acceptodds
Under review as a conference paper at ICLR 2027

Learning Interpretable Tumor Microenvironment Representations by Fitting Pan-Cancer Cell State–Niche Correlation

Abstract

In the tumor microenvironment, a cell’s state is influenced by cell–cell interactions (CCIs) with neighboring cells in its niches. Identifying dysregulated CCIs that are associated with pathogenic processes pinpoints targets for drug discovery. Imaging-based spatial transcriptomics and single-cell RNA sequencing provide, respectively, single-cell spatial information and transcriptome-wide measurements needed to study CCIs, but neither modality provides both. Existing spatial transcriptomics foundation models also cannot effectively learn from spatially resolved single-cell data with full-transcriptome coverage, explicitly infer the CCI mechanisms driving cell state–niche associations, or be interpretable enough to support direct biological interpretations. Here, we present GITIII-scale, a hierarchical, interpretable pan-cancer spatial transcriptomics foundation model for TME representation learning that investigates cell state–niche associations and their underlying ligand–receptor (LR) signaling pathways. GITIII-scale uses transformers to model interactions between pairs of cells at defined spatial distances, an interpretable single-layer graph transformer without a feed-forward network to decompose how each gene in a receiver cell is influenced by each neighboring sender cell, and a graph transformer to generate cellular-neighborhood embeddings. Trained on our assembled pan-cancer database of specimen-matched scRNA-seq and imaging-based spatial transcriptomics datasets, GITIII-scale generated TME embeddings that recovered niche-associated state changes more accurately than existing spatial transcriptomics foundation models in cancer types unseen during training. A case study of an unseen breast cancer dataset further demonstrated the model’s interpretability by identifying potentially drug-targetable LR pathways associated with endothelial overgrowth and tumorigenesis.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Foundation models, graph transformers, spatial transcriptomics, tumor microenvironment (TME), cell–cell interactions (CCI), cell state–niche correlation, ligand–receptor signaling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rTkdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents GITIII-scale, a scaled and pretrained extension of GITIII for modeling cell state–niche associations in the tumor microenvironment. The model is trained on a pan-cancer collection of specimen-matched imaging-based spatial transcriptomics and scRNA-seq data, with spatial profiles imputed to a shared 20,000-gene vocabulary. Architecturally, it combines an interpretable additive module that decomposes receiver-cell state predictions into sender-by-gene contributions with a deeper graph transformer that produces neighborhood-level TME embeddings.

The model is evaluated on held-out cancer types and platforms. Compared with several spatial foundation models and graph-based baselines, GITIII-scale achieves better prediction of within-cell-type expression residuals from local niches. The authors further demonstrate its interpretability using a breast cancer case study, where the model recovers KDR/VEGFR2 signaling associated with endothelial proliferation.

Critical strengths

The main strength is the two-level architecture that separates interpretability from representational capacity. The first-level additive head preserves an exact sender-by-gene decomposition of the predicted receiver-cell state, while the second-level graph transformer captures higher-order neighborhood effects and produces a more expressive TME representation. This is a sensible design choice because it avoids forcing a single module to simultaneously maximize predictive flexibility and strict decomposability. Among the methodological extensions over GITIII, this is the most convincing and potentially reusable contribution.

Critical weaknesses
  1. The “foundation model” and scaling claims are not convincingly supported. GITIII-scale is pretrained on only 10 specimens, and its purported OOD evaluation still trains a supervised probe on cells from each target dataset. This shows representation transfer rather than true zero-shot generalization. More fundamentally, as a scaled successor to GITIII, the paper does not clearly demonstrate that scaling enables a qualitatively new scientific capability rather than simply improving a surrogate prediction task.

  2. The validation is misaligned with the paper’s claimed biological contribution. Predicting within-cell-type expression residuals is only a proxy for the real goal of CCI/TME modeling: identifying biologically meaningful sender–receiver interactions and signaling mechanisms. Yet these attributions are not systematically benchmarked against established CCC/CCI methods. Moreover, “interpretability” is overstated: exact decomposition of a model prediction does not establish correct biological attribution or causality. The 20,000-gene “full-transcriptome spatial” setting is largely imputed rather than measured, and its added biological value is unclear. Finally, the main case study only recovers the canonical VEGF–KDR endothelial proliferation axis, demonstrating consistency with textbook biology rather than novel mechanistic discovery.

Questions for authors

Clarification requests

  1. What is the key capability that makes GITIII-scale a foundation model rather than a pretrained transfer model? In particular, how should the reader interpret this claim given that pretraining uses only 10 specimens and downstream evaluation relies on target-dataset supervised probes rather than zero-shot prediction?

  2. What evidence supports the claim that the inferred sender-by-gene contributions are biologically interpretable, rather than simply an exact decomposition of the model prediction? How should these attributions be distinguished from associative CCI signals already modeled by other methods, including GITIII?

  3. What is the biological advantage of imputing all samples to a shared 20,000-gene vocabulary? Specifically, how robust are the inferred gene-level influences and LR pathways when they involve genes that were not spatially measured, and how much of the reported gain remains when evaluation and interpretation are restricted to measured genes only?

  4. Can the authors provide a more stringent test of transferability, such as leave-one-sample/patient/cancer-type/platform-out evaluation without fitting a supervised probe on the target specimen? This would materially affect whether the current OOD and foundation-model claims are convincing.

  5. For the biological discovery claim, how does GITIII-scale compare with established CCC/CCI methods on the same mechanistic tasks? In particular, can the authors show examples where GITIII-scale identifies a biologically supported interaction or pathway that is missed by existing methods, rather than recovering a canonical pathway such as VEGF–KDR?

