CRISPR screens stay inside the tissue section: Perturb-Tags ties perturbation identity, transcriptome, chromatin and coordinates to one nucleus
Synopsis
Pooled CRISPR screens usually dissociate tissue into single cells, losing the spatial relationships needed to see how a perturbation in one cell affects its neighbors. Perturb-Tags first stamps nuclei in a tissue section with photocleavable spatial barcodes, then reads spatial coordinates, guide RNAs, transcriptome-wide expression and optionally chromatin accessibility from the same nucleus on a 10x Genomics platform. In human embryoid bodies screened for ligands, receptors and kinases, cells around a perturbation showed distance-dependent shifts in expression, ligand–receptor signaling, chromatin state and differentiation.
Interpretation
Pooled CRISPR screening can now read perturbation identity and molecular state while keeping each cell's position in the tissue, without first dissociating it into suspension. Earlier methods such as Perturb-seq and CROP-seq depend on tissue dissociation; they capture cell-autonomous perturbation responses but lose the spatial relationships needed to determine how a perturbation affects neighboring cells. In the first ligand-and-receptor experiment, 4,115 cells were recovered with a median of 4,084 transcripts and 2,600 genes per cell, and 65.7% of cells had detectable guide RNAs; transcripts per nucleus were comparable to reported conventional single-nucleus RNA sequencing of human embryoid bodies and substantially higher than previous spatial perturbation approaches.
The method writes spatial barcodes into nuclei first, then performs ex situ sequencing on a standard single-cell platform. Spatial barcodes are released by ultraviolet light from beads on a Trekker slide and transferred to nuclei in the overlying tissue section, so a barcode sequence indicates a nucleus's original coordinates in the tissue. Nuclei are then extracted and loaded onto the 10x Genomics Chromium system, where transcripts, guide RNAs and spatial barcodes are captured in the same droplet; with a modified 10x Multiome workflow plus a custom reverse-transcription primer and splint oligonucleotide, one nucleus yields RNA, ATAC and guide RNA, and a single Trekker slide produced about 26,000 nuclei, 46% of which could be assigned one or more guide RNAs.
Around perturbed cells, neighboring cells change state with distance across expression, signaling, chromatin and differentiation. Cells not carrying the perturbation were treated as receiver cells and grouped by distance to the nearest perturbed cell; single perturbations and multi-guide combinations both showed distance-dependent expression differences. Across the screen, 34 direct ligand–receptor interactions and 25 compensatory interactions changed significantly with distance; in the Multiome data, accessibility at representative regions increased with distance after PDGFC knockdown, and SEMA6D knockdown produced distance-dependent changes in both accessibility and transcription-factor motif activity.
A graph attention network takes perturbation identity, cell type, spatial position and neighbors as input and predicts expression and chromatin state across an entire tissue. A non-spatial multilayer perceptron first learns the cell-autonomous effects of guide RNA and cell type, then a distance-aware graph attention network learns a spatial correction from neighboring cells, and the two are summed into the final prediction. Evaluated on 10 held-out embryoid bodies, adding spatial information gave a Pearson correlation of 0.652 for predicting RNA counts, better than the non-spatial multilayer perceptron; ATAC loss was also better than the non-spatial model, and guide-averaged expression profiles were predicted with a correlation of 0.801.
Perspective
The demonstrated scope is the human embryoid body, an early-development model: the ligand and receptor screen covers 26 targets, the kinase screen covers about 300 kinases expressed in embryoid bodies across 70,235 cells and 806 embryoid bodies, and the Multiome dataset contains about 26,000 nuclei. The authors propose extending the experiments across developmental time, additional perturbation types and more complex human tissue models to distinguish transient signaling responses from persistent changes in cell identity.
The authors note that variation in spatial clone size may reflect developmental timing, stochastic differences in clonal expansion, or perturbation-specific effects on proliferation or survival, and the data cannot separate these. Embryoid bodies in the kinase screen were generally smaller than in the ligand and receptor experiments, which the authors attribute to broad effects of kinase perturbation on growth, survival and differentiation. The model's effective radii describe the spatial range of predicted effects, not a physical boundary of signaling.
