Scored donors. Table 11 lists the classes and donor counts for each evaluation. Across the five repeats, CoState’s disease classification scored 72 Test, 136 validation, and 43 external donors at least once. Cross-study retrieval scored 47, 75, and 19 donors, respectively (see the revised manuscript, Appendix H.1, lines 2164–2169 and Table 11).

Uncertainty. We added paired donor-bootstrap 95% confidence intervals for selected retrieval and disease kNN comparisons (see the revised manuscript, Appendix H.1, lines 2205–2210 and Table 13).

State codes and cell availability. Using 16 state codes gave higher disease kNN scores on Test and validation, while 8 gave a slightly higher score on external data. We compared 125 and 250 cells per donor view and separately tested the effect of excluding rare cell types. Sensitivity to cell counts within each individual type remains open (see the revised manuscript, Appendix H.3, lines 2327–2332 and Table 9).


W1 — What drives the improvement?

W1.1 · Does CoState benefit from access to more labels?

As shown in Q1, CoState’s validation macro-F1 was 0.145–0.218 higher than the better of joint-label scANVI and ProtoCell4P on each donor readout. All three methods use both disease and cell-type labels during training. CoState also scored higher on all seven cell-level metrics (see the revised manuscript, Tables 1–2).


W1.2 · Donor-loss and encoder ablations

Removing only the disease contrastive loss reduced kNN to the composition-only scores in all three cohorts (see the ablation table in Q1). Q1 also links to the MLP encoder comparison with the same donor head and training objectives.

W1.3 · What does the Hellinger geometry contribute?

Proposition 1 gives an exact split of the donor distance into composition and within-type state contributions (§3.3, lines 195–214). With the learned distributions fixed, squared Hellinger scored higher than total variation, Euclidean, and cosine distances on kNN in each cohort. We added the distance comparison in Appendix H.3, Table 16. Its scope is the choice of distance at evaluation. In a separate comparison with the same training objectives, trained mean pooling gave similar Test and validation retrieval, while CoState scored higher externally (0.673 vs. 0.355 across 19 queries, Appendix H.2, lines 2284–2289).


W1.4 · How does CoState compare with UCE and pseudobulk + PCA on external data?

CoState scored higher than both baselines on all three external classification readouts. The kNN gap was 0.006 for each comparison, and both kNN intervals include zero (see the revised manuscript, Tables 1 and 13).

MethodExternal kNNExternal Proto.External Linear
CoState0.8400.8870.823
Frozen UCE0.8340.8020.806
Pseudobulk + PCA0.8340.8740.678

W2 — Gene selection and evaluation scope

W2.1 · How was the original gene panel selected and used?

The 82 internal Test donors contributed to gene selection and were held out from CoState training. We selected the original panel before splitting, using expression data and disease labels from all 817 main-cohort donors (see the revised manuscript, §4, lines 276–279 and Appendix E.1, lines 1124–1133 and 1147–1154).

CoState and all panel-based baselines used the same 2,000 genes. The frozen foundation models used full-transcriptome inputs (see the revised manuscript, Appendix E.4, lines 1406–1413).


W2.2 · Does CoState still perform well with a train-only panel?

All three internal Test scores exceeded the original run means in the train-only retraining run (see the table in Q2).


W2.3 · What kind of generalization does the external cohort test?

Among studies held out from training, the retrieval analysis covers healthy controls and SLE (see Q3). The separate six-class analysis includes studies represented in training (see the revised manuscript, Appendix H.1, lines 2212–2213 and 2234–2236).

AuthorsOpenReview ↗

Response to Reviewer M3Zf (3/3)

W3 — Headline result, class counts, and uncertainty

W3.1 · What did the approximately 16% improvement refer to?

The original 16% relative gain came from external retrieval on 19 donors across two classes, healthy controls and SLE. The revised abstract highlights validation cross-study retrieval, where CoState achieved 0.984 macro-F1 versus 0.902 for MultiMIL. This evaluation covered 75 donors across two classes, healthy controls and COVID-19, from studies represented in training (see the revised manuscript, Abstract, lines 024–027).