13 comments
AuthorsOpenReview ↗

## W1 and Q1: Foundation-model definition and transferability (Part I)

W1 and Q1: Foundation-model definition and transferability

W1 (critical weakness): “The ‘foundation model’ and scaling claims are not convincingly supported. GITIII-scale is pretrained on only 10 specimens, and its purported OOD evaluation still trains a supervised probe on cells from each target dataset. This shows representation transfer rather than true zero-shot generalization. More fundamentally, as a scaled successor to GITIII, the paper does not clearly demonstrate that scaling enables a qualitatively new scientific capability rather than simply improving a surrogate prediction task.”

Q1 (author question): “What is the key capability that makes GITIII-scale a foundation model rather than a pretrained transfer model? In particular, how should the reader interpret this claim given that pretraining uses only 10 specimens and downstream evaluation relies on target-dataset supervised probes rather than zero-shot prediction?”

Response:

We address the conceptual capability, the paired-data resource, and the two evaluation settings separately.

(1) What capability makes GITIII-scale a foundation model in this study?

GITIII-scale is a hierarchical representation-learning model designed to learn both pairwise sender–receiver interactions and the cellular neighborhood as a whole. Its representation learning operates at two corresponding levels:

  • Pairwise interaction representation: For each sender–receiver pair, the pair encoder jointly processes the sender and receiver expression profiles, their spatial distance, and their direct-contact feature. It produces a 256-dimensional pair embedding eije_{ij} that represents the specific cellular pair in its spatial context. The first head uses these pair representations to estimate gene-specific sender contributions and provides an additive decomposition of the predicted receiver-cell state.
  • Neighborhood-level representation: A deeper graph transformer takes the set of sender–receiver pair representations within a receiver's neighborhood and models their higher-order context. It produces a 1,024-dimensional TME embedding that represents the cellular neighborhood as a whole.

Thus, GITIII-scale does not only encode individual cells or tissue domains. It learns representations of who interacts with whom, under what spatial relationship, and how the collection of pairwise interactions forms a receiver's cellular niche. The model is pretrained once on the pan-cancer corpus and then reused with frozen parameters on unseen specimens. This reusable pairwise and neighborhood-level representation is the foundation-model capability we claim.

(2) How should the reader interpret pretraining on only 10 specimens?

Specimen-matched imaging-based spatial transcriptomics (IST) and scRNA-seq data are relatively scarce because both assays must be available for the same biological specimen, a requirement also reflected in recent paired-data imputation frameworks (Haviv et al., 2025; Li et al., 2025; Yang et al., 2026). This requirement is important: IST supplies single-cell spatial positions, whereas scRNA-seq supplies transcriptome-wide information, and the matched measurements allow us to combine these complementary modalities.

Our assembled corpus contains 10 paired samples spanning four tumor contexts and two imaging platforms, comprising 4,254,069 spatial cells. We distinguish the number of independent paired samples from the number of cells: the corpus contains 10 biological samples, while the millions of cells provide substantial cell-level training data within those samples. We therefore define the scope of the foundation-model claim in terms of this multi-sample, multi-context, multi-platform paired resource, rather than claiming universal coverage of all cancer types or platforms.

(3) Does the primary benchmark demonstrate zero-shot generalization when a probe is fitted on each target specimen?

No. The primary benchmark in Table 1 demonstrates frozen-representation transfer, not zero-shot prediction. The pretrained GITIII-scale encoder is kept frozen, and a 1,024-unit downstream probe is fitted using training-block cells from each target specimen. The probe does not update the pretrained encoder. Performance is then evaluated on spatially held-out validation blocks. This tests whether the pretrained representation captures transferable cell-state–niche information that can be read out from a limited amount of target-specimen training data.

AuthorsOpenReview ↗

W1 and Q1: Foundation-model definition and transferability (Part II)

(4) Is there evidence of prediction without a target-specimen probe?

Yes. Table S6 reports the direct outputs of both pretrained heads without fitting a probe or any model parameters on the target specimen. Head 1 directly predicts cell-state residuals by summing its sender-specific influences. Its within-cell-type PCC on all genes is 0.155, 0.115, and 0.204 for Xenium Breast, MERFISH CRC, and MERFISH Melanoma, respectively; on measured genes, the corresponding values are 0.141, 0.095, and 0.193.

These direct outputs should be interpreted as probe-free within-cell-type ranking performance, not as calibrated absolute-expression predictions. As explained in Section 2.3, absolute expression scales are affected by panel sensitivity and batch effects, which differ across specimens and platforms. Consequently, a single generalized prediction head may not be optimally calibrated to the scale and offset of every target dataset. We therefore report within-cell-type PCC for the direct outputs, which is invariant to positive per-gene rescaling; we do not report target-specific R² for these direct outputs because calculating R² would require fitting a scale and offset on the target specimen. The downstream probe in Table 1 provides a lightweight way to adapt the readout to the target specimen while keeping the pretrained encoder frozen.

(5) What does scaling add beyond improving the surrogate prediction task?

The demonstrated benefit of scaling is not only improved cell-state prediction. It enables one pretrained model to learn and reuse the two-level representation described above: pairwise sender–receiver interaction embeddings and a neighborhood-level TME embedding. The pairwise representation supports sender- and gene-resolved decomposition, while the neighborhood embedding captures higher-order context across multiple senders.

