Large-Scale Analysis Of The Molecular Anatomy Of The Dorsolateral Prefrontal Cortex (DLPFC) Through The Use Of Unsupervised Methods With Spatial RNA-seq

Author(s): Abby Spangler, Nicholas J. Eagles, Kelsey Montgomery, Madhavi Tippani, Heena R. Divecha, Stephanie Hicks, Keri Martinowich, Kristen R. Maynard, Leonardo Collado Torres

Affiliation(s): Lieber Institute for Brain Development



The molecular anatomy of the cortex is well known in the context of histologic layers. However, emerging spatially-resolved transcriptomic approaches have enabled unbiased classification of novel spatial domains based on molecular signatures. Here we generated a large-scale spatially-resolved transcriptomic dataset from post-mortem human dorsolateral prefrontal cortex (DLPFC) with the 10x Genomics Visium platform. The dataset contains 30 samples from 10 neurotypical donors (n=3 tissue sections per donor) making it a valuable resource for studies investigating the DLPFC in the context of neuropsychiatric disorders. We evaluated recently published unsupervised clustering methods such as BayesSpace (Zhao et al. 2021) and spaGCN (Hu et al. 2021) in comparison to more commonly used graph-based clustering methods. Unlike published comparisons, we clustered across samples, instead of clustering one sample at a time, which allowed us to identify common clusters across all samples. We used an Adjusted Rand Index to compare predicted clusters to manual annotations (Maynard et al. 2021), which we considered a gold standard of histological cortical layers. We determined that algorithms utilizing spatial information generated the most robust spatial domains. We found that application of BayesSpace on batch-corrected data with Harmony produced clusters most similar to manually annotated classic histological layers. Using BayesSpace, we explored different levels of clustering resolution across sections to identify spatial domains similar to classic histological layers as well as novel domains with distinct expression profiles.To determine an optimal clustering resolution (k number of clusters), we evaluated an unsupervised method that computes a discordance internal validity metric implemented in the fasthplus package (Dyjack et al, 2022). We then used spatialLIBD (Pardo et al, 2021) to correlate enriched genes in BayesSpace clusters to enriched genes from the gold standard layers, which allowed us to make inferences about how BayesSpace clusters compare to classical histological layers. We found that this semi-supervised approach worked best with our complex dataset where gray and white matter, then histological layers, and finally cell types have a large effect on gene expression variability. Finally, we identified differentially enriched genes across our data-driven molecularly-defined spatial domains. In summary, this work expands our molecular mapping of the DLPFC and is a resource for investigating unsupervised methods to define novel spatial patterns within tissue architecture.