acceptodds
Under review as a conference paper at ICLR 2027

SILTA: Reference Learning and Target Adaptation for Spatial Transcriptomic Deconvolution

Abstract

Most sequencing-based spatial transcriptomics platforms pool multiple cell-types per capture location, requiring deconvolution to estimate cell-type proportions. Existing methods suffer from spot compositions that are sparse and long-tailed: one type often dominates, yet many methods predict smooth mixtures, and similar gene signatures blur related types. Although neighboring spots share spatial context, spot-wise inference is sensitive to local noise, and these errors accumulate into sample-level composition drift. Platform differences complicate applying composition mappings learned from scRNA-seq references to unlabeled ST targets. To this end, we propose a Two-stage learning model SILTA, Spatial Inference by combining composition Learning and Target Adaptation, which jointly tackles issues above. Stage I learns fine-type boundaries from structured, diffusion-augmented pseudo-spots with the spatial encoder frozen. Stage II refines target proportions from expression and tissue geometry through hierarchical parameter updates, gradual unfreezing, and reference replay. Expression-spatial fusion combines reference composition with target context, while qNB and post-hoc calibration align predictions with measured expression and correct global composition bias. SILTA achieves the best mean for all four primary metrics among evaluated methods on seqFISH, osmFISH and cross-animal Moffitt. We verify both stages' roles through perturbation experiments. Without ground-truth proportions, SILTA demonstrates strong performance in marker-gene concordance across three datasets and in pathology-region classification on HBC, extending its gains to real tissues across assay platforms. The code is at: https://anonymous.4open.science/r/SILTA-SpaDeconv-A38C/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, Cell-type deconvolution, Target adaptation, Two-stage learning, Reference knowledge preservation

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 2 reviews, above 27% of the others.
Reviewer_c7f8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

SILTA proposes a two-stage reference-based spatial transcriptomics deconvolution framework. Stage I learns cell-type composition mappings using structured and diffusion-augmented pseudo-spots, while Stage II adapts the learned representations to unlabeled target tissues through spatial information, expression alignment, and reference replay. A quasi-negative-binomial objective and hierarchical post-hoc calibration are used to improve composition estimation. Compared with the other submissions, this paper provides relatively extensive experimental evidence, including matched ablations, perturbation recovery, reference-retention analysis, and downstream validation.

Critical strengths
  1. The problem is defined fairly accurately. Given the sparse, long-tail composition of cell types and the shift in the reference-to-target distribution, the challenges identified are genuine and significant.

  2. The experiments are fairly comprehensive. The method is compared with 10 other approaches, and four key metrics are reported for Moffitt, osmFISH, and seqFISH.

Critical weaknesses
  1. Although the approach is effective overall, it is overly complex, resulting in a lack of clear attribution to the core innovative contributions. While each module is valid in its own right when considered in isolation, the authors have not sufficiently explained which one constitutes the true major algorithmic breakthrough. In particular, the matched experiment in Table 1 shows that, on the Moffitt dataset, the structured curriculum yields only a limited improvement over random aggregation. In contrast, the effects of qNB and Stage II adaptation, as described in the appendix, are more significant. The authors need to clarify what the truly indispensable core innovation is.

  2. Cell-type deconvolution is inherently an ill-posed problem. For example, OSM contains 33 genes and 31 cell types. For cell subtypes with similar expression profiles, this setup may suffer from a severe lack of distinguishability. While a seemingly reasonable prediction rate can be achieved using only a sparse composition prior, spatial smoothing, and qNB reconstruction, this does not guarantee that different cell types will be correctly separated.

  3. Appendix A.2 clearly states that the known proportion assessments for OSM and seqFISH are based on constructing spots by aggregating imaged cells into spatial bins. While this approach is reasonable for obtaining ground-truth composition, its data generation mechanism is not equivalent to that of actual technical platforms such as Visium. Real sequencing-based spots also involve complex factors such as capture efficiency, RNA diffusion, variations in sequencing depth, and platform-specific biases. Therefore, although the authors provide marker concordance and pathology-region classification in real tissue, these are indirect validations and cannot substitute for quantitative cell proportion assessments based on real sequencing-based spots.

