acceptodds
Under review as a conference paper at ICLR 2027

ReDiM: Decodable Multi-view Representation Learning of RNA–Protein Discordance with Spatial and Metabolic Context

Abstract

Spatial multi-omics connects transcriptomic, proteomic, and metabolic organization within intact tissue, but spatial proteomics remains more limited in molecular coverage and availability than spatial transcriptomics. Predicting spatial protein abundance from RNA is essential for bridging this measurement gap. However, RNA is not a quantitative proxy for protein: translation, degradation, and metabolic state create context-dependent discordance across tissue regions. Existing predictors typically learn a direct RNA-to-protein mapping and absorb such deviations into undifferentiated prediction error. We introduce ReDiM, a decodable representation-learning model that casts context-dependent RNA–protein correction as a supervised, low-dimensional learning problem. ReDiM freezes a reference RNA-to-protein predictor, defines correction as the difference between observed and reference-expected protein abundance, and represents it in a training-fitted 16-dimensional protein basis. It predicts correction coordinates from three RNA-accessible views: reference protein expectation, spatially contextualized RNA, and RNA-derived reaction-level flux enrichment scores (FES). FES maps RNA onto a genome-scale metabolic network, enabling FES-associated corrections to be traced across protein, reaction, and pathway levels. A fixed decoder maps corrections to a common 31-protein space. Across held-out liver sections, amyotrophic lateral sclerosis (ALS) donors, and brain regions, training-derived bases reconstruct held-out discordance, and ReDiM improves prediction beyond the frozen reference. Its two contextual inputs play distinct roles: spatial RNA drives the recovery of protein rank structure, whereas FES further refines absolute prediction error, with larger gains in target spots that differ more from the training data. To characterize this metabolic contribution, we analyze FES attribution in ALS and identify protein- and pathway-level correction patterns that recur across donors. By learning transferable RNA–protein corrections from complementary spatial and metabolic context, ReDiM improves protein prediction in unseen biological domains. ReDiM thus recasts spatial RNA-to-protein prediction as learning a transferable, decodable correction rather than treating discordance as undifferentiated error.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spatial multi-omics, Multi-view representation learning, cross-modal learning, domain generalization, Interpretable representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_v1RNOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

ReDiM decomposes RNA-to-protein prediction into a reference prediction and a low-dimensional residual correction, incorporating spatial RNA and RNA-derived metabolic network features. The problem is relevant, but the method primarily combines a frozen predictor, PCA, graph networks, and metabolic features. The experiments show some predictive gains, yet do not establish a compelling methodological contribution. In particular, the emphasis on spatial information, shared geometry, and biological interpretation exceeds what the current evidence supports.

Critical strengths
  • Domain-level holdouts, training-domain PCA, and separate spatial graphs provide a reasonably credible evaluation framework.
  • Direct prediction, full-dimensional residual, mean-correction, capacity-matched, and shuffled-FES controls support some predictive value. The authors also acknowledge several important limitations.
Critical weaknesses
  1. The necessity of the core design remains unproven. The comparison between the frozen reference predictor and M1 simultaneously introduces a trained correction network, RNA inputs, and spatial neighborhoods. It therefore does not isolate the contribution of spatial context. Without a capacity-matched nonspatial correction model trained under the same supervision, the main gains could simply reflect an additional RNA prediction network. This omission substantially weakens the argument for the proposed spatial design.

  2. The external comparisons do not yet establish a clear competitive advantage. STProtein, DGAT, and totalVI are relevant baselines, but their access to the paired reference data used by ReDiM is unclear. The Brain experiments also rely on region-matched reference data for the held-out target region. Consequently, their transfer setting is narrower than the phrase “held-out region” might suggest. Until reference-data access and output-scale transformations are documented, the rankings in Table 3 are insufficient to establish a clear methodological advantage.

  3. The shared-geometry and attribution analyses lack convincing calibration. Predicted corrections are constrained by a fixed PCA decoder, so cross-domain subspace similarity cannot directly establish that the model learns meaningful shared residual structure. Unconstrained random subspaces do not adequately control for this architectural restriction. The directional-agreement metric in Figure S2 is also ambiguous. For example, ITGAX has two positive values but is labeled as having 1/3 agreement. Acknowledging these limitations does not substitute for validation. These analyses currently provide descriptions of model outputs rather than strong evidence for the representation or its biological interpretation.

