acceptodds
Withdrawn from ICLR 2027

STTT: Test-Time Training for Cross-Domain Spatial Transcriptomics Prediction from Histology

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High costs and limited availability of spatial transcriptomics data motivate predicting spatial gene expression directly from routine H&E histology. However, variations in staining, scanners, cohorts, platforms, and organs create substantial domain shifts. Most histology-to-spatial-transcriptomics predictors rely on static inference, leaving unlabeled target slides unable to guide adaptation at prediction time. We propose Spatial Test-Time Training (STTT), a framework for spatially structured adaptation at inference time. STTT combines masked-self reconstruction and bidirectional center–context reconstruction to adapt using unlabeled target-slide morphology and spatial neighborhoods. Low-rank query/value updates, representation damping, and slide-wise reset constrain this label-free adaptation. Across four cross-domain settings, STTT consistently improves mean PCC over source-trained SpaFactor and achieves the highest mean four-task PCC among seven compared TTT and TTA baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Test-time training, spatial transcriptomics, computational pathology, domain adaptation, self-supervised learning.

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_cAGXOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a test-time training framework for cross-domain spatial gene-expression prediction from histology images. Building upon SpaFactor, the method adapts to unlabeled target slides through self-supervised masked feature reconstruction and bidirectional spatial context reconstruction. It employs LoRA for parameter-efficient updates and evaluates performance across four domain-transfer settings. The authors report modest improvements in Pearson correlation over source-only SpaFactor and several test-time adaptation baselines.

Critical strengths

This paper addresses timely problems. Cross-domain generalization is an important challenge in histology-to-transcriptomics prediction.

Critical weaknesses

The authors appropriately acknowledge several limitations, including the evaluation on a single backbone, inconsistent improvements in absolute-error metrics, and additional test-time computational costs. However, these limitations are particularly consequential given the modest empirical gains and limited methodological novelty.

My primary concern is whether the paper convincingly demonstrates the value of the proposed test-time training framework. First, the connection between self-supervised histology reconstruction and improved gene-expression prediction under different types of domain shift remains insufficiently justified. Second, the main comparison is confounded by additional source-stage meta-training, making it difficult to isolate the actual contribution of test-time adaptation. Finally, the ablation results provide mixed evidence for the proposed spatial reconstruction objective.

Although the authors acknowledge the limited generalizability of their experiments, the current evidence remains insufficient to establish a substantial methodological or empirical contribution.

Questions for authors
  1. What is the underlying justification for expecting self-supervised histology reconstruction to improve gene-expression prediction under different types of domain shift?

  2. Can the authors isolate the contribution of test-time adaptation by comparing the same source-trained checkpoint with and without test-time updates across all four settings?

  3. Why is bidirectional spatial context reconstruction necessary when removing this component improves MAE and MSE in the reported ablation?

  4. Does STTT generalize to other histology-to-ST prediction backbones, and how does it compare with stronger recent predictors under matched experimental settings?

  5. What is the additional computational cost per slide, and is the modest improvement sufficient to justify test-time optimization?

Reviewer_xui3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose Spatial Test-Time Training (STTT), a framework aimed at improving the cross-domain generalization of spatial transcriptomics prediction models based on H&E histology images. The framework combines masked self-reconstruction and bidirectional center–context reconstruction to adapt using morphological and spatial information from unlabeled target slides. It improves the model’s adaptation capability through source-stage adaptation-aware expression meta-training and updates only the Q/V LoRA parameters at test time, with representation damping and slide-wise reset constraining adaptation. The authors apply STTT to the SpaFactor predictor and compare it with seven TTT/TTA methods across four cross-domain tasks. The results show that STTT improves mean PCC on all four tasks and achieves the highest four-task macro PCC, although improvements in MAE and MSE are inconsistent.