The quantitative results support the transfer benefit of this representation. Frozen GITIII-scale outperforms GITIII trained separately on each target specimen on the measured-gene benchmark: R² values are 0.150 versus 0.129 for Xenium Breast, 0.074 versus 0.046 for MERFISH CRC, and 0.118 versus 0.081 for MERFISH Melanoma. Under matched optimization steps, all-gene R² on Xenium Breast rises from 0.115 to 0.158 and 0.170 as the pretraining corpus grows from approximately 2.35 million to 3.04 million and 4.25 million cells. In addition, the GITIII-scale architecture does not converge when trained from scratch on an individual evaluation specimen.

We have therefore scoped the claim carefully: GITIII-scale is a pan-cancer pretrained representation-learning model that transfers pairwise interaction and neighborhood representations to unseen specimens. The primary probe-based benchmark is frozen-representation transfer, while Table S6 separately demonstrates probe-free prediction.

Changes in the manuscript:

  • Introduction: Clarification of the paired-data resource and the two levels of representation learning.
  • Section 2.2: Description of the pairwise sender–receiver encoder, additive head, and neighborhood-level TME embedding.
  • Section 3.2–3.4: Frozen-representation and probe-free evaluation.
  • Tables 1 and 2: Transfer and corpus-scaling results.
  • Appendix S9: Dataset-specific GITIII comparison.
  • Appendix S12: Direct and probe-based head evaluations.
  • Appendix S1: Scope of the foundation-model and transferability claims.

W2a and Q5: Comparison with established CCI methods and biological discovery

W2 (critical weakness): “The validation is misaligned with the paper’s claimed biological contribution... novel mechanistic discovery.”

Q5 (author question): “For the biological discovery claim..”

Response:

We answer the comparison and biological-discovery aspects separately. We also clarify the exact task used for the quantitative benchmark.

(1) How does GITIII-scale compare with established CCC/CCI methods, and what task do the reported results measure?

We added HoloNet (Li et al., 2023) and GITIII (Xiao et al., 2026) to the quantitative benchmarking, alongside NCEM-GCN (Fischer et al., 2023) and GAT (Veličković et al., 2018). The comparison is performed on a shared cell-state prediction task: predict the receiver cell's within-cell-type expression residual from its cellular neighborhood, using the same-type masking and spatial block evaluation protocol. GITIII, HoloNet, NCEM-GCN, and GAT are trained on the training blocks of each target specimen for this task. GITIII-scale uses its frozen pan-cancer-pretrained encoder and the same 1,024-unit downstream probe on the training blocks.

The following values are from Table 1, measured genes, evaluated on the spatially held-out validation blocks. The three columns correspond to the three specific evaluation specimens: Xenium Breast, MERFISH colorectal carcinoma (CRC), and MERFISH Melanoma. Targets are the original measured-panel assay values (313, 500, and 500 genes, respectively); R² is pooled over genes, and values are mean ± standard deviation over five seeds.

MethodXenium Breast R²MERFISH CRC R²MERFISH Melanoma R²
GITIII-scale (frozen encoder + probe)0.150 ± 0.0020.074 ± 0.0040.118 ± 0.001
GITIII (trained on target specimen)0.129 ± 0.0120.046 ± 0.0110.081 ± 0.001
NCEM-GCN0.120 ± 0.0040.033 ± 0.0010.070 ± 0.008
HoloNet0.096 ± 0.0010.029 ± 0.0040.071 ± 0.005
GAT0.101 ± 0.0040.016 ± 0.0100.041 ± 0.007

GITIII-scale achieves the highest measured-gene R² across all three held-out specimens. These results show that its learned representations are more effective for the shared quantitative cell-state prediction task than the evaluated dataset-specific CCI models. They are not, by themselves, a direct benchmark of the causal correctness of each method's mechanistic LR rankings.

(2) Can we show an example of a biologically supported interaction that GITIII-scale identifies?

The breast cancer case study prioritizes SEMA6D→PLXNA1–KDR, rather than the canonical VEGF-A→KDR interaction implied by the original description. The selected pair is top-ranked across all five selection-stability repetitions. Its score achieves a PCC of 0.230 with measured MKI67 in spatially held-out Xenium Breast endothelial cells, compared with 0.147 for KDR expression alone and 0.104 for unweighted ligand–receptor co-expression.

The receptor complex PLXNA1–KDR has been reported to mediate Sema6D-induced endocardial cell migration during cardiac development (Toyofuku et al., 2004). This provides biological context for the candidate association, but does not establish that the same mechanism drives proliferation in breast tumor endothelium. We therefore describe it as a testable hypothesis.

(3) Can we claim that this interaction is missed by existing CCI methods?

We do not make that claim. The observational datasets do not provide ground truth for the causal correctness of inferred LR-to-target relationships, so we cannot establish that another method misses the true interaction. The added HoloNet, GITIII, NCEM-GCN, and GAT comparisons evaluate the shared cell-state prediction task, while the receptor-only and unweighted co-expression comparisons evaluate the predictive contribution of the selected LR score. These are complementary forms of evidence, not a direct causal ranking benchmark.

The revised manuscript therefore makes two distinct claims: GITIII-scale improves cell-state prediction relative to the evaluated CCI models, and its LR analysis can prioritize spatially informed candidate signaling hypotheses. It does not claim to establish causal mechanisms or to show that other methods fail to identify them.