  4. The limited gains do not yet justify the additional pipeline. FES produces small, dataset-dependent RMSE improvements. On ALS, the full model increases RMSE from the reference predictor’s 0.307 to 0.332, despite improving correlation. The manuscript does not adequately establish the stability of these gains or report the actual costs of reference training and spFBA preprocessing. Given the additional paired-data requirements and processing stages, the practical value of this pipeline remains unclear.

Questions for authors
  • How much benefit does the spatial branch retain against a capacity-matched nonspatial correction model trained under identical supervision?
  • Does ReDiM retain its advantage when all methods receive the same paired reference data?
  • Would random latent predictions passed through each fold’s fixed decoder produce comparable cross-domain subspace similarity?
  • Could the authors provide domain-level paired differences, complete attribution definitions, reproducible FES configurations, and end-to-end computational costs to justify the additional complexity?
6 comments
AuthorsOpenReview ↗

W1 / Q1 — Isolating the spatial contribution, W2 / Q2 — Equal paired-data access and Brain evaluation scope

We sincerely thank the reviewer for the careful reading and insightful feedback.

W1 / Q1 — Isolating the spatial contribution

The spatial branch supplies tissue-neighborhood context to ReDiM’s decodable correction representation, supporting protein prediction while allowing protein-specific correction patterns to be examined in their spatial biological context. The relevant question is whether this neighborhood context contributes beyond what a spot-level correction network can already learn.

To isolate this contribution, we added and trained nonspatial controls matched to M1 and M3 in inputs, supervision, data splits, seeds, and optimization settings. In these controls, inter-spot edges were replaced with self-loops (NS1 and NS3), retaining each spot’s own features while removing information exchange between neighboring spots.

M1 exceeded NS1 in mean protein-wise Spearman correlation by 0.021 in ALD, 0.041 in ALS, and 0.014 in Brain, corresponding to a domain-equal mean gain of 0.025, and in 26/27 domain-seed comparisons. M3 exceeded NS3 by a domain-equal mean of 0.020 in mean protein-wise Spearman correlation and in 8/9 held-out domains (Table A10). Thus, the rank-recovery benefit of neighborhood aggregation was observed both without FES (M1 versus NS1) and with FES (M3 versus NS3).

We have added these controls to Table A10 and incorporated the results into Section 4.2. These results support a specific role for the spatial branch: neighboring-spot context contributes to predicting relative protein abundance across spots, beyond the correction learned from spot-level inputs alone. This predictive contribution complements ReDiM’s decodable representation, through which the resulting protein-specific corrections can be examined spatially.

W2 / Q2 — Equal paired-data access and Brain evaluation scope

The comparison uses matched paired supervision. We agree, however, that the Brain setting evaluates correction transfer with anatomy-matched reference supervision, rather than transfer to unseen anatomy. ReDiM, STProtein, DGAT, and totalVI received the same paired reference observations and paired outer-training target observations, and were evaluated on the same held-out units and common 31-protein CLR evaluation panel. Held-out proteins were unavailable to every method during training, adaptation, checkpoint selection, and model selection (Tables 3 and A13). To make these supervision boundaries explicit, we expanded Table A1 with reference-cohort composition and paired-spot counts, added the method-access matrix in Table A13, and clarified the comparison in Section 4.3. The access matrix distinguishes paired training supervision from RNA-only query adaptation.

Table A1 reports the reference cohorts: 6,714 spots from two non-ALD liver sections; 72,953 spots from nine control cervical-spinal-cord sections; and four Brain anchors fitted using three nAD donors per region. Each evaluated Brain donor-region was excluded from reference fitting, correction training, graph construction, checkpoint selection, and model selection, while its anatomy was represented in the corresponding nAD reference for every method. For ReDiM, the held-out target was represented by a separate graph constructed from its RNA and spatial coordinates without protein labels. We clarified this evaluation scope in Section 3.1 and Appendix A and revised the Table 3 caption to identify the region-matched reference anchor.