Critical strengths

The research problem has practical value. The paper offers an interesting perspective on cross-domain generalization in histology-to-spatial-transcriptomics prediction. To address distribution shifts across cohorts, organs, and platforms, the authors explore test-time adaptation using unlabeled target slides to mitigate the effects of cross-domain distribution shifts on gene expression prediction performance.

Critical weaknesses
  1. The effectiveness and necessity of bidirectional center–context reconstruction are insufficiently validated. Table 3 shows that removing center–context reconstruction decreases PCC by only 0.18 percentage points while improving both MAE and MSE. The existing ablation results are insufficient to establish that bidirectional reconstruction provides consistent performance gains or to adequately justify using both reconstruction directions simultaneously.

  2. The existing experiments do not sufficiently establish the independent contribution of test-time adaptation. In the main experiments, full STTT introduces both additional source training and test-time updates compared with source-only SpaFactor, so the overall performance improvements cannot be directly attributed to test-time updates. Although Table 3 includes a No update ablation, it is evaluated only on Compound–Breast and does not establish whether test-time updates are effective on the other cross-domain tasks.

Questions for authors
  1. Why are both context→center and center→context reconstruction used rather than unidirectional reconstruction or standard masked self-reconstruction alone? Could the authors establish their necessity through separate ablations and experiments on the other cross-domain tasks?
  2. Could the authors conduct four controlled experiments combining the presence or absence of adaptation-aware expression meta-training with the presence or absence of test-time updates across the four cross-domain tasks, while keeping the test-time update strategy the same, to quantify the independent gains from test-time adaptation and determine whether meta-training enhances its effectiveness?
  3. What additional computational overhead does STTT introduce compared with static SpaFactor and the other TTT/TTA methods? Could the authors report the adaptation time, total inference time, and peak GPU memory usage per slide?
  4. What accounts for the differences in STTT results between Tables C.4 and E.2 on Compound–Breast under the same random seed (e.g., seed 42)? What adaptation steps, learning rates, and damping coefficients were used for the four cross-domain tasks, and how were they selected?
Reviewer_Z74tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spatial Test-Time Training (STTT) for adapting histology-to-spatial-transcriptomics predictors to unlabeled target slides. It combines masked-self reconstruction with bidirectional center–context reconstruction to exploit target-slide morphology and spatial neighborhoods. Low-rank query/value updates, representation damping, and slide-wise reset are used to constrain adaptation. Across four cross-domain settings, STTT improves mean PCC over a source-trained SpaFactor model and is competitive with or better than several test-time training and adaptation baselines. The paper also reports mechanism and sensitivity ablations. The results are promising, but the evidence is concentrated on one predictor and a limited set of transfer settings.

Critical strengths
  • STTT addresses a practical deployment problem: adapting to a new histology slide without collecting target-slide expression labels. Its use of spatial neighborhoods is well matched to the structure of spatial transcriptomics.
  • The paper compares against multiple TTT/TTA objectives, reports matched-seed results, and includes ablations of the reconstruction mechanisms and adaptation settings. The reported gains across four transfer settings make the approach worth further consideration.
Critical weaknesses
  • Reconstructing frozen GigaPath embeddings and local spatial context on the target slide only adapts visual-spatial representations, whereas cross-organ (such as liver to breast) and cross-platform shifts fundamentally alter the morphology-to-gene mapping and expression scale. Consequently, in Table 1, adaptation increases MAE on the organ-shift task and worsens both MAE and MSE on the platform-shift task, while the correlation gains on the remaining two shifts are well within a single standard deviation across seeds.
  • The paper's own mechanism ablation (Section 5.3) undermines the proposed architecture. Removing the bidirectional center-to-context spatial reconstruction barely changes correlation while achieving the lowest MAE and MSE among all variants, and removing representation damping or updating the full transformer yields higher correlation than the proposed constrained LoRA update.
  • Empirical evaluation is restricted to a single base predictor (SpaFactor) with frozen GigaPath features across four hand-picked transfer pairs, which is too narrow to establish a general test-time training framework for spatial transcriptomics.
Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.