Models that predict spatial gene expression from ordinary microscope slides are scored almost entirely by per-spot agreement: how well, spot by spot, the predicted level of each gene correlates with the measured level. Cosmologists hit the same trap when they emulated the large-scale structure of the universe. A fast approximate simulation can match the matter density nearly point by point and still get the clumpy, filamentary small-scale structure badly wrong. The field responded by grading emulators on scale-by-scale statistics: the two-point correlation at each distance, three-point correlations, skewness, and environment-weighted ("marked") correlations. This note proposes to bring that grading discipline to tissue expression prediction and then to train against it. The core intuition is simple. A regression model that is unsure where a tissue-restricted gene switches on will predict a blurred average. That blur costs little on per-spot correlation, but it erases the sharp, sparse, skewed patterns that make expression tissue-specific in the first place.
The concrete opportunity is on HEST-1k, a benchmark pairing H&E (hematoxylin and eosin, the standard pink-purple stain) whole-slide images with spot-level expression across ten organs. There, the best reported model reaches only $R^2 = 0.34$. We propose a suite of spatial summary statistics computed on the irregular spot lattice, an audit of current predictors with it, and a loss that matches statistics, scale by scale. The audit tests whether these predictors systematically lose small-scale variance and higher-order structure. The loss tests whether that loss can be repaired without giving up per-spot accuracy.
Core Claim
Per-spot $R^2$ and Pearson correlation cannot tell apart a model that captures a gene's spatial organization from one that outputs a smoothed version of it. We expect current histology-to-expression predictors to behave like second-order perturbation theory (2LPT, an analytic approximation to gravitational clustering) behaves against full simulations. They should be reasonable at large spatial scales. At small scales they should be suppressed, with too little variance and skewness and too weak a marked correlation, and this suppression should be worst for the most tissue-restricted genes. Two things would be new. The first is the measurement itself: a scale-resolved "transfer function" and higher-order moment profile for predicted expression fields across organs. The second is a training objective that matches these statistics per slide, plus evidence that the higher-order terms add something beyond ordinary two-point spatial autocorrelation.
Research Program
The cosmology work supplies the method. Learning to Predict the Cosmological Structure Formation trained a 3D U-Net to map a crude approximation to full gravitational evolution. It judged success by power spectrum agreement (0.4% at intermediate scales) and three-point residuals (0.79% versus 7.82% for 2LPT), not only by point-wise displacement error. That pattern is the template for our audit. Dark Energy Survey Year 3 results: cosmology with moments of weak lensing mass maps showed that second and third moments of mass maps smoothed at several scales carry up to 25% more constraining power than two-point statistics alone. We borrow its smoothed-moment construction directly. Optimal marked statistics from the Effective Field Theory of Large Scale Structure showed that weighting a correlation function by a polynomial "mark" of the local density substantially tightens parameter constraints. It also showed the information is unchanged under affine rescaling of the mark. We reuse the mark idea, with the local abundance of an expression program as the mark, and we exploit the affine invariance so that the statistic does not reward a model for simply rescaling its output. These papers motivate and supply the statistics. They study gravitational fields with known physics, whereas we have no generative theory of tissue and use the statistics only as measurements and losses.
On the biology side, HiST: A Hierarchical Sparse Transformer for Cross-Modal Spatial Transcriptomics Modeling is both the strongest predictor and the system we audit. It computes only on occupied tissue spots, uses a multi-scale encoder-decoder, and calibrates per-slide staining with a learned token. It reports only $R^2$ and Pearson. Its authors note a "morphological ceiling": some expression variation is invisible in histology. This constrains our claims. Where a pattern is not visible in the image, no loss can recover it deterministically, so we must separate "suppressed but recoverable" from "unrecoverable". The biological rationale for focusing on restricted genes comes from the gene-family and splicing literature. Global analysis of exon creation versus loss and the role of alternative splicing in 17 vertebrate genomes argues that new exons survive by being used in narrow tissue niches. Catalogs such as the stigma-restricted CsCCD2 in Full-length transcriptome sequencing provides insights into the evolution of apocarotenoid biosynthesis in Crocus sativus show expression concentrated in small compartments. In field language, such genes are sparse, heavy-tailed fields, exactly the regime where two-point statistics and per-spot correlation say least.
The statistics are built as follows. For gene $g$ on slide $s$, let $x_g(i)$ be the standardized measured expression at spot $i$ and $\hat x_g(i)$ the prediction. We estimate the isotropic two-point correlation $\xi_g(r)$ in distance bins $r$ from spot pairs on the lattice, and define the transfer function
$$T_g(r) = \sqrt{\hat\xi_g(r)/\xi_g(r)},$$
which equals 1 when predicted and measured fields have the same spatial coherence at separation $r$ and falls below 1 when the prediction is too smooth. We add cross-correlations between gene pairs, smoothed-field variance, skewness and kurtosis at three or four Gaussian smoothing radii, and a marked correlation in which pairs are weighted by a polynomial of the local level of a co-expressed program. Because the lattice is irregular and tissue edges cut correlations, every estimator uses pair counts against random points placed within the tissue mask, the standard cosmological correction for survey boundaries.