W3.2 · Which classes and donors contributed to each evaluation?

Table 11 of the revised manuscript lists the scored classes and per-class donor counts. We summarize the donor totals in Q4.


W3.3 · How uncertain are the reported differences?

The paired 95% confidence interval for the validation retrieval difference between CoState and MultiMIL was [0.024, 0.159] macro-F1. The interval for the external comparison with ProtoCell4P included zero (see Q3 and Table 13 of the revised manuscript).


W4 — What the learned states explain

W4.1 · Do the learned codes identify distinct biological states?

The codes capture disease-related variation within each cell type. The donor losses train the state head, and reconstruction and cell-type supervision shape its input features. We treat the biological annotations of the codes as exploratory (see the revised manuscript, §3.4, lines 260–265 and Appendix H.4, lines 2357–2360).


W4.2 · What does the distance decomposition reveal about a donor match?

Proposition 1 decomposes each donor distance into differences in cell-type proportions and differences in states within each type. Comparing these terms across two references shows which differences favored the match (see the revised manuscript, §3.3, lines 195–214).

In COMBAT, a held-out COVID-19 donor was closer to the COVID-19 reference than to the non-COVID sepsis reference (squared Hellinger distances of 0.083 and 0.264). Differences in monocyte states accounted for 75.2% of the distance margin. One state received 85.1% of the average monocyte assignment weight in the donor, 82.5% in the COVID-19 reference, and 10.0% in the sepsis reference (see the revised manuscript, §4.5, lines 471–477 and Appendix H.4, lines 2335–2347).

A follow-up analysis across COMBAT donors highlighted IFI27 in the monocyte state comparison. The median within-donor expression contrast was +0.87 in 17 COVID-19 donors and −0.29 in eight sepsis donors (q = 0.023 after correction across 2,000 genes). Saichi et al. independently reported elevated IFI27 RNA in CD14⁺ monocytes from patients with moderate COVID-19. This example shows how CoState can trace a donor match to differences in cell states and highlight genes that help interpret those differences (see the revised manuscript, §4.5, lines 471–477 and Appendix H.4, lines 2348–2360).


W4.3 · How sensitive are the results to state count, cell-type labels, and cell availability?

  • Number of state codes. Sixteen codes gave higher disease kNN scores on Test and validation, and eight scored slightly higher on external data (see Q4 and the revised manuscript, Table 1 and Appendix H.3, lines 2327–2328).
  • Cell-type labels. Using predicted cell types to build donor profiles changed mean disease macro-F1 by at most 0.010 (see the revised manuscript, §3.1, lines 160–161 and Appendix E.1, lines 1155–1165 and Table 4).
  • Cell availability. We compared 125 and 250 cells per donor view and tested excluding rare cell types. The effect of varying cell counts within a given type remains open (see Q4 and the revised manuscript, Appendix H.3, lines 2327–2332).

Minor — Training configuration

We added a training configuration table to the revised manuscript (Appendix E.4, Table 6).

  • Training. We used AdamW for 24,000 steps, with 48 donors per batch and 256 cells per donor. The peak learning rate was 10⁻⁴ for the encoder and decoder and 10⁻³ for the donor head, with a 500-step warm-up followed by cosine decay.
  • Losses. All four losses had a nominal weight of 1, and the disease contrastive temperature was 0.1. Table 6 gives the schedule for ramping up the loss terms.
  • Model size and compute. The cell model has 755,488 parameters and the donor head has 23,936. Each run used one NVIDIA H800. The two loss-ablation runs took approximately 6.2–6.5 hours each.
Reviewer_WRn8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this work, the authors propose a representation learning framework, CoState, that learns joint donor-level and cell-level representations. In particular, they focus on the interplay between these two distributions and how each representation improves the quality of the other. The authors present results on disease retrieval, cell type preservation, and training scaling to demonstrate the quality of their representations.

