- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes ATLAS S2V, a reference-conditioned method for estimating 3D cell-density and cell-type distributions from one observed 2D tissue section and a spatially registered organ atlas. It calibrates atlas abundances to the observed specimen, forms a spatial residual relative to the reference, and transfers that residual with a distance-dependent two-expert gate. The authors report brain reconstruction on held-out mice, organ-specific extensions to lung, breast, and kidney, and separate atlas-conditioned readouts for sparse-panel gene imputation and cell typing. The output consists of virtual, grid-based distribution estimates rather than the actual positions of unseen cells.
Critical strengths
- Predicting unobserved tissue distributions when only one target section is available is relevant to valuable or archived spatial-omics specimens. An external anatomical atlas is a sensible prior for this underdetermined problem, and the paper acknowledges that exact unseen cells cannot be recovered. 2.The abundance calibration, residual correction, and distance-aware mixture are simple and understandable. The paper reports comparisons with central-copy, atlas-only, nearest-reference, and U-Net alternatives, as well as held-out-animal and multiple-organ experiments. The sparse-panel analyses also offer evidence that anatomical atlas context can be helpful.
Critical weaknesses
- The added specimen-specific 3D information beyond the atlas is not convincingly isolated. The results suggest that the reference atlas may explain much of the depth-dependent structure. In Table 1, the class Pearson of ATLAS S2V and atlas-only is identical at 2,000 micrometers (0.553) and nearly identical at 800 micrometers (0.539 versus 0.538). In Table 3, the full model achieves 0.5983, compared with 0.5737 without residuals and 0.5953 with fixed 1:1 fusion. Although the 60-mouse density study reports a stronger routing benefit, the primary reconstruction results leave the individual-specific contribution and necessity of the proposed gate uncertain. The 3D visualization in Fig. 2 does not clearly separate atlas-derived anatomy from newly inferred individual structure. The manuscript should quantify the incremental spatial information, provide atlas-minus-prediction and prediction-error maps, and avoid implying recovery of a uniquely determined 3D specimen.
- Important claims are insufficiently separated from reference availability, registration assumptions, and the evidence provided by the downstream tasks. The main brain benchmark uses supplied CCF pose information (Sec. 3.2), while the released count-field interface assumes that the observed section has already been registered and converted into cell-count fields; raw H&E is not itself an input to this interface. The main results therefore do not establish robustness to realistically estimated translation, rotation, or local registration errors. In addition, the downstream sparse-panel studies use separately fitted atlas-conditioned readouts. Some auxiliary experiments consume virtual 3D features, but the main downstream evidence does not by itself show that reconstructed hidden-section volumes improve the task. The more fundamental issue is the lack of published, domain-standard external baselines. The current comparisons are largely author-defined controls, including panel-only inputs, 2D versus 3D atlas variants, shuffled coordinates, central extrusion, and simple linear readouts. These controls are useful for isolating the contribution of atlas information, but they do not establish competitiveness against existing spatial-omics methods. The authors should compare with the strongest relevant published baselines for each downstream task, including Nicheformer where its input, supervision, and prediction target can be matched; where a direct comparison is infeasible, they should explain the task mismatch explicitly. Finally, the organ studies use different task-specific fitting protocols, so they support adaptability of the general principle more than cross-organ generalization of one fixed model.
Questions for authors
Q1. What unique, specimen-specific information does ATLAS S2V recover beyond a calibrated atlas? Please report gains over atlas-only at each held-out animal or section and distance, uncertainty across the four outer leave-one-animal-out folds, and residual or difference maps against matched ground truth. Q2. How would the required section-to-atlas pose be obtained for a new specimen without supplied CCF coordinates? How sensitive are the main predictions to realistic translation, rotation, and local registration errors? Q3. Which downstream experiments directly consume the predicted virtual 3D fields, and which use only observed-section or atlas-derived features? Please distinguish evidence that atlas context assists downstream inference from evidence that the reconstructed volume itself improves downstream performance. Q4. Which published methods provide the strongest relevant baselines for the hidden-gene imputation and panel cell-typing tasks? Please compare with established spatial-omics models, including Nicheformer where its task contract is compatible, using matched specimen-level splits, observed inputs, target genes or classes, and evaluation metrics. If direct comparison is not possible, please explain the mismatch in input modality, spatial resolution, supervision, or prediction target. Q5. Does the anonymous repository include the data-preparation scripts, checkpoints, and evaluation code needed to reproduce the principal results? Please identify which source datasets and intermediate caches must still be downloaded or rebuilt, and report the compute requirements for the main experiments. Q6. The released composition checkpoint appears to have its neural composition correction initialized to zero, leaving the effective mixture weight as a fixed function of absolute distance. Is the reported composition result therefore an analytic distance-gated expert mixture rather than an adaptive learned router? Please compare it with this exact analytic gate and clarify the claims about learned or specimen-conditioned routing.