Questions for authors

see the weaknesses

2 comments

We thank the reviewer for recognizing the importance of the problem and the breadth of the experimental evaluation, and for thoroughly reviewing our paper. We hope to address each of your questions.

W1. What is SILTA’s indispensable core contribution?

Thanks for your valuable suggestion! SILTA's core algorithmic contribution is to construct complementary training signals from reference single-cell data and target spatial data, and to coordinate target adaptation and reference supervision through role-specific parameter updates.

Compared with adaptation methods such as AddaGCN (Ding et al., 2026), which use domain-adversarial feature alignment, SILTA uses qNB on the target side to measure the discrepancy between expression explained by the predicted composition and observed expression. Together with spatial information, this guides proportion predictions to adapt to the target tissue. On the reference side, pseudo-spots labelled with cell-number fractions provide direct composition supervision. Parameter protection, differentiated branch updates and configuration-dependent reference replay control how target expression and spatial information modify the reference-learned proportion mapping, thereby implementing target adaptation constrained by reference composition supervision.

Matched experiments validate the roles of target correction and reference supervision.

Dataset / comparisonMetricBeforeAfterTarget accuracy
Moffitt: Stage IITarget RMSE ↓0.05310.0408RMSE decreases
External-reference seqFISH: replayReference RMSE ↓0.1024500.064365Similar target accuracy

This shows that reference supervision can continue to constrain the expression-to-proportion mapping during target adaptation, and its role needs evaluation separately from target accuracy.

Structured pseudo-spots strengthen reference supervision in this framework by expanding labelled composition coverage. They actively increase training examples of balanced, near-pure and confusable compositions, providing more complete proportion supervision for Stage I to learn different types and composition boundaries, and continuing to participate in adaptation when replay is enabled. Matched experiments show gains on seqFISH.

seqFISH structured-sampling endpointRelative change
Rare dominant-type F138.86% increase
Proportion-vector RMSE at rare-type-rich locations6.31% reduction

Please refer to the blue-highlighted content in Section 1, Tables 1 and 16, and Appendices F.8–F.9 in the revised manuscript.

W2. Can SILTA distinguish similar and low-abundance cell types under a limited gene panel?

Thanks for your valuable suggestion! Direct evaluation against target cell-composition ground truth supports improved recovery of similar and low-abundance types. We evaluate the ability to distinguish cell types using per-type spatial PCC, rare-type average precision, and ordering accuracy for reference-defined similar-type pairs.

DatasetPer-type spatial PCC ↑ (rank)Rare / low-abundance type AP ↑ (rank)Similar-type ordering accuracy ↑ (rank)
OSM0.683330 (1)0.634364 (1)0.859779 (1)
Moffitt0.714091 (1)0.542434 (1)0.828850 (1)
seqFISH0.804538 (1)0.688331 (2)0.878911 (1)

These results support improved discrimination of similar and low-abundance cell types under the measured gene panels.Pathology-region classification also supports the utility of these predictions.

Please refer to the blue-highlighted content in Section 3.1 and Appendix F.7 (Table 48) in the revised manuscript.

W3. Does SILTA have quantitative validation on real sequencing-based spatial spots?

Thanks for your valuable suggestion! We evaluate quantitative deconvolution on real sequencing spots from the DBiTplus human lymph-node dataset and visualize the results.

DBiTplus evaluation itemSetting
Model inputRaw counts and coordinates from 50 μm sequencing spots
ReferenceExternal lymphoid scRNA reference
Composition referenceSame-section CODEX cell segmentation and annotations
Evaluation coverage2,244 common spots; 19 cell types

We evaluate deconvolution on this dataset, and SILTA achieves the best mean on each of the four metrics compared with ten methods.This experiment directly compares sequencing-based proportion estimates with imaging-derived compositions. Independent scRNA-to-MERFISH and scRNA-to-osmFISH experiments test cross-platform transfer; controlled expression, RNA and capture perturbations test recovery under observation changes, and input-gene exclusion tests recovery with less expression information.

Please refer to the blue-highlighted content in Section 3.2 and Appendices F.2 and F.9–F.10 (Tables 36–38 and 50–52) in the revised manuscript.