Critical strengths
  1. The paper is well structured and clearly motivated. The proposed method by decomposing donor-level variation into cell type composition and within-type states differences provides an intuitive and potentially useful framework for studying patient heterogeneity.
  2. The Hellinger-based formulation is well defined and interpretable distance measure. The evaluations are well structured and comprehensive in terms of number and variety of baseline methods the authors compared.
Critical weaknesses
  1. The main concern is that CoState uses both disease labels and cell-type annotations during training, and requires cell-type annotations at inference, while most baselines do not have access to the same information. There is also a potential data leakage issue when the gene selection procedure includes internal test donors and uses disease-specific information. The authors should clarify its potential impact and ideally perform gene selection using training data only.
  2. The current ablations do not fully isolate the contributions of individual components. For example, removing the donor-level losses also changes the donor representation to mean pooling. It would be more informative to remove each loss separately while keeping the remaining architecture unchanged.
Questions for authors
  1. How much of the improvement comes from the additional supervision rather than the proposed method? This is particularly concerning for the cell-type clustering and prediction results. Although scANVI is included, the authors should also compare with a disease label supervised method that using cell type specific information.
  2. When there are new testing single-cell datasets, they typically lack ground-truth cell-type annotations. How well would CoState perform using predicted cell-type labels instead of the provided annotations?
  3. It is unclear the contribution of Hellinger geometry to the model’s performance, given the emphasis on it throughout the manuscript. Is it the properties of the Hellinger distance that allows decomposition of the donor distances, or is it more accurate in recovering said decomposition? If it is the former, then this needs to be clarified in the main text. If it is the latter, then an ablation with other decomposition methods as well as distance metrics would be useful to distinguish the importance of Hellinger distance.
  4. Could the authors provide additional evidence that donor-level learning improves cell-level representations, for example, by comparing held-out gene-expression reconstruction performance?
  5. Could the authors evaluate whether CoState can distinguish multiple diseases across independent studies, particularly diseases with overlapping transcriptional signatures?
  6. In Equation (5), the authors average the predicted state probabilities across cells to represent the within-type state distribution. While this is a reasonable summary, it may lose information about cell-level heterogeneity. For example, a mixture of cells confidently assigned to different states could produce the same average as cells with uncertain state assignments. How does CoState distinguish these situations? It would be helpful to discuss this limitation or evaluate how much information is lost through probability averaging.
  7. Additional methodological clarifications
  • In Section 3.2, the notation “b” is introduced without prior explanation.
  • In Figure 3, the claim is made that CoState better preserves donor distance in the representation space. This is done through a visualization with classical MDS. While classical MDS does preserve distances as best as possible in the visualization, a quantification comparing the distances in the representation spaces would greatly support the claim being made.
4 comments
AuthorsOpenReview ↗

Response to Reviewer WRn8 (1/4)

Thank you for the careful review. We have revised the manuscript in response to your comments. In the revised version, we compare CoState with baselines that also use both disease and cell-type labels during training. We also report ablations that remove one donor loss at a time and evaluate a gene panel selected using training donors alone.


W1 — Supervision, cell-type annotations, and gene selection

Supervision. We added joint-label scANVI, which uses the disease and cell-type combination as its training label, and ProtoCell4P, which uses cell-type labels to learn prototypes and disease labels to train a donor classifier. We also added disease-supervised MultiMIL. The revised manuscript reports these comparisons in Table 1 and describes the training protocols in Appendix E.4, lines 1434–1447. Bold marks the highest score among the methods shown.

MethodSupervisionTest (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)
CoStatedisease + cell type0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
scANVI, joint disease × cell type (new)disease + cell type0.735 / 0.741 / 0.6560.690 / 0.636 / 0.5700.866 / 0.771 / 0.762
ProtoCell4P (new)disease + cell type0.829 / 0.862 / 0.6600.715 / 0.689 / 0.6490.788 / 0.758 / 0.629
MultiMIL (new)disease0.923 / 0.860 / 0.8440.676 / 0.704 / 0.6860.760 / 0.744 / 0.677
PaSCientdisease0.825 / 0.855 / 0.6620.706 / 0.737 / 0.5630.762 / 0.739 / 0.574
scANVI (disease)disease0.519 / 0.539 / 0.5250.598 / 0.521 / 0.4800.740 / 0.629 / 0.514
scANVIcell type0.402 / 0.444 / 0.4210.385 / 0.369 / 0.3070.604 / 0.505 / 0.442

CoState remains competitive when other methods also use disease and cell-type labels.

