acceptodds
Under review as a conference paper at ICLR 2027

BASILISC: Bayesian Alignment of Siamese-Integrated Latents for Inferring Single-cell Concordance

Abstract

Single-cell assays are the established method to study the genome one cell at a time. Unimodal assays measure one molecular layer, and the newer multimodal assays measure several. Most data are single-modality, so integrating them relies on models trained on the multimodal assays, where the cell correspondence across modalities is known. These models return a matching, but rarely a confidence measure. We present BASILISC (Bayesian Alignment of Siamese-Integrated Latents for Inferring Single-cell Concordance), an uncertainty-aware siamese network for matching unimodal single-cell embeddings. It places the uncertainty on the correspondence itself: for a cross-modal pair of cells, it predicts a distribution over whether the two share a cell type, and at the cell level over which cell is the partner. To our knowledge, this is the first uncertainty score on a learned cross-modal correspondence evaluated against a known ground-truth pairing. BASILISC learns a biologically meaningful similarity across modalities that reflects cell lineage: an unseen cell type is matched to its lineage relative, matching still works on such held-out cell types (balanced accuracy 0.85 on unseen vs. 0.93 on seen types), and the predictive uncertainty rises on these out-of-distribution cells. The uncertainty score also selects high-quality matches for downstream multimodal analysis: discarding the least certain fifth of the pairs raises accuracy from 0.93 to 0.98. At the cell level, a deep ensemble of the architecture reaches 0.062 matching accuracy on the benchmark test set, above the competition winner, and the posterior read on that assignment flags the wrong matches that cross a cell-type boundary better than a deterministic read of the same scores (AUROC 0.68 against 0.64).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
uncertainty quantification, single-cell multimodal integration, computational biology, Bayesian neural networks, cross-modal similarity learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_X3NTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BASILISC estimates uncertainty in cross-modal single-cell matching using Bayesian siamese networks. It evaluates both same-cell-type prediction and exact cell matching, comparing variational inference, Monte Carlo dropout, ensembles, and deterministic controls. The results show benefits for detecting unseen cell types and some cross-type matching errors, alongside important limitations.

Critical strengths

Uncertainty in cell correspondence is a meaningful practical problem, beyond simply improving matching accuracy.

The experimental design includes deterministic controls, multiple uncertainty methods, held-out cell types and families, and replication across assay types.

The paper clearly reports negative results, including reduced matching accuracy with variational inference and limited novelty detection at the cell level.

Critical weaknesses

The practical benefit of abstention is unclear. Keeping the most certain half removes only 32.8% of cross-type errors. As written, this implies that the retained matches have a higher cross-type error rate than the original set, despite outperforming the deterministic control.

The added value of Bayesian inference is modest. At the cell-type level, error detection essentially ties the deterministic margin. At the cell level, the overall AUROC improvement is only 0.013, and the best matching accuracy comes from the deterministic ensemble.

Questions for authors

Can you report cross-type error rates among retained matches across coverage levels, including comparison with random rejection?

Can a simple combination of deterministic confidence and embedding-based novelty scores match the proposed uncertainty methods?

Were held-out cells excluded from Liam pretraining, and were all thresholds and hyperparameters selected without test labels?

Does uncertainty identify meaningful within-type mismatches, and is the proposed association with transitional states supported by independent biological evidence?

2 comments
AuthorsOpenReview ↗

Response to Reviewer X3NT

We are grateful for the careful review and the valuable feedback. We attempted to address each analysis raised in the review and answer your points in order; the full analyses are in the revised Supplement.

W1 / Q1. We agree with your point on filtering at the cell level. In the cell-level model, filtering on any posterior uncertainty score does not lower the cross-type error share below random at any coverage; the deterministic margin is slightly below it, and filtering on the matcher's own similarity of the assigned pair does better than random on lowering cross-type errors. The reason is what we have already reported in our manuscript: cross-type errors are the confident assignments. We discarded the claim about filtering at the cell level in the manuscript (abstract, contributions, Section 5.2, Discussion; Supplement E.1, Table S12, Figure S11) and reported our findings as a diagnostic of which errors the posterior uncertainty separates from correct matches (Table 1).

At the cell-type level, abstention removes most of the errors at 80% coverage, compared with a fifth for random rejection (Section 5.1, number corrected), and query-level label transfer keeps its error below random from 20% coverage up (below).