Dear Reviewer c7f8,

Thank you for your constructive comments and suggestions. Our revised manuscript clarifies SILTA’s core contribution and includes quantitative deconvolution experiments on real sequencing-based spatial spots. We also provide additional evidence supporting SILTA’s ability to correctly distinguish similar cell types.

We would appreciate your review of our responses and revised manuscript when you have a chance. Please let us know whether any concerns remain; we welcome further questions and suggestions.

Thank you for your time and consideration.

Best regards,

The Authors

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

This paper introduces SILTA, which is a two-stage framework for reference-based spatial transcriptomics deconvolution. In Stage I, the model learns cell type proportions from synthetic pseudo-spots generated using annotated scRNA-seq references. The authors propose a composition-aware sampling strategy and diffusion-based expression augmentation to improve coverage of rare and closely related cell types. In Stage II, the model adapts to unlabeled spatial transcriptomics data with a spatial encoder and reference-preserving parameter updates.

Critical strengths

SILTA achieves competitive or leading results against ten established deconvolution methods, including RCTD, cell2location, CARD, and Tangram.

Critical weaknesses
  1. The proposed method, especially stage 1, assumes that a mapping learned from reference-derived pseudo-spots can recover cell-type proportions in real spatial tissues. However, the underlying inverse problem is generally ill-posed. Spatial gene expression depends not only on cell-type proportions but also on cell numbers, cell-type-dependent RNA abundance, and spatially varying transcriptional states within each cell type.
  2. The model predicts cell-type proportions without explicitly characterizing uncertainty arising from ambiguous mixtures, missing reference cell types, or spatially varying expression programs. Evaluations involving missing reference cell types, novel cell states, and systematic within-type expression shifts would help determine the robustness of SILTA.
  3. Structured pseudo-spot sampling's improvement over random aggregation is marginal as shown in table 1. Similarly, distance bias and the auxiliary dominant-type embedding both seem to have extermely limited contribution to the performance gain. Without further justification, it is hard to determine if these modules are meaningful.
Questions for authors
  1. Can the authors provide a controlled experiment in which cell-type proportions remain fixed while within-type gene expression varies spatially, and evaluate whether the inferred proportions remain stable?
  2. How robust is SILTA when cell types differ substantially in total RNA abundance or capture efficiency?
  3. How well does SILTA generalize when the reference and target are independently collected using different measurement platforms, or when the reference is not perfect i.e., missing cell types in the reference?
  4. Can the authors clarify the necessity of the modules mentioned in W3?
3 comments

We thank the reviewer for carefully reviewing the paper and recognizing SILTA's performance relative to existing methods. We hope to address each of your questions.

W1 / Q1. How does SILTA address the ill-posed mapping, and how stable are predictions under spatial expression changes at fixed composition?

Thanks for your valuable suggestion! We conduct fixed-composition spatial-expression perturbation experiments on Moffitt, OSM and seqFISH. We keep cell labels, cell numbers, spot membership and coordinates unchanged, and redistribute expression through spatially varying type-specific gene programs while preserving each cell's total count. All methods receive the same perturbed inputs, and we compare their final predictions after training and adaptation.

Changes in proportions need to be interpreted against composition ground truth: clean predictions already contain error, so a change can either correct or worsen that error; small degradation can also accompany poor initial accuracy. We therefore evaluate prediction changes, relative degradation and absolute recovery accuracy together.

We use the mean per-spot L1 distance between clean and perturbed proportion vectors to calculate the absolute differences between the proportions predicted before and after perturbation.

DatasetPrediction change L1 ↓Stability rankRMSE rankJSD rankSSIM rankPCC rank
Moffitt0.1382863/111111
OSM0.1913222/111111
seqFISH0.6415523/111111

These results directly quantify the stability of predicted proportions under perturbation across the three datasets. Although our model does not show the smallest change in predicted proportions, its stability remains competitive; together with its strong predictive performance, these results further demonstrate SILTA's stability.

Please refer to the blue-highlighted content in Section 3.2 and Appendix F.2 (Table 36) in the revised manuscript.

W2. How robust is SILTA when under uncertain factors such as ambiguous mixtures, missing reference types, novel states and within-type expression changes?