Cell-type annotations at inference. CoState can infer cell types and use them to construct profiles for new donors. The revised manuscript states this in §3.1, lines 160–161, and reports the evaluation in Appendix E.1, lines 1160–1165 and Table 4. See Q2 for the results.

Gene panel. We ranked highly variable genes within each disease group and combined the rankings by voting to obtain a single panel of 2,000 genes shared across donors and cohorts. The original selection used expression data and disease labels from all 817 main-cohort donors, including the 82 internal Test donors. Validation and external donors did not contribute. The revised manuscript states the Test exposure in §4, lines 276–279, and describes the panel and retraining check in Appendix E.1, lines 1124–1133 and Table 3.

Gene panelGenes shared with original panelTest (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)
Original panel (mean over 3 seeds)2,0000.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
Panel selected on 735 training donors only (1 retraining run)1,940 (97%)0.941 / 0.915 / 0.8970.840 / 0.875 / 0.8590.893 / 0.800 / 0.733

All three internal Test scores in the train-only retraining run exceeded the original three-seed means. The panels share 1,940 genes (97%). One retraining run combines the effects of panel selection and training variability.

AuthorsOpenReview ↗

Response to Reviewer WRn8 (2/4)

W2 — Donor losses and readout

We agree that the original Figure 5 ablation changes two things at once. It removes both donor losses and uses mean cell embeddings for the donor readout. To isolate the disease contrastive loss, we removed it while keeping the Hellinger donor head, the same-donor consistency loss, and disease-balanced sampling.

Variant and donor readoutTest (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)
CoState with both donor losses and the Hellinger head0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
Disease contrastive loss removed, Hellinger head retained (new)0.204 / 0.248 / 0.2400.209 / 0.236 / 0.2200.463 / 0.658 / 0.480
Both donor losses removed, mean cell embedding used for readout (Figure 5, w/o donor losses)0.637 / 0.622 / 0.5570.583 / 0.501 / 0.4290.807 / 0.777 / 0.607

The full CoState model scored higher than both variants on every metric shown. The full-model and both-losses-removed rows are means over three runs. With the Hellinger head retained, removing the disease contrastive loss reduced kNN to the composition-only scores in all three cohorts. This suggests that the loss helps the model capture disease information in the states within each cell type.

The revised manuscript clarifies the original ablation in §4.5, lines 434–441 and Table 14. The revised manuscript reports the fixed-head kNN results in Appendix H.2, lines 2264–2267, and the separate consistency ablation at lines 2277–2280. The prototype and linear scores for the fixed-head run are reported here.


Q1 — Cell representations and baseline comparisons

We added joint-label scANVI and ProtoCell4P to Table 2. Both use disease and cell-type labels during training. §4.3, lines 380–383 describes the comparison and notes that Leiden resolution was selected using cell-type labels.

Method (Test donors)SupervisionNMI_LARI_LNMI_kARI_kkNN F1
CoStatedisease + cell type0.7960.7620.7030.5200.846
scANVI, joint disease × cell type (new)disease + cell type0.5580.2320.4890.2650.828
ProtoCell4P (new)disease + cell type0.6230.5260.4210.2570.584
PaSCientdisease0.5700.3780.4270.2180.683
scANVI (disease)disease0.5680.2670.5650.3130.788
scANVIcell type0.6230.3680.5670.3150.832

CoState scored higher than joint-label scANVI and ProtoCell4P on all five metrics shown. Q4 reports the comparison within CoState.