W2. We agree that the gain from MFVI is modest, and the revised text states this where it applies. However, where the gain is positive, it has practical use. At the cell-type level, a single sampled score can flag both the model's errors and the cells of a type that was not seen during the matcher's training, whereas no single deterministic score does both as often (Section 5.1). In query-level label transfer, the uncertainty score is higher on the cells whose label is incorrect (Supplement E.4). The manuscript also reports where MC dropout and the deep ensemble perform better than MFVI. For the matching task itself, our ensemble is in the range of CLUE and MatchCLOT, which we trained with their own code and evaluated on the same test set (Supplement E.9). In small-data settings, when we trained cell-level models on a training set that is one tenth of the size of the original one, BASILISC largely retains its accuracy while MatchCLOT's accuracy drops sharply (Supplement E.10). In summary, BASILISC performs comparably to other state-of-the-art methods on the matching task, while outperforming them in low-data regimes in this application, and additionally provides an uncertainty score on the cross-modal correspondence, which other methods do not. This work also presents a comprehensive evaluation protocol that shows where the uncertainty score is useful and where it is not.

Q2. We evaluated the fusion score you suggested: a rank average of the deterministic margin and the embedding-based cosine novelty score, once with equal weights and once with the weight chosen leave-one-type-out. The fusion score detects both the model's errors and the unseen types in as many runs as the mean-field posterior, but in fewer than MC dropout or the deep ensemble. Therefore, we have updated the text to say that this fusion of the two deterministic scores reaches MFVI, but not the other two approximations (Section 5.1; Supplement E.3, Table S14).

Q3. We agree and have clarified the scope: "unseen" refers to the matcher's supervision and not Liam's. Liam was trained on all cell types but in an unsupervised manner, therefore the embeddings are label-free (Supplement B.1). We ran the controls you suggested. (a) Inductive: on a PCA/LSI backbone fit without the held-out type, the gain of psdp_{\mathrm{sd}} over the margin stays positive in three of four types but shrinks; (b) on Liam itself refit without the type, the posterior still improves on the margin for two of the four types, but not for the other two, while the same refit on every cell reproduces the released backbone, so we state the result as transductive (Supplement E.4, Tables S16–S17). (c) Query level without label screening: filtering on psdp_{\mathrm{sd}} removes about twice the random share of held-out cells at 80% coverage, more than the margin, and keeps the transfer error below random from 20% coverage up, but only when several cells of the novel type are in the query set (Supplement E.4, Table S18). In addition, the Sinkhorn temperatures are now chosen on the validation, and Table 1 does not change (Supplement E.5, Table S20).

Q4. We separated the two types of error as they matter for different use cases of a cross-modal matcher. For label transfer and any analysis at the cell-type level, a same-type partner carries the correct label. For matching at the level of cell state, every error counts, and there the posterior's improvement over the deterministic control is below our predefined threshold. The statement on cells in transition is only kept as a hypothesis in the Discussion (Supplement E.7, Table S22).

Dear Reviewer X3NT, As the discussion period comes to its end, we would like to briefly draw your attention to our recent interaction with Reviewer RmTW. Several of the key technical questions you raised overlapped with the concerns raised by Reviewer RmTW, specifically regarding the cross-type error rates during filtering and whether held-out cells were excluded from the Liam pre-training. In our rebuttals, we provided the requested controls, baselines, and clarifications to address these shared points. Reviewer RmTW has since reviewed these additions and noted that the new controls and reported outcomes successfully answered those technical questions.
We have integrated all of these analyses into the revised manuscript and supplement. We hope our responses and the updated results can address your concerns as well.

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

This paper studies uncertainty in cross-modal correspondences in single-cell multi-omics. It distinguishes two tasks: determining whether two cells belong to the same type and retrieving the same cell from another modality. The cell-type-level model uses a Siamese network on frozen unimodal Liam embeddings. A Bayesian classifier head predicts the same-type probability, and the standard deviation across sampled predictions provides an uncertainty score. The cell-level model uses Liam-initialized encoders, a shared projection, and symmetric contrastive learning. It obtains a one-to-one matching from the mean similarity matrix, then evaluates support for the final assignment using sampled local matching votes and transport plans. The comparisons include mean-field variational inference (MFVI), MC dropout, deep ensembles, deterministic models with the same architecture, and simple representation baselines.

