Research map / Image generation / Autoregressive and latent generation

Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Retrieval and matching with generative features · 0 citations (OpenAlex)

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 image-based differential expression ranking (IDER), evaluating and training histology-to-ST predictors so that predicted profiles preserve contrast-specific DEG rankings, using a differentiable Mann-Whitney surrogate trained on morphology-derived K-means proxy contrasts.

Problem

Histology-based spatial gene expression prediction is trained and evaluated by per-gene spatial reconstruction, which is misaligned with downstream differentially expressed gene (DEG) discovery that ranks genes across tissue contrasts.

Approach

Define IDER as preserving the gene ranking induced by one-vs-rest Mann-Whitney U statistics, and optimize a differentiable sigmoid-relaxed U-statistic with a gene-axis Pearson correlation loss, using CONCH-clustered morphology groups as proxy contrasts so no biological labels are needed.

Key results

• On Lymph Node cluster-based DEG setting, the proposed objective reaches SCC-DEG 0.676 and nDCG-DEG@200 0.643 versus 0.650 and 0.548 for the best baseline (MSE & PCC).
• On Her2st real-annotation setting, average SCC-DEG improves to 0.353 from 0.239 (best baseline MSE) and nDCG-DEG@200 to 0.247 from 0.109 (STRank).
• Pathway Jaccard overlap on Lymph Node rises to 0.571 versus 0.443 for the best baseline (NB).

Surprising findings

• Methods with higher per-gene spatial PCC do not necessarily achieve better DEG-ranking agreement or pathway overlap, and the proposed method is not best under per-gene PCC.

Limitations

• Evaluation of DEG rankings relies on one-vs-rest contrasts and a single top-k cutoff set, and the Her2st real-annotation evaluation excludes several tissue classes due to small sample sizes.
• Uses frozen CONCH features with only a three-layer MLP trained, so the gains depend on the feature extractor.
• Per-gene spatial PCC is not best, indicating a trade-off with conventional reconstruction quality.

Experimental setup

Datasets: HEST-1K (Bowel, Ovary, Lymph Node) Her2st

Benchmarks: HEST-1K Bowel HEST-1K Ovary HEST-1K Lymph Node Her2st

Metrics: SCC-DEG (Spearman) nDCG-DEG@50 nDCG-DEG@100 nDCG-DEG@200 Jaccard pathway overlap per-gene PCC

Baselines: MSE PCC MSE & PCC Poisson NB STRank

Builds on: STRank ST-Net TRIPLEX CONCH Mann-Whitney U test

Closest papers by meaning

• Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.91)
• STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.89)
• M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images (AAAI 2025 · similarity 0.89)
• MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images (CVPR 2025 · similarity 0.89)
• HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis (NeurIPS 2024 · similarity 0.88)
• Multi-modal Topology-embedded Graph Learning for Spatially Resolved Genes Prediction from Pathology Images with Prior Gene Similarity Information (CVPR 2025 · similarity 0.88)