Multimodale analyse van celvrij DNA voor vroege en gevoelige kankerdetectie

My research aims to make multimodal analysis of circulating cell-free DNA (cfDNA) from liquid biopsies more sensitive and reliable for non-invasive cancer detection. The key idea is to integrate genomic alterations (e.g., copy-number signals), DNA methylation, and fragmentomic patterns (fragment sizes, end motifs, jagged ends), since these modalities carry complementary cancer and tissue-of-origin information. A major challenge is that early cancers produce very weak signals, and fragmentomic biomarkers are strongly affected by wet-lab and preanalytical factors, which can cause systematic differences between datasets and hurt reproducibility. To address this, I developed methods that either implicitly or explicitly model and correct for these variations: (1) a semi-supervised deconvolution approach that estimates cell-type composition and supports tissue-of-origin inference while handling shifts between reference atlases and test samples; (2) a domain-adaptation method to correct protocol-driven shifts in sequencing-derived profiles, improving robustness across laboratory settings; and (3) a kernel-based multimodal learning framework that combines modalities effectively in limited-sample regimes and improves detection compared to single-modality models. Finally, I studied how modalities are biologically coupled through endonuclease activity, treating it both as