VISTA

Spatial transcriptomics resolves gene expression in situ, but the assays remain costly enough that they are research instruments rather than routine clinical measurements. Hematoxylin-and-eosin-stained histopathology slides are the opposite: cheap, standard, and already archived for nearly every cancer resection and biopsy. VISTA, from the AI for Science Interdisciplinary Research Center at Northwestern Polytechnical University with clinical collaborators at Qilu Hospital of Shandong University, predicts spot-level spatial gene expression directly from those slides, so existing pathology cohorts can be analyzed as though they had been assayed.

Its central argument is about the scale of context available to each prediction. Earlier histology-to-expression methods embed each spot's image tile independently, which caps the receptive field at a few hundred microns and discards the tissue architecture a pathologist reads across an entire slide. VISTA instead composes two frozen pathology foundation models — UNI2 for tile-level morphology and TITAN for whole-slide context — and connects them with a global-to-local cross-attention path, letting every spot query the slide-level representation. A parallel branch models variation inside a spot, which pools transcripts from several cells.

Released as a preprint in August 2026, VISTA joins a growing virtual spatial transcriptomics literature that includes Phoenix and Path2Space. Its distinguishing move is procedural as much as architectural: one model is trained on a HER2-positive breast cancer cohort and then applied without refitting to independent cohorts on a different capture platform, to tumor slides with no paired spatial data at all, and to a second cancer type.

#Key Features

• Frozen foundation-model backbones: Tile and slide encoders are used as fixed feature extractors, so the trainable parameters sit entirely in the fusion and prediction head — a practical choice given how little paired image-expression data exists.
• Global-to-local cross-attention: Spot embeddings act as queries against the slide-level representation as keys and values, expanding each spot's effective receptive field to the whole section.
• Spatial-position-aware branch: Three ConvNeXt branches with 3x3, 5x5, and 7x7 receptive fields are combined by a GatedFusion mechanism that weights the most informative scale per region, then given learnable x and y coordinate embeddings.
• Correlation-regularized objective: Training combines reconstruction loss with a correlation term to absorb the gene-level sequencing-depth differences that otherwise dominate cross-slide variability.
• Virtual spots on unassayed slides: Whole-slide images without spatial ground truth are tiled into background-filtered virtual spots, extending inference to cohorts that were never profiled.

#Technical Details

Tiles of 224x224 pixels centered on each spot pass through the frozen UNI2 encoder; the embeddings are projected to 512 dimensions and aggregated by a multi-head graph attention network over a neighborhood graph whose edges combine k-nearest-neighbor spatial distance with feature cosine similarity. Slide-level context comes from the frozen TITAN encoder applied to the projected spot embeddings and their pixel-level center coordinates. The cross-attention output and the spatial-position-aware output are fused by learned attention weights and decoded by a three-layer MLP. On leave-one-section-out cross-validation over 32 sections from 8 patients in a HER2-positive breast cancer dataset, VISTA reached gene-level and spot-level Pearson correlations of 0.196 and 0.629 against 0.149 and 0.610 for the strongest of seven baselines, and an SSIM of 0.505, 7.91% above the next method. Trained on that cohort and evaluated on five independent 10x Genomics breast cancer datasets over 1,000 shared highly variable genes, it scored 0.552 and 0.374 on the paper's composite spot- and gene-level metrics — a transfer across sections averaging 3,477 spots from training sections averaging 361.

#Applications

The concrete use is retrospective: applying spatial inference to slide archives that will never be assayed. On 79 TCGA breast cancer samples, pseudobulk profiles aggregated from VISTA predictions correlated at 0.48 with matched bulk RNA-seq, and a Cox model over the top five inferred survival-associated genes separated risk groups with a C-index of 0.9 (log-rank p = 0.002); spatial-domain clustering split the cohort into two subtypes with distinct survival (p = 0.0044). The authors further report preserved tumor-normal organization and the spatial biomarkers CLDN4 and CYP3A4 in an intrahepatic cholangiocarcinoma cohort, and prediction of pathological response to neoadjuvant trastuzumab-based therapy in a HER2-positive breast cancer cohort.

#Impact

VISTA is a preprint awaiting peer review, and no code, weights, or hosted inference has been released, so its results cannot yet be reproduced independently; the manuscript carries a CC-BY-ND license. Read as a design study, it is a clear demonstration that composing existing pathology foundation models with a modest trained head outperforms task-specific architectures trained from limited spatial transcriptomics data, and that the gain survives a change of capture platform, spot density, and tumor type. The absolute gene-level correlations remain low, which is characteristic of the task rather than of this model, and downstream conclusions rest on aggregate and spatial-domain signals rather than per-gene accuracy.