AuthorsOpenReview ↗

W3 / Q3 — Decoder calibration and corrected FES attribution

W3 / Q3 — Decoder calibration and corrected FES attribution

Decoder calibration. We agree that predicted-subspace similarity alone cannot establish learned shared geometry when folds have related decoders and overlapping training domains. We therefore withdrew that interpretation and replaced the original Figure 2B comparison with a direct held-out correction-recovery test, calibrated using fold-specific Haar orientation randomizations with each fitted decoder held fixed. Predicted–observed correction-coordinate correlations were 0.421 for ALD, 0.361 for ALS, and 0.350 for Brain and exceeded the null in all nine held-out units (Figure 2B). After decoding into the common 31-protein space, recovery also exceeded the corresponding 95% null intervals in all nine units (Figure S1). We revised Section 4.1 and Appendix C to distinguish this evidence for correction predictability from the original cross-domain affinity interpretation.

Attribution definition and calibration. The ITGAX discrepancy required clarification and correction of the aggregation and agreement convention. The previous panel used a seed-42 rerun and compared donor signs with a pooled cross-donor direction, producing the ambiguous ITGAX 1/3 result. The corrected analysis first averages seeds 42–44 within each donor. Agreement is defined as the fraction of the three donor means sharing the majority nonzero sign; effects with absolute magnitude ≤10⁻⁸ are treated as zero and do not count toward agreement. ITGAX now shows 2/3 agreement; eight proteins show 3/3 (Figure S2). We replaced Figure S2 and specified the aggregation and agreement rules in Appendix C.2. We also expanded Section 4.5 with shuffled-control calibration: seven proteins show consistent signs in at least 8/9 donor-seed combinations with real FES, compared with none for shuffled FES, and mean pairwise donor-effect cosine similarity is 0.39 versus 0.04 (Figure 4). These comparisons calibrate the reproducibility of decoded model effects, not causal regulation.

AuthorsOpenReview ↗

W4 / Q4 — Practical value, robustness, FES configuration, and computational cost

W4 / Q4 — Practical value, robustness, FES configuration, and computational cost

The practical significance of ReDiM combines controlled predictive refinement with an inspectable representation of context-dependent RNA–protein correction. The reaction-annotated FES view is linked to decoded protein corrections, allowing its contribution to be tested at both protein and tissue levels. Its value should therefore be assessed through these capabilities as well as endpoint error.

The predictive trade-offs remain explicit. In ALS, M3 improved Spearman correlation from 0.286 to 0.318 but increased RMSE from 0.307 to 0.332 relative to the frozen reference (Table 2). To assess the consistency of the additional FES contribution, we added domain-level paired results, showing that M3 reduced RMSE relative to M1 in 6/9 domains (mean difference -0.0038; Table A7), and real FES outperformed the capacity-matched shuffled control in 7/9 domains (mean difference -0.0024; Table A8). Together with the capacity-matched RNA controls in Table A5, these comparisons argue against added parameter count alone as a sufficient explanation.

We also added a 16D/31D robustness comparison that separates endpoint performance from across-seed prediction variability (Figure S4), and a within-checkpoint FES-assignment permutation analysis for the sl5 ITGAX result. The observed discordance–benefit association (ρ=0.258) exceeded all 1,000 assignment permutations (empirical p=0.001; Figure S5). Thus, the interpretation analysis is supported by a test of assignment-specific predictive benefit, not only by visualization of model outputs. These additions are incorporated into Sections 4.1 and 4.6, Appendix C.2, and Figures S4–S5.

We also expanded Appendix B and Tables A3 and A11 to document the FES configuration and computational cost. The recorded configuration includes MAGIC, ENGRO2 gene–protein–reaction aggregation, FVA-constrained GLPK optimization, 20 Raspy CBS solutions per spot, section standardization, and fold-specific filtering and PCA. Across the 32 spatial samples used in this study (4 ALD, 12 ALS, and 16 Brain), recorded FES sampling and statistics aggregation required 42,533.34 seconds (11.81 hours) in total. FES was computed once per spatial sample and cached for reuse across correction-model runs. A representative ALD profile required approximately 21 seconds for reference-predictor fitting; the available ALS and Brain profiles included data loading and did not isolate fitting time. With prepared FES and spatial graphs, median M3 training times were 2.0, 10.2, and 13.4 seconds per fold and seed for ALD, ALS, and Brain, respectively, with at most 1.2 GB GPU memory. Cached inference required 0.006, 0.020, and 0.020 seconds per held-out unit, respectively. Because the FES timings were recovered from historical executions using heterogeneous sampling modes, they are reported as an aggregate one-time preprocessing cost rather than as a controlled cross-dataset benchmark. These measurements characterize stage-specific costs and do not constitute complete end-to-end deployment latency; accordingly, we do not claim an end-to-end efficiency advantage.

