Attack of the Clones: Mapping melanoma clonal architecture using deep learning and single-cell multi-omics analysis

Abstract

Melanoma is an aggressive skin cancer marked by extreme genomic instability and resistance to therapy. As the disease progresses, clonal evolution becomes increasingly complex, driven by diverse genetic and epigenetic alterations. Traditional lineage-tracing approaches—such as copy number variation (CNV)-based phylogenetics—struggle to resolve this complexity, particularly in the context of spatial heterogeneity and convergent evolution. This thesis presents a new computational framework for clonal inference in melanoma, using mitochondrial DNA (mtDNA) mutations as stable, high-resolution markers of lineage. In order to address the limitations of existing methods, I apply and adapt ReDeeM, a mitochondrial variant calling and lineage tracing pipeline, to melanoma. I further introduce a modified version of ReDeeM that incorporates UV mutation signature filtering, positional bias correction, and graph-based connectivity filters to reduce false positives. In parallel, I develop a graph neural network (GNN) model that constructs a cell–cell graph based on shared mtDNA, CNV, and chromatin accessibility features and helps infer downstream phylogenetic relationships through learned latent embeddings. Compared to CNV-based clustering, mtDNA-based clustering yields higher silhouette scores (0.46 vs. 0.32) and greater subclonal resolution. ReDeeM consistently identifies a broader set of informative variants, including low-frequency mutations missed by mgatk and other conservative callers. The pipeline was applied to a single metastatic melanoma biopsy of over 9,000 cells. Using read-depth filtering and variant-level quality control, I identified robust mtDNA-based clusters and constructed phylogenetic trees with high internal consistency. Despite the tissue’s mutational burden and technical artifacts common in solid tumors, the approach revealed structured subclonal relationships with lineage coherence across multiple modalities. Modified ReDeeM and GNN methods produced consistent tree topologies and performed at least comparably to baseline methods in Adjusted Rand Index and phylogenetic distance metrics. Taken together, this work demonstrates that mtDNA-based clonal inference—augmented by careful filtering and graph-based modeling—may be able to resolve evolutionary structure in melanoma where other single-cell methods fall short. These results provide a foundation for tracking clonal selection under therapy and, ultimately, for anticipating resistance before it emerges.