Research map / Image generation / Autoregressive and latent generation

Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images

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 CELLO, an end-to-end framework that predicts single-cell gene expression from H&E images using one pathology foundation model forward pass per patch, grid-sampled cell queries, and distance-decay cross-attention, evaluated on 52 Xenium-H&E pairs from HEST-1k.

Problem

Predicting single-cell spatial gene expression from H&E histology is hard: spot-level methods blur cellular heterogeneity, while per-cell PFM cropping is computationally prohibitive and segmentation-based models lack strong visual representations.

Approach

Run a pathology foundation model once per patch, bilinearly sample a feature for each cell location from the token grid, and refine it with cross-attention using 2D RoPE and a Gaussian distance-decay bias over local patch tokens, trained end-to-end on paired single-cell Xenium data.

Key results

• On HVG PCC-50 in-distribution, CELLO averages 0.4390 vs 0.2773 for DeepSpot2Cell and 0.3700 for Grid, and improves over DeepSpot2Cell by 58.5% on Breast/ID-level comparisons reported in the text.
• OOD HVG PCC-50 average of 0.2801 for CELLO vs 0.1635 for DeepSpot2Cell (71.3% relative improvement on HVG PCC-50).
• CELLO is 14.0x faster than DeepSpot2Cell on average excluding cell segmentation (947.2 s vs 67.5 s per slide), but comparable to UNet3+ at similar cost.

Surprising findings

• Adding the global CLS token context consistently degraded performance, suggesting local morphology drives single-cell expression prediction.
• Grid-only querying without cross-attention is a strong baseline that performs close to the full model on several OOD tissues, and is better than CELLO on some OOD metrics (e.g., Kidney HVG-50).

Limitations

• Speed advantage over UNet3+ is marginal; the main gain is accuracy at similar cost rather than speed.
• DeepSpot2Cell used a frozen PFM, so the performance gap partly reflects end-to-end fine-tuning rather than the architecture alone.
• Several OOD tissues (single WSI each, e.g., Bone, Heart) show low absolute PCC and mixed results across gene sets.

Experimental setup

Datasets: 10X-Xenium-52 (from HEST-1k)

Benchmarks: 10X-Xenium-52

Metrics: Pearson correlation coefficient (PCC) inference time

Baselines: UNet3+ DeepSpot2Cell GHIST Grid

Builds on: DeepSpot2Cell GHIST Virchow2 RoPE HEST-1k

Closest papers by meaning

• Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.91)
• Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.90)
• GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow (NeurIPS 2025 · similarity 0.89)
• M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images (AAAI 2025 · similarity 0.89)
• HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis (NeurIPS 2024 · similarity 0.89)
• STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.88)