Thanks for your valuable suggestion! We examine the factors raised in your comment through missing reference types, incomplete within-type state coverage, spatial expression-program changes and input-gene exclusion. Cells entering new states can express different combinations of genes, making some reference-derived expression cues less reliable; we exclude part of the input genes to introduce missing expression information and test recovery with fewer available cues. On Moffitt, we separately remove common, rare or confusable reference types. OSM additionally removes the highest-scoring quarter of reference cells within each type according to an expression-program score, directly testing performance when the reference incompletely covers target expression variation.

We retrain all methods using the same remaining panels.

DatasetExcluded genes / full panel
Moffitt27/135
seqFISH70/351
OSM7/33

Together, these experiments test SILTA's robustness under the examined sources of uncertainty, with SILTA achieving the best mean on each of the four metrics across these experiments.

Please refer to the blue-highlighted content in Section 3.2 and Appendices F.2–F.4 (Tables 36 and 39–41) in the revised manuscript.

Q2. How robust is SILTA to differences in RNA abundance and capture efficiency?

Thanks for your valuable suggestion! We perturb expression counts while keeping cell labels, cell numbers, spot membership and coordinates unchanged. Type-specific multipliers change effective RNA contributions, and binomial thinning simulates capture-efficiency loss.

DatasetRNA-contribution perturbationCapture condition
MoffittApproximately 0.861–1.162×Global: 50% of counts retained
OSMApproximately 0.50–2.01×Global: approximately 50% of counts retained
seqFISHStrength 0.7Global: 30% of counts retained
seqFISH (type-specific capture)—Type-specific capture, strength 0.7

A separate seqFISH type-specific capture experiment samples and normalizes random coefficients for each type to simulate differences in capture efficiency between types (Table 38(b)).

Across these RNA-contribution and capture experiments, SILTA achieves the best mean on each of the four metrics, supporting proportion recovery under the tested perturbations.

Please refer to the blue-highlighted content in Section 3.2 and Appendix F.2 (Tables 37–38) in the revised manuscript.

Q3. How well does SILTA generalize across independently measured platforms and incomplete references?

Thanks for your valuable suggestion! We now use independent scRNA references to evaluate both imaging and sequencing targets.

DatasetTransfer / evaluation setting
MoffittGSE113576 scRNA-seq → MERFISH; six shared coarse types
OSMGSE60361 → osmFISH
DBiTplusReal sequencing spots and same-section CODEX-derived composition references

Across the three experiments comparing eleven methods, SILTA achieves the best mean on each of the four metrics.

We conduct the following missing-reference-type experiments: on Moffitt, we separately remove common, rare or confusable reference types; On OSM, we remove common or rare-confusable reference types and reduce within-type state coverage by excluding the highest-scoring quarter of cells within each type, directly testing performance when the reference incompletely covers target expression variation. SILTA achieves the best mean on each of the four metrics in all six Moffitt and OSM reference-deletion/state-coverage conditions, supporting recovery with incomplete references.

Please refer to the blue-highlighted content in Section 3.2 and Appendices F.3 and F.9–F.10 (Tables 39–40 and 50–52) in the revised manuscript.

W3 / Q4. What value do structured pseudo-spots, distance bias and the auxiliary dominant-type embedding provide?

Thanks for your valuable suggestion! Cell-type deconvolution is inherently an ill-posed problem. For example, OSM contains 33 genes and 31 cell types. This setup may provide insufficient discrimination between subtypes with similar expression profiles. A seemingly reasonable prediction rate does not guarantee correct distinction between different cell types. RMSE and the other metrics capture only one aspect of this design. We compare the improvement in fine-type discrimination from structured versus random pseudo-spots across three datasets.

DatasetSource macro-type RMSETarget Rare-F1 (identification of rare dominant types)Rare-rich spot RMSE (quantitative deconvolution in mixed tissue enriched for rare types)
Moffitt3.44% improvement3.57% improvement1.03% improvement
OSM1.85% improvement6.79% improvement0.13% improvement
seqFISH5.75% worsening38.86% improvement6.31% improvement

This trade-off is consistent with structured sampling changing the training composition distribution, creating a trade-off between fitting ordinary reference mixtures and recovering rare types in the target. Overall, the effect is more positive. These results support its contribution to target-side rare-type recovery.

