acceptodds
Under review as a conference paper at ICLR 2027

TSNR: Protein-Specific Signals with Broad Molecular Context for Spatial Protein Prediction

Abstract

Spatial transcriptomics measures transcriptome-wide gene expression while preserving tissue organization, enabling prediction of spatial protein abundance beyond directly measured antibody panels. Such prediction draws on protein-specific transcript signals at the center spot and broader molecular context from the surrounding tissue. Existing spatial methods increasingly use context, but what molecular information the context should retain and how reliably it transfers across tissue environments remain unclear. We compare predictors that share a protein-specific center representation but differ in contextual molecular scope. Their transferred risk difference admits a value-mismatch characterization: broader context is favored when its target-side value advantage exceeds the corresponding change in source-to-target transfer mismatch. For convex prediction objectives, we further relate mismatch to cross-environment estimating-score variation; squared loss yields an explicit covariance geometry. This analysis motivates broad molecular context together with bounded-score robust regression. We develop Target-Specific Niche Regression (TSNR), which combines a protein-specific center representation, compact broad molecular context, and a protein-specific Huber regressor. On the six-section Spatial-CITE-seq benchmark, TSNR achieves the highest mean macro Spearman correlation among the reproduced methods under matched biological holdout evaluation. Controlled analyses support both broader context and robust regression. Across two independent external cohorts, protein-level broad-context gains are positively rank-correlated with the value-mismatch balance in all 11 held-out biological folds.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, Spatial protein prediction, RNA-to-protein prediction, Representation learning, Contextual transfer, Distribution shift

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qdc3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether spatial context should use the same protein-specific genes as the center spot. TSNR combines selected center genes with a PCA representation of neighborhood-wide gene expression and fits a Huber regressor for each protein. It also decomposes the transferred objective difference between broad and restricted context into target-side value and transfer mismatch. Experiments include a six-section Spatial-CITE-seq benchmark, context and estimator ablations, and retrospective analyses on two external cohorts.

Critical strengths
  1. TSNR is a simple and inexpensive method with a useful result on the primary benchmark: mean macro Spearman is 0.4391, compared with 0.4164 for DGAT. The reported CPU runtime also makes the method attractive as a practical baseline.
  2. The context ablations are informative. Broad context improves over both center-only and target-restricted inputs, while shuffled correspondence and random-gene controls do not recover the gain (Table 2). Most of the benefit remains after excluding the selected center genes from context.
Critical weaknesses
  1. My main concern is how much the value–mismatch analysis adds to the empirical result. Eq. (3) is an exact rearrangement, and Eq. (4) follows directly from first-order convexity. In Eq. (8), substituting the definitions of value and mismatch causes all terms involving the target-reference fits to cancel. The resulting balance is simply the target Huber objective of the source-trained restricted-context model minus that of the source-trained broad-context model. Its association with Spearman gains therefore shows agreement between two measures of predictive performance. I appreciate the separate component analysis and the spatially disjoint validation in Appendix C.2. The latter addresses reuse of the same evaluation spots. However, I still do not see what additional conclusion about transfer is established by the decomposition. The bounds also do not establish that Huber reduces mismatch relative to other estimators. This leaves the theoretical contribution, and its role in motivating the method, limited.

  2. The evidence for the method's advantage across environments is incomplete. Comparisons with DGAT, SPINE and SR2P are confined to the six-section primary benchmark. On the external cohorts, broad context improves over restricted context by 0.01570.0157 on GSE289326 but loses 0.01420.0142 on CytAssist (Sec. 4.4). The authors acknowledge this reversal, but explaining it retrospectively does not tell us when the fixed broad-context design should be used. External comparisons against the main baselines would help establish whether TSNR remains competitive. There is also limited evidence that bounded residual scores explain the primary-benchmark gain: with the same PCA64 context, Elastic Net and HistGradientBoosting obtain 0.43980.4398 and 0.44040.4404, respectively, versus 0.43910.4391 for Huber (Table 8). These small differences do not establish that either alternative is better.

