Boston University
Bayesian methods for improved scientific discovery with mutational signatures
Abstract
dc:description.abstractCancer development is driven by diverse mutational processes that leave distinctive patterns of DNA alterations, known as mutational signatures, in tumor genomes. Identifying these signatures from sequencing data supports etiological interpretation, prognosis, and treatment selection. The Bayesian Poisson non-negative matrix factorization (NMF) model is a standard tool for signature discovery. However, it has three limitations: it cannot easily combine prior knowledge of established signatures with the discovery of novel ones in a single cohort; it is restricted to the tri-nucleotide sequence context, which omits biologically informative signal at extended flanking positions; and it conflates sub-processes that share similar tri-nucleotide profiles into single signatures. This dissertation develops three Bayesian NMF models that address these limitations. To address the first limitation, we propose a joint learning model that fixes known signatures at reference values and estimates novel ones from the same cohort. A sparsity-inducing prior on the exposure matrix selects the number of active novel signatures from the data, and a power posterior is incorporated for robustness to likelihood mis-specification. To address the second limitation, we introduce a Markovian Bayesian framework that extends signature analysis to flanking positions beyond the tri-nucleotide context. The factorization conditions each flanking base on its adjacent inner neighbor and the central substitution type, reducing the parameter count per signature from exponential to linear in the number of flanking positions. This makes inference tractable at extended-context scales where direct category expansion is computationally infeasible. To address the third limitation, we develop a downstream mixture model that decomposes each Markovian-recovered signature into sub-process motifs. A bootstrap-based stability criterion selects the number of motifs per signature, retaining only sub-process patterns that recur across resamples. We validate our findings through carefully designed experiments on synthetic data and on real cancer cohorts.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Mathematics & Statistics
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Feng, Haoyue
- Advisor dc:contributor.advisor
-
- Huggins, Jonathan H.
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53674