4 comments
Y2. Pose estimation and registration sensitivity
Response. Across 191 sections, DeepSlice has a median 3D center error of 137.8 µm, with 94.8% within 400 µm. In separate controlled tests, AtlasS2V retains 80.8–100% of its unperturbed Pearson correlation under tested single-axis translations of 0–400 µm.
Obtaining pose without supplied coordinates. Appendix G uses the H&E image, pixel size and Allen CCFv3 template. We search coronal AP planes and in-plane rotations/reflections, refine the best candidates by constrained affine alignment of tissue outlines and internal structure, and freeze the pose before loading expression. Coronal orientation is specified from metadata or anatomical inspection; no point landmarks are supplied. Images determine pose, while molecular measurements supply the count-field adapter.
1. Registration accuracy. We evaluated all 191 sections from four datasets. Center and anterior–posterior (AP) errors are in µm; the last column gives the proportion with 3D center error ≤200 µm. Overall section-normal error was 2.88° (median) and 6.64° (P90).
| Dataset | Sections | Center median | Center P90 | AP median | ≤ 200 µm |
|---|---|---|---|---|---|
| Myelin | 41 | 123.0 | 246.3 | 73.9 | 82.9% |
| PITX3 | 32 | 211.8 | 423.0 | 108.2 | 43.8% |
| Calb1 | 58 | 96.1 | 181.8 | 59.0 | 94.8% |
| Amyloid | 60 | 164.1 | 359.8 | 120.7 | 63.3% |
| Overall | 191 | 137.8 | 297.1 | 77.7 | 73.8% |
2. Translation sensitivity. We tested ±100–1000 µm shifts in 100 µm steps along AP, dorsal–ventral (DV) and folded medial–lateral (ML). The 0–400 µm range covers 94.8% of measured center-error magnitudes; 500–1000 µm shifts probe larger errors.
All perturbation tests use one independent mouse, the frozen model and 193 section pairs, with fixed targets and masks and no retraining or tuning. We average equally over four target distances and both perturbation signs; zero perturbation is counted once. Both sensitivity tables report AtlasS2V Pearson. Retention is Pearson divided by the unperturbed value, 0.5983.
| Translation magnitude (µm) | AP | DV | Folded ML |
|---|---|---|---|
| 0 | 0.5983 | 0.5983 | 0.5983 |
| 100 | 0.5833 | 0.5916 | 0.5979 |
| 200 | 0.5491 | 0.5720 | 0.5776 |
| 300 | 0.5126 | 0.5421 | 0.5478 |
| 400 | 0.4833 | 0.5087 | 0.5113 |
| 500 | 0.3463 | 0.4765 | 0.4744 |
| 600 | 0.3263 | 0.4466 | 0.4341 |
| 700 | 0.3097 | 0.4203 | 0.3972 |
| 800 | 0.2932 | 0.3984 | 0.3661 |
| 900 | 0.3211 | 0.3800 | 0.3464 |
| 1000 | 0.3092 | 0.3641 | 0.3304 |
Pearson retention is 80.8–100% across tested translations up to 400 µm, and 91.8–100% up to 200 µm. At 400 µm, retention is 80.8% (AP), 85.0% (DV) and 85.5% (folded ML).