  3. The baseline comparison is too thin for the state-of-the-art claim. Table 1 has only three baselines. Established methods cited in Sec. 2.1 (e.g., cTP-net, sciPENN, scLinear) and simple linear baselines are missing. Since the paper's own center-only linear variant (0.4148) already matches DGAT (0.4164), the gain in Table 1 is hard to interpret.

Questions for authors
  1. Why are the comparisons with DGAT, SPINE and SR2P limited to Spatial-CITE-seq? How do these methods compare with TSNR on GSE289326 and CytAssist under the same biological holdout splits?

  2. Table 8 shows that Elastic Net and HistGradientBoosting perform similarly to Huber on the primary benchmark. How do these estimators compare on the two external cohorts when using the same TSNR representation?

1 comment
AuthorsOpenReview ↗

Response to Reviewer qdc3: Clarifications and Additional Experimental Evidence

We thank the reviewer for the constructive feedback. The revision adds external comparisons, broader baselines, and controlled analyses.

SR2P correction

We identified an inconsistency between SR2P's original training-response preprocessing and the common evaluation targets. We corrected this mismatch and re-evaluated SR2P with source-only feature selection and out-of-fold stacking under the same biological holdout protocol. Its primary-benchmark mean increases from 0.2464 to 0.4040; TSNR remains highest at 0.4390 (revised Table 1; Appendix B.5).

Q1. External spatial baselines

Direct comparisons with DGAT, SPINE, and SR2P now cover two external cohorts under matched biological holdout protocols (revised Table 2):

MethodGSE289326CytAssist
DGAT0.18270.2448
SPINE0.21270.2321
SR2P0.06380.1611
TSNR0.28470.3236

TSNR leads 10/11 biological holdouts. Under the same zero-shot protocol as DGAT on 15 external records (revised Table 3), TSNR achieves the highest mean among the four spatial predictors (0.2089 vs. DGAT 0.1804, SR2P 0.1685, and SPINE 0.1189), outperforming DGAT on 14/15 records.

Q2. External estimator comparison

With the TSNR representation fixed, Huber leads both refitted cohorts (revised Table 4): 0.2847/0.3236, versus ElasticNet 0.2489/0.2830 and HistGradientBoosting 0.2171/0.2512 (GSE289326/CytAssist).

Huber outperforms both in all 11 holdouts, with mean paired gains of +0.0380/+0.0698, respectively. It also exceeds penalty-matched and source-tuned Ridge in 17/17 biological holdouts (Table 5). These results support Huber as a robust empirical choice across independently refitted environments (Appendices C.4–C.5).

W3. Baseline coverage and spatial contribution

The primary comparison now includes 11 baselines, including the requested cTP-net, sciPENN, and scLinear, plus RNA2000 OLS/Ridge (Table 1). TSNR leads in mean macro Spearman (0.4390 vs. DGAT 0.4176 and RNA2000 Ridge 0.3774). It requires 126.8 s per holdout on CPU, compared with 229.3 s for DGAT on an A100 GPU, under their respective hardware configurations (Table 8).

Dimension-matched controls (Table 4) show zero-shot gains from spatial neighborhoods over broad own-spot features at 320 dimensions (+0.0233), and over extra center features (+0.0214) or shuffled neighborhoods (+0.0218) at 384 dimensions, each on 15/15 records.

These results provide direct evidence that spatial neighborhoods contribute predictive information beyond increased feature dimensionality or broader center features, particularly under zero-shot transfer.

W1. Theoretical contribution and contextual transfer

The exact value–mismatch decomposition separates target-side contextual value from transfer mismatch. The score-shift bound provides sufficient conditions for favorable transfer, while the squared-risk specialization characterizes mismatch through target feature covariance (Section 3.3; Appendix A).

Using the Huber objective, the complete balance achieves a mean Spearman association of 0.577 with protein-level contextual gains across 17 biological holdouts, versus 0.108 for value alone and 0.467 for negative mismatch. Associations are positive in 17/17 holdouts and 33/34 spatially disjoint evaluations (Tables 6–7).

Although the balance algebraically equals a difference in target Huber objectives, its association with protein-level Spearman improvements is not implied by this identity. The spatially disjoint evaluation further examines this relationship using separate target regions for diagnostic fitting and evaluation.

