DGAT

Spatial transcriptomics measures genome-wide mRNA in tissue context, but transcript abundance is an imperfect proxy for the proteins that execute cellular function. Antibody-based spatial proteomics — spatial CITE-seq and similar assays that measure a panel of surface proteins alongside mRNA on the same section — remains expensive, panel-limited, and rare relative to the large archive of transcriptomics-only Visium and legacy ST datasets. DGAT (Dual-Graph Attention Network) closes that gap by learning the RNA-to-protein mapping from paired spatial multi-omic sections and then imputing protein abundance for sections where only transcriptomes were measured.

Developed by the Osmanbeyoglu lab at the University of Pittsburgh and posted as a preprint in July 2025, DGAT was published in Nature Communications in May 2026. Its core idea is that spatial position and molecular similarity are complementary sources of structure: two spots can be neighbors because they sit next to each other in the tissue, or because their expression profiles resemble one another. DGAT encodes both relationships simultaneously in a heterogeneous graph, one per modality, and uses graph attention to decide how much each neighbor should contribute.

This places DGAT alongside other cross-modal spatial imputation models such as SpatialEx, which predicts single-cell multi-omics from histology images, and ST-Painter, which imputes genome-wide transcripts from sparse panels. DGAT differs in targeting the protein layer specifically, using paired RNA–protein measurements rather than images as the supervisory signal.

#Key Features

• Dual-graph construction: Each modality gets an adjacency matrix that sums a spatial graph (each spot linked to its six nearest physical neighbors) with a k-nearest-neighbor molecular graph (k = 10) built in PCA space, so attention operates over both tissue geometry and expression similarity.
• Shared latent space with task-specific decoders: Separate GAT encoders for mRNA and protein map into a common 1024-dimensional representation; one decoder reconstructs mRNA while a second predicts protein abundance, with an alignment loss tying the two views together.
• Protein-specific decoder branches: The protein decoder uses a shared two-layer trunk followed by one branch per protein, letting individual markers with different RNA–protein coupling learn their own readout.
• Pretrained checkpoint for protein-free data: Released weights trained on the six paired samples can be applied without retraining to transcriptomics-only sections, which is how the authors analyzed lymph node, triple-negative breast cancer, melanoma, and prostate cancer datasets.
• Reproducible release: The repository ships separate training, prediction, and pretraining notebooks plus scripts that regenerate every figure in the paper.

#Technical Details

Each encoder is a stack of three graph attention layers with layer normalization and skip connections; the mRNA decoder is a residual feedforward network. Training optimizes a weighted sum of five terms — mRNA reconstruction, protein reconstruction, cross-modal alignment, mRNA prediction, and protein prediction — with the alignment weight adjusted dynamically once alignment loss falls below 0.015. The training corpus is six spatial CITE-seq samples profiled on 10x Genomics Visium CytAssist (two tonsil, one estrogen-receptor-positive breast cancer, one glioblastoma, two malignant mesothelioma), covering 11,535 genes and 31 proteins.

Evaluation used leave-one-sample-out testing, holding out each section in turn. On tonsil1, median Spearman correlation between predicted and measured protein reached ρ = 0.610 ± 0.230 versus ρ = 0.308 ± 0.219 for the matched-RNA baseline, with RMSE falling from 1.61 to 0.226; across the other held-out samples median correlation ranged from 0.658 (tonsil2) down to 0.283 (glioblastoma). DGAT achieved the highest average rank score against cTP-net, scLinear, sciPENN, and Seurat v4. Downstream, imputed proteins improved germinal-center detection in lymph node (F1 0.823 vs. 0.575 for transcriptome-only) and spatial domain recovery in prostate cancer (ARI 0.683). Ablations confirm each component contributes: removing the spatial graphs alone dropped tonsil1 correlation from 0.610 to 0.589.

#Applications

DGAT is aimed at labs with existing transcriptomics-only spatial datasets who want protein-level readouts without rerunning the assay — annotating immune phenotypes by surface markers, resolving tumor microenvironment architecture, and clustering spatial domains on a signal closer to cell state than mRNA alone. It slots into standard single-cell workflows as an imputation step before clustering and differential analysis, letting retrospective cohorts be reanalyzed at the protein layer.

#Impact

DGAT demonstrates that a modest amount of paired spatial multi-omic data can be leveraged to add a protein channel to a much larger body of transcriptomics-only tissue data. Its honest limits are worth stating: the checkpoint predicts only the 31 proteins in its training panel, it was trained on six samples spanning four tissue types, and the authors note that RNA–protein relationships vary substantially by tissue and cell state, so tissues unlike the training set — hepatic, neural, or fibrotic — may require retraining or fine-tuning. Accuracy is also uneven across tissues, with glioblastoma and mesothelioma notably harder than tonsil. The GitHub repository carries no license file, which leaves reuse terms for the code and weights unspecified.