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

ReDiM predicts spatial protein abundance by adding a learned correction to a frozen RNA-to-protein reference model. The correction is represented in a 16-dimensional PCA basis fitted to training-domain residuals and decoded into a common 31-protein panel. Its predictor combines the reference protein expectation, spatially contextualized RNA, and RNA-derived reaction-level flux enrichment scores (FES). The evaluation holds out two ALD sections, three ALS donors, and four brain regions in separate experiments. It examines residual reconstruction, protein prediction, component and representation ablations, and exploratory FES attribution. Spatial RNA provides most of the rank-correlation improvement, while FES gives smaller, dataset-dependent reductions in absolute error. The paper distinguishes RNA-derived metabolic features from measured flux and model attribution from causal biological regulation.

Critical strengths
  1. The correction representation is explicit, inspectable, and evaluated beyond endpoint accuracy. Equations (1)–(3) make the reference-relative target and fixed decoder precise. Training-derived 16-component bases reconstruct held-out residuals with mean R² of 0.750, 0.888, and 0.649 for ALD, ALS, and Brain, respectively (p. 6, Figure 2A). The paper correctly treats these as representation diagnostics rather than prediction results. Whole-domain testing, separate validation/test graphs, and domain-equal aggregation provide a coherent evaluation of correction transfer within the stated datasets (Appendices A and C).

  2. The ablations address important alternative explanations. Table 1 compares direct protein prediction, full-dimensional residual prediction, and a mean-residual correction under matched inputs. Appendix D additionally includes equal-capacity RNA and metabolic-gene controls, shuffled FES, and end-to-end rank sensitivity. For example, Table A5 matches the three learned third-view models at 131,248 parameters. These controls are more persuasive than a comparison against the frozen reference alone and support a modest contribution from the specific FES transformation.

Critical weaknesses
  1. The reference-supervision boundary is not sufficiently specified to audit the external comparison or the strongest transfer interpretation. Section 3.1 and Appendix A describe ALD/ALS reference data as “designated” references without identifying their sample composition and counts or fully specifying reference-model training. Brain uses a region-matched reference model trained on nAD donors from the corresponding anatomy (p. 13), so the held-out region is unseen to the correction learner but not necessarily to the overall RNA-to-protein pipeline. Appendix E does not establish whether all external baselines have access to the paired reference observations used to train these anchors; totalVI is explicitly described as training on the outer-training domains. This is a reporting gap, not demonstrated leakage or unfairness, but it matters because extra reference supervision could explain part of Table 3’s advantage.

  2. The practical gain is narrower than a general improvement in protein prediction. The paper appropriately acknowledges the trade-offs, but they limit the strength of the contribution. In ALS, M3 improves correlation from 0.286 to 0.318 while worsening RMSE from 0.307 to 0.332 relative to the frozen reference (Table 2); its RMSE is also worse than all three external baselines in Table 3. FES improves dataset-average RMSE by 0.0008, 0.0057, and 0.0039 and improves six of nine domains (p. 7; Figure S3). The low-rank formulation also has small margins over full residual prediction in ALS and Brain. The contribution is consequently best supported as rank recovery with modest, context-dependent error refinement.

Questions for authors

Reference-supervision clarification: What exact samples/donors and paired observations train each reference anchor, how are their preprocessing and checkpoints selected, and which of those observations are available to STProtein, DGAT, and totalVI? In Brain, please explicitly distinguish the held-out correction region from the region-matched reference supervision. An access-matched comparison, or confirmation that one is already reported, would materially strengthen my assessment of Table 3.

3 comments
AuthorsOpenReview ↗

W1 / Q1 — Reference supervision, access matching, and Brain evaluation scope

We sincerely thank the reviewer for the careful reading and constructive feedback.

