DPAS-Graph

Spatial transcriptomics reads RNA in place, but protein is what most tissue phenotyping needs, and the assays that capture both in one section — spatial CITE-seq, antibody-panel Visium CytAssist, spatial molecular imaging — stay costly enough that most spatial archives carry RNA and coordinates and nothing else. Predicting protein from those inputs turns on how spots relate. Physical adjacency says two touching spots share a microenvironment; transcriptomic similarity says two spots far apart are in the same cell state. At a tissue boundary or in a mixed-cellular region the two disagree, and a model that collapses them into one fixed edge weight before learning begins has already decided which of them wins.

DPAS-Graph — Dual-graph Protein prediction with Adaptive Spatial-feature learning — declines to decide in advance. Per section it builds one shared edge support, the union of a k-nearest-neighbour graph over the 2D coordinates and one over PCA-reduced expression, and hangs two relation-weight channels on every edge, one for proximity and one for similarity, with a zero wherever an edge came from only one source. Its Niche-Coupled Field Encoder rewrites those weights at each layer from the current node states, propagates each channel through its own GATv2 operator, and reconciles the two messages by sequential cross-branch residual correction rather than a static sum.

Three researchers at Yan'an University in Shaanxi, China, developed the model and published it in Briefings in Bioinformatics in 2026. Its nearest relative in the catalog is DGAT, a separate dual-graph spatial RNA-to-protein model from the University of Pittsburgh, which serves here as the strongest baseline.

#Key Features

• Two relation channels on a shared edge support: Proximity and similarity are separate per-edge weights on one merged graph, so their relative influence changes per edge and per layer instead of being fixed by a fusion step.
• Layer-wise edge recalibration: A descriptor built from cosine similarity, difference magnitude and representation scale feeds an edge MLP that rescales each channel, letting one edge mean different things at different depths.
• Asymmetric channel refinement: The spatial branch prunes its globally weakest edges, while the feature branch gets a non-negative additive patch expressing increments rescaling alone would miss.
• Sequential cross-branch correction: The spatial message is corrected against its discrepancy with the feature message, then the feature message against the corrected spatial one, keeping the channels distinct rather than averaging them.
• Cross-dataset checkpoint: Training is leave-one-dataset-out over whole sections, and inference on a held-out or external section consumes only RNA and coordinates.

#Technical Details

DPAS-Graph is a 71.5 M-parameter graph neural network: three encoder layers with branch-specific GATv2 operators and skip-connected LayerNorm/LeakyReLU updates, feeding a protein abundance head plus an auxiliary mRNA-reconstruction head used only as a training regulariser. Seven paired spatial multi-omics sections supply supervision — three 10x Visium CytAssist gene-and-protein libraries (Tonsil, Tonsil Add-on, Breast Cancer) and four sections from GEO series GSE263617 — after a harmonisation step that projects RNA into a shared feature space and aligns the antibody panels onto one 31-protein target vector, so the head is panel-fixed rather than open-vocabulary. Across the seven folds the macro averages are RMSE-Global 0.703, MAE-Global 0.405 and RMSE-Pro 0.573, the lowest errors among the methods compared, with PCC-Spot 0.913 against DGAT's 0.904. Rankings are not uniform: TotalVI leads the protein-wise correlations (PCC-Pro 0.505, SP-Pro 0.523 versus 0.452 and 0.461) while losing every error metric. An ablation dropping cross-branch correction raises RMSE-Global from 0.706 to 0.729 and cuts PCC-Pro from 0.445 to 0.363. Inference takes 0.128 s per section, at 33.8 GB peak reserved GPU memory.

#Applications

The intended use is the backlog of spatial transcriptomics slides never assayed for protein. The authors apply the trained checkpoint unchanged to two such sections — a 10x human lung cancer FFPE slide and a tonsil section (esvq52_nluss5) from a public tonsil atlas — and generate virtual protein maps for KRT5, ACTA2, MS4A1 and CD3E, recovering epithelial/basal partitioning in the tumour and non-overlapping B- and T-cell compartments in the tonsil. Neither section carries paired protein measurements, and the authors present this as a qualitative demonstration with no protein-level quantitative validation: it nominates regions and markers worth an antibody assay rather than replacing one.

#Impact

The contribution is a design argument more than a leap in accuracy: the ablations indicate the gain comes less from having two relation sources than from updating and coordinating them under the prediction objective, a claim that carries beyond protein prediction to any spatial task that sums a proximity graph and a similarity graph. The caveats are equally clear: gains concentrate in error-based metrics and spot-level concordance rather than protein-wise correlation; a new antibody set requires redefining the harmonised target space and retraining; and the 33.8 GB footprint, which the authors flag as motivation for sparse edge operations and block-wise inference, makes the implementation awkward on commodity hardware. The reference code is released under the MIT licence, with no trained checkpoint alongside it.