In CytAssist, positive contextual value (+0.037) is outweighed by increased mismatch (+0.052), consistent with the observed negative contextual gain. These results support an empirically grounded characterization of contextual transfer. The diagnostic uses target protein measurements retrospectively and does not affect TSNR predictions.

Reviewer_BYGoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Target-Specific Niche Regression (TSNR), an interpretable, training-free framework for spatial protein prediction from transcriptomics. It combines protein-specific center gene signals with broad neighborhood molecular context (via PCA) using robust Huber regression.

Critical strengths
  • Interpretable and efficient design
  • Rigorous empirical controls
Critical weaknesses

W1. Limited technical and theoretical novelty: Methods rely on off-the-shelf tools (correlation selection, PCA, Huber regression). The value–mismatch formula and bounds follow directly from standard definitions and convexity, offering limited new theory.

W2. Retrospective theoretical validation: The value–mismatch diagnostic requires ground-truth target protein labels and reduces algebraically to target objective differences, limiting its utility for predicting optimal context in unobserved target domains.

W3. Confounded spatial vs. transcriptomic scope: Improvements from broad context may stem from including non-target center genes rather than spatial neighbor information. A broad center-spot baseline without spatial neighbors is needed.

W4. Inconsistent external generalization: Main baseline comparisons (e.g., DGAT, SPINE) are omitted on external cohorts. Broad context fails to improve over restricted context on the CytAssist dataset, highlighting generalizability limits.

Questions for authors

Provided above

1 comment
AuthorsOpenReview ↗

Response to Reviewer BYGo: Additional Evidence and Clarifications

We thank the reviewer for recognizing TSNR's interpretable and efficient design. The revision adds dimension-matched spatial controls, external baseline comparisons, and broader transfer analyses.

SR2P correction

We identified an inconsistency between SR2P's original row-centered CLR training responses and the common CLR evaluation targets. We corrected the response mismatch and re-evaluated SR2P with source-only feature selection and out-of-fold stacking under the same biological holdout protocol. Its revised primary-benchmark mean is 0.4040, compared with the original 0.2464; TSNR remains highest at 0.4390 (revised Table 1; Appendix B.5).

W3. Spatial contribution beyond center transcripts

We added dimension-matched controls comparing spatial neighborhoods with broad center-spot features (revised Table 4).

Although broad own-spot features nearly match TSNR on the primary benchmark (0.4386 vs. 0.4391), spatially matched neighborhoods provide consistent gains under zero-shot transfer. At 320 dimensions, neighborhoods outperform equally broad own-spot features by +0.0233. At 384 dimensions, they outperform additional center features (+0.0214) and shuffled neighborhoods (+0.0218). All three comparisons show improvements on 15/15 external records.

These controlled comparisons provide evidence that spatial neighborhoods contribute predictive information beyond broader center transcripts or increased dimensionality, particularly under zero-shot transfer.

W4. External generalization and CytAssist

We added direct comparisons with DGAT, SPINE, and SR2P under matched biological holdout protocols (revised Table 2).

MethodGSE289326CytAssist
DGAT0.18270.2448
SPINE0.21270.2321
SR2P0.06380.1611
TSNR0.28470.3236

TSNR achieves the highest cohort means and leads in 10/11 patient/specimen holdouts. Under the same zero-shot protocol as DGAT, it also achieves the highest mean among the four spatial predictors (0.2089 vs. DGAT 0.1804), outperforming DGAT on 14/15 records (Table 3).

The negative broad-minus-restricted gain on CytAssist (−0.0142) is consistent with the value–mismatch analysis: positive target-side contextual value (+0.037) is outweighed by increased transfer mismatch (+0.052). This characterizes the observed reversal without assuming that broader context always improves transfer (Section 4.4).

W1–W2. Methodological and theoretical contribution

TSNR investigates an asymmetric molecular-scope design, combining protein-specific center signals with broader neighborhood information. The method uses established components to implement this representation efficiently. With the representation fixed, Huber also outperforms ElasticNet and HistGradientBoosting across 11/11 external biological holdouts (Table 4).