3. Orientation sensitivity. We tested ±5–45° rotations in 5° steps about AP (in-plane), DV (tilt in the original folded coordinates) and ML (tilt in the AP–DV plane). The pivot is the centroid of each observed section's valid spatial grid. Measured normal-angle errors inform the tilt tests; they do not measure in-plane rotation.
| Angle | In-plane: AP | Tilt: DV | Tilt: ML |
|---|---|---|---|
| 0° | 0.5983 | 0.5983 | 0.5983 |
| 5° | 0.5953 | 0.5874 | 0.5727 |
| 10° | 0.5543 | 0.5374 | 0.5209 |
| 15° | 0.5070 | 0.4891 | 0.4766 |
| 20° | 0.4657 | 0.4451 | 0.4403 |
| 25° | 0.4304 | 0.4084 | 0.4120 |
| 30° | 0.4032 | 0.3815 | 0.3867 |
| 35° | 0.3829 | 0.3613 | 0.3686 |
| 40° | 0.3686 | 0.3438 | 0.3523 |
| 45° | 0.3569 | 0.3307 | 0.3377 |
At 5°, Pearson retention is 99.5% for in-plane rotation and 98.2%/95.7% for DV/ML tilt. Larger rotations progressively reduce performance.
4. Optional serial-section constraints. On the same 159 Myelin, Calb1 and Amyloid sections, combining angles and spacing reduces median center error from 126.1 to 111.0 µm, AP error from 75.9 to 41.3 µm and normal error from 2.88° to 1.36°. PITX3 has two cutting blocks and retains its single-section results.
| Registration input | Center median | AP median | Normal median | ≤ 200 µm |
|---|---|---|---|---|
| Independent single sections | 126.1 | 75.9 | 2.88° | 79.9% |
| Multi-section angle integration | 124.7 | 68.4 | 1.36° | 73.6% |
| Angles + spacing constraints | 111.0 | 41.3 | 1.36° | 83.0% |
Positional errors are in µm; the last column refers to 3D center error. These optional constraints are not used in the 191-section single-section results.
Failure regime. At negative AP shifts of 500 µm and beyond, atlas totals become zero at some central locations, causing excessively large calibration ratios. Sign-averaged WAPE reaches approximately 3.34×10⁹–4.84×10⁹.
These tests assess global translations and rotations. Local nonrigid registration errors have not been evaluated; the Appendix G pipeline uses coronal-plane and affine alignment.
Anonymous GitHub Code: AtlasS2V-code/benchmarks/rebuttal/registration_sensitivity_0_400.zip
Y3. Atlas context versus predicted-volume context
Response. We now evaluate downstream use of the predicted 3D volume for 34-class cell typing with 10, 25 and 50 measured genes. The new readout uses context derived from the predicted volume. The table labels distinguish this evaluation from the original direct-atlas-context readout.
| Method / downstream input | 10 genes | 25 genes | 50 genes |
|---|---|---|---|
| Genes only | 0.2090 | 0.4716 | 0.6366 |
| + 2D context | 0.3058 | 0.5665 | 0.6713 |
| + 3D atlas context (original) | 0.4707 | 0.6620 | 0.7094 |
| + Atlas and predicted-volume context (new) | 0.4616 | 0.6547 | 0.6997 |
Mean macro-F1; bold highlights the new readout.
The first two rows use measured genes alone or genes plus 2D context. The original 3D readout receives atlas context directly; the new readout additionally uses context computed from the predicted volume. Thus the table identifies the source of downstream information for each model.
The new readout improves over both gene-only and 2D inputs at every tested gene budget. At 10, 25 and 50 genes, gains over 2D context are 0.1558, 0.0882 and 0.0284, and gains over genes alone are 0.2526, 0.1831 and 0.0631. Direct 3D atlas context remains higher by 0.0091, 0.0073 and 0.0097.
Both the original and new readouts use atlas context as an additional input, which is consistent with their small performance differences. The new evaluation shows that the predicted volume can supply useful downstream context within the atlas-informed pipeline. This follows the central premise of AtlasS2V: using a biological atlas as a predictive reference when target-specimen measurements are limited, while incorporating observed-section calibration and spatial residuals.
We will distinguish the original atlas-conditioned readouts from this new volume-augmented evaluation in the text, tables and contribution statement. The largest gains over gene-only and 2D inputs occur at ten genes, where measured molecular information is most limited. This gene-budget comparison directly addresses whether the complete volume-augmented readout remains useful when few markers are measured. Its absolute macro-F1 rises from 0.4616 to 0.6997 as the gene budget increases, while the gain over 2D context decreases from 0.1558 to 0.0284.
