Research
Tumor evolution and subclonal architecture
Cancer is a disease of evolution. Cells divide, acquire mutations, and the descendants of an early mutation form a subclone that carries it. From a single bulk sequencing sample you can see the variant allele frequency of every mutation, and from those frequencies you can work backwards to how many distinct cell populations there are and what fraction of the tumor each represents.
The bottleneck for a long time was not the statistics but the compute. Subclonal reconstruction methods were slow enough that running them on a large cohort was impractical, which meant nobody could ask whether subclonal structure related to clinical outcome across many cancers at once.
CliPP, short for Clonal structure Identification through Pairwise Penalization, attacks that with a regularized likelihood model that clusters subclonal mutations under a pairwise penalty. It processed whole-genome and whole-exome data from more than 12,000 tumor samples in under 24 hours, which made the cohort-scale question answerable.
I designed and ran the pan-cancer analysis that followed, covering 7,827 tumors across 32 cancer types. The finding was that subclonal mutation fraction (sMF), essentially a measure of how much evolutionary time a tumor has had, was associated with better outcomes in 16 cancer types with low to moderate tumor mutation burden.
Transcriptomic deconvolution
Deconvolution recovers component-specific expression from a mixed sample. The established methods work well when counts are dense, but important data types break that assumption, such as microRNA sequencing, where the transcriptome is small and counts are low, and spatial transcriptomics, where each spot has few reads. Under sparsity the Gaussian and log-normal assumptions underneath most deconvolution models stop being reasonable.
DeMixNB models the mixture as a sum of negative binomial distributions, embedded so that the tumor-specific component can be estimated semi-reference-based. To show the problem was real rather than theoretical, we generated a mixed small-RNA benchmark dataset with known ground truth and demonstrated where existing methods fail on it.
Applied to miRNA-seq from 856 breast cancer patients and 3,755 spatial spots from lung tumors, the method produced clinical and mechanistic findings about tumor cell plasticity that the mixed signal had been hiding. The paper is in Genome Biology.
Mutation calling and genomic infrastructure
Studying cancer evolution by tracing somatic mutations depends on somatic variant calls being right. I have worked on the production side of that: optimizing a mutation calling pipeline across 7,000+ whole exomes, and co-authoring the methods chapter for MuSE, which models sample-specific error rather than assuming a shared error profile across a cohort.
Additionally, my work has included multi-omic integration of ATAC-seq with RNA-seq to identify tumor signature genes, single-cell RNA-seq analysis linking immune composition to tumor evolutionary state, and managing large genomics datasets on HPC infrastructure with attention to access control, versioning, and reproducibility.
Earlier work: multiple sclerosis genomics
Before cancer I spent three years on multiple sclerosis genetics at BYU in the lab of Dr. Mary Davis. That included Mendelian randomization studies of candidate causal factors, GWAS of variants influencing MS comorbidities, and a random forest classifier for MS subtype built on routine blood panel data.
Hematological profile in MS → Viral infection and cancer risk in SLE and MS →