W1 / Q1 — Reference supervision, access matching, and Brain evaluation scope

The Table 3 comparison used the same paired reference observations and outer-training target observations for ReDiM, STProtein, DGAT, and totalVI. The original description did not make these two supervision sources sufficiently explicit. We addressed this reporting gap by expanding the reference-cohort description in Table A1 and adding the method-access matrix in Table A13.

We now report reference-cohort composition and paired observation counts in Table A1. ALD used two non-ALD donors and liver sections containing 6,714 paired spots, while ALS used nine control donors and cervical-spinal-cord sections containing 72,953 paired spots. Brain used four anatomy-specific reference anchors, each fitted using the corresponding region from three nAD donors: 35,643 spots for fcx, 35,103 for hipp, 37,199 for pcx, and 42,114 for tcx. These are unique paired observations, not counts multiplied by folds or optimization seeds. Evaluated target-domain sizes are reported separately in Table A2.

Reference-anchor fitting was separate from correction learning, and the fitted anchors were frozen before training the correction model.  RNA preprocessing comprised library-size normalization, log transformation, a 2.5% detection filter, standardization, clipping, and PCA64; all transformation parameters were fitted using reference spots only. Validation used 20% of occupied blocks from a 5×55\times5 spatial grid. The MLP was trained with full-batch Smooth-L1 loss and AdamW for up to 300 epochs, with the best epoch selected by reference-validation loss. A newly initialized model was then refitted on all reference spots for the selected number of epochs and frozen. No target protein was used in preprocessing, parameter fitting, early stopping, or refitting. Complete settings are reported in Appendix Table A4.

STProtein and DGAT used the available paired observations through their native prediction objectives; totalVI used the paired reference and outer-training target observations for base-model fitting and RNA-only query adaptation for the held-out target. All methods were evaluated on the same held-out units and ordered 31-protein CLR panel. Held-out target proteins were unavailable during training, adaptation, checkpoint selection, and model selection. We clarified these method-specific procedures in Appendix E and their data-access boundaries in Table A13.

For Brain, the evaluated donor-region was excluded from reference fitting, correction training, training-graph construction, checkpoint selection, and model selection. For ReDiM, an independent held-out graph was constructed from target RNA and spatial coordinates without protein labels. However, its anatomical category was represented by the corresponding nAD reference observations for every method. We therefore describe this setting as a held-out correction region with a region-matched nAD reference anchor, rather than as unseen-anatomy transfer. We made this distinction explicit in Section 3.1, the Table 3 caption, and Appendix A.

AuthorsOpenReview ↗

W2 — Scope of the predictive contribution

W2 — Scope of the predictive contribution

A key advantage of ReDiM is its decodable correction representation, which connects RNA–protein prediction with the interpretation of spatial and FES-associated contributions. It separates a frozen RNA-conditioned protein expectation from the remaining context-dependent correction, and integrates an RNA-derived FES view linked to reaction pathways into correction representation learning. A fixed decoder maps predicted coordinates to individual protein corrections, allowing the contribution of this annotated view to be examined in protein space and across tissue.

We strengthened the interpretation analysis with explicit controls. Section 4.5 and Figure 4 now compare real and shuffled FES: seven proteins show consistent effect signs in at least 8/9 donor-seed combinations with real FES, compared with none for shuffled FES; mean pairwise donor-effect cosine similarity is 0.39 versus 0.04. We also added 1,000 within-checkpoint FES-assignment permutations among the 1,203 sl5 anterior-horn spots. The observed ITGAX discordance–benefit association was ρ=0.258, compared with a null 95th percentile of 0.0396; the upper-quartile benefit contrast was 0.0076, compared with a null 95th percentile of 0.0015. Both empirical p-values were 0.001 (Figures 5 and S5). These additions, described in Section 4.6 and Appendix C.2, test an assignment-specific predictive contribution rather than relying only on attribution visualization.

The endpoint benefits nevertheless remain metric-dependent. Relative to P₀, M3 improves mean protein-wise Spearman correlation in ALD (0.434→0.499), ALS (0.286→0.318), and Brain (0.276 to 0.325). RMSE decreases in ALD (0.400→0.361) and Brain (0.513 to 0.477), but increases in ALS (0.307 to 0.332); totalVI achieves ALS RMSE 0.298 (Tables 2-3). We revised Sections 4.2–4.3 to distinguish rank recovery from absolute-error calibration.