Experiments use bone marrow single-cell multi-omics data. The cell-type-level analysis covers seen types, leave-one-cell-type-out, leave-one-lineage-family-out, and few-shot recovery, with replication on gene expression–surface protein data. On gene expression–chromatin accessibility data, type-level classification achieves AUROC 0.976. Removing the most uncertain 20% of pairs raises retained-set accuracy from 0.932 to 0.975, although the deterministic confidence margin gives essentially the same benefit. In the cross-batch leave-one-type-out experiments, the default MFVI method achieves novel-type detection AUROC 0.619, compared with 0.564 for the same-architecture deterministic model and 0.707 for MC dropout (main text, pp. 4–8; Tables S9–S10).

The cell-level results use 20,009 test nuclei. Exact matching accuracy is 0.0624 for the five-model deterministic deep ensemble and 0.0478 when averaging five MFVI seeds. The paper compares these results with published scores for CLUE (0.058) and MatchCLOT (0.0755). For detecting cross-type errors against exact matches, the posterior transport-plan score achieves AUROC 0.679 versus 0.635 for the Sinkhorn control using the same mean matrix. The paper's main value is its analysis of correspondence reliability and differences across error types. Its practical filtering benefit and the scope of novel-type detection remain less clear (pp. 8–9, Table 1).

Critical strengths
  1. The problem is practically important, and the reliability analysis uses ground-truth matching and meaningful error strata.

Cross-modal integration requires both a matching and an indication of which assignments are trustworthy. The paired test data allow the authors to distinguish exact matches from same-type mismatches, cross-type mismatches within a lineage family, and mismatches across families. This is more informative than reporting an overall matching score. Table 1 shows that posterior scores and deterministic transport weights have different relative performance across these groups. For cross-type errors versus exact matches, the posterior transport-plan score reaches AUROC 0.679, compared with 0.635 for the deterministic control. These conditional results do not yet establish deployment-time filtering benefits, but they provide concrete evidence about which errors the uncertainty scores can identify. This supports the value of studying correspondence uncertainty (main text, pp. 3, 8–9, Table 1; Figures S5–S7).

  1. The comparisons cover several uncertainty methods and report their limitations as well as their gains.

The paper compares MFVI, MC dropout, deep ensembles, same-architecture deterministic models, and simple baselines. It also examines posterior width and placement, type and family holdouts, and few-shot recovery. The cell-type-level experiments are replicated on a second modality pair. The authors report that seen-type error detection is nearly matched by the confidence margin, default MFVI underperforms MC dropout for novel-type detection, the overall cell-level detection gain falls below the prespecified +0.02 threshold, and MFVI reduces exact matching accuracy. These controls help avoid attributing ordinary classifier confidence effects to a Bayesian posterior. The negative results also make the paper useful as an empirical study of uncertainty (main text, pp. 6–9; Supplement B.5–B.6, Tables S7–S10).

Critical weaknesses
  1. The conditional advantage in detecting cross-type errors has not been shown to reduce harmful errors in the full set of retained assignments.

The cross-type AUROC in Table 1 compares cross-type errors with exact cell matches, excluding 16,852 same-type mismatches. Among the 20,009 assignments, 956 are exact matches and 2,201 are cross-type errors. The authors treat same-type mismatches as approximately harmless for downstream analyses at the cell-type level. If the practical objective is to filter out cross-type errors, the comparison class should therefore include all non-cross-type assignments, not just exact matches. The conditional AUROC is a valid diagnostic, but it evaluates a different task from that filtering objective (main text, p. 9, Table 1; Supplement B.5, p. 23).

The rejection result raises a more direct concern. The text states that retaining the most certain 50% of assignments according to the top-1 vote score u_i removes 32.8% of cross-type errors, compared with 25.1% for the Sinkhorn control (pp. 8–9). The improvement over Sinkhorn is useful, but random rejection at the same coverage would remove 50% of those errors in expectation. If the denominator is all cross-type errors, the reported result implies a retained-to-original cross-type error-rate ratio of (1 - 0.328) / 0.5 = 1.344.

Using the counts in Table 1, the cross-type error rate would rise from 2,201/20,009, approximately 11.0%, to approximately 14.8%, a relative increase of 34.4%. This is a calculation from the paper's reported counts and removal rate, not an additional experiment reported by the authors. The text also acknowledges that every score rates cross-type errors as more certain than same-type errors, with a posterior AUROC of 0.35 on that contrast (p. 9), consistent with this concern.