The matched HBC sampling comparison also improves marker concordance for the macrophage/DC/monocyte population on the Marker-A panel. This demonstrates that structured pseudo-spots expand labelled coverage of balanced, near-pure and confusable-type compositions; bounded diffusion adds expression variation while preserving proportion labels.

Distance bias. We agree that distance bias adds only a small gain to the four primary metrics and is not our main contribution. Distance bias is the default way for E2 to receive coordinates; the E2/g pathway itself contributes most to target adaptation. The model also applies smoothness and BATV constraints on the spatial graph, through which geometric information can enter the model, so these mechanisms overlap.

Auxiliary dominant-type embedding. This module has two roles. First, the representation learned by m itself contains type information: a linear probe trained on m alone gives the following dominant-type balanced accuracy (Table 28).

Datasetm-only probeRandom level
Moffitt0.5590.022
OSM0.4280.032

Auxiliary dominant-type supervision trains E1/h to learn type boundaries during Stage I and is an important training objective and information source for that stage. The revision therefore describes the direct fusion path of m as optional and auxiliary supervision as a Stage I training objective. Second, the path by which m directly enters fusion is optional (Section 2.6; disabled when λ_m = 0). Its direct effect is indeed small, mainly because the first role already supplies the main information, leaving a limited additional gain.

Please refer to the blue-highlighted content in Sections 1 and 3.1, Appendix F.8 (Table 49), and Tables 22–23, 27–29 and 33 in the revised manuscript.

Dear Reviewer dZjL,

Thank you for your thoughtful comments and suggestions. Our responses and revised manuscript include controlled robustness experiments and cross-platform evaluations addressing your concerns about proportion recovery under expression changes and incomplete references. We also expand the discussion of structured pseudo-spots and auxiliary modules.

We would appreciate your review of our responses and revised manuscript when you have a chance. Please let us know if we have adequately addressed your concerns. We are always open to further suggestions.

Thank you for your time and consideration.

Best regards,

The Authors

Comments on the paper

7 comments

We sincerely thank the reviewer for recognizing the model design and for the detailed suggestions on methodological details and evaluation protocols. We hope to address each of your questions.

W1. Why do update schedules, replay, bilinear fusion and structured sampling have dataset-dependent effects?

Thanks for your valuable suggestion! Differences in cell-type heterogeneity, expression states, gene panels and capture characteristics motivate multiple Stage II update schedules within our unified framework. These schedules balance target adaptation with reference retention, so their benefits vary across datasets and evaluation endpoints.

On Moffitt, the update strategies differ mainly in reference retention, whereas OSM also shows a clearer difference in target accuracy. On OSM, spatial-only adaptation also improves target RMSE relative to direct joint adaptation.

DatasetMetricDirect joint adaptationReplaySpatial-only interval + replay
MoffittReference RMSE ↓0.05360.04470.0448
MoffittTarget RMSE ↓0.0407890.0408030.040804
OSMReference RMSE ↓0.076640.073460.07358
OSMTarget RMSE ↓0.068630.066360.06650

On OSM, direct joint adaptation already reduces reference RMSE from 0.07962 to 0.07664, and replay further improves reference retention while allowing joint adaptation. Adding a spatial-only interval to replay provides no additional mean improvement in this setting. A protected interval temporarily restricts reference-path updates, whereas replay supplies reference supervision during joint updates; these mechanisms address related needs, so their benefits need not add together.

In contrast, ordinary joint adaptation on seqFISH causes clear reference forgetting; adding a protected interval reduces reference RMSE. Adding protection to replay also reduces target RMSE and improves the means of the other three metrics.

seqFISH comparisonMetricBefore protectionWith protection
Ordinary joint adaptation → protected intervalReference RMSE ↓0.083270.06752
Replay → replay + protectionTarget RMSE ↓0.068630.06828

Rare-F1 combines precision and recall for identifying rare types as the dominant type at target spots. In a separate matched, uncalibrated seqFISH comparison (three splits × three seeds), bilinear fusion increases rare dominant-type F1, while the four overall fraction metrics, rare-type AP and rare-rich spot fraction error show no mean improvement, and the F1 gain is not significant after Holm correction.

seqFISH metricAdditive fusionBilinear fusionRelative change
Rare dominant-type F1 ↑0.2761560.296302+7.29%; all splits
seqFISH structured-sampling endpointRelative improvement
Target Rare-F1 (rare dominant-type identification)38.86%
RMSE of the full proportion vector at rare-type-rich locations6.31%

Reference retention and rare-type identity recovery thus complement overall proportion accuracy.

