RARE-19. NETWORK AND DEEP LEARNING INFERENCE IN SINGLE CELL RNA SEQUENCING REVEAL DETAILED TRANSCRIPTIONAL SIGNATURES CONGRUENT WITH MOLECULAR UNDERSTANDING OF ADAMANTINOMATOUS CRANIOPHARYNGIOMA. (1st June 2021)

• Record Type: Journal Article
• Authors: Prince, Eric; Trudeau, Tammy; Chatain, Oscar; Chee, Keanu; Vijmasi, Trinka; Staulcup, Susan; Donson, Andrew; Foreman, Nicholas; Hankinson, Todd
• Is Part Of: Neuro-oncology. Volume 23(2021)Supplement 1
• Page Start: i44
• Page End: i45
• Publication Date: 2021-06-01
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noab090.180
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: Adamantinomatous Craniopharyngioma (ACP) is a highly morbid, cellularly heterogeneous pediatric tumor arising in the sellar/suprasellar region of the brain. This cellular heterogeneity makes ACP an ideal candidate for study using single-cell RNA-sequencing (scRNA-seq). We collected a 10, 000 cell scRNA-seq dataset on the 10X v3 platform, from 6 unique patients. Using the industry standard Seurat software package, we identified 34 unique cell clusters. By crossing the results of two separate expert curated cellular reference atlases (Azimuth and scHCL), we determined that 33 of these cell types were immune-related (e.g., T cells, monocytes, etc.) or histologically related (e.g., glial cells). The remaining 2, 048 cells were inferred to be ACP driver cells. Rigorous statistical testing of third-generation graph topology-based network enrichment methods utilizing the Reactome database supported this conclusion. In order to identify effective antitumor therapies, it is critical to understand the temporal evolution of tumor cell behavior. Computational solutions that describe the potential lifecycle of tumor cells have been derived using scRNA-seq datasets. Using a well-established method, Monocle3, we generated a potential model of temporal evolution of the ACP driver cell population. To identify a specific transcriptional "point-of-no-return" for ACP driver cells, which may help define a rational target for intervention, we created a custom probabilistic Deep Learning framework in the form of a Convolutional Variational Autoencoder (CVAE). By applying this CVAE to our data, we identified 31 anomalous transcripts, each of which was aberrantly active at all times or demonstrated a temporal pattern of anomalous activity. Strikingly, this small list – representing roughly 0.15% of the protein coding genome – aligns closely with extant data describing the molecular behavior of ACP. This work provides a novel transcriptome benchmark for comparison of in vitro models, a deeper understanding of ACP heterogeneity, as well as a generalizable approach for scRNA-seq analysis.