Research map / Image generation / Autoregressive and latent generation
Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images
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.
Stem applies a conditional DiT-based diffusion model to predict spatially resolved gene expression from H&E patches, conditioned on pooled pathology foundation model embeddings, and introduces a relative variation distance metric.
Problem
Predicting spatially resolved gene expression from H&E histology is usually cast as deterministic regression, which ignores one-to-many image-to-expression relationships and is evaluated mainly with Pearson correlation that can overstate performance.
Approach
Model the conditional distribution of gene expression given an H&E patch with a DDPM whose backbone is a DiT adapted to gene tokens, with conditioning from pooled CONCH+UNI foundation model embeddings via adaLN, and evaluate with a new relative variation distance (RVD) capturing gene variance fidelity.
Key results
- On Kidney Visium HMHVG, Stem reaches PCC-10 0.5893 and RVD 0.0751 versus TRIPLEX PCC-10 0.4654 and RVD 0.5871.
- On HER2ST HMHVG, Stem reaches PCC-10 0.8298 and RVD 0.0693 versus BLEEP PCC-10 0.7727 and RVD 0.6025.
- On the mouse brain dataset, Stem reaches PCC-10 0.4908 versus BLEEP 0.3419; TRIPLEX could not be run due to GPU memory limits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Stem instead of BLEEP | Kidney Visium HMHVG · Kidney Visium HMHVG | RVD: 0.2170 -> 0.0751 | Helped |
| use Stem instead of BLEEP | HER2ST HMHVG · HER2ST HMHVG | RVD: 0.6025 -> 0.0693 | Helped |
| use Stem instead of TRIPLEX | Human prostate Visium HMHVG · Human prostate Visium HMHVG | PCC-10: 0.6173 -> 0.6103 | Hurt |
| combine CONCH and UNI embeddings instead of using CONCH alone (ablation) | Stem · Kidney Visium HMHVG; no image augmentation | PCC-10: 0.3662 -> 0.4817 | Helped |
| combine CONCH and UNI embeddings instead of using UNI alone (ablation) | Stem · Kidney Visium HMHVG; no image augmentation | PCC-10: 0.4690 -> 0.4817 | Helped |
| increase image augmentation ratio from 2:1 to 1:4 (ablation) | Stem · Kidney Visium HMHVG | PCC-10: 0.4843 -> 0.5485 | Helped |
| hold out 21-0057 (CKD) instead of 20-0038 (AKI) (ablation) | Stem · Kidney Visium HMHVG | PCC-10: 0.5893 -> 0.7026 | Helped |
| use sample median instead of sample mean to summarize generated samples (ablation) | Stem · HER2ST 1000 DEGs | PCC-10: 0.8279 -> 0.8222 | Hurt |
| use sample median instead of sample mean to summarize generated samples (ablation) | Stem · HER2ST 1000 DEGs | MAE: 0.5649 -> 0.5616 | Helped |
| replace sample mean with sample mode to summarize generated samples (ablation) | Stem · HER2ST 1000 DEGs | RVD: 0.1208 -> 0.0928 | Helped |
| use Virchow-2 instead of CONCH + UNI as image encoder (ablation) | Stem · Kidney Visium HMHVG | PCC-10: 0.5893 -> 0.5404 | Hurt |
| use H-Optimus-0 instead of CONCH + UNI as image encoder (ablation) | Stem · Kidney Visium HMHVG | PCC-10: 0.5893 -> 0.5435 | Hurt |
Surprising findings
- The mean prediction across genes can yield surprisingly high log-space PCC, so PCC alone can overstate model quality.
- Retrieval with ResNet18 or UNI embeddings alone performs competitively on PCC, and larger foundation models (Virchow-2, H-Optimus-0) did not beat CONCH+UNI.
Limitations
- Relies on a single sample mean over generated samples; other statistics were not consistently better.
- Evaluated on a small number of held-out slides, and TRIPLEX could not be run on the mouse brain dataset.
- Cross-slide and cross-dataset generalization is shown only on limited holdouts; the paper notes improved sample statistics remain future work.
Experimental setup
Datasets: Kidney Visium HER2ST PRAD Visium (HEST-1k) healthy mouse brain Visium (HEST-1k)
Benchmarks: Kidney Visium HER2ST PRAD Visium healthy mouse brain Visium
Metrics: PCC-k MAE MSE RVD
Baselines: HisToGene BLEEP TRIPLEX ST-Net Hist2ST
Builds on: DDPM DiT adaptive LayerNorm (adaLN-Zero) BLEEP TRIPLEX
Closest papers by meaning
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.92)
- HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis (NeurIPS 2024 · similarity 0.92)
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · similarity 0.91)
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.91)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.91)
- GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow (NeurIPS 2025 · similarity 0.90)
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.