GHIST+

Spatial transcriptomics resolves gene expression inside intact tissue, but profiling whole slides across patient cohorts remains prohibitively expensive. Most studies fall back on tissue microarrays (TMAs) or sparse profiling designs, where measurements cover only a small fraction of the section — sometimes under 10% of the paired whole-slide image — and different cohorts use only partially overlapping gene panels. GHIST+ reframes H&E-to-expression modelling from a per-slide regression task into a reconstruction task: sparsely profiled tissue cores act as molecular anchors from which single-cell expression is propagated across the surrounding unprofiled tissue, a setting the authors term "core-to-WSI".

GHIST+ was developed by the Sydney Precision Data Science group at the University of Sydney with collaborators at the National University of Singapore, and posted to bioRxiv in June 2026. It succeeds GHIST (Nature Methods, 2025) from the same group, extending single-slide, single-panel prediction to joint training across slides, cancer types and heterogeneous panels. Where the earlier model learned its histology encoder from scratch, GHIST+ sits on a frozen general-purpose pathology foundation model — the UNI2-h checkpoint from UNI — and applies pretrained checkpoints to new tissue, including pan-cancer-assisted zero-shot inference on unseen slides.

#Key Features

• Core-to-WSI anchoring: Measurements from selected TMA cores are supplied to a pretrained model to reconstruct expression in withheld regions of the same slide. Increasing anchor area from 0.5 mm² to 3.5 mm² raised correlation on the top 20 spatially variable genes by 14.5%.
• Edge-Conditioned Residual Mixer (ECRM): A graph module that refines each cell using neighbours selected by edge features encoding spatial distance, relative position, morphology similarity and cell-type agreement, gated so the cell's own morphology signal is retained.
• Vector-quantised tissue prototypes: Pooled tile features are matched to a compact codebook of recurring tissue states, which supplies tile-level context during prediction and identifies representative regions for TMA core selection.
• Mask-Aware Gene Conditioned Imputer (MAGCI): An observed-gene mask excludes unmeasured genes from the training target while the model learns to recover genes absent from a given panel, letting heterogeneous panels share one output space.
• Multi-task supervision: Cell-type heads driven by histology and by predicted expression are aligned to each other, with tile-level composition supervised alongside per-cell expression.

#Technical Details

Each 224×224 H&E patch passes through the frozen UNI2h encoder; per-cell embeddings come from mean-pooling encoder features within HoVer-Net nuclei masks and concatenating them with a globally pooled tile representation. A two-step ECRM operates over a centroid-based k-nearest-neighbour cell graph (k = 12 within a patch, augmented with k = 8 cross-patch edges), pooled tile features are discretised into 64 prototype codes, and all slides map into a shared union gene space with a binary gene mask and slide identity. The loss combines zero-aware and masked-correlation expression terms with cell-type, composition and auxiliary imputation/VQ terms. Training used 13 10x Xenium slides spanning breast, colorectal, lung and pancreatic tissue — 2,785,236 cells and 1,275 unique genes after preprocessing — for 30 epochs at batch size 64 on a single 48 GB NVIDIA RTX A6000, with Macenko stain normalisation and HED colour augmentation.

On within-slide evaluation of Xenium BreastCancer2, GHIST+ reached a median PCC of 0.744 across the top 20 spatially variable genes versus 0.714 for GHIST and 0.521 for SpatialEx, and 0.653 on an independent slide. For unmeasured-gene recovery across five Xenium breast slides with synthetically masked genes, it achieved the highest median PCC (0.743, 0.747) and lowest median RMSE (2.904, 3.010) against PASTA, SpaGE, SpaIM, stImpute, Tangram and SpatialEx, with correlation matrix distance below 0.15 and a TISSUE calibration score of 0.343. Ablating the VQ prototypes, ECRM and the foundation encoder cut top-20 SVG correlation by 5.5%, 11.5% and 16.9% respectively.

#Applications

GHIST+ targets cohort-scale studies where spatial profiling budgets cover only a fraction of the tissue. Groups running TMA-based experiments can extend a handful of profiled cores into whole-slide molecular maps, and use prototype-based region selection to choose which cores to profile — useful because visually similar regions can carry distinct immune, apoptosis and hormone-response programmes. Applied to 77 normal GTEx female breast sections, it recovered epithelial, stromal and adipose compartments and BANKSY-defined niches, and detected age-associated declines in B- and T-cell abundance alongside rising ECM-remodelling scores in a mixed epithelial-immune niche.

#Impact

GHIST+ pushes histology-to-expression modelling past the per-slide, per-panel setting, showing that pan-cancer training yields representations transferable to unseen cancers and to normal tissue: median PCC on held-out pancreatic adenocarcinoma rose from roughly zero for single-slide models to 0.236 under pan-cancer-assisted zero-shot inference. Limits remain — it reconstructs targeted Xenium-scale panels rather than the full transcriptome, requires paired nuclei segmentation and stain normalisation, and anchored reconstruction trades accuracy against how much tissue is profiled. The code is GPL-3.0 and the preprint CC BY-NC-ND 4.0; the work awaits peer review.