Source-linked AI summary
Benchmarking Active Spot Selection for Cost-Efficient Spatial Transcriptomics
Zheyu Zhu, Junchao Zhu, Fengbei Liu, Tianyuan Yao, Gelei Xu, John Cannon, Haichun Yang, Yuankai Huo, Mert R. Sabuncu, Ruining Deng
TL;DR
Dense spatial-transcriptomics capture is costly, and evidence is limited on active selection for continuous expression vectors and spatially correlated candidates. This paper retrospectively benchmarks four active strategies against Random across two cohorts and several budgets. Most active strategies do not consistently improve on Random at small budgets under the reported horizons, and rankings depend on cohort and measure.
Problem
The paper asks whether selected spots can provide sufficient supervision for morphology-to-expression learning, given costly dense acquisition and spatially correlated expression.
Source: §1 · pp. 1–2Method
The authors retrospectively mask labels in two fully profiled cohorts and compare four active strategies with Random under matched budgets, predictors, and schedules.
Source: §1 · p. 2Results
Takeaways & Limitations
At small budgets, the evaluated active strategies show no consistent advantage over Random, while larger-budget advantages depend on cohort and evaluation measure.
Source: §7 · p. 8, §8 · p. 8Comparisons are descriptive: four overlapping folds and one seed do not separate acquisition from training or sampling variation.
Source: §6 · p. 7
Abstract
from arXiv · showhide
Spatial transcriptomics (ST) measures gene expression in tissue context, but dense capture grids can be costly and may repeatedly sample morphologically similar regions. Most active learning strategies were developed for categorical labels and independent samples. We conduct a retrospective pool-based benchmark of active learning versus uniform Random sampling for ST, where expression vectors are high-dimensional and continuous and candidates are spatially correlated. Using two fully profiled public ST cohorts, we mask candidate expression vectors and simulate multi-round selection with uncertainty-based Monte Carlo dropout (MC-dropout) and temporal output discrepancy (TOD), and diversity-based CoreSet and TypiClust-inspired selection. We compare 160 completed configurations at 5%, 10%, 30%, and 50% of the fold-wide training spot pool under patient-level cross-validation, with a separate full-label reference. Within each budget, strategies share the selection schedule, morphology-to-expression predictor, and optimization protocol. We assess mean per-gene within-slide Pearson correlation coefficient (PCC), expression-cluster agreement, and Moran's I fidelity. On HER2-positive breast cancer, pooled mean PCC differences from Random across the four active strategies were -0.0176, -0.0117, +0.0056, and +0.0057 at 5%, 10%, 30%, and 50%, respectively. On cutaneous squamous cell carcinoma (cSCC), three strategies were below Random at 5%, and all four were below Random at 10%. On HER2-positive breast cancer, CoreSet and MC-dropout had lower PCC but higher expression-cluster agreement than Random at the two smallest budgets; this pattern did not reproduce on cSCC. Under the reported fixed training horizons, the evaluated active strategies do not consistently improve on Random at small budgets, and rankings depend on the evaluation measure.
1. INTRODUCTION
Dense spatial-transcriptomics capture can be costly and may sample morphologically similar regions, motivating tests of whether selected spots provide sufficient supervision. This benchmark compares standard active-selection strategies with Random under matched conditions and multiple budgets and evaluation measures.
1. INTRODUCTION
- Spatial-transcriptomics acquisition is costly, and spots in morphologically homogeneous regions may have similar expression, potentially limiting the number of patients and sections profiled.These properties motivate testing measurement redundancy rather than assuming it.Source: §1 · p. 1, Figure 1 · p. 2
- The benchmark tests whether selected spot subsets can provide sufficient supervision for learning the morphology-to-expression relationship, compared with uniform subsets of the same size.The premise is tested retrospectively before customized capture regions or region-selective assays are developed.Source: §1 · p. 2
- The benchmark compares four uncertainty- and diversity-based strategies with Random under shared predictors and schedules across four budgets and three evaluation measures.The measures are mean per-gene within-slide PCC, expression-cluster agreement, and Moran’s I fidelity.Source: §1 · p. 2
2. METHODS
The retrospective benchmark reveals expression labels for selected spots from a fold-wide pool, then trains one shared morphology-to-expression predictor. It compares Random with four active-selection strategies under matched budgets and protocols.
2.1 Retrospective Active Spot-Selection Benchmark
- Labels are masked in fully profiled cohorts and revealed for selected spots; the pool spans training-fold patients and slides without enforcing equal slide coverage.Strategies can use candidate images and revealed expression, but not labels for unselected spots or data from held-out patients.Source: §2.1 · pp. 2–3
- Six-round runs use budget-specific quotas and shared schedules; the first batch uses Random, and the full-label reference is contextual rather than a schedule-matched upper bound.The predictor continues training between rounds; the full-label reference has no acquisition step and is trained for 50 epochs.Source: §2.1 · p. 3
- The benchmark selects individual spots, not prospective sequencing regions, and does not test expert-guided acquisition or region-selective assay feasibility.Strategies omit explicit pathologist-defined morphology, tissue-compartment constraints, and clinical knowledge.Source: §2.1 · p. 3
2.2 Acquisition Strategies
- MC-dropout ranks aggregated predictive variance across genes, while TOD scores temporal changes in candidate expression predictions and uses Random when no prior cache is available.The MC-dropout aggregate can be influenced by gene scale because targets are not standardized.Source: §2.2 · p. 3
- CoreSet greedily selects candidates farthest from labeled centers in learned image-feature space; TypiClust-inspired selection chooses typical members of clusters formed from task-predictor features.The TypiClust-inspired implementation sets its cluster count to the round quota and uses Random in its first round.Source: §2.2 · p. 3
2.3 Shared Morphology-to-Expression Predictor
- All strategies share an ST-Net-style predictor with a trainable DenseNet-121 trunk and regression head, controlling architecture within this benchmark.The shared predictor does not establish that acquisition rankings transfer to other backbones or objectives.Source: §2.3 · p. 3
3. DATA AND EXPERIMENTAL DESIGN
The benchmark uses two public legacy Spatial Transcriptomics cohorts and budget-specific training settings. It evaluates pointwise accuracy, expression-cluster agreement, spatial-expression fidelity, and acquisition-scoring time, with descriptive strategy comparisons.