This affects the main application claim. Outperforming a deterministic score does not by itself establish that retaining high-confidence assignments reduces harmful-error risk. The paper needs detection results for cross-type errors against all non-cross-type assignments and a cross-type risk–coverage curve, with random rejection and deterministic controls evaluated on the same assignments. This can largely be done using existing predictions. Evidence of risk reduction under explicit conditions would strengthen the acceptance case. Otherwise, the practical contribution should be limited to the conditional error-diagnosis setting that the results actually support.

  1. Novel-type detection is evaluated with transductive representations and on ground-truth same-type pairs, which limits the unknown-type interpretation.

Supplement B.1 states that the type-level multiome Liam embeddings are trained on all 69,249 nuclei. The unimodal models used for cell-level initialization are trained on 62,501 cells, explicitly including test nuclei. LOCTO and LOFO remove types from the downstream matcher's training and validation pairs. Held-out types therefore do not enter matching supervision, but their observations have already entered representation pretraining. This supports a transductive analysis of types unseen in matching supervision. It does not establish inductive generalization to types or observations unseen throughout representation learning (Supplement B.1, B.4, pp. 21–23).

The main novelty AUROC is also calculated only on ground-truth same-type pairs. This is a useful controlled diagnostic. However, when a query type is unknown or absent from a reference atlas, these pairs cannot first be selected using true labels. The analysis of unseen-type negatives and closely related types provides useful clues, but the paper does not yet give a complete evaluation of query-level novelty decisions constructed from all candidate correspondences without known query labels (main text, pp. 6–7; Supplement B.5, Table S10).

Unlabeled transductive pretraining can be reasonable, and the authors disclose it. I am not alleging label leakage or concealment. The issue is the scope of statements about types the model has never seen and label transfer for unknown types. The manuscript should specify the training stage and pair population to which those statements apply. A representative control fitting the encoder only on the retained training cells, or a query-level evaluation without ground-truth positive-pair screening, would help establish how far the novelty result extends.

Questions for authors
  1. Practical filtering benefit, using existing predictions.

Please confirm whether the 32.8% removal rate at 50% coverage uses all cross-type errors as its denominator. If so, the reported figures imply that retained-set cross-type risk rises from about 11.0% to 14.8%. For all 20,009 assignments, please treat cross-type errors as positive and all other assignments as negative, and report detection AUROC, AUPRC, and cross-type risk–coverage curves. Include random rejection and the same-matrix Sinkhorn comparison. At which coverage or under which use conditions does posterior scoring reduce this risk? This would directly affect my assessment of the main application benefit (main text, pp. 8–9, Table 1; Supplement B.5).

  1. Scope of the novel-type result.

Does unseen in LOCTO/LOFO refer only to matching supervision, while unimodal pretraining still includes the held-out cells? If Liam is fitted only on the retained training cells, does the novelty-detection gain remain in at least one representative holdout? Also, without selecting same-type pairs using true labels, how would a query-level novelty score be constructed and validated from the candidate pairing scores? If the intended claim is transductive diagnosis conditional on same-type pairs, please state that scope in the main text (Supplement B.1, B.4–B.5).

  1. Interpretation of the uncertainty scores.

In Supplement B.3, u_i compares each sampled matrix's row-wise argmax with the final global Hungarian assignment. It does not compare a Hungarian assignment for each sample with the final assignment. Even if all sampled matrices are identical, local and global optima can differ, giving nonzero u_i. Similarly, v_i = 1 - E[P_i,π(i)] can remain nonzero under a degenerate posterior because the entropic transport plan is dispersed. How do these scores separate parameter-sampling variation, local–global assignment conflicts, and ambiguity in the matching itself? Would support for the final assignment be a more accurate interpretation than posterior disagreement alone? A small degenerate-posterior control or a comparison using sample-wise global assignments could clarify the source of the gain and the role of the Bayesian mechanism (main text, pp. 1, 4–5; Supplement B.3, p. 22).

  1. Replication scope and data protocol.

Supplement A mentions the GEO release with 90,261 cells for CITE-seq cell-level replication, but Table S6 lists only multiome cell-level experiments. Section D and Table S10 report cell-type-level results. Where are the CITE-seq cell-level results? If only the type-level replication is reported, please clarify that scope. Does the CITE-seq test-set entry described as the same as multiome in Table S1 refer to held-out sites and batches, or to the number of cells? Please give the data version and train/validation/test cell counts for each task, including the relationship between the full data, competition-phase files, and pretraining inputs. This is needed to establish which two-level, three-modality conclusions have direct evidence and whether the published baseline scores use the same protocol (Supplement A, B.1, B.4, D, pp. 16–23, 25–26; Tables S1, S4–S6, S10).