Changes in the manuscript:

  • Section 3.3 and Table 1: Quantitative comparison with HoloNet, GITIII, NCEM-GCN, and GAT.
  • Appendix S7: Baseline implementation and training protocols.
  • Section 3.5 and Figure 2: Revised SEMA6D→PLXNA1–KDR case study.
  • Appendix S15: Held-out LR baseline comparisons.
  • Appendix S1: Scope of biological interpretation and lack of causal ground truth.
Reviewer_V2ExOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces GITIII-scale, a model for learning associations between cell states and their spatial neighborhoods in the tumor microenvironment. Cell state is defined as the deviation from the mean expression of the corresponding cell type within a sample. The model encodes sender–receiver pairs using expression and spatial features, then combines an explicitly decomposable prediction head with a second head that produces neighborhood representations. It is pretrained on approximately 4.25 million spatial cells, with specimen-matched scRNA-seq used to impute a shared 20,000-gene vocabulary. Frozen-representation evaluations on three samples excluded from pretraining show improvements over the reported baselines, including on measured genes. A breast cancer case study identifies a KDR-associated ligand–receptor axis correlated with endothelial MKI67 expression.

Critical strengths
  1. Separating cell-type means from cell-state residuals, and masking the receiver and same-type neighbors, addresses the shortcut of copying nearby cells of the same type. The two heads provide both an explicit sender-wise decomposition of model predictions and a representation for downstream prediction. The summation identity in Appendix S10 is valid and offers useful computational transparency.
  2. Table 1 reports measured-panel R2R^2 values of 0.168, 0.059, and 0.112, compared with 0.146, 0.041, and 0.072 for the strongest reported spatial foundation model in each dataset. The advantage persists with the smaller probe in Table S1. Breast cancer, melanoma, and the MERFISH platform provide meaningful shifts relative to pretraining, while the scaling and geometry ablations add useful evidence. Thus, the empirical contribution does not depend solely on treating imputed expression as ground truth.
Critical weaknesses
  1. Equation 8 combines Pearson correlation with an error term between separately z-scored predictions and targets. Both terms are invariant to a positive affine transformation of the prediction, z^′=az^+b\hat{z}' = a\hat{z} + b, with a>0a > 0. Indeed, mean squared error between standardized vectors is proportional to 1−ρ1 - \rho. Equation 9 nevertheless uses residuals on the original output scale to weight the LR regression, while Section 2.5 fits a distance scaler from estimated influences. For the additive head, transforming the values as vij′=avij+bv'_{ij} = a v_{ij} + b preserves the summation structure but changes individual influences to Iij′=aIij+bαijI'_{ij} = a I_{ij} + b\alpha_{ij}. I could not identify an output calibration step in Appendices S5–S10 that resolves this mismatch. The exact decomposition remains valid, but it does not establish that the residual weights or influence magnitudes have a consistent interpretation. This concern follows from the written equations; an additional calibration step in the implementation could materially change the assessment. (Eq. 8, p. 5; Eq. 9 and Section 2.5, p. 6; Table S2; Appendix S10.)

  2. Appendix S7 specifies 10 neighbors for TERRA, compared with 49 for GITIII-scale, alongside different graph definitions for other methods. This matters quantitatively: in Table 2, GITIII-scale's breast R2R^2 falls from 0.1684 at k=49k = 49 to 0.1398 at k=25k = 25 and 0.1022 at k=10k = 10; the latter values are below TERRA's reported 0.146. These ablations are not a fully controlled comparison, but they demonstrate sensitivity to a variable that is not matched across methods. Using native configurations is reasonable for a system-level benchmark, yet it does not support the claim that the representation is the only variable. Moreover, probes use random cell splits within each evaluation sample. Variation across five seeds does not quantify generalization across independent patients or spatial regions, and same-type masking does not eliminate all dependence from overlapping neighborhoods. This limits the interpretation of the results without establishing direct test-label leakage. (Section 3.2, p. 7; Tables 1–2, p. 8; Appendix S7, pp. 18–20.)

Questions for authors
  1. Are predictions and targets separately standardized in Eq. 8? Before Eq. 9 and distance-scaler fitting, is the first head calibrated for scale and offset, and on which cells? If not, how is the positive affine ambiguity handled, and how sensitive are the LR rankings to equivalent output transformations?
  2. Do the gains persist with matched neighbor counts or spatial information budgets? Are spatial-block or independent-sample probe results available to distinguish spatial interpolation from transfer to independent regions? These controls would directly affect the interpretation of the reported advantage.
  3. For the SEMA6D→PLXNA1–KDR axis shown in Figure 2c, which relevant genes and targets are measured rather than imputed? Please clarify LR feature standardization and what varies across the five repetitions. Does a held-out or spatially controlled analysis, together with a simpler LR-scoring baseline, show that the learned distance and residual weights add information beyond conventional association scoring?
5 comments
AuthorsOpenReview ↗

W1 and Q1: Scale ambiguity and calibration of influence magnitudes (part I)

W1 and Q1: Scale ambiguity and calibration of influence magnitudes

Reviewer’s weakness (W1):

Equation 8 combines Pearson correlation with an error term between separately z-scored predictions and targets. Both terms are invariant to a positive affine transformation of the prediction [the mathematical expression is not present in the supplied review text], with [the condition is not present in the supplied review text]. Indeed, mean squared error between standardized vectors is proportional to [the expression is not present in the supplied review text]. Equation 9 nevertheless uses residuals on the original output scale to weight the LR regression, while Section 2.5 fits a distance scaler from estimated influences. For the additive head, transforming the values as [the transformation is not present in the supplied review text] preserves the summation structure but changes individual influences to [the resulting transformation is not present in the supplied review text]. I could not identify an output calibration step in Appendices S5–S10 that resolves this mismatch. The exact decomposition remains valid, but it does not establish that the residual weights or influence magnitudes have a consistent interpretation. This concern follows from the written equations; an additional calibration step in the implementation could materially change the assessment. (Eq. 8, p. 5; Eq. 9 and Section 2.5, p. 6; Table S2; Appendix S10.)

