Date of Award
2026
Degree Type
Thesis
Degree Name
Master of Science in Biological and Environmental Sciences (MSBES)
Specialization
Cell & Molecular Biology
Department
Biological Sciences
First Advisor
Marta Gomez-Chiarri
Abstract
Bivalve transmissible neoplasia (BTN) is a contagious cancer that has caused significant mortality in Mercenaria mercenaria aquaculture. Current BTN diagnostics rely on cytology and histology and are therefore limited to cases in which morphologic changes are already established, allowing undiagnosed carriers to continue to spread the disease. To uncover the molecular signatures of early BTN infection and to develop a framework for pre-symptomatic diagnosis, we designed and trained Support Vector Machine (SVM) classifiers on single-cell RNA sequencing data of clam hemocytes from two independent co-habitation transmission trials. The analysis of 61,658 single-cell transcriptomes achieved near-complete separation of pool-labelled neoplastic versus pool-labelled normal hemocytes (AUC = 0.9985 with the FS-HVG5k feature set; we refer to this configuration as the reference model, Ref-SVM). A robustness framework using three 70/30 SVM configurations (FS-HVG3k, FS-HVG5k, FS-ALL) produced a candidate pool of 406 unique genes (GS-406); 46 of these were recovered in all three configurations (GS-46). The top 20 features of Ref-SVM (GS-20) were carried forward as the primary candidate diagnostic panel; downstream logistic-regression models (LR-TopN) showed diminishing returns beyond approximately 7 to 10 genes, suggesting that the transcriptional distinction is concentrated in a compact feature set. Functional annotation mapped 17 of the 20 annotated genes onto at least six hallmarks of cancer, including apoptosis evasion, proliferative signaling, and immune modulation, providing biological evidence that the classifier captured genuine neoplastic biology. Convergence analysis showed that 18 of the 20 GS-20 genes (90%) were also differentially expressed between PoolLabel-Neo and PoolLabel-Norm cells (padj < 0.01, |log2 FC| > 0.5), providing orthogonal validation. Using the GS-20 panel in a neoplastic-likeness scoring model (ENet-Score), trained on a class-balanced down-sample of PoolLabel-Neo and PoolLabel-Norm cells, approximately 7% of hemocytes collected 5 days and approximately 6% of hemocytes collected 12 days after naïve clams were exposed to neoplastic donors exhibited a neoplastic-like transcriptional expression profile above the scoring threshold — an eight-fold enrichment over the uninfected baseline. We interpret these cells as carrying a neoplastic-like transcriptional signature rather than as confirmed neoplastic cells, pending independent cell-level validation. Together, these results yield informatically derived gene candidates with biologically plausible relationships to BTN progression, supporting the feasibility of a compact, pre-symptomatic diagnostic panel for bivalve transmissible neoplasia.
Recommended Citation
Paz, Alberto, "MACHINE LEARNING AS A TOOL FOR UNCOVERING TRANSCRIPTIONAL SIGNATURES OF BIVALVE TRANSMISSIBLE NEOPLASIA IN MERCENARIA MERCENARIA" (2026). Open Access Master's Theses. Paper 2704.
https://digitalcommons.uri.edu/theses/2704