GenAR

A spatial transcriptomics spot reports how many mRNA molecules of each gene were captured there — a non-negative integer, often between zero and a few thousand. Almost every method that predicts those profiles from an H&E image first applies a log transform, turning the counts into floating-point values somewhere in the range 0–15, and then regresses each gene with its own output unit. Two things are lost in that move. The prediction is no longer a count, so it cannot be handed to a differential expression or pathway enrichment tool that expects one; and genes that sit in the same co-expression module are predicted as though they had nothing to do with each other.

GenAR, from Jiarui Ouyang and colleagues at the Hong Kong University of Science and Technology, keeps the counts discrete and predicts them jointly. Genes are first clustered by their spatial expression patterns into a coarse-to-fine hierarchy, from a single global group down to individual genes. The model then generates the profile one scale at a time, each scale conditioned on every coarser prediction already made — the next-scale autoregressive scheme introduced for image generation, applied to a target that is already discrete. Because expression counts are integers to begin with, no vector-quantized autoencoder is needed to discretize them first: GenAR is codebook-free and trains end to end, avoiding the reconstruction loss a learned codebook introduces.

The setup is deliberately narrow: GenAR predicts a fixed 200-gene panel chosen per dataset, and the released checkpoints are one per cohort rather than a single general-purpose predictor.

#Key Features

• Discrete count generation: Expression is emitted as integer tokens over a fixed vocabulary covering the observed count range, so predictions stay on the scale that downstream count-based analyses assume.
• Hierarchical gene ordering: k-means on Z-score-normalized expression profiles splits the 200 genes into four major clusters, then into subgroups of roughly twelve, giving the autoregressive factorization a biologically grounded coarse-to-fine order.
• Coarse-to-fine decoding: Six scales of 1, 4, 8, 40, 100 and 200 groups run in sequence, beginning with one token for global transcriptional activity and ending at individual genes.
• Histology and position fused as one condition: Patch features from a frozen encoder and a sinusoidal encoding of the spot's coordinates are projected into a single 768-dimensional embedding that conditions every scale.
• Gene-identity modulation: Learned per-gene embeddings produce scale and shift parameters that modulate the decoder's hidden states, injecting gene-specific inductive bias into a shared output head.

#Technical Details

Each 224×224-pixel patch, one per spatial spot, is encoded by UNI into a 1024-dimensional feature vector; a fusion module with GELU-activated linear layers combines it with the encoded coordinates into the 768-dimensional histological embedding. A causal Transformer decoder consumes the token sequence for the current scale and is conditioned on that embedding through adaptive layer normalization. Intermediate scales are supervised with a KL divergence against temperature-smoothed group targets from adaptive average pooling; the final scale uses a heteroscedastic Gaussian likelihood whose variance scales with the predicted mean, σ² = αμ + β, matching the mean-variance relationship of count data. Training used Adam at a learning rate of 1e-4, batch size 64, on H100 GPUs. Evaluated on HEST-1k cohorts against BLEEP, TRIPLEX, M2OST and the diffusion-based STEM, GenAR reaches PCC-10 / PCC-200 of 0.842 / 0.663 on the HER2ST breast cancer slides, 0.702 / 0.512 on prostate cancer (PRAD), 0.589 / 0.354 on kidney, and 0.568 / 0.367 on healthy mouse brain, with the lowest MSE and MAE on each.

#Applications

Spatial transcriptomics runs cost hundreds to thousands of dollars per sample, while H&E slides are produced routinely and archived in volume. GenAR is aimed at that gap: recovering spot-level expression for a defined gene panel from imaging a pathology lab already has, whether to prioritize which blocks are worth sequencing or to add a molecular read-out to a retrospective cohort. Five checkpoints are released — prostate (PRAD), breast (HER2ST), kidney, clear cell renal cell carcinoma and mouse brain — each paired with the ordered gene list it was trained on, and inference runs from a checkpoint and a slide identifier without a training step.

#Impact

GenAR's contribution is a reframing rather than a new encoder: treating histology-to- expression as generation over a discrete target, and letting the co-expression structure of the panel supply the autoregressive order. That the target is natively discrete is what makes the codebook-free formulation possible, and it is the part most likely to carry over to other count-valued biological readouts. The limits are the ones the authors set: each checkpoint is tied to its cohort's gene panel, cross-cohort transfer is not demonstrated, and a new tissue or panel requires its own training run, including a re-run of the gene clustering. The code is MIT-licensed; the released weights are restricted to non-commercial academic research, and the UNI encoder they depend on carries separate terms.