5 comments
AuthorsOpenReview ↗

Response to Reviewer RmTW

We are grateful for the detailed review and the valuable feedback. We attempted to address each analysis raised in the review and answer your points in order; the full analyses are in the revised Supplement.

W1 / Q1. You are right, and your calculation holds. In the cell-level model, filtering on any posterior uncertainty score does not lower the cross-type error share below random at any coverage; the deterministic margin is slightly below it, and filtering on the matcher's own similarity of the assigned pair does better than random on lowering cross-type errors. The reason is what you mentioned in your review and we have also reported in our manuscript: cross-type errors are the confident assignments. We discarded the claim about filtering at the cell level in the manuscript (abstract, contributions, Section 5.2, Discussion; Supplement E.1, Table S12, Figure S11) and reported our findings as a diagnostic of which errors the posterior uncertainty separates from correct matches (Table 1).

At the cell-type level, abstention removes most of the errors at 80% coverage, compared with a fifth for random rejection (Section 5.1, number corrected), and query-level label transfer keeps its error below random from 20% coverage up (below).

To answer your question on how to reduce the share of cross-type (i.e., harmful) errors at the cell level, we adopted a cell-type agreement approach as a post-processing step for the cell-level matching. To this end, we first run the cell-level inference and obtain a matching score for each pair, then check in two ways if the assigned pair is of the same cell type: 1. the BASILISC cell-type-level model, 2. a kk-NN cell-type label transfer, both fitted on the training cells only. The pairs that are likely not sharing a cell type are marked as "not a match" at the cell level. This approach does help reduce the share of cross-type errors but the posterior does not bring an advantage in this case (Supplement E.2, Table S13; CITE-seq in D.2, Table S11).

W2 / Q2. We agree and have clarified the scope: "unseen" refers to the matcher's supervision and not Liam's. Liam was trained on all cell types but in an unsupervised manner, therefore the embeddings are label-free (Supplement B.1). We ran the controls you suggested. (a) Inductive: on a PCA/LSI backbone fit without the held-out type, the gain of psdp_{\mathrm{sd}} over the margin stays positive in three of four types but shrinks; (b) on Liam itself refit without the type, the posterior still improves on the margin for two of the four types, but not for the other two, while the same refit on every cell reproduces the released backbone, so we state the result as transductive (Supplement E.4, Tables S16–S17). (c) Query level without label screening: filtering on psdp_{\mathrm{sd}} removes about twice the random share of held-out cells at 80% coverage, more than the margin, and keeps the transfer error below random from 20% coverage up, but only when several cells of the novel type are in the query set (Supplement E.4, Table S18).

Q3. We ran the three controls you suggested on the same assignment: the vote of the mean score matrix (uiu_i under a degenerate posterior), the vote of the five seeds' point estimates, and a one-to-one assignment solved per posterior draw. On cross-type errors against exact matches, the per-draw assignment ranks best of all, ahead of uiu_i. Solving the one-to-one assignment for each draw removes the disagreement between local and global optimum and ranks cross-type errors better than uiu_i; this disagreement is therefore real, but it adds noise to the signal of uiu_i rather than producing it. We adopted your phrase, "the draws' support for the final assignment", and report the per-draw assignment with uiu_i (Section 5.2; Supplement E.5, Tables S19–S20). In addition, the Sinkhorn temperatures are now chosen on the validation nuclei, and Table 1 does not change.

Q4. We apologize that the CITE-seq cell-level results were mentioned but not included; they are now in the revised Supplement D.2. Accuracy and the direction of the cross-type diagnostic here are similar to the multiome results. However, when filtering on the uncertainty score, CITE-seq results are different: with its finer annotation, cross-type errors make up a quarter of the assignments and are not the most confident ones, so filtering on every score slightly lowers the share of harmful errors, but none performs better than the deterministic control. We also added a data-protocol table per task as requested (Supplement E.8, Table S23; Supplement A, Table S1).

