AntiSplodge: a neural-network-based RNA-profile deconvolution pipeline designed for spatial transcriptomics

Abstract

With the current surge of spatial transcriptomics (ST) studies, researchers are exploring the deep interactive cell-play directly in tissues, in situ. However, with the current technologies, measurements consist of mRNA transcript profiles of mixed origin. Recently, applications have been proposed to tackle the deconvolution process, to gain knowledge about which cell types (SC) are found within. This is usually done by incorporating metrics from single-cell (SC) RNA, from similar tissues. Yet, most existing tools are cumbersome, and we found them hard to integrate and properly utilize. Therefore, we present AntiSplodge, a simple feed-forward neural-network-based pipeline designed to effective deconvolute ST profiles by utilizing synthetic ST profiles derived from real-life SC datasets. AntiSplodge is designed to be easy, fast and intuitive while still being lightweight. To demonstrate AntiSplodge, we deconvolute the human heart and verify correctness across time points. We further deconvolute the mouse brain, where spot patterns correctly follow that of the underlying tissue. In particular, for the hippocampus from where the cells originate. Furthermore, AntiSplodge demonstrates top of the line performance when compared to current state-of-the-art tools. Software availability: https://github.com/HealthML/AntiSplodge/.

Figure 1.

Overview of the AntiSplodge pipeline. Dotted boxes are outside the AntiSplodge regime, while full drawn boxes are the steps of AntiSplodge. (1) First the scRNA data along with cell type information are added. (2) The data is then split into three chunks, stratified by cell type information, to make train, validation, and test dataset splits. (3) For each of the dataset splits, synthetic spot profiles are generated based on the cells from the scRNA data. (4) The AntiSplodge neural network is initialized and trained using the synthetic profiles, where the ground truth is known. (5) Once the network has been trained, the spatial transcriptomics data is loaded. (6) With the trained network, the spot profiles are passed through the network to deconvolute each profile. (7) The output of the network is a proportion vector indicating the percentage of each cell type for each spot profile, these can then be used in downstream analysis tasks.

Figure 2.

Results for performance compared across the models tested, based on synthetic data generated from single-cell data from the Heart Cell Atlas, measured by Jensen–Shannon divergence (JSD). Ordered by their mean JSD, from left to right: Baseline (mean: 60.5%), Regression-based Random Forest (mean: 35.1%), SPOTlight (mean: 24.8%) Stereoscope (mean: 20.6%), Cell2Location (mean: 20.6%), DSTG (mean: 10.0%), and our method AntiSplodge (mean: 7.6%). P-values at the top of the plot are Bonferroni adjusted t-test P-values, comparing the distributions of JSDs for each model against AntiSplodge.

Figure 3.

Results for the AntiSplodge 10-fold cross-validation. Each fold contains 2500 synthetic test profiles with a cell density of 10. From left to right, we have a mean JSD of: Fold 1: 8.64% (SD: 7.97%), Fold 2: 8.27% (SD: 7.11%), Fold 3: 9.18% (SD: 8.43%), Fold 4: 10.05% (SD: 8.07%), Fold 5: 8.67% (SD: 8.28%), Fold 6: 7.78% (SD: 7.74%), Fold 7: 8.11% (SD: 7.83%), Fold 8: 8.16% (SD: 8.14%), Fold 9: 7.72% (SD: 8.15%), Fold 10: 9.37% (SD: 8.53%). Red line denotes the mean JSD across all folds with 8.59%.

Figure 4.

Results for ST samples from the developing human heart (N = 19). (A) Loss over epochs, best loss found is indicated with a green dot (epoch = 92, loss = 0.0017). (B) Jensen–Shannon divergence box plot for the test samples (mean JSD = 12.9%). (C) Distributions of cell types across samples, ordered by PCW. (D) Mean proportion of each cell type for each PCW. (E) Zoom of the fourth 6.5 PCW sample. First four plots are the top four cell types ordered by their cell type abundance across ST spots (highest first), ventricular cardiomyocytes, erythrocytes, smooth muscle cells/fibroblast-like, and atrial cardiomyocytes. Fifth plot is the corresponding plot in (F). Last plot is the cell type distribution across spots. (F) Map of maximum cell types for each spot for each of the samples. Samples are ordered as; first four is 4.5–5 PCW (black), next nine is 6.5 PCW (green), and last six is 9 PCW (purple). (G) Legend for cell types and their associated colors.

Figure 5.

Results for ST samples from the developing human heart (N = 5). (A) Loss over epochs, best loss found is indicated with a green dot (epoch = 390, loss = 0.0006). (B) Jensen-Shannon Divergence box plot for the test samples (mean JSD = 12.12%). (C) Predicted cell types for layers of the mouse brain cortex in the sagittal anterior samples, different types are found in the different cortex layers. (D) Predicted distributions of cell types for CA1, CA2, CA3 (cornu ammonis) and DG (dentate gyrus). Note: In the Allen mouse brain atlas, the cell subclass types are defined by the regions in which they were found. (D*) Hippocampus regions according to the Allen mouse brain atlas. (E) Predicted cell types in the mouse brain for the first sagittal posterior sample. (F) Predicted cell types in the mouse brain for the second sagittal posterior sample. (G) Pairs of the tissue and deconvoluted spots, for the mouse brain ST samples. (H) Legend for cell types and their associated colors, for the figures in (G).