Reviewer’s question (Q1):

Are predictions and targets separately standardized in Eq. 8? Before Eq. 9 and distance-scaler fitting, is the first head calibrated for scale and offset, and on which cells? If not, how is the positive affine ambiguity handled, and how sensitive are the LR rankings to equivalent output transformations?

Response:

We answer each part of the question below.

(1) Are predictions and targets separately standardized in Equation 8?

Yes. Predictions and targets are independently standardized to zero mean and unit variance for each gene within a batch of receiver cells of the same cell type. Therefore, Equation 8 is invariant to a positive affine transformation of the prediction:

\mapsto a_g\hat z_i^{(g)}+b_g, \qquad a_g > 0.$$ Writing the separately standardized prediction and target as $\widetilde{\hat z}_i^{(g)}$ and $\widetilde z_i^{(g)}$, respectively, their mean squared difference is $$\frac{1}{|B|} \sum_{i\in B} \left( \widetilde{\hat z}_i^{(g)} -\widetilde z_i^{(g)} \right)^2 = 2\left(1-\rho^{(g)}\right),$$ where $\rho^{(g)}$ is the Pearson correlation across receiver cells in batch $B$. Thus, Equation 8 learns relative cell-state–niche relationships and does not identify the absolute scale or offset of the pretrained prediction. **(2) Is the pretrained head-1 output calibrated before Equation 9 and distance-scaler fitting? If so, on which cells?** **The pretrained head-1 output is not directly calibrated or used for the LR analysis.** Instead, we introduce a separate expression-scale additive readout in Section 2.5, Step (0). This readout applies a newly fitted MLP to each frozen sender–receiver pair representation, together with the receiver's cell-type embedding: $$\widetilde I_{ij}^{(g)} = f_{\psi} \left( [e_{ij};\operatorname{Emb}(t(i))] \right)^{(g)}, \qquad \widetilde z_i^{(g)} = \sum_{j\in\mathcal N_i} \widetilde I_{ij}^{(g)}.$$ The readout is fitted against the unstandardized cell-state target using ordinary mean squared error and only the receiver cells in the 20 training blocks of the target specimen. The pretrained pair encoder and both prediction heads remain frozen. This is a newly fitted mapping, not an affine rescaling of the pretrained head-1 influence tensor in Equation 6. **(3) How is the positive affine ambiguity handled?** The downstream LR analysis uses the expression-scale contributions $\widetilde I_{ij}$, not the pretrained head-1 influence values $I_{ij}$. The distance scaler is fitted using $\widetilde I_{ij}$, and the residual-based weights in Equation 9 are calculated using the corresponding readout prediction $\widetilde z_i$. Consequently, the scale ambiguity of the pretrained head-1 output does not enter these downstream calculations. The weights are dimensionless because they are defined as a normalized residual: $$\omega_i = \exp\left( -\frac{ \left|z_i^{(g^{\ast})}-\widetilde z_i^{(g^{\ast})}\right| }{ \displaystyle \max_{i'} \left|z_{i'}^{(g^{\ast})}-\widetilde z_{i'}^{(g^{\ast})}\right| } \right),$$ The additive readout also retains an exact sender-wise decomposition: each contribution is computed from an individual sender–receiver pair before aggregation, and the contributions sum exactly to the readout prediction. This is described in Appendix S10F.
AuthorsOpenReview ↗

W1 and Q1: Scale ambiguity and calibration of influence magnitudes (part II)

(4) How sensitive are the LR rankings to equivalent output transformations?

The original head-1 output is not transformed and reused in the LR analysis; therefore, an affine transformation of that unused output does not alter the LR analysis. For the distance scaler used to construct LR signals, we establish invariance to positive rescaling: multiplying the distance scaler by a positive constant multiplies every LR signal by that constant. Since each LR signal is standardized across training-block receiver cells before LASSO, this positive rescaling leaves the standardized LR features and their ranking unchanged.

We do not claim invariance to an arbitrary additive shift of the distance scaler. Instead, the distance scaler is fitted from the separately trained expression-scale readout, whose scale and offset are determined by fitting the unstandardized cell-state target.

The held-out prediction performance of this additive readout is reported in Table S6. Its all-gene PCC is 0.338, 0.256, and 0.300 for breast, CRC, and melanoma, respectively; the corresponding measured-gene PCC is 0.290, 0.222, and 0.288.

Changes in the manuscript:

  • Section 2.3: Clarification of the affine invariance of Equation 8.
  • Section 2.5, Step (0): Introduction of the expression-scale additive readout.
  • Section 2.5, Steps (1)–(4): Use of the readout for distance scaling, residual weighting, and LR analysis.
  • Appendix S10F: Exact decomposition of the additive readout.
  • Appendix S12: Description and evaluation of the different model readouts.

W2 and Q2: Neighborhood size and spatial generalization

Reviewer’s weakness (W2):

Appendix S7 specifies 10 neighbors for TERRA, ... This limits the interpretation of the results without establishing direct test-label leakage. (Section 3.2, p. 7; Tables 1–2, p. 8; Appendix S7, pp. 18–20.)

Reviewer’s question (Q2):

Do the gains persist with matched neighbor counts or spatial information budgets? Are spatial-block or independent-sample probe results available to distinguish spatial interpolation from transfer to independent regions? These controls would directly affect the interpretation of the reported advantage.

