Spatial transcriptomics skips dense grids: sparse 3D sampling plus three-way prediction reaches 0.361 PCC on HER2

A 3D spatial transcriptomics volume can be measured at only a subset of spots and completed into a full expression map, instead of sequencing every position on a fixed grid. Conventional acquisition relies on fixed-grid sampling, which produces redundant measurements in 3D tissue while informative regions stay undersampled; existing predictors are mostly 2D and single-paradigm. Across HER2, Breast, cSCC and a private Kidney cohort, STRATA has the highest PCC and lowest MSE; on HER2 its PCC is 0.361 and MSE 0.356, and because the strongest-baseline figures stated in the text disagree with the table's baseline values, that gap is not treated as a confirmed comparison.

A reinforcement-learning agent chooses spots round by round, scoring candidates by morphological sparsity and transcriptomic informativeness and updating its policy as new measurements arrive. Scoring runs on a graph built by registering sections into a volume and by cellular-state prototypes from a single-cell reference, so cross-section continuity and cell-type rarity enter the decision. At a 10% budget on HER2 and cSCC the policy gives the highest PCC under three predictors; with ST-Net on cSCC, PCC rises from 0.075 to 0.085.

Completion comes from three parallel branches, retrieval, regression and 3D cross-layer imputation, whose outputs a lightweight fusion head weights per spot. Each single path has a gap: retrieval is bounded by reference coverage, regression lacks cross-layer structural constraints, and imputation depends on nearby measured spots. In the ablation, single-branch regression reaches 0.328 PCC on HER2 and all three branches reach 0.361; on cSCC all three reach 0.228. The text says adding 3D consistency supervision moves HER2 PCC from 0.314 to 0.328, while the table lists the regression-plus-3D branch at 0.338, so that increment is not quoted here.

Evaluation covers the public HER2, Breast and cSCC cohorts and a private Kidney cohort under patient-level cross-validation, using each patient's middle slice for training and 3D imputation anchoring and the remaining slices for evaluation at budget ratios including 10%; the intended reader works in spatial transcriptomics or computational pathology.

The low-budget advantage with ST-Net on cSCC is clearest in the 10–15% budget range; branch effectiveness varies by cohort, with single-branch regression best on HER2, so fusion weights must be learned per cohort. On HER2 the strongest-baseline figures stated in the text disagree with the table's baseline values, and the text says adding 3D consistency supervision moves PCC from 0.314 to 0.328 while the table lists the regression-plus-3D branch at 0.338, so both comparison sizes need checking against the original data before they are quoted.