The exact value–mismatch decomposition separates target-side contextual value from transfer mismatch. The score-shift bound provides sufficient conditions for favorable transfer, while the squared-risk specialization characterizes mismatch through target feature covariance (Section 3.3; Appendix A).

Empirically, the complete Huber balance achieves a mean Spearman association of 0.577 with protein-level contextual gains, versus 0.108 for value alone and 0.467 for negative mismatch. Associations are positive in 17/17 holdouts and 33/34 spatially disjoint evaluations (Tables 6–7).

Although the balance algebraically equals a difference in target Huber objectives, its association with protein-level Spearman improvements is not implied by this identity. Spatially disjoint evaluation further examines this relationship using separate target regions for diagnostic fitting and evaluation.

The diagnostic uses target protein measurements retrospectively and does not affect TSNR predictions. Together, the controlled ablations and external evaluations provide empirical support for the asymmetric signal–context design, while the retrospective analysis helps characterize its transfer behavior across biological environments.

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

This paper proposes TSNR for predicting spatial protein abundance from spatial transcriptomics. The method combines protein-specific gene selection at the central spot, broad molecular context obtained through neighborhood averaging and PCA, and protein-specific Huber regression. The authors also introduce a value–mismatch characterization to analyze the trade-off between contextual predictive value and cross-environment transfer mismatch.

Experiments on the six-section Spatial-CITE-seq benchmark show that TSNR achieves a better macro Spearman correlation than the strongest reproduced baseline. Additional ablations and external-cohort analyses investigate contextual representations, robust regression, and the proposed theoretical characterization.

Critical strengths

The proposed method is relatively simple and computationally efficient. It achieves competitive performance against several existing spatial protein prediction methods while requiring substantially fewer computational resources. The leave-one-section-out evaluation and controlled ablation studies also provide useful empirical evidence.

Critical weaknesses
  • The central value–mismatch characterization is essentially an algebraic decomposition, while the subsequent bounds largely follow standard convexity arguments. More importantly, the analysis relies on target-environment labels and provides mainly retrospective explanations rather than actionable guidance for improving cross-environment prediction.
  • TSNR mainly combines established techniques, including Pearson-based gene selection, neighborhood averaging, PCA, and Huber regression. The experiments do not adequately isolate the contribution of spatial context. In particular, a key baseline using broad molecular features from the center spot is missing, making it unclear whether the improvements arise from spatial information or simply additional transcriptomic features.
  • The improvement over the strongest baseline is relatively small, with only six biological holdout sections in the primary benchmark. Moreover, broad context produces negative gains on the external CytAssist cohort. Alternative estimator and representation combinations also achieve comparable or slightly better performance, raising questions about the necessity of the proposed design and the strength of its generalization claims.
Questions for authors
  • Beyond the algebraic value–mismatch decomposition and standard convexity-based bounds, what new theoretical insight does the proposed characterization establish? Can it provide a practically useful prediction or contextual-representation selection criterion without access to target-environment protein labels?
  • Have the authors compared the proposed broad neighborhood representation with an equally broad representation extracted directly from the center spot? How can the observed improvements be attributed specifically to spatial context rather than additional transcriptomic information?
  • How should the negative broad-context gain on CytAssist be interpreted in relation to the paper's claims about broad-context transfer? Can the authors provide direct external-cohort predictive comparisons against the main baselines, rather than primarily reporting retrospective value–mismatch associations?
  • Given that alternative estimators and representations achieve comparable or slightly better performance in Appendix Table 8, what evidence establishes the necessity of Huber regression as a central component of TSNR?
2 comments
AuthorsOpenReview ↗

Response to Reviewer KgsY: Experimental Evidence and Clarifications

We thank the reviewer for recognizing TSNR’s efficiency and controlled experiments. We further examine whether spatial context adds information beyond center transcripts, how its value changes across environments, and whether the estimator generalizes.

SR2P correction

We identified a mismatch between SR2P’s original row-centered CLR training responses and the common CLR evaluation targets. After correction, SR2P was re-evaluated with source-only feature selection and out-of-fold stacking under the same biological holdouts. Its mean increases from 0.2464 to 0.4040; TSNR remains highest at 0.4390 (Table 1; Appendix B.5).