Response:

We answer the two parts separately.

(1) Do the gains persist with matched neighbor counts or spatial information budgets?

Yes, in the direct matched-count comparison with TERRA. We evaluated the final pretrained GITIII-scale model using only its 35, 25, or 10 nearest neighbors at inference, without retraining. With k=10k=10, GITIII-scale achieves measured-gene R² values of 0.134, 0.052, and 0.097 on breast, CRC, and melanoma, respectively, compared with 0.129, 0.040, and 0.077 for TERRA (Birk et al., 2026) using its native k=10k=10 neighborhood.

GITIII-scale also exceeds NCEM-GCN (Fischer et al., 2023) and GAT (Veličković et al., 2018) on all three specimens. These two graph baselines retain their native graph definitions, so the matched-neighbor-count comparison is specifically with TERRA rather than a fully matched spatial-information budget across every baseline.

The truncated GITIII-scale model was pretrained with 49 neighbors. Thus, this comparison matches the number of neighbors supplied at inference, while retaining the pretrained model's original training context.

Why do the revised neighborhood-size results differ from Table 2 in the original submission? The original ablation retrained a separate model for each neighborhood size, whereas the revised analysis uses one fixed model pretrained at k=49k=49 and truncates its input only at inference, without retraining. The evaluation procedure and statistic also changed: the original Table 2 used random cell splits and gene-averaged R² on measured-panel genes; the revised Table 2 uses spatial block splits with an exclusion buffer and pooled R² on all genes. Consequently, the original breast k=10k=10 value (R² = 0.1022) cannot be directly compared with the revised all-gene k=10k=10 value (R² = 0.155). On the revised measured-gene benchmark, the fixed model at k=10k=10 reaches R² = 0.134 (Appendix S13, Table S7). This revised control evaluates performance when the spatial information supplied at inference is restricted; it does not claim to match the models' pretraining neighborhood budgets.

(2) Do the gains persist under spatial-block evaluation, and do the results distinguish spatial interpolation from transfer to independent regions or specimens?

The gains persist under spatial-block evaluation with an exclusion buffer. Each tissue section is partitioned into 25 squares, five composition-matched squares are held out, and the remaining 20 are used for training. A validation cell is retained only if none of its 49 nearest neighbors belongs to a training block. The same spatial partition and exclusion rule are applied to all methods.

The evaluation specimens are entirely excluded from pretraining. The probe is fitted using training-block cells from each target specimen, and performance is evaluated on spatially held-out validation cells. This reduces direct spatial-neighborhood overlap between training and evaluation cells and assesses transfer of the frozen representation to held-out regions within unseen specimens.

The probe is nevertheless fitted separately for each target specimen; cross-specimen transfer of a probe was not evaluated. To provide a complementary probe-free assessment, Table S6 reports the direct outputs of the pretrained prediction heads without target-specimen fitting.

Breast and melanoma are cancer types absent from the pretraining corpus, and MERFISH is an imaging platform absent from pretraining.

Changes in the manuscript:

  • Sections 3.2–3.4: Spatial block evaluation and neighborhood-size comparisons.
  • Tables 1, 2, and S1: Updated results.
  • Appendix S11: Spatial partitioning and exclusion-buffer procedure.
  • Appendix S13: Matched-neighbor comparison with TERRA.
  • Appendix S1: Clarification of the evaluation scope.
Reviewer_uuJUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

this paper proposes GITIII-scale, a pretrained model for learning relationships between cell states and their spatial tumour microenvironments. It integrates specimen-matched spatial transcriptomics and scRNA-seq through imputation and trains on 4.25 million spatial cells. The architecture combines an additive sender–gene influence head with a neighbourhood embedding head. Evaluation reports improved expression prediction on three specimens excluded from pretraining, supported by ablations and an endothelial proliferation case study.

Critical strengths

The paper presents a well-motivated approach to learning how cellular neighbourhoods relate to cell-state variation. I particularly value the alignment between the biological question, the model design, and the evaluation objectives. However, following strengths provide a credible basis for the contribution and support its potential significance.

The within-cell-type prediction objective targets a meaningful biological question and reduces the dominance of cell-type identity. Masking the receiver and same-type neighbours thoughtfully addresses a potential expression-copying shortcut. The directional pair encoder and additive head provide an explicit sender–gene decomposition of the first head’s predictions. The curated corpus and consistent measured-panel improvements across three held-out specimens provide encouraging evidence of representation transfer. Also, the data, capacity, geometry, neighbourhood-size, and probe-width ablations help explain which components contribute to performance.

Critical weaknesses

My main concerns are whether the gains reflect the proposed architecture and whether the learned influences support the biological interpretations. Both require stronger evidence before I could support acceptance.

Equation 8 does not calibrate prediction magnitudes: squared error between separately z-scored predictions and targets is proportional, leaving the objective invariant to positive affine transformations. Without additional calibration, the interpretation of raw influence values and the error-based weights in Equation 9 remains questionable. Also, The generalisation assessment is also limited. Although evaluation specimens are excluded from pretraining, probes are trained on random cell subsets within each specimen, allowing spatial dependence between training and test sets. Spatial block splits with exclusion buffers are needed.

Baseline comparisons introduce another confound: GITIII-scale uses 49 neighbours, TERRA uses 10, and NCEM/GAT use different graphs. Reducing GITIII-scale to 10 neighbours lowers breast R2 from 0.1684 to 0.1022, demonstrating the need for comparisons with matched spatial information.