Y4. Additional benchmark against Nicheformer and other methods
Response. We also completed a matched 10-gene, 34-class cell-typing benchmark against the Nicheformer foundation-model readout and four other published cell-typing baselines. The new AtlasS2V readout outperforms all five external baselines on both macro-F1 and accuracy.
| Method | Macro-F1 ↑ | Accuracy ↑ |
|---|---|---|
| Genes only | 0.2090 | 0.5481 |
| Nicheformer (frozen + linear readout) | 0.1573 | 0.4963 |
| CellTypist | 0.2077 | 0.5465 |
| SingleR | 0.1007 | 0.2533 |
| scmap-cluster | 0.1162 | 0.3756 |
| TACCO (abundance adaptation) | 0.1348 | 0.4012 |
| AtlasS2V: atlas + predicted-volume context | 0.4616 | 0.6512 |
| Direct 3D atlas context (original control) | 0.4707 | 0.6832 |
The comparison uses four-animal leave-one-animal-out evaluation and three seeds, with the same classifier training/test cells. Scores are averaged over seeds within each animal and then equally across animals. Rejected or unassigned cells remain in the evaluation. External methods receive no additional explicit 3D context. Nicheformer uses frozen representations and a linear readout trained to convergence; this benchmark does not evaluate end-to-end fine-tuning.
AtlasS2V's macro-F1 of 0.4616 exceeds Nicheformer (0.1573) by 0.3043 and the strongest external baseline, CellTypist (0.2077), by 0.2539. The direct-atlas control remains close and slightly higher, consistent with both readouts using atlas context. These results show a substantial advantage for the complete atlas-informed pipeline under this limited-gene protocol.
Reporting both metrics separates average per-class performance from overall cell-level correctness: macro-F1 gives equal weight to each class, whereas accuracy gives equal weight to each evaluated cell. AtlasS2V also improves accuracy over Nicheformer by 0.1549 and CellTypist by 0.1047, so the advantage is present under both summaries.
The tables complement each other: Y3 compares context sources at three gene budgets, while Y4 places the complete pipeline against external methods at ten genes. The direct-atlas control remains visible in both tables to make the contribution of atlas information explicit. Nicheformer is the foundation-model baseline; CellTypist, SingleR, scmap-cluster and TACCO are additional cell-typing methods.
Anonymous GitHub Code: AtlasS2V-code/benchmarks/rebuttal/downstream_matched_04616/
Y5. Reproducibility and compute
Response. The code and frozen artifacts for the principal experiments are already available in the anonymous AtlasS2V-code repository. The relevant components can be located as follows.
scripts/andsrc/contain the data-preparation, atlas-construction, model-fitting, inference, and evaluation implementations.docs/experiments.mdmaps the principal mouse-brain, BrainCAD, organ-specific, and downstream experiments to their preparation and evaluation commands.results/unified_mouse_brain_predictor_fidelity_v2/contains the frozen composition checkpoint and composition reference.results/strict_expression_router_20260921/contains the separately trained expression checkpoints. Artifact identities and provenance are recorded indocs/artifacts.json.docs/data.mdandconfigs/data/specify source datasets, expected directory layouts, and specimen splits. Raw microscopy and expression datasets are obtained from their original public sources. The supplied preparation scripts generate the per-animal reference fields, native-section maps, and large expression/training caches required by the corresponding experiments.requirements-tested.txt,docs/validation.json, andscripts/verify_release.pyprovide tested dependency versions, recorded validation results, and distributed-file checksum verification. Additional released rebuttal experiments, result tables, and reproduction instructions are organized underbenchmarks/rebuttal/.
The count-field API takes registered per-class count maps. Registration and dataset-preparation scripts are in scripts/. docs/model.md specifies the composition inference rule, with numerical parity checks against the supplied PyTorch checkpoint.
Computational resources are now documented in docs/compute.md and docs/compute_downstream.md, with per-job accounting evidence and machine-readable tables under benchmarks/compute_resources/. Together, these records cover 44 workload stages and 243 successful jobs. They report available CPU allocations, runtime GPU models, wall-clock times and host-memory MaxRSS, with workload scopes identifying prepared inputs and reused artifacts. The four production stages for 25/50-gene Nicheformer embeddings and readouts sum to 12.64 wall-clock job-hours across their parallel jobs. The README and experiment guide link directly to these resource summaries.