Please refer to the blue-highlighted content in Tables 14, 16, 18–19 and 21–24, Appendix F.8 (Table 49), and Appendix F.5 (Tables 43–45) in the revised manuscript.

Q1. What are the training, replay, calibration and reference-selection rules, and how do you select adaptation settings without target labels?

Thanks for your valuable suggestion! The training procedure combines reference validation with held-out target-expression selection. E1 uses multi-head self-attention without distance bias. The graph-masked variant restricts attention to expression-graph edges, while the dense variant attends to all input pairs. E2 uses dense multi-head attention with learnable Gaussian distance bias. Distances are normalized by the median nearest-neighbor distance. Stage I uses fraction KL, enabled identity supervision and enabled Gaussian regularization. Stage II uses qNB, spatial smoothness, boundary-aware total variation and enabled regularization, while replay supplies reference proportion supervision.

Each training step performs one optimization update, with losses and trainable parameters determined by the stage and event type. Stage I freezes the spatial branch at g = 0 and trains the expression encoder and corresponding identity or proportion readout using reference supervision. For example, during seqFISH's 300-step joint adaptation, 38 replay events occur uniformly within the final 150 steps; each adds reference proportion KL with weight 0.1 to the current target loss. Reference KL constrains the expression-side mapping while the target loss continues to train the spatial branch.

After Stage II, we freeze the backbone and fit calibration parameters using reference pseudo-spots and their known proportions, comparing no calibration, ungrouped type-specific biases and type-specific biases with group-shared temperatures. Across datasets, the final calibrated proportions follow:

pikcal=[softmax(Cend(ℓi)/τinf)]k.p_{ik}^{\mathrm{cal}} = [\mathrm{softmax}(C_{\mathrm{end}}(\ell_i)/\tau_{\mathrm{inf}})]_k.

Here, ℓik\ell_{ik} denotes the uncalibrated logit for cell type kk at spot ii. The reference-fitted affine map CendC_{\mathrm{end}} combines group-shared temperatures and the enabled type-specific corrections; τinf\tau_{\mathrm{inf}} is the fixed inference temperature.

Here, c(k)c(k) follows preset biological groups from reference cell-type annotations. Calibration uses 300 training steps at a learning rate of 5×10−45\times10^{-4}, with proportion KL, type-weighted Huber loss, overall type-proportion error and identity regularization without target cell-proportion labels.

The unified Stage I specification ranks checkpoints separately by reference RMSE, JSD, SSIM, PCC and Rare-F1, then selects the valid checkpoint with the lowest mean rank across the five metrics. Ties favor lower RMSE and then lower JSD. Validity requires finite ranks and a pseudo-prediction standard deviation of at least 0.001. Adaptation comparisons start from their respective previously trained common source checkpoints.

Stage II holds out approximately 20% of target genes, stratified by reference expression, and excludes these columns from target inputs, preprocessing, expression graphs and adaptation losses. We score held-out expression using unweighted qNB negative log-likelihood, with fixed reference signatures and dispersion parameters and gene-response multipliers fixed at one. We average scores across three seeds and select the lowest-scoring configuration for each fold.

DatasetRMSE of the selected configuration ↓
Moffitt0.038068
seqFISH0.065997
OSM0.064570

On Moffitt, the rule identifies a configuration that improves on the previous result in all four mean metrics; on seqFISH and OSM, it stays very close to their strong fixed configurations. These results support the reliability of our practical configuration-selection procedure without target proportion labels.

Please refer to the blue-highlighted content in Sections 2.3–2.7 and 3.2, Table 4, Appendix B, and Appendix F.1 (Table 35) in the revised manuscript.

Q2. How are biological endpoints and marker panels chosen, and are they fixed before performance evaluation?