Imputation further complicates interpretation. Panels of 312 or 500 measured genes are expanded to 20,000, so “All genes” largely measures agreement with SpaIM-derived targets. The authors should confirm that measured-panel targets remain original assay values and assess sensitivity to imputation.

The confounds also limit the comparison with GITIII. Table S5 changes training data, architecture, scale, and potentially gene coverage simultaneously, while gains over this closest predecessor are modest, particularly for colorectal cancer. Comparisons using the same corpus and input coverage, alongside targeted architectural ablations, are needed to establish the specific methodological contribution.

Questions for authors
  1. is there an unstandardised loss or calibration step omitted from Equation 8? How are influence magnitudes and Equation 9 weights calibrated?
  2. Are “Measured” targets original assay values? When are imputation, cell-type means, and normalisation computed relative to the evaluation splits?
  1. Do the gains persist under spatial block splits with exclusion buffers and matched neighbourhood sizes across methods?
  2. How does the method compare with GITIII trained on the same corpus and gene coverage? Can direct head-one and head-two performance be reported separately on all three specimens?
  3. Were LR selection and testing separated? Does the selected association survive spatially separated evaluation and improve upon receptor-only or simple LR-expression baselines?
7 comments

Response to Reviewer uuJU

We thank the reviewer for recognizing the potential of our approach and for identifying important issues concerning prediction calibration, spatial evaluation, baseline comparisons, and biological interpretation. We respond only to the comments raised in this review below.

W1 and Q1: Scale ambiguity and calibration of influence magnitudes

W1 (critical weakness): “Equation 8 does not calibrate prediction magnitudes: squared error between separately z-scored predictions and targets is proportional, leaving the objective invariant to positive affine transformations. Without additional calibration, the interpretation of raw influence values and the error-based weights in Equation 9 remains questionable.”

Q1 (author question): “is there an unstandardised loss or calibration step omitted from Equation 8? How are influence magnitudes and Equation 9 weights calibrated?”

Response:

We answer the two parts of the question separately.

(1) Is an unstandardized loss or calibration step omitted from Equation 8?

No unstandardized calibration loss is omitted from pretraining. Equation 8 intentionally optimizes relative cell-state–niche relationships rather than absolute expression magnitudes. Predictions and targets are independently normalized to zero mean and unit variance for each gene within a batch of receiver cells of the same cell type. Consequently, the objective is invariant to a positive affine transformation of the predictions:

\qquad a_g > 0.$$ Because both vectors are standardized, their mean squared difference satisfies $$\frac{1}{|B|}\sum_{i\in B} \left( \widetilde{\hat z}_i^{(g)}-\widetilde z_i^{(g)} \right)^2 = 2\left(1-\rho^{(g)}\right),$$ where $\rho^{(g)}$ is the Pearson correlation for gene $g$ across the receiver cells in $B$. Thus, the second term is a rescaled correlation objective and does not independently constrain prediction magnitude. We now state this explicitly in Section 2.3. **(2) How are the influence magnitudes and Equation 9 weights calibrated?** The pretrained head-1 influence tensor is not directly used for the expression-scale LR analysis. Instead, Section 2.5, Step (0), introduces a separate expression-scale additive readout. A newly fitted MLP is applied to each frozen sender–receiver pair representation, together with the receiver's cell-type embedding: $$\widetilde I_{ij}^{(g)} = f_{\psi}\\!\left([e_{ij};\operatorname{Emb}(t(i))]\right)^{(g)}, \qquad \widetilde z_i^{(g)} = \sum_{j\in\mathcal N_i}\widetilde I_{ij}^{(g)}.$$ This readout is fitted against the unstandardized cell-state target using ordinary mean squared error and only the receiver cells in the 20 training blocks of the target specimen. The pretrained pair encoder and both prediction heads remain frozen. The readout is a newly fitted mapping, not an affine rescaling of the pretrained head-1 influence tensor in Equation 6. The distance scaler is fitted using the expression-scale contributions $\widetilde I_{ij}$, and Equation 9 uses the residuals between the target $z_i$ and the readout prediction $\widetilde z_i$. The resulting weights are dimensionless because they are obtained by exponentiating a normalized residual: $$\omega_i = \exp\\!\left( -\frac{ \left|z_i^{(g^{\ast})}-\widetilde z_i^{(g^{\ast})}\right| }{ \displaystyle\max_{i'} \left|z_{i'}^{(g^{\ast})}-\widetilde z_{i'}^{(g^{\ast})}\right| } \right),$$ The additive readout preserves an exact sender-wise decomposition: each contribution is computed from an individual sender–receiver pair before aggregation, and the contributions sum exactly to the readout prediction. This is described in Appendix S10F. Finally, each LR signal is standardized across training-block receiver cells before LASSO. Therefore, a positive rescaling of the fitted distance scaler does not change the standardized LR features or their ranking. The held-out prediction performance of the additive readout is reported in Appendix S12. Its all-gene PCC is 0.338, 0.256, and 0.300 for breast, CRC, and melanoma, respectively; the corresponding measured-gene PCC is 0.290, 0.222, and 0.288. **Changes in the manuscript:** - Section 2.3: Clarification of the affine invariance of Equation 8 and the relationship between its two loss terms. - Section 2.5, Step (0): Introduction of the expression-scale additive readout. - Section 2.5, Steps (1)–(4): Use of the new readout for distance scaling and residual weighting. - Appendix S10F: Mathematical justification of the additive decomposition of the readout. - Appendix S12: Additional description and evaluation of the different model readouts.

