DIRAC folds spatial multi-omic integration and annotation into one model, and finds a microvascular proliferation domain other methods missed in glioblastoma
Synopsis
Spatial multi-omic data pose two jobs at once: merging molecular layers such as RNA and protein measured on the same tissue section, and transferring cell-type labels from a reference dataset onto a new sample. DIRAC treats both as domain adaptation, using a graph neural network and an adversarial discriminator to learn embeddings that are insensitive to batch and modality but discriminative for cell type. On simulated and real DBiT-seq data it beat comparison methods on domain recovery and mixing, and in glioblastoma tissue it identified a microvascular proliferation domain the other methods missed.
Interpretation
Two jobs previously handled by separate tools in spatial multi-omics, fusing several molecular layers of one tissue section into a single embedding and transferring labels from a labeled reference onto new spatial data, are now done by one model. Earlier vertical integration methods came from single-cell multi-omics and ignored spatial position; two spatially aware methods did not explicitly correct batch effects across omics layers, and one of them grew its memory footprint quadratically with cell or spot number. Horizontal annotation methods were mostly built for single cells and ignored spatial neighborhood, and the one that used spatial information handled only a single modality. On trimodal structured data simulated with NSF, DIRAC reached an adjusted Rand index of 0.956 ± 1.64 × 10−3, better than the second-best method; on a real DBiT-seq mouse E10 embryo dataset, across ten random seeds, DIRAC significantly outperformed the second-best method on both omics mixing and biological preservation.
The model turns each molecular layer into a graph whose node weights are molecular features and whose edge weights are spatial proximity, then uses a graph neural network to aggregate information from each cell and its neighbors. At its core sits an adversarial domain discriminator that estimates which modality or dataset an input came from, while a gradient reversal layer pushes the feature extractor toward inputs the discriminator cannot tell apart, keeping cell-type signal and removing modality and batch differences. Vertical integration adds multimodal consistency and maximum mean discrepancy losses to map layers into one latent space; horizontal integration uses pseudo-label alignment to iteratively calibrate class distributions between reference and query, and splits the full graph into dense subgraphs so training scales to millions of cells.
On real tissue, this integration produced spatial domains that other methods did not produce. On a 10x Visium glioblastoma slide, DIRAC annotated nine spatial niches, with C1 labeled a microvascular proliferation domain supported by CD31 and αSMA signals and by elevated endothelial and mural cell scores in Tangram deconvolution; MaxFuse and SpatialGLUE domains straddled tumor and normal boundaries, and MISO domains were fragmented. On Stereo-seq data from prefrontal cortex in severe Alzheimer's disease, DIRAC still resolved cortical layers including layers 2 and 4, whereas SCC could not resolve layers 2 and 3 and Seurat could not resolve layer 4 in control and moderate cases; independent pathologist annotation of H&E stains agreed with DIRAC.
Perspective
The demonstrated scope covers simulated data, a real DBiT-seq mouse embryo dataset, glioblastoma and Alzheimer's disease brain tissue, mouse embryo seqFISH and Stereo-seq data, and newly generated mouse thymus and spleen Stereo-CITE-seq data; cell-type annotation relies on external reference datasets, and the method runs on cellularly resolved data.
Cell-type annotation depends on external references; segmentation inaccuracies may affect annotation, and the authors propose using horizontal integration to iteratively refine cell segmentation. The authors also note that boundary smoothness should be interpreted cautiously because it may not reflect multi-omic integration performance.