Thank you for running all the requested controls and reporting the outcomes so openly in Supplement D.2 and E.1-E.8. Confirming that cell-level posterior filtering does not beat random or deterministic controls on Multiome or CITE-seq (and adds no gain in cell-type post-processing), that novel-type gains are transductive, and that per-draw Hungarian matching improves over u_i answers all four of my technical questions.

Given these revisions, deterministic similarity/margin or MC dropout now matches or beats Bayesian MFVI in almost all practical filtering and abstention tasks, leaving Table 1's conditional post-hoc diagnosis as the main positive signal. Could you briefly clarify whether there is still a label-free use case where BASILISC's posterior sampling has a clear practical edge, or if you now frame the paper mainly as a diagnostic and negative-result study?

Thank you for your reply. We frame the paper as a method together with an evaluation protocol with a defined scope, not as a negative result study. MC dropout and the deep ensemble are two of the three posterior approximations of BASILISC, so the claim concerns sampling-based uncertainty quantification on the correspondence rather than MFVI in particular. Further, the observed benefit of uncertainty quantification differs between the cell-type level and the cell level.

At the cell-type level, we have demonstrated that a single posterior uncertainty score has practical impact in two settings: when the model's prediction is wrong, and for cells of a type unseen in the matcher's training. With MC dropout and the ensemble, this holds for both settings in 88% and 82% of the cross-batch leave-one-cell-type-out runs. The deterministic margin reaches 53%, and a fusion of two deterministic scores reaches 70% (Tables S9 and S14). This evaluation uses ground-truth positive pairs. In a setting without labels, the query-level label transfer, a cell of an unseen type always receives a wrong label, because its type is missing from the reference. Filtering on psdp_{sd} removes 41% of these wrong transfers at 80% coverage, while the confidence margin removes 33–35% and random rejection removes 20%. This holds when several cells of the novel type are in the query set (Table S18).

At the cell level, we agree that the posterior is only diagnostic (Table 1), and the type-agreement check is the practical filter. This filtering result was only one part of our results: the uncertainty score on the correspondence, the evaluation protocol, the cell-type-level results, and the comparison with published matchers, especially the advantage of our model when little training data is available, do not depend on it.

Reviewer_Bx6xOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses diagonal integration of single-cell modalities and, to the reviewer's knowledge, is the first to place an uncertainty score on a learned cross-modal correspondence evaluated against a known ground-truth pairing. BASILISC is built at two granularities: a Bayesian siamese classifier over frozen Liam embeddings at the cell-type level, and a Bayesian contrastive matcher trained end-to-end from raw counts at the cell level (Hungarian assignment on the mean score matrix). Contributions: (i) uncertainty on the correspondence itself; (ii) a controlled demonstration across three modalities of what the score delivers — error detection, abstention, and moderate novelty detection — each compared against a deterministic control with a pre-stated bar; (iii) evidence that the uncertainty reflects biology, following cell lineage and possibly marking cells in transition. Validation uses controlled data constructions (mix, LOCTO, LOFO, few-shot), five seeds, a pre-registered analysis log, and replication on CITE-seq.