W2: Spatial dependence between training and evaluation cells

W2 (critical weakness): “Also, The generalisation assessment is also limited. Although evaluation specimens are excluded from pretraining, probes are trained on random cell subsets within each specimen, allowing spatial dependence between training and test sets. Spatial block splits with exclusion buffers are needed.”

Response:

Thank you for this suggestion. We have replaced the original random cell-level splitting procedure with a spatial block evaluation protocol applied consistently to all methods.

Specifically, each tissue section is partitioned into 25 approximately equal spatial squares. Five squares are selected as spatially held-out validation regions, with their cell-type composition matched as closely as possible to that of the entire slide. The remaining 20 squares are used for training.

To further reduce spatial dependence, we introduce an exclusion buffer based on the same 49-nearest-neighbor definition for every method. A validation cell is retained for evaluation only if none of its 49 nearest neighbors belongs to a training block. This prevents an evaluated cell’s defined 49-neighbor neighborhood from directly containing training-block cells.

The same spatial partition and exclusion rule are used for all methods, including pretrained foundation models and dataset-specific baselines. The five repetitions vary model initialization rather than the spatial partition.

Under this spatially separated evaluation, GITIII-scale maintains the best performance across the three held-out specimens (Table 1). We refer to these regions as spatially held-out validation blocks because they are also used for early stopping and performance reporting.

Changes in the manuscript:

  • Section 3.2: Replacement of random cell splitting with spatial block evaluation.
  • Tables 1 and S1: Updated benchmarking results.
  • Appendix S11: Detailed spatial partitioning and exclusion-buffer procedure.
  • Appendix S1: Clarification of the evaluation protocol and its scope.

W3 and Q3: Neighborhood-size differences between methods

W3 (critical weakness): “Baseline comparisons introduce another confound: GITIII-scale uses 49 neighbours, TERRA uses 10, and NCEM/GAT use different graphs. Reducing GITIII-scale to 10 neighbours lowers breast R2 from 0.1684 to 0.1022, demonstrating the need for comparisons with matched spatial information.”

Q3 (author question): “Do the gains persist under spatial block splits with exclusion buffers and matched neighbourhood sizes across methods?”

Response:

We answer the two parts separately.

(1) Do the gains persist under spatial block splits with exclusion buffers?

Yes. We replaced random cell-level splitting with a spatial block protocol applied consistently to all methods. Each tissue section is partitioned into 25 approximately equal spatial squares. Five composition-matched squares are held out, and the remaining 20 squares are used for training. A validation cell is retained only if none of its 49 nearest neighbors belongs to a training block. This exclusion rule is applied consistently across methods.

Under this protocol, GITIII-scale achieves the highest measured-gene performance across all three held-out specimens (Table 1).

(2) Do the gains persist when neighborhood sizes are matched?

Yes, in the direct matched-neighbor-count comparison with TERRA. We evaluated the final pretrained GITIII-scale model using only its 35, 25, or 10 nearest neighbors at inference, without retraining. With k=10k=10, GITIII-scale achieves measured-gene R² values of 0.134, 0.052, and 0.097 on breast, CRC, and melanoma, respectively, compared with 0.129, 0.040, and 0.077 for TERRA (Birk et al., 2026) using its native k=10k=10 neighborhood.

Measured-gene R² on the spatially held-out validation blocks (Xenium Breast, MERFISH CRC, and MERFISH Melanoma):

MethodXenium BreastMERFISH CRCMERFISH Melanoma
GITIII-scale, k=49k=490.1500.0740.118
GITIII-scale, k=10k=100.1340.0520.097
TERRA, native k=10k=100.1290.0400.077
NCEM-GCN0.1200.0330.070
GAT0.1010.0160.041

The targets are the original measured-panel assay values; results are averaged over five training seeds.

GITIII-scale also exceeds NCEM-GCN (Fischer et al., 2023) and GAT (Veličković et al., 2018) on all three specimens, although these two methods retain their native graph definitions. Thus, the matched-neighbor-count comparison is specifically with TERRA, rather than a fully matched spatial-information budget across every baseline.

Why do the neighborhood-size ablation values differ from the original submission? The two tables report different experiments. In the original submission, each neighborhood-size setting used a separately retrained model, and the reported breast value of R² = 0.1022 for k=10k=10 came from that retrained model. In the revision, we instead keep the same final model pretrained with k=49k=49 and supply only its 35, 25, or 10 nearest neighbors at inference, without retraining. This isolates the effect of restricting the inference-time neighborhood of a fixed pretrained model and is directly relevant to comparison with TERRA's native k=10k=10 inference setting.

The evaluation protocol and reported statistic also changed: the original Table 2 used random cell splits and gene-averaged R² on measured-panel genes, while the revised Table 2 uses spatial block splits with an exclusion buffer and pooled R² on all genes. The original 0.1022 and revised 0.155 are therefore not directly comparable. Under the revised protocol, the fixed model's all-gene R² decreases from 0.170 at k=49k=49 to 0.155 at k=10k=10 (Table 2); its measured-gene R² at k=10k=10 is 0.134 (Appendix S13, Table S7). This is an inference-time robustness comparison; it does not imply that a model newly trained with k=10k=10 would achieve the same score.

Changes in the manuscript:

  • Section 3.2: Spatial block evaluation and exclusion buffer.
  • Section 3.3: Addition of the matched-neighbor comparison.
  • Section 3.4 and Table 2: Revised neighborhood-size ablation using inference-time truncation.
  • Appendix S7: Detailed descriptions of NCEM-GCN and GAT graph construction.
  • Appendices S11 and S13: Spatial partitioning and matched-neighbor comparison.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.