Q2 — Predicted cell-type labels

We predicted cell types with 15-NN on frozen CoState embeddings, using annotated training cells as references. The predictions matched 91.8% of the manual Test annotations. We then used predicted types to rebuild donor profiles and repeated the disease readouts. The revised manuscript reports the results in Appendix E.1, lines 1160–1165 and Table 4.

Cell types used to build the donor vectors (CoState, 3 seeds)Test (kNN / Proto. / Linear)Val. (kNN / Proto. / Linear)Ext. (kNN / Proto. / Linear)
Atlas labels, as in Table 1 (Test: manual; Val. and Ext.: scANVI label transfer)0.891 / 0.900 / 0.8490.860 / 0.888 / 0.8670.840 / 0.887 / 0.823
Predicted by CoState itself (15-NN on its cell embedding)0.884 / 0.910 / 0.8420.857 / 0.884 / 0.8640.832 / 0.891 / 0.825

The largest absolute change in mean disease macro-F1 was approximately 0.010 across the nine comparisons.


Q3 — Role of the Hellinger geometry

Squared Hellinger gives an exact, nonnegative decomposition into differences in cell-type proportions and abundance-weighted differences in states within each type. We clarified the construction and attribution in §3.3, lines 195–214 and Proposition 1. In the cross-study donor pairs we examined, within-type states accounted for 74–78% of the distance between donors with different diseases (Appendix G.4, lines 1989–1995 and Figure 9). With the encoder fixed, retaining only composition reduced external kNN from 0.840 to 0.463 (§4.5, lines 464–469 and Figure 5).

With the learned distributions fixed, squared Hellinger scored higher than total variation, Euclidean, and cosine distances on kNN in each cohort. The revised manuscript reports this comparison in Appendix H.3, Table 16. The comparison evaluates the choice of distance at inference.

AuthorsOpenReview ↗

Response to Reviewer WRn8 (3/4)

Q4 — Evidence on cell representations

We compared cell-type clustering and prediction with and without the donor losses. With the donor losses, CoState scored higher on all five metrics averaged across runs. We added these results to §4.5, lines 434–435 and Appendix H.2, Table 15.

Cell embedding (3 seeds)NMI_LARI_LNMI_kARI_kkNN F1
CoState0.7960.7620.7030.5200.846
Same model trained without donor losses0.7510.5980.6510.4190.833

As a separate check, we used covarying neighborhood analysis (CNA) to identify SLE-associated monocytes in the same expression data. The median within-donor AUROC for ranking these cells was 0.659 with the donor losses and 0.548 without them, across 28 donors. We added this analysis to Appendix H.2, lines 2297–2302.


Q5 — Multiple diseases across studies

We evaluated cross-study retrieval on 214 donors across six classes: healthy controls (92), COVID-19 (43), SLE (38), Crohn’s disease (19), sepsis (16), and cirrhosis (6). Each query takes the label of its nearest donor from a different study. The evaluation includes studies represented in training. We report the class counts and results in Appendix H.1, Table 11 and lines 2212–2213 and 2234–2236.

CoState achieved the highest accuracy among the main benchmark configurations (62.6%), followed by MultiMIL (55.3%).


Q6 — Averaging in Equation (5)

We agree that the two cases can yield the same average. Equation (5) uses that average to summarise the relative proportions of the learned states within each cell type.

To see what the averages reflect in practice, we examined the five most abundant cell types. We report the effective number of states, exp(H), for individual cells and for the average within each donor and cell type, aggregating across donors, cohorts, and seeds. A value of 1 means that all probability is on one state. A value of 16 means the probability is spread evenly across all 16 states.

Cell typeStates per cell, exp H(α_i)States per donor, exp H(q_dc)Mean cell entropy as a share of H(q_dc)
Monocytes1.072.238%
B cells1.092.4110%
CD4 T, naive1.122.6012%
CD8 T, non-naive1.152.6415%
CD4 T, non-naive1.182.8816%
These five types1.122.5513%

In these five cell types, each cell’s probability was typically concentrated on one state, while cells of the same type within a donor occupied several states. This suggests that the average mostly reflects a mixture of cells assigned to different states. The revised manuscript summarises this check in Appendix H.3, lines 2320–2323.