Y6. Reference-fitted analytic distance gate
Response. The released composition predictor is an analytic distance-gated mixture with zero neural composition correction. Its fusion policy is learned from reference data during fitting and frozen at inference.
For independent-mouse evaluation, each of four reference mice is held out in turn: three build the atlas; the fourth supplies observed-center/hidden-target pairs. We compare a calibrated-atlas expert with an expert that additionally transfers the center's spatial residual. Reference targets determine the mixture weight minimizing normalized MSE at each distance; weighted least squares then fits a smooth, monotone sigmoid. Writing its intercept and positive slope as ,
For C57BL6J-638850, and . Residual-expert weights at 200/400/1,000 µm are approximately 0.674/0.291/0.0032. The independent test mouse's hidden sections do not enter fitting.
This is reference-data-driven adaptation during fitting: expert reliability across distances determines the fusion policy. At test time, the gate is shared across positions and classes at a given distance. Specimen information enters through central-section abundance calibration and spatial residuals, changing the experts and final prediction. We will distinguish this from specimen-adaptive neural routing.
An exact analytic implementation with the same inputs and parameters is this predictor and should reproduce its outputs; released numerical parity checks compare it with the PyTorch checkpoint.
Organ/task fitting differs:
| Task | Weight selection |
|---|---|
| Mouse-brain composition | Reference-animal leave-one-out fitting of the distance sigmoid |
| BrainCAD density | Separately trained gate with marker-specific parameters |
| Kidney | Distance-specific fusion coefficients fitted using designated training donors |
| HER2ST breast | Reference-patient validation selects the residual model and ; is constant within each outer fold, while distance enters the residual predictor |
Code (under AtlasS2V-code/):
scripts/run_allen_langlieb_external_distance_gate.py:fit_distance_gate().scripts/align_unified_composition_fidelity.py:migrate_bundle()stores fitted parameters and zeroes neural correction.results/unified_mouse_brain_predictor_fidelity_v2/freeze_manifest.json: splits, per-distance weights and fitted parameters.docs/model.md: inference rule.
Response to Reviewer yY5R
Thank you for your questions about specimen information, registration, downstream inputs and reproducibility.
Y1. Specimen-specific information beyond a calibrated atlas
Response. The original “Atlas-only” already used the observed central section for abundance calibration; we have renamed it Calibrated atlas. AtlasS2V additionally transfers the spatial residual between the observed section and its calibrated reference. Its contribution is strongest nearby; convergence toward the calibrated atlas at greater depths is expected by design.
Class Pearson on independent mouse C57BL6J-638850, rounded to three decimals.
New Atlas-only uses the reference alone, without central-section calibration. Four-animal leave-one-animal-out evaluation excludes each test animal from atlas construction, fitting and selection. Means equally weight animals and the shared 200/900-µm offsets.
Using an atlas reference for single-section 3D prediction is itself our core innovation. The model is simple, but effective. Its contribution includes reference anatomy, specimen-specific calibration and spatial residual transfer. The gain over an already calibrated atlas measures the residual-transfer component, rather than the full benefit of introducing atlas guidance.
At shared offsets, Pearson gains over calibrated atlas are +0.0065, +0.0190, +0.0314 and -0.0230 for ABCA-1–4: 0.0085 ± 0.0233 (mean ± between-animal sample SD from rounded fold means; Appendix C.1, Table 10). Pearson improves in three animals; normalized count error decreases in all four. JSD changes from 0.0383 to 0.0399. The SD describes animal variation, not a confidence interval. Separately, removing residual transfer lowers Pearson from 0.5983 to 0.5737.
The router serves as an organ-specific adapter, as emphasized in Fig. 2. Fixed 1:1 mixing is already competitive in mouse brain. In breast, anatomical variation can limit information from a shared atlas, motivating greater reliance on the observed section. Calibrating residual strength improves density/composition Pearson from 0.8093/0.6516 to 0.8258/0.6820 over fixed 1:1 mixing. This interpretation concerns task-specific implementations across organs.
We will add matched prediction-minus-atlas, truth-minus-atlas and prediction-error maps with the same sections, classes, masks and color scales, including near/far targets and ABCA-4. These maps will show where transferred deviations agree with the measured target.