Research map / Probabilistic and scientific modeling / Molecules and proteins
Learning Relative Gene Expression Trends from Pathology Images in Spatial Transcriptomics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STRank, a count-based probabilistic ranking loss (binomial pairwise, multinomial listwise) that learns relative gene expression trends from pathology images instead of absolute expression values, to be robust to batch effects and stochastic noise in spatial transcriptomics.
Problem
Estimating gene expression from pathology images with point-wise losses (MSE) is hurt by batch effects and stochastic count noise in spatial transcriptomics, so absolute expression values are hard to estimate accurately.
Approach
Model the observed expression counts of paired or listwise patches within a tissue as binomial (pairwise) or multinomial (listwise) draws parameterized by softmax over predicted rank scores, and train by negative log-likelihood, so the loss weights pairs by count magnitude and becomes a noise-aware relaxation of the standard ranking loss.
Key results
- On real HEST-1k datasets with a frozen CONCH encoder and trained FC head, ListSTRank (0.345) and PairSTRank (0.343) achieve the best average SCC versus MSE (0.306) and PCC (0.326).
- On the synthetic 1D batch-effect task, the proposed STRank losses consistently outperform MSE, Poisson, NB, Rank and PCC baselines in uniform and imbalanced settings.
- STRank did not consistently beat baselines on real data; on the multi-cohort COAD Visium set with 250-2000 genes, it underperformed MSE and PCC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ListSTRank instead of conventional losses | CONCH expression head · Real datasets average | SCC: 0.326 -> 0.345 | Helped |
| Use STRank rather than pointwise losses | Expression estimation on real HEST-1k datasets · IDC, PAAD, ccRCC, IDC-L | SCC | Helped |
| Use STRank rather than conventional losses | Expression estimation on real HEST-1k datasets · PRAD, COAD, READ | SCC | Hurt |
| Use ListSTRank instead of PairSTRank | Synthetic data with batch effects · Synthetic uniform and imbalanced settings | SCC | Helped |
| Set ListSTRank list size N^k to 8 rather than 2 (ablation) | ListSTRank · Uniform synthetic | SCC: 0.907 -> 0.958 | Helped |
| Set ListSTRank list size N^k to 128 rather than 2 (ablation) | ListSTRank · Imbalanced synthetic | SCC: 0.818 -> 0.845 | Helped |
| Set ListSTRank list size N^k to 32 rather than 2 (ablation) | ListSTRank · B Xenium | SCC: 0.447 -> 0.463 | Helped |
| Vary ListSTRank mini-batch sampling from Default to Intra-tissue (ablation) | ListSTRank · 250 and 1000 gene sets | SCC: 0.175 -> 0.177; 0.110 -> 0.105 | Mixed |
| Vary ListSTRank mini-batch sampling from Default to Inter-tissue (ablation) | ListSTRank · 250 and 1000 gene sets | SCC: 0.175 -> 0.173; 0.110 -> 0.105 | Mixed |
| Use PairSTRank and ListSTRank instead of Rank and PCC | Pairwise/listwise expression estimation · Downsampled real data at p=0.01 | SCC | Helped |
| Increase target gene-set size while using STRank (ablation) | HER2ST gene expression estimation · HER2ST, 50–9385 genes | SCC | Mixed |
| Use STRank on multi-cohort COAD with larger gene sets | COAD Visium expression estimation · COAD, 250–2000 genes | SCC | Hurt |
Surprising findings
- On multi-cohort COAD with large gene sets, STRank underperformed MSE and PCC, and ListSTRank degraded as gene count grew.
- Mini-batch sampling strategy did not change ListSTRank performance, which the authors attribute to possible numerical effects.
Limitations
- STRank does not consistently outperform baselines on real datasets and the reasons are not fully explained.
- Converges more slowly than PCC on single-patient data.
- Performance degrades in multi-cohort settings with many target genes, where cohort-specific noise is not modeled.
Experimental setup
Datasets: HEST-1k (IDC, PRAD, PAAD, COAD, READ, ccRCC, IDC-LymphNode) HER2ST synthetic 1D negative binomial data
Benchmarks: HEST-1k benchmark datasets (IDC, PRAD, PAAD, COAD, READ, ccRCC, IDC-LymphNode) HER2ST synthetic 1D NB data
Metrics: Spearman correlation coefficient
Baselines: MSE Poisson loss Negative Binomial loss Rank loss PCC loss
Builds on: learning to rank RankNet-style pairwise ranking loss CONCH
Closest papers by meaning
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.89)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.88)
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.88)
- M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images (AAAI 2025 · similarity 0.88)
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.87)
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.