To quantify the incremental FES contribution, we added domain-level paired summaries. M3 reduced RMSE relative to M1 in six of nine domains, with a mean paired RMSE difference (M3 - M1) of -0.0038 (Table A7). Dataset-mean RMSE reductions were 0.0008 in ALD, 0.0057 in ALS, and 0.0039 in Brain. In a separate matched shuffle rerun, real FES outperformed the capacity-matched shuffled-FES control in seven of nine domains, with a mean paired RMSE difference (real - shuffled) of -0.0024 (Table A8). Seeds were averaged within each domain before domain-equal aggregation. These results support modest, context-dependent refinement beyond the presence of an additional branch, not uniform improvement across domains.

As a representation diagnostic, the 16D representation provided a compact correction space while retaining endpoint performance comparable to direct prediction in the full 31D residual space. Mean Spearman correlations for the 16D and 31D models were 0.499/0.489 for ALD, 0.318/0.322 for ALS, and 0.325/0.322 for Brain (Table 1 and Figure S4A). Corresponding RMSE values were 0.361/0.372, 0.332/0.334, and 0.477/0.485. We added deterministic robustness reruns with seeds 42–44: prediction variability across seeds was lower for 16D in ALD (0.049 vs. 0.059) and Brain (0.061 vs. 0.073), and nearly unchanged in ALS (0.039 vs. 0.040; Figure S4C). We clarified in Section 4.1 that lower prediction variability is distinct from uncertainty in performance differences and does not imply uniform accuracy gains.

We further trained M1 and M3 directly in the full 31-protein correction space. Adding FES reduced RMSE in ALS (0.341→0.334) and Brain (0.486→0.485), but increased it in ALD (0.366→0.372; Table A9). FES-associated refinement can therefore occur without the 16D projection, although the benefits remain dataset-dependent.

Thank you again for your time and constructive feedback on our submission. As the discussion period is drawing to a close, we wanted to kindly follow up and ask whether our response and the revised manuscript have addressed your concerns regarding reference supervision, the Brain evaluation scope, and the predictive and interpretive contribution of FES.

If any concerns remain or further clarification would be helpful, we would be grateful for the opportunity to address them before the discussion closes. Thank you again for your consideration.

Reviewer_vUN7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReDiM is a decodable representation-learning model that casts context-dependent RNA–protein correction as a supervised, low-dimensional learning problem. It freezes a reference RNA-to-protein predictor, defines correction as the difference between observed and reference-expected protein abundance, and represents it in a training-fitted 16-dimensional protein basis. It predicts correction coordinates from three RNA-accessible views: reference protein expectation, spatially contextualized RNA, and RNA-derived reaction-level flux enrichment scores. A fixed decoder maps corrections to a common 31-protein space. Across held-out liver sections, ALS donors, and brain regions, training-derived bases reconstruct held-out discordance, and ReDiM improves prediction beyond the frozen reference.

Critical strengths
  1. The correction formulation is useful and inspectable. Separating a frozen reference prediction from a decoded residual makes the prediction process easier to examine. The fixed decoder also provides an explicit connection between latent predictions and individual protein corrections.

  2. The component controls address relevant alternatives. Comparisons against direct protein prediction, full-dimensional residual prediction, and capacity-matched RNA branches help assess what the proposed parameterization and metabolic transformation contribute. Holding out complete biological domains is also more informative than random spot-level evaluation.

Critical weaknesses
  1. The practical gains are small and inconsistent. Adding FES yields modest RMSE reductions and does not consistently improve performance across held-out domains. In ALS, the full model has worse RMSE than both the frozen reference and totalVI. Given the limited number of evaluated domains and the lack of uncertainty estimates for paired improvements, the evidence does not yet establish a reliable benefit from the additional complexity.

  2. The shared-geometry interpretation is not adequately separated from architectural constraints. Predicted corrections are constrained by training-derived PCA decoders, while ALS and Brain folds share training domains. High affinity between decoded prediction subspaces could therefore partly reflect the decoders rather than successful correction learning. The random-subspace comparison in Figure 2B does not preserve these constraints, weakening its support for the claim that ReDiM recovers recurring correction geometry.