To test the effect on disease prediction, we kept the cell assignments fixed and changed how we constructed the donor vectors.

Donor vector built fromTest kNNVal. kNNExt. kNN
Average of the cells' α_i (Eq. 5; CoState as in Table 1)0.8910.8600.840
Each cell counted once, in its most likely state0.8950.8570.849
Mean per-cell assignment entropy for each cell type0.2760.2860.253

Counting each cell in its most likely state changed kNN macro-F1 by at most 0.009 across the three cohorts. Using only the mean assignment entropy gave much lower scores. These results support averaging the probabilities to summarise state proportions for this prediction task. We report these results in Appendix H.3, lines 2324–2325, and explain this choice and its limitation in §3.2, lines 189–192.

AuthorsOpenReview ↗

Response to Reviewer WRn8 (4/4)

Q7 — Notation and Figure 3

We now define the three feature blocks at their first use in Section 3.2, lines 169–176. Equation (3) reads

[bdicell;bdipat;bdires]:=Fθ(xdi).[b_{di}^{cell}; b_{di}^{pat}; b_{di}^{res}] := F_\theta(x_{di}).

For cell i from donor d, each of the four readout MLPs combines two attention-query summaries with a learned projection of the cell’s expression profile. The first MLP produces the cell block, the middle two outputs form the patient block, and the last MLP produces the residual block (Appendix C.1, lines 965–977).

To complement the MDS visualisation in Figure 3, we quantified the pattern using distances in the original representation space. For each of the 19 COVID-19 donors, we checked whether the nearest COVID-19 donor from the other study was closer than the nearest healthy donor. For CoState and scVI, we report the mean percentage across runs, using 16 state codes for CoState. Q5 reports the broader comparison across six classes.

Method shown in Figure 3Training labelsCOVID-19 donors meeting the distance criterion (n = 19)
CoStateDisease and cell type96.5%
scVINone10.5%
PCANone5.3%
HarmonyNone (study correction)42.1%

With CoState, 96.5% of COVID-19 donors on average were closer to a COVID-19 donor from the other study than to any healthy donor. The revised manuscript describes this comparison in Section 4.2, lines 357–359 and reports the results in Appendix H.1, lines 2254–2257.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their thorough and constructive feedback. Reviewers WRn8 and M3Zf highlighted the value of separating differences in cell-type composition from differences in cell states within each type. Both noted the breadth of the benchmark, and Reviewer M3Zf highlighted CoState’s performance against disease-supervised baselines on the validation cohort. Reviewer 2MkJ noted that the downstream applications and interpretability analyses offer insights into learning donor representations.

In response to the reviewers’ suggestions, we revised the manuscript and appendix, with changes marked in blue. The main additions are summarized below.

  • New baselines. We added joint-label scANVI and ProtoCell4P, both with access to the same labels as CoState, and the STATE foundation model. CoState had the highest point estimate on most donor classification readouts and all cell-level readouts (Tables 1–2). Please see our responses to Reviewers WRn8, M3Zf, and 2MkJ.
  • New ablations. With the Hellinger readout held fixed, separate ablations showed that the disease contrastive loss drove donor kNN performance and that the same-donor consistency loss stabilized embeddings across cell subsamples (Appendix H.2). Please see our responses to Reviewers WRn8 and M3Zf.
  • New biological interpretation. In a worked COMBAT example, CoState’s Hellinger decomposition attributed 75.2% of the distance margin favoring the COVID-19 reference over the sepsis reference to within-type monocyte states. A complementary expression analysis highlighted IFI27 in the overrepresented monocyte states of COVID-19 donors (§4.5). Please see our responses to Reviewers M3Zf and 2MkJ.
  • Training-only gene panel. Selecting genes using training donors alone recovered 97% of the disease-informed panel. In a single retrain, all three internal Test classification scores exceeded the corresponding three-seed means with the disease-informed panel (Appendix E.1). Please see our responses to Reviewers WRn8 and M3Zf.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.