Critical strengths
  1. Novel and consequential problem framing: the first uncertainty score placed on the cross-modal correspondence (a relational prediction) and validated against known ground-truth pairings. Single-cell integration in practice urgently needs a "which matches are trustworthy" signal for downstream analysis and label transfer; the paper fills this gap and shows the uncertainty is biologically meaningful (lineage structure, raised on unseen types), not merely an engineering metric.
  2. Exemplary experimental discipline: every analysis is paired with a deterministic control of the identical architecture, comparisons carry pre-registered bars, confidence intervals and multi-seed statistics are reported across many controlled data constructions, and the authors proactively report where the score fails (ties the margin in-distribution; cell-level novelty detection breaks down). This honest mapping of the method's boundary substantially strengthens the claims and gives the community a reusable evaluation protocol.
Critical weaknesses
  1. The practical gains from Bayesian uncertainty estimation are limited. For error detection on seen cell types, the posterior score performs almost identically to the deterministic confidence margin (0.856 vs. 0.858). For novelty detection, MFVI (0.619) performs worse than MC dropout (0.707) and the simple cosine baseline (0.667). Although the Bayesian approach can address both tasks with one uncertainty score, it is unclear whether a combination of deterministic scores could achieve similar results.
  2. Cell-level matching performance and practical utility remain limited. MFVI achieves 0.0478 accuracy, below the deterministic ensemble's 0.0624, and the best result remains below MatchCLOT's benchmark score of 0.0755. The overall improvement in error detection is small. Although cross-type error detection improves, treating within-type errors as relatively harmless may not be justified for accurate cellular-state matching. Direct downstream validation is missing.
  3. Generalization and biological interpretation are insufficiently validated. Experiments focus on bone marrow data, without evaluation on other tissues or independent datasets. The observed lineage structure may largely originate from pretrained Liam embeddings. The hypothesis that highly uncertain cells represent transitional states also lacks direct evidence. In addition, cell-level novelty detection is affected by cell-type abundance.
Questions for authors
  1. MC dropout and deep ensembles outperform MFVI in cell-type-level novelty detection, while MFVI also incurs a noticeable matching accuracy penalty at the cell level. Could the authors clarify the motivation for using MFVI as the default Bayesian approximation? A comparison of training time, inference cost, and uncertainty detection performance across MFVI, MC dropout, and deep ensembles would help establish their practical cost-benefit trade-offs.
  2. A central claim is that a single posterior-based uncertainty score can support both error detection and novelty detection, whereas the deterministic confidence margin and embedding cosine similarity are individually effective for different tasks. Have the authors considered combining these deterministic signals into a single confidence score and comparing it against the posterior-based score? For example, fusion weights could be selected on validation data, followed by comparisons using dual-task AUROC and risk-coverage curves. This would help isolate the unique contribution of posterior sampling beyond inexpensive deterministic alternatives.
  3. The Sinkhorn baseline temperature is selected using the test set. Although this makes the comparison conservative in favor of the deterministic baseline, it still introduces test-set feedback into model evaluation. Would the reported posterior advantages remain consistent if the temperature were selected exclusively on held-out validation data? Additionally, have the authors considered normalizing the cell-level novelty score for the number of competing candidate matches or cell-type abundance to address the observed type-size dependence?
1 comment
AuthorsOpenReview ↗

Response to Reviewer Bx6x

We are grateful for the thorough review and the valuable feedback. We attempted to address each analysis raised in the review and answer your points in order; the full analyses are in the revised Supplement.

W1 / Q2. We revised the text to say that the gain is limited wherever the posterior ties the margin. We ran the fusion you suggested: a rank average of the deterministic margin and the embedding-based cosine novelty score, once with equal weights and once with the weight chosen leave-one-type-out. The fusion flags both the model's errors and the unseen types in as many runs as the mean-field posterior, but in fewer than MC dropout or the deep ensemble. So a fusion of two deterministic scores reaches MFVI and not the two other approximations, and we adjusted the text accordingly (Section 5.1; Supplement E.3, Table S14).

Q1. The focus of the paper is uncertainty on the correspondence and not MFVI, and we initially chose MFVI because it is the one approximation with an explicit and controllable posterior width (Table S7). The revised text now says clearly that MC dropout is the best novelty detector of the three and that the ensemble model has the higher cell-level accuracy. We also add the requested cost table (see Supplement E.3 and Table S15).

W2. At the cell level we compete with the best matchers, while being able to provide UQ to the matching. To evaluate BASILISC further, we now train CLUE and MatchCLOT with their own code and evaluate them on the same test cells with our metrics (Supplement E.9, Table S24): on the rank of the true partner our ensemble is in their range, slightly behind MatchCLOT and ahead of CLUE. The competition score has known limitations (Rautenstrauch, 2026, PhD dissertation), and we postulate that differences of this size do not meaningfully separate matchers. BASILISC provides an explicit uncertainty on each match, which their plan weights do not, and it keeps most of its accuracy with a tenth of the data, which we do not see for the competing approaches (Supplement E.10). On within-type errors we separate the two kinds because they matter for different uses: a same-type partner still carries the right label, for state-level matching every error counts, and there the posterior's gain is below the threshold we set in advance. Two downstream analyses are added: query-level label transfer, where the uncertainty flags the cells whose transferred label is wrong (E.4, Table S18), and a cell-type agreement check on the assigned pair, which lowers the cross-type share among the retained assignments (E.1–E.2, Tables S12–S13). The cell-level filtering claim is removed.

W3. We agree that a single tissue is a limitation, and we have stated it in the Discussion. The point that Liam carries some lineage information appears correct: the model preserves the embedding's lineage structure and does not create it, and we state this in Section 5.1. However, what BASILISC adds is the cross-modal correspondence: for a type it never saw, its placement next to its lineage relative in the other modality follows from the learned similarity, not from either embedding alone. Also, the claim about cells in transition states is only kept as a hypothesis in the Discussion (Supplement E.7, Table S22).