Q1. Theoretical insight and practical relevance

The exact value–mismatch decomposition separates target-side contextual value from transfer mismatch. The score-shift bound gives sufficient conditions for favorable transfer; squared-risk analysis characterizes mismatch through target covariance (Section 3.3; Appendix A).

Using the Huber objective, the complete balance achieves a mean Spearman association of 0.577 with protein-level contextual gains across 17 holdouts, versus 0.108 for value alone and 0.467 for negative mismatch. Associations are positive in 17/17 holdouts and 33/34 spatially disjoint evaluations (Tables 6–7).

Although the balance algebraically equals a difference in target Huber objectives, its association with protein-level Spearman gains is not implied by this identity. Spatially disjoint evaluation uses separate target regions for diagnostic fitting and evaluation.

The diagnostic requires target protein measurements and is retrospective. It does not affect context selection or TSNR predictions, which use source-training data.

Q2. Spatial contribution beyond center transcripts

We added dimension-matched neighborhood and center-spot controls (Table 4).

Broad own-spot features nearly match TSNR on the primary benchmark (0.4386 vs. 0.4391). Under zero-shot transfer, spatially matched neighborhoods outperform equally broad own-spot features by +0.0233 at 320 dimensions.

At 384 dimensions, matched neighborhoods outperform additional center features (+0.0214) and shuffled neighborhoods (+0.0218). All three comparisons improve across 15/15 external records.

Broad neighborhood context also improves over target-restricted context (0.4391 vs. 0.4241), retaining most of the gain after excluding selected center genes.

These controls provide evidence that spatial neighborhoods add predictive information beyond broader center transcripts or increased dimensionality. The benefit is most evident under zero-shot transfer, where matched neighborhoods consistently outperform dimension-matched controls.

Q3. External generalization and CytAssist

We added two independently refitted external cohorts, comparing DGAT, SPINE, and SR2P under matched biological holdouts (Table 2).

TSNR achieves the highest mean macro Spearman on GSE289326 (0.2847) and CytAssist (0.3236), leading 10/11 patient/specimen holdouts. Under DGAT’s zero-shot protocol, TSNR leads the four spatial predictors (0.2089 vs. DGAT 0.1804) and outperforms DGAT on 14/15 records (Table 3).

On CytAssist, broad-minus-restricted gain is negative (−0.0142), despite TSNR’s overall lead. Retrospective analysis finds positive contextual value (+0.037) outweighed by increased mismatch (+0.052), consistent with this reversal.

TSNR requires 126.8 seconds per biological holdout on CPU (Table 8).

Q4. Evidence for Huber regression

Huber’s bounded residual score motivates its robust estimation. We evaluated alternatives with the TSNR representation fixed (Tables 4–5).

Although ElasticNet and HistGradientBoosting slightly outperform Huber on the primary benchmark, Huber leads both across 11/11 external holdouts, with mean paired gains of +0.0380 and +0.0698. It also exceeds penalty-matched and source-tuned Ridge in 17/17 biological holdouts.

Under 5% training-response contamination, Huber’s primary mean decreases by 0.0009, versus 0.0091 for Ridge (Appendix C.5).

These results support Huber as an empirically robust choice across independently refitted environments.

Overall, controlled ablations support the additional predictive value of spatial neighborhoods, particularly under transfer. External evaluations establish TSNR’s effectiveness, while value–mismatch analysis characterizes variation in contextual gains across environments.

Thank you for the detailed response and the additional experiments. I appreciate the authors' efforts to address my concerns, particularly the dimension-matched spatial controls, external baseline comparisons, and further validation of Huber regression. These additions strengthen the empirical evaluation.

However, my main concerns regarding the limited methodological and theoretical novelty remain. The proposed framework still primarily combines established techniques, while the value–mismatch analysis remains retrospective and offers limited additional insight beyond the observed predictive performance differences. Moreover, the broad own-spot control nearly matches TSNR on the primary benchmark, making the contribution of spatial context less compelling in that setting.

Overall, while I acknowledge the improved experimental evidence and practical efficiency of the method, I do not find that the rebuttal sufficiently addresses my central concerns. I therefore maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.