Spectral Prompting: Unsupervised Recovery of Human Hair Follicle Cell-Type and Multiscale Systems Architecture from Bulk and Single-Cell RNA-Seq Datasets via Single-Gene Seeded Spectral Unfolding

Abstract

Bulk RNA sequencing datasets are assumed to carry minimal resolvable programmatic and cell type biological information; as such, in the absence of single-cell resolution, researchers prioritise data analysis approaches based on differential expression, or rely on deconvolution and co-expression methods that require external reference panels, large multi-sample cohorts, or prior single-cell data to resolve cell-type structure. Here I describe the recovery of specialised cell-type and systems gene expression architecture resolved from a static gene expression dataset of untreated cultured human hair follicles (pooled from N=12 patients) isolated from scalp skin. To achieve this, I used graph theoretic methods to mathematically transform gene expression data into a latent space of relational structure, which was spectrally organised into coarse- and fine-grained modes and partitioned using a purpose-built computational algorithm. This permitted the synthesis of a computational Spectral Prompting system, whereby a single gene can be seeded to "unfold" to reveal associated partners across manifold projections in gene expression space. Individual projections across the manifold can reveal rich individual gene expression programmes, which can then be aggregated to identify core-associated genes for a given spectral gene prompt, both within the manifold analysed and across >1 manifold constructions. With this, I recover hitherto unresolved gene expression programmes from bulk data, including, but not limited to, epithelial hair follicle stem cell (eHFSC), hair shaft, dermal papilla and endothelial gene expression signatures. Focusing on querying KRT15, a human anagen bulge eHFSC and progenitor marker, raw output from individual spectral prompts during testing recovered known eHFSC-associated genes including LGR5, LHX2 and CXCL14, and discovered new candidate human eHFSC and progenitor cell-associated markers, such as RGMA and MUCL1 which were validated in situ. Finally, I show a brief demonstration that the technique can be similarly applied to single-cell data (GSE129611), whereby a KRT15 gene prompt from a combined expression matrix was mapped to a KRT15+/CXCL14+/LHX2+/DIO2+/SFRP1+ cell population (31/6000 cells) independent of standard clustering tools. Moving forward, from this foundation, the method will be developed to study how latent gene expression space shifts following perturbation or pathology.