Questions for authors
  1. Which donors, sections, and numbers of paired observations were used to train the reference predictors? Were these same reference observations available to all external baselines, particularly for the region-matched Brain setting?

  2. How much of Figure 2B’s affinity remains when predicted latent coordinates are randomized while retaining each fitted decoder? How do predicted subspaces align with observed held-out residual subspaces?

  3. Can the authors report paired domain-level differences and seed variability for the low-rank versus full-residual comparison and for M3 versus M1? This would help determine whether the small average gains are stable.

3 comments
AuthorsOpenReview ↗

W1 / Q3 — Predictive gains, context-dependent correction, and stability

We sincerely thank the reviewer for the careful reading and insightful feedback.

W1 / Q3 — Predictive gains, context-dependent correction, and stability

A key advantage of ReDiM is its decodable correction representation, which connects protein prediction with inspection of spatial and RNA-derived, reaction-annotated FES contributions. The predictive gains are metric- and context-dependent; our case for the additional view rests on both matched predictive comparisons and tests of the specificity of its decoded effects.

To quantify consistency across held-out units, we added paired-domain RMSE summaries. M3 reduced RMSE relative to M1 in six of nine domains, with a mean paired RMSE difference (M3 - M1) of -0.0038 and domain-level differences ranging from −0.0135 to +0.0024 (Table A7). In a separate matched shuffle rerun, real FES achieved lower RMSE than a capacity-matched shuffled-FES control in seven of nine domains, with a mean paired RMSE difference (real - shuffled) of -0.0024 (Table A8). Seeds were averaged within each held-out unit before domain-equal aggregation. These results support a spot-specific FES contribution under matched capacity, not a uniform improvement across domains. In ALS, M3 improved Spearman correlation from 0.286 to 0.318 but increased RMSE from 0.307 to 0.332 relative to P₀; totalVI achieved RMSE 0.298 (Tables 2–3). We clarified these metric-dependent trade-offs in Sections 4.2–4.3.

Under matched inputs and held-out folds, the 16D model achieved endpoint performance comparable to direct 31D residual prediction. Mean Spearman correlations for the 16D and 31D models were 0.499/0.489 for ALD, 0.318/0.322 for ALS, and 0.325/0.322 for Brain (Table 1 and Figure S4A). Corresponding RMSE values were 0.361/0.372, 0.332/0.334, and 0.477/0.485. We also added deterministic robustness reruns with seeds 42, 43, and 44. In deterministic robustness reruns, across-seed prediction SD was lower for 16D in ALD (0.049 vs. 0.059) and Brain (0.061 vs. 0.073), and nearly unchanged in ALS (0.039 vs. 0.040; Figure S4C). This prediction-variability diagnostic is distinct from uncertainty in paired performance differences. Our evidence for the predictive contribution of FES therefore rests on the matched held-out-domain comparisons in Tables A7–A8, not on this prediction SD. Those comparisons average seeds within each domain and do not establish seed-wise stability of the M3–M1 performance difference. We revised Section 4.1 and Figure S4 to distinguish compactness, endpoint performance, and prediction variability, rather than infer uniform superiority or preservation of all full-space information.

We further trained M1 and M3 directly in the full 31-protein correction space. Adding FES reduced RMSE in ALS (0.341→0.334) and Brain (0.486→0.485), but increased it in ALD (0.366→0.372; Table A9). Thus, FES-associated refinement can also occur without the 16D projection, but its benefit remains dataset-dependent. The fixed decoder also permits protein- and tissue-level inspection of the correction. To test interpretation beyond visualization, we added 1,000 within-checkpoint FES-assignment permutations among the 1,203 sl5 anterior-horn spots, holding model parameters, non-FES inputs, observed proteins, and the FES-zero prediction fixed. The observed association between absolute ITGAX discordance and local FES benefit was ρ=0.258, versus a permutation-null 95th percentile of 0.0396. The mean-benefit contrast between the upper discordance quartile and remaining spots was 0.0076, versus a null 95th percentile of 0.0015; both empirical p-values were 0.001. We incorporated this analysis into Section 4.6, Appendix C.2, and Figures 5 and S5. It supports an assignment-specific predictive contribution in this donor, not a causal metabolic mechanism.