Q3. We now choose the Sinkhorn temperatures on the validation nuclei (cf. Supplement E.5, Table S20). For abundance, we evaluate the cell-level novelty of each held-out type against only the seen types with a comparable number of competitors. Please see Supplement E.6, Table S21.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers and summary of changes for AC

We thank the three reviewers for their time and care in their feedback. We're glad that all three reviewers found our work useful and the presentation clear. We have replied to each point below and marked the changes in blue in the manuscript.

BASILISC provides meaningful uncertainty estimates on the correspondence between two modalities, at the cell-type level and the cell level. Our results concern two aspects: First, how well the model matches, which is seen at the cell type level and, compared to current approaches namely CLUE and MatchCLOT, at the cell level - for every cell type, with a cell type or a lineage held out and with less training data. Second, what the uncertainty of each approximation (MFVI, MC dropout, the deep ensemble) provides as a tangible advantage. We evaluate this against several controls: a known ground truth, a deterministic control of the same architecture for every number, random rejection, and explicit baselines.

Our additional evaluation of cell-level BASILISC against the published matchers CLUE and MatchCLOT showcases the advantages of BASILISC: in low-data regimes, and in flagging the cell types not seen in training through uncertainty quantification (Supplement E.10). Trained on a tenth of the data or less, BASILISC largely retains its accuracy while MatchCLOT is strongly affected; on the full training set our ensemble is in the range of the other models (Supplement E.9).

We did our best to provide all additional analyses the reviews asked for, each with defined baselines. In addition to the model, the reviewers appreciated that we define an evaluation protocol: a way to evaluate uncertainty in a field where the data are noisy and ground truth is rarely available.

We hope the revised manuscript can convince the reviewers to raise their scores. We chose ICLR as we believe the challenges that single-cell biology poses to representation learning deserve the attention of this community, and our work can spark curiosity in the CS audience to explore problems in biology. We would appreciate the opportunity to do so.

Changes made to the manuscript in the revised version:

  • New evaluations at the cell-level show BASILISC's superiority over the state-of-the-art published matchers in low-data regimes. On the full dataset our ensemble is in the same range as CLUE and MatchCLOT. Yet, with a tenth of the training data our model is far ahead of MatchCLOT, and it provides an uncertainty score that is higher for cells of unseen types, which can help to separate them from the seen types (E.9, E.10).
  • Additional analyses the reviews asked for. A fusion of deterministic scores is evaluated as an additional uncertainty score and a cost table for different UQ methods is reported (E.3); an inductive backbone, Liam refit without the held-out cell type, and query-level label transfer (E.4); Sinkhorn temperatures chosen on validation (E.5); size-matched novelty (E.6); the rank of the true partner as an additional metric for assessing the cell-level model (E.9); the CITE-seq cell level experiments (D.2); a data-protocol table (E.8).
  • Claims clarified to be in line with the evidence. We now examine cell-level filtering as a diagnostic, cell-type level filtering benefits are clearly explained, and deterministic checks that lower the cross-type share are added (E.1–E.2); the definition of "unseen" is clarified to concern the matcher's supervision and not Liam's (Section 4); the claim about cells in transition states is rephrased as a hypothesis in the Discussion (E.7).
  • Corrections. Abstention removes 70% of the errors at 80% coverage (the submission's 63% was the drop in the error rate); the cell-level test set corresponds to the competition set of 20,009 nuclei (E.8).

Dear Reviewers and Area Chair,

As the discussion period comes to its end, we would like to briefly highlight a positive update regarding the reviewers' technical concerns. We have provided all the additional controls, baselines, and clarifications requested by the reviewers. Reviewer RmTW has explicitly confirmed that these new experiments and reported outcomes successfully answer all four of their technical questions. Because Reviewer X3NT shared very similar technical concerns, we hope the comprehensive updates to our manuscript and supplement effectively resolve their core critiques as well.

Importantly, we wish to reiterate that BASILISC demonstrates clear predictive superiority in low-data regimes and provides a practically useful uncertainty score for cross-modal matching—a capability no other current model offers.

We are grateful for the reviewers' feedback, which has genuinely strengthened the paper, and we thank you for your time in handling our submission.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.