Benchmarking Spot Deconvolution Methods in the Human Dorsolateral Prefrontal Cortex
Author(s): Nicholas J Eagles,Louise A. Huuki-Myers,Abby Spangler,Kelsey D. Montgomery,Sang Ho Kwon,Boyi Guo,Melissa Grant-Peters,Heena R. Divecha,Madhavi Tippani,Chaichontat Sriworarat,Annie B. Nguyen,Prashanti Ravichandran,Matthew N. Tran,Arta Seyedian,Thomas M. Hyde,Joel E. Kleinman,Alexis Battle,Stephanie C. Page,Mina Ryten,Stephanie Hicks,Keri Martinowich,Leonardo Collado Torres,Kristen R. Maynard
Affiliation(s): Lieber Institute for Brain Development
Spatial transcriptomics is an increasingly popular field of study that allows the measurement of gene-expression information along with spatial coordinates. The Visium platform is a popular choice involving a grid of discrete “spots”, typically larger than individual cells, where genes bind to unique barcodes to allow tracing back expression to a fairly precise location. Because these spots may potentially include multiple cells, researchers have been interested in “spot deconvolution” methods, which quantify cell-type composition and/or abundances within each spot, further refining the precision of spatial transcriptomics experiments. Single-nucleus RNA-sequencing (snRNA-seq) data provide a set of gene-expression measurements for individual cells of known type, which can be compared against gene-expression patterns in spatial data to infer cell-type composition in Visium spots. We used Visium to study gene expression in the dorsolateral prefrontal cortex (DLPFC) in post-mortem human brains from 10 neurotypical donors. In four of these donors, we also applied Visium-SPG, a platform generating both spatial gene-expression data and immunofluorescence images, which we marked for lipofuscin, NeuN, OLIG2, TMEM119, and GFAP, to detect background fluorescence, neurons, oligodendrocytes, microglia, and astrocytes, respectively. We show that these Visium-SPG images can be directly leveraged to accurately provide cell-type counts within Visium spots, forming a ground-truth for benchmarking spot-deconvolution methods that integrate snRNA-seq and spatial information. We benchmarked three such software tools against the Visium-SPG: Tangram, Cell2location, and SPOTlight. Software-estimated and image-derived counts can be compared using correlation and root mean squared error (RMSE). Alternatively, cell compositions can be treated as probabilities, and two composition estimates can be compared with measures such as Kullback-Leibler divergence. Depending on the metric used, Tangram or Cell2location perform best; the appropriate tool likely depends on the downstream analysis use-case.