SVC treats each gene's intracellular spread as an image patch: other genes in the same cell supply the context for predicting where unmeasured transcripts sit

SVC yields cell-specific intracellular transcript distributions for genes that were not directly measured, not just cell-level expression totals. Earlier subcellular spatial transcriptomics methods focused on specific analyses, such as finding genes with distinctive localization patterns or detecting co-localized gene pairs, and offered no unified framework for characterizing and predicting transcript localization in a cell-specific, context-aware manner. On seqFISH+ mouse embryonic fibroblast data, batch 3 was held out as the test set and 1,000 genes were randomly split into 20 groups for 20-fold masked gene prediction; across six metrics including RMSE, PCC and cosine similarity, SVC outperformed all four baselines for at least 97.3% of genes.

The model encodes each gene's within-cell spatial expression as an image patch, then infers a masked patch from the remaining genes in the same cell. This reuses the Vision Transformer's self-attention machinery but replaces image-patch positional encoding with gene functional embeddings, because gene order carries no biological meaning. During training, 10% of gene images in each cell are randomly masked and a decoder reconstructs them under a pixel-level negative binomial loss plus a cell-level expression loss; inputs also include cell and nuclear morphology, neighboring-cell expression and optional cell type or state labels.

The learned representations also transfer to cell- and tissue-level tasks and support in silico prediction of localization changes after gene perturbation. Cell-level gene imputation is not the model's primary design target, yet it still achieved the lowest median RMSE and highest median PCC against ENVI, Tangram, gimVI and uniPort. On the seqFISH+ data, SVC's cell-level median RMSE was 5.12 and median PCC 0.50; the source lists median RMSE of 6.33, 6.47, 5.64 and 5.66 for uniPort, ENVI, Tangram and gimVI in that order, with corresponding median PCC values of 0.45, 0.43, 0.42 and 0.47.

The demonstrated scope is subcellular-resolution spatial transcriptomics data from cultured cell lines and tissue sections, including seqFISH+ mouse embryonic fibroblasts, MERFISH human osteosarcoma U2-OS, Xenium mouse brain and human breast cancer sections, plus Xenium 5K mouse brain hemisphere and P301S tauopathy model sections; the model is evaluated on held-out batches or sections and tested for transfer across datasets and tissues.

The model is trained only on static data and cannot yet capture dynamic single-cell responses over time; training focuses on targeted gene panels with predesigned probes, mainly covering moderately to highly expressed genes; the data are observational, so perturbation effects are in silico predictions rather than measured causal relocalization. Errors in upstream cell segmentation and transcript assignment remain a consideration, although sensitivity analysis suggests the model tolerates moderate segmentation differences.