acceptodds
Under review as a conference paper at ICLR 2027

VAST: Verified Asymmetric Supervision Transfer for Histology-Based Spatial Transcriptomics

Abstract

Predicting spatial gene expression from routine histology offers a scalable comple- ment to spatial transcriptomics, yet most models optimize only a small evaluation panel and discard measurements for thousands of other genes. Although recent work uses these genes as auxiliary supervision, it remains unclear whether en- larging the supervised panel consistently benefits the target genes. We study this question as gene-supervision scaling. Our analysis separates average improvement into gene-level positive and negative transfer and indicates that scaling behavior varies across architectures: adding supervision can improve the aggregate score while degrading a nontrivial subset of target genes. Existing auxiliary-supervision methods give auxiliary labels direct optimization access to the target predictor. Our results challenge the premise behind this design: more supervision does not neces- sarily provide more useful information for every target. We argue that the missing component is a mechanism that controls whether newly added supervision is al- lowed to modify the final target predictor. We therefore formulate auxiliary-gene learning as asymmetric supervision transfer and introduce VAST. For each target gene, VAST first establishes and freezes a target-only predictor as a panel-invariant reference, lets expanded supervision propose a correction to that reference, and ap- plies the correction only to the extent supported by pre-update, slide-level evidence computed exclusively from the training set. This two-stage protect–propose–verify interface implements a simple principle: auxiliary supervision should modify target predictions only through a rejectable transfer interface

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial transcriptomics, computational pathology, histology, gene expression prediction, multi-task learning, auxiliary supervision, representation sharing, negative transfer, scaling

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

This paper studies auxiliary gene supervision scaling in spatial gene expression prediction from H&E images. The authors show that expanding supervised gene panels through multi-task learning does not always improve target gene prediction and may introduce negative transfer. They formulate gene-supervision scaling by gradually increasing auxiliary supervision while keeping target genes fixed, and propose VAST (Verified Asymmetric Supervision Transfer), which uses a protect–propose–verify mechanism to preserve a target-only predictor, learn auxiliary-induced corrections, and selectively apply transfers based on training-slide evidence. Experiments on 15 HEST-1k and STimage datasets demonstrate improved target gene PCC and reduced negative transfer. Overall, the paper identifies an important problem and proposes an interpretable transfer control framework; however, its novelty beyond existing multi-task transfer control methods and the reliability of the verification mechanism require further validation.

Critical strengths

1.This paper does not simply assume that “more gene supervision is always better”, but systematically analyzes the impact of gene-supervision scaling on fixed target genes. The authors show that expanding the supervision panel may improve average performance while also degrading some target genes. This is a practically meaningful observation.

2.Compared with conventional multi-task learning where all tasks directly update shared parameters, VAST restricts auxiliary information through a rejectable transfer interface using a target-only anchor, auxiliary proposal, and evidence-based verification.

Critical weaknesses

1.The methodological novelty of this work is limited and highly related to existing negative transfer control methods in multi-task learning. The core idea of VAST is to preserve a target task model, use auxiliary tasks to generate residual corrections, and use a gate to decide whether to accept the transfer. This idea has strong similarities with existing multi-task learning approaches such as task weighting, gradient surgery, and mixture-of-experts gating. Although the paper applies this idea to the spatial transcriptomics gene supervision scaling scenario, it does not sufficiently clarify the fundamental difference between VAST and existing negative transfer control methods.

2.VAST uses training-only slide-level evidence to estimate the transfer weight α. However, the relationship between this evidence and final test-time generalization performance has not been sufficiently demonstrated.

3.The paper compares with methods such as STNet and BLEEP, but lacks more direct transfer control baselines, such as GradNorm, PCGrad, uncertainty weighting, and learned residual adapters. These methods also aim to address multi-task negative transfer. Without these comparisons, the source of VAST’s advantages remains unclear.

4.VAST assumes that useful information exists in the expanded gene panel and mainly focuses on “how to safely transfer” auxiliary supervision. However, in practical applications, a key question is: which auxiliary genes are worth including? Although the paper discusses gene-supervision scaling, it does not deeply investigate the effects of auxiliary gene quality, gene correlation, and biological pathway relevance on VAST.

