Spatial transcriptomics, which reads gene activity while preserving each cell's location in tissue, is slow and expensive; the field has long tried to predict it from routine pathology slides.
For decades, tissue diagnosis has begun with a familiar, cheap tool: the H&E (Hematoxylin and Eosin) stained whole-slide image sitting in nearly every clinical archive. Spatial transcriptomics, which reads gene expression while preserving each cell's location in tissue, is the gold standard for understanding tumor microenvironments and tissue architecture, but it remains slow and expensive. For years, the field has tried to teach a model to bridge the two: read the molecular layer from the cheap slide. The results have looked numerically solid in benchmarks. To biologists, the predictions still feel hollow, numerically close to ground truth but biologically unfaithful, a failure a new ICML 2026 paper calls the 'Gene Dimension Curse.' A team from the Shanghai Academy of AI for Science (上海科学智能研究院, known in shorthand as 上智院) with Shanghai Jiao Tong University and Fudan University has named this failure mode and built a framework around it: FLAG (Foundation-model representation with Latent diffusion Alignment via Graph), a structure-aware latent diffusion model that reframes spatial gene-expression prediction from per-gene pointwise regression to structured distribution modeling, preserving the gene-gene regulatory relationships and spatial autocorrelation that standard loss functions average away.
The team's arXiv preprint and HTML version describe why the Gene Dimension Curse emerges as a model is asked to predict more genes at once: empirical correlations among genes concentrate, forcing the joint denoising objective to approximate a high-curvature score field, and producing a theoretically grounded lower-bound gap between joint and per-node losses. In practice, predictions become numerically close to ground truth, with high per-gene correlation (PCC) and low mean-squared error (MSE), but biologically unfaithful.
This is the implicit critique the team levels at prior work: results that look numerically similar but lack biological substance, a pattern biologists describe as hollow even when per-gene scores look solid. It is also why benchmarks optimized on PCC and MSE have not translated into trustworthy starting points for downstream analyses like differential gene-expression (DEG) calling, spatial-domain identification, or pathway recovery.
FLAG's recipe has three explicit pieces. First, a fixed spatial graph encoder enforces topological consistency, keeping neighboring tissue spots close in feature space regardless of how gene-level predictions shift. Second, a conditional latent diffusion Transformer generates expression inside that fixed spatial scaffold. Third, and only at training time, the gene latent space is aligned with a frozen gene foundation model such as Geneformer, scGPT, or CellPLM, so that gene-gene fidelity is preserved without forcing the deployed model to depend on those larger networks at inference.
To measure whether structure actually survives prediction, the authors propose two new metrics: Gene Structural Correlation (GSC) and Spatial Structural Correlation (SSC), which capture gene-gene and spatial autocorrelation patterns that pointwise PCC and MSE average away. On the HEST-1k benchmark (HER2ST, KIDNEY, PRAD) and DLPFC, they report GSC and SSC leadership against baselines including STFlow, with competitive pointwise error. These are the authors' own benchmark numbers, with no third-party reproduction yet cited, and GSC and SSC are newly proposed in the same paper, so field-wide adoption is the next test rather than settled fact.
The stake is methodological, not clinical. Models that minimize per-gene numerical error, the authors argue, average away the very structures (regulatory networks, spatial autocorrelation) that downstream biological analyses actually consume. A framework that preserves those structures gives researchers more trustworthy starting points, even when its per-gene score trails a regression-only baseline by a fraction. The team has released code at darkflash03/FLAG, with reported training cost of roughly 35 seconds per epoch on a single NVIDIA H800 at about 4.5 GB peak memory, supported by the 星河启智 (Xinghe Qizhi) scientific intelligence open platform.
What to watch next: independent benchmarks against the new GSC and SSC metrics on HEST-1k and DLPFC; whether downstream tools (spatial-domain callers, DEG pipelines) recover known biology more reliably on FLAG outputs than on PCC-tuned baselines; and whether other groups adopt or refute the Gene Dimension Curse framing. Giving a failure mode a peer-reviewed name is, by itself, a handle for the field to argue with.