Thanks for your valuable suggestion! In this type of experiment, we generally choose cell populations according to the tissue-specific research question, then define marker panels using the literature, reference annotations and gene availability. For Melanoma, we focus on the spatial distributions of macrophages and T cells and use existing literature to determine the relevant markers. For HBC and DLPFC, we select five genes per type using gene-detection rates in the reference and type-versus-rest expression thresholds, then check whether those genes occur in the target data. Target data are used only to confirm gene availability, not to select markers according to method scores; the panels are therefore fixed before examining method performance.

Please refer to the blue-highlighted content in Appendices A.9 and A.11, Appendix F.7 (Table 48), and Appendix F.8 in the revised manuscript.

Q3. Do evaluation genes enter training inputs or reference signatures, and does relative performance persist after their exclusion?

Thanks for your valuable suggestion! The original measured genes are model inputs and reference features. We retrain all compared methods after excluding the same fixed genes within each fold. SILTA still achieves the best mean on each of the four metrics in the reported comparisons.

DatasetRetained genes / full panel
Moffitt108/135
seqFISH281/351
OSM26/33

We also conduct an evaluation-gene exclusion experiment on a Melanoma section. We remove 14 evaluation genes covering all three cell types from the fitting features of SC, ST and pseudo-spots. We access their observed ST expression only at evaluation, calculate per-gene PCC and Spearman correlations with the predicted cell fractions, and then average equally within each cell type. After exclusion, SILTA still exceeds all ten baselines in both PCC and Spearman correlation for Endo, NK and T cells under the same evaluation conditions.

Please refer to the blue-highlighted content in Appendix F.4 (Tables 41–42) and Appendices A.9 and A.11 in the revised manuscript.

Q4. When do qNB gene-response parameters β_j update, including the spatial-only interval, and how sensitive are results to b_max?

Thanks for your valuable suggestion! When the qNB gene-response parameter β is enabled, we initialize it at one, so that it initially leaves reference gene expression unchanged. β is shared across all spatial spots in the same target dataset and learns through the qNB loss on observed target expression without using ground-truth target cell fractions. When enabled, β is optimized by the target-expression qNB loss during Stage II, including the spatial-only interval. β uses a separate optimizer. A bounded log response and regularization constrain its deviation from initialization, and reference replay does not directly update β. After training, β remains fixed during post-hoc calibration and inference.

DatasetBound settingsRelative changes in four primary mean metrics
Moffitt0, 0.05, 0.10, 0.25, 0.50Below 0.17%, both before and after calibration
seqFISH0.1 and 0.3All four metrics slightly degrade; changes below 0.6%

Please refer to the blue-highlighted content in Appendices A.5 and B.4 and Appendix F.6 (Tables 46–47) in the revised manuscript.

Q5. How are calibration groups defined, and how does calibration perform without grouping or with incorrect groups?

Thanks for your valuable suggestion! Calibration groups are manually defined from reference cell-type annotations. Ungrouped calibration already improves all four mean metrics on Moffitt, OSM and seqFISH.

DatasetCell typesGroups
Moffitt457
OSM317
seqFISH219

With the same frozen backbone and reference fitting budget, adding group-shared temperatures further improves RMSE, SSIM and PCC on all three datasets. JSD improves slightly on seqFISH and increases slightly on Moffitt and OSM.

OSM comparisonRMSEJSDSSIMPCC
Ungrouped biases: paired mean improvement3.04% improvement0.40% improvement1.99% improvement2.07% improvement
Additional group temperatures2.19% improvement0.66% increase0.29% improvement0.54% improvement

Incorrect-group experiments preserve group counts, group sizes, parameter counts and fitting budgets; we average results over five mappings at each perturbation level.

DatasetReassigned typesEffect
OSM16/31All four means slightly improve; changes below 0.2%
Moffitt14/45Each mean changes by less than 0.35%
seqFISH10/21Somewhat more sensitive: RMSE increases by 2.05%; RMSE, SSIM and PCC remain better than without calibration

These results support group-shared temperatures as a useful parameter-sharing design that retains calibration benefits under the tested grouping errors.

Please refer to the blue-highlighted content in Appendix F.5 (Tables 43–45) in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.