5.The authors emphasize that verification does not use validation/test labels, but the evidence is obtained from outer-training slides. For small-scale spatial transcriptomics datasets, the number of training samples is often limited and slide-level variance can be large. Therefore, the stability of training-only estimation requires further analysis.

Questions for authors

1.What is the fundamental difference between VAST and existing negative transfer control methods in multi-task learning, such as gradient surgery, task weighting, and mixture-of-experts gating? Does the main novelty come from the PPV framework or its application to spatial transcriptomics?

2.How can the authors demonstrate that training-only verification evidence can reliably predict transfer quality during testing? Could the authors provide a correlation analysis between α and the actual gene-level transfer gain?

3.Could the authors include more direct multi-task negative transfer baselines, such as PCGrad, GradNorm, uncertainty weighting, and Mixture-of-Experts?

4.How does VAST handle low-quality or irrelevant auxiliary genes? Can it automatically identify biologically irrelevant auxiliary supervision?

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

VAST address the negative transfer issue from auxiliary supervision in histology-based spatial gene-expression prediction and introduce Protect-Propose-Verify framework. It first freezes a target only predictor, then learns auxiliary-informed corrections, and finally verifies utility of auxiliary-informed corrections before integration. VAST is evaluated across 15 datasets from HEST-1k and Stimage and reports competitive results.

Critical strengths
  • The Protect-Propose-Verify design is well motivated. VAST shows that increasing auxiliary-gene supervision does not necessarily improve target-gene prediction.
  • VAST is evaluated, using 15 datasets from HEST-1k and Stimage, and ablation studies are performed evaluating components of protect-propose-verify interface.
Critical weaknesses
  • For CCRCC dataset, Table 4 appears difficult to reconcile with the ablation results in Table 2. Variant C (alpha=1) and Full VAST have PCCs of 0.2536 and 0.2891, respectively, implying an improvement of 0.0355, whereas the gene-quartile ΔPCC values reported in Table 4 average to approximately 0.0066. Could the authors clarify the aggregation or experimental differences between these results and provide the underlying per-gene results?
  • Given that the authors observe that VAST provides greater benefits on larger cohorts, how sensitive is the verification mechanism to the number of training slides, and is there a minimum cohort size required to obtain reliable transfer estimates?
  • There are several abbreviations, including PCC, LCB, MOB, and DkGSB, are used without being explicitly defined in the manuscript or supplementary material.
Questions for authors
  • How do the gene-quartile results in Table 4 relate to ablation presented in Table 2, please see the weakness section which addresses numerical discrepancy in more detail?
Reviewer_nYSSOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates gene-supervision scaling in histology-based spatial transcriptomics prediction, examining when adding auxiliary training genes helps or harms a fixed set of target genes. Rather than updating a single shared predictor, VAST freezes a target-only base model and linearly blends it with an expanded-panel proposal model using a per-gene shrinkage weight derived from training-slide residual utility and cross-slide variance. Across multiple HEST-1k and STImage cohorts, this gated residual formulation curbs gene-level negative transfer and outperforms unconstrained multi-task training and auxiliary-selection baselines.

Critical strengths
  1. The paper examines gene-supervision scaling at the individual target-gene level rather than looking only at panel averages. Showing that aggregate correlation gains under shared multi-task training often mask per-gene degradation on roughly half of the targets is a useful empirical finding.
  2. The experimental controls are well designed. The authors tune supervision panel sizes across the full scaling grid for each baseline, include a matched two-stage target-only control to rule out capacity confounding, and run inner cross-fitting checks.
Critical weaknesses

Conceptually, the method does not resolve cross-gene gradient interference within the representation itself, but bypasses it by training two separate models and applying linear residual shrinkage (analogous to post-hoc stacking) per target gene. While effective as a safeguard, this doubles the training and inference footprint without offering a single unified encoder.

The verification rule (Eqs. 8 to 12) estimates utility on the exact same training slides used to update the Stage-II network and relies on cross-slide standard deviation to penalize unreliable corrections. Because Stage-II parameters fit the training slides more closely over time, in-sample utility is optimistically biased, and on small HEST cohorts with only one or two training slides per fold, the cross-slide variance estimate degenerates.

Questions for authors

In cohorts where an outer training fold contains only one or two slides, how is the cross-slide uncertainty penalty in Eq. 10 computed, and does the lower confidence bound collapse to the unpenalized training utility?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.