3.1 Datasets and Experimental Setup
- The benchmark uses two public legacy HEST-1k cohorts: HER2-positive breast cancer and cutaneous squamous cell carcinoma.Source: §3.1 · p. 4
- Training uses budget-specific horizons, and optimizer exposure differs across budgets, so cross-budget comparisons do not match optimization or convergence.The reported step counts are 715 at 5% and 3,156 at 50%.Source: §3.1 · p. 4
3.2 Evaluation and Reporting
- PCC averages within-slide, per-gene correlations; ARI compares independently clustered predicted and measured expression profiles, not histological domain recovery.Slides with fewer than 2k spots are excluded from the ARI measure.Source: §3.2 · pp. 4–5
- Moran’s I fidelity is the across-gene PCC between measured and predicted Moran’s I; active-minus-Random differences and fold sign counts are reported descriptively.Overlapping folds, a shared Random comparator, and one seed motivate not reporting pooled p-values; acquisition timing excludes training.Source: §3.2 · p. 5
4. RESULTS
Active strategies did not consistently outperform Random, with paired PCC comparisons varying by cohort and budget. Spatial rankings also depended on the measure, and PCC increased with budget while fold variation exceeded strategy differences.
4.1 Gene-Expression Accuracy Across Budgets
- PCC increased with budget in both cohorts, but fold variation exceeded strategy mean differences; Random at 50% reached 98% and 92% of full-label means on HER2-positive cancer and cSCC, respectively.These ratios are descriptive because the full-label reference used 50 epochs, versus 60 epochs for the 50% runs.Source: Table 3 · p. 5, §4.1 · p. 5
4.2 Paired Comparison With Random
- At 5%, 10%, 30%, and 50%, HER2-positive pooled PCC differences from Random were −0.0176, −0.0117, +0.0056, and +0.0057; cSCC strategies mostly trailed at low budgets.On cSCC, all four strategies slightly exceeded Random at 50%.Source: Table 4 · p. 5, §4.2 · p. 5, Table 5 · p. 6
4.3 Spatial-Expression Structure
- At 5% and 10% on HER2-positive cancer, CoreSet and MC-dropout had lower PCC but higher ARI than Random; this ARI reversal did not occur on cSCC.On cSCC, every active strategy had lower ARI than Random at both budgets.Source: Table 6 · p. 7, §4.3 · p. 7
- Moran’s I fidelity had no consistently best active strategy: Random led at all HER2-positive budgets except the small TypiClust-inspired mean differences, while the cSCC leader varied with budget.Source: Table 7 · p. 7, §4.3 · p. 7
6. DISCUSSION AND FUTURE WORK
Under the reported training horizons, most active strategies trailed Random at small budgets, while gains at larger budgets were small. The comparisons remain descriptive, and prospective or contiguous-region acquisition was not tested.
6. DISCUSSION AND FUTURE WORK
- Under the reported horizons, most active strategies trailed Random at 5% and 10%, while mean PCC gains at larger budgets were small.Selected regions were not analyzed, so their possible contribution remains uncertain; unequal optimization exposure is an alternative explanation.Source: §6 · p. 7
- Four overlapping folds and one seed do not separate acquisition effects from training or sampling variation, making the comparisons descriptive.Small-budget models were not converged, targets were unnormalized log(1 + cig) counts, and the full-label reference was not schedule-matched.Source: §6 · p. 7
- Prospective and contiguous-region acquisition remain untested, so the observed crossover is not a general budget threshold.The benchmark also covers two cohorts on one platform, one backbone, and four automated model-based strategies.Source: §6 · pp. 7–8
7. NEW OR BREAKTHROUGH WORK TO BE PRESENTED
The study benchmarks four active-selection strategies against Random under matched spot budgets and a shared predictor. Most active strategies underperform Random at small budgets, while small larger-budget advantages depend on cohort and evaluation measure.
7. NEW OR BREAKTHROUGH WORK TO BE PRESENTED
- The benchmark compares four standard active-selection strategies with Random under matched spot budgets and a shared predictor.It reports both pointwise and spatial-expression measures.Source: §7 · p. 8
- At small budgets, most active strategies underperform Random under the reported training horizons.Source: §7 · p. 8
- At larger budgets, small active-strategy advantages depend on cohort and evaluation measure.Source: §7 · p. 8
8. CONCLUSION
Under matched conditions, the retrospective evidence does not show a consistent small-budget advantage for the evaluated active strategies. Rankings can vary by evaluation measure, and future region-selective studies should report key settings and validate computational selection in a physically realizable assay.
8. CONCLUSION
- The retrospective evidence does not show a consistent advantage for active strategies at small budgets.At 5% and 10% of the fold-wide training pool, most active-strategy PCC means are below Random; at 30% and 50%, mean differences are positive but small relative to fold-to-fold variation.Source: §8 · p. 8
- On HER2-positive breast cancer, low-budget rankings differ between PCC and expression-cluster agreement, but this reversal does not reproduce on cSCC.Source: §8 · p. 8
- The full-label reference contextualizes budgeted results but is not schedule-matched.Source: §8 · p. 8
- Future region-selective ST studies should report acquisition budget, training horizon, and evaluation measure, and validate computational selection in a physically realizable assay.Source: §8 · p. 8