The audit is run on all ten HEST-1k organs. We stratify genes by a tissue-specificity score computed from training slides, for instance spot-level kurtosis or the fraction of tissue where the gene is detected. The training intervention adds to the regression loss a per-slide term $\lambda \sum_{k} w_k\,(\hat S_k - S_k)^2$. Here $S_k$ are the statistics above computed on measured expression, $\hat S_k$ the same computed on predictions, $w_k$ inverse-variance weights estimated across training slides, and $\lambda$ a single trade-off coefficient. All estimators are differentiable sums over spot pairs, so they slot into HiST's training without architectural change.
Baselines
The naive reference point is a predictor that outputs each gene's training-set mean, and a second that outputs the per-slide mean. These fix the zero of both metric families: the mean predictor has $T_g(r)=0$ by construction. We also include a constructed oracle-blur reference. This is the measured expression smoothed with a Gaussian kernel of increasing width. It shows directly how much per-spot $R^2$ survives while small-scale structure is destroyed, which is the claim the whole proposal rests on.
The strongest existing method is HiST, which on HEST-1k under 4-fold cross-validation split by slide ID reports $R^2 = 0.34$ and Pearson $0.59$. Its predecessor STFlow reports $0.27$ and $0.54$ (HiST). For transfer functions, moments or marked correlations on HEST-1k, I found no reported result on Lacuna for any method, so the audit numbers will be new. We retrain HiST from its published configuration on the identical folds and gene panel, with the same compute and hyperparameter budget given to every variant. Beating it means improving the statistic-matching error, especially $T_g(r)$ at the smallest two separation bins for the top specificity quartile, while keeping $R^2$ within one fold-to-fold standard deviation of 0.34. A gain on statistics that costs substantial $R^2$ would count as a trade-off, not a win.
The key ablation separates two-point from higher-order matching. One HiST variant matches only $\xi_g(r)$ and cross-correlations. A second also matches the moments and the marked correlation. If the second adds nothing on held-out higher-order statistics or downstream tasks, the cosmology-specific contribution is void and the result reduces to "use a spatial autocorrelation loss."
Why not just…?
Why not just report Moran's I or a two-point spatial autocorrelation for each gene? This is the most serious threat. Two-point statistics may capture most of what matters, and DES found the higher moments gave "up to 25%" more, not an order of magnitude. Our two-point-only ablation exists to settle this. We will report how much unexplained discrepancy remains in skewness and marked correlation once $\xi_g(r)$ is matched. If that residual is negligible, the note's recommendation becomes the simpler two-point metric, and we will say so.
Why not just compute Pearson correlation on spatially variable or tissue-specific genes only? Restricting the gene set sharpens per-spot metrics but keeps their blindness to scale. The oracle-blur experiment tests this. If blurred ground truth keeps high Pearson on restricted genes while its small-scale transfer function collapses, gene selection cannot substitute for scale-resolved statistics.
Why not just train a generative model, such as diffusion, which samples sharp fields instead of averages? Large, fast and accurate HI intensity maps with latent overlap diffusion shows that a conditional diffusion model, a generator that learns by gradually denoising random noise into samples, can reproduce small-scale power to within 10% up to high wavenumbers. A generative predictor is therefore a real competitor on our statistics. It is not a competitor on our audit, which is needed to evaluate it at all. The issue is that sampling can match statistics with patterns placed in the wrong locations. We therefore include a conditional diffusion head on HiST features at matched compute. We report its per-spot $R^2$, which we expect to drop, alongside its statistics. The question becomes empirical: whether a deterministic statistic-matched predictor sits at a better point on the $R^2$-versus-transfer-function frontier than a sampler.
Proposed Experiments
The first decisive check is the audit itself. Across HEST-1k folds, we compute $T_g(r)$ and moment ratios for HiST, STFlow, and the mean baselines, stratified by organ and gene specificity. The hypothesis fails if HiST's small-scale transfer function is near 1 or does not worsen with specificity. In that case per-spot metrics are adequate and the proposal stops there.
The second check is the oracle-blur curve. We report, for each smoothing width, the per-spot $R^2$ and the statistic errors of blurred ground truth. This calibrates how strongly per-spot metrics reward smoothing on real tissue.
The third check is the training intervention. We compare HiST, HiST with two-point matching, HiST with full matching, and the diffusion head on the same folds. For each we report $R^2$, Pearson, held-out statistics not used in the loss (we hold out one smoothing radius and one mark polynomial), and stability across $\lambda$.
The fourth check is about relevance. Matching statistics matters only if it changes biological readouts, so we test spatial domain recovery and boundary sharpness. We cluster predicted expression and compare against clusters of measured expression on the same slide using adjusted Rand index. We also measure the spatial width of transitions for organ-restricted marker genes. Restricted-gene recovery near the morphological ceiling is analyzed separately on slides from Xenium, the higher-resolution platform. There, finer measurement tells us whether suppressed structure was ever visible in the image.
References
- Learning to predict the cosmological structure formation
- Dark Energy Survey Year 3 results: Cosmology with moments of weak lensing mass maps
- Optimal marked statistics from the Effective Field Theory of Large Scale Structure
- HiST: A Hierarchical Sparse Transformer for Cross-Modal Spatial Transcriptomics Modeling
- Global analysis of exon creation versus loss and the role of alternative splicing in 17 vertebrate genomes
- Full-length transcriptome sequencing provides insights into the evolution of apocarotenoid biosynthesis in Crocus sativus
- Large, fast, and accurate H <scp>i</scp> intensity maps with latent overlap diffusion