AuthorsOpenReview ↗

W2 / Q2 — Decoder-controlled correction recovery , Q1 — Reference supervision and paired-data access

W2 / Q2 — Decoder-controlled correction recovery

The fixed decoder restricts the space of possible corrections, but does not by itself establish that predicted corrections recover the observed held-out discordance. Our revised evidence addresses this predictive alignment directly, rather than using similarity between predicted subspaces as independent evidence of learned shared geometry.

We replaced the original affinity comparison in Figure 2B with a predicted–observed recovery analysis calibrated against a fixed-decoder Haar null that randomizes latent orientation while preserving each fold’s training-fitted decoder. Actual and randomized coordinates were evaluated using the same recovery statistic. In the learned correction-coordinate space, predicted–observed weighted Spearman correlations were 0.421 for ALD, 0.361 for ALS, and 0.350 for Brain, exceeding the corresponding null in all nine held-out units (Figure 2B).

We additionally evaluated predicted–observed correction alignment after decoding into the common 31-protein space. Spearman correlations were 0.35-0.36 for ALD, 0.30-0.31 for ALS, and 0.14-0.23 across Brain regions, and exceeded the corresponding 95% fixed-decoder null intervals in all nine units (Figure S1). This shows that the prediction–observation alignment is retained in decoded protein space and is not confined to the latent-coordinate metric. We revised Section 4.1 and Figure 2B, incorporated the decoded-space check into Figure S1, and described the calibration protocol in Appendix C. The revised claim is that ReDiM predicts held-out corrections within a training-derived representation—not that predicted-subspace affinity alone establishes a learned shared biological geometry.

Q1 — Reference supervision and paired-data access

ReDiM, STProtein, DGAT, and totalVI received the same paired reference observations and paired outer-training target observations, and were evaluated on the same held-out units and common 31-protein CLR panel. To make this supervision boundary auditable, we expanded Table A1 with reference-cohort composition and paired-spot counts and added a method-access matrix in Table A13.

The reference cohorts and paired observation counts are reported in Table A1. ALD used two non-ALD liver sections from two donors (GSE278662; 6,714 paired spots); ALS used nine control cervical-spinal-cord sections from nine donors (GSE288365; 72,953 paired spots); and Brain used four anatomy-matched nAD reference anchors (GSE263079), each fitted using the corresponding region from three nAD donors: fcx, 35,643 paired spots; hipp, 35,103; pcx, 37,199; and tcx, 42,114. These counts refer to unique paired observations and are not multiplied by folds or optimization seeds.

ReDiM fitted and froze its reference predictor before correction learning, whereas the external baselines used the same available paired observations through their native learning objectives. Held-out target proteins were unavailable during training, adaptation, checkpoint selection, and model selection. Baseline-specific training and query-adaptation procedures are described in Appendix E.

For Brain, the evaluated donor-region was excluded from reference fitting, correction training, training-graph construction, and model selection. A separate held-out graph was constructed from its RNA and coordinates without protein labels. However, the corresponding anatomy was represented in the nAD reference data available to all methods. We therefore clarified in Section 3.1 and Appendix A that Brain evaluates held-out donor-region correction with an anatomy-matched reference anchor, not transfer to unseen anatomy.

Thank you again for your time and constructive feedback on our submission. As the discussion period is drawing to a close, we wanted to kindly follow up and ask whether our response and the revised manuscript have addressed your concerns regarding the stability and scope of the predictive gains, decoder-imposed geometry, and access to paired reference supervision.

If any concerns remain or further clarification would be helpful, we would be grateful for the opportunity to address them before the discussion closes. Thank you again for your consideration.

Comments on the paper

1 comment
Area_Chair_j8vYOpenReview ↗

Please check authors' rebuttal and participate in the discussion

The discussion phase is coming to an end. Please read the authors’ rebuttal and post a brief follow-up in the discussion thread. Let us know which of your concerns the response addresses and which remain open.

Please post any follow-up questions in the thread visible to the author as soon as possible, so that the author has time to respond and provide clarification. I also encourage you to engage with one another’s assessments—especially where your perspectives differ—and to note whether any counterarguments or discussions lead you to revise your own assessment.

Thank you for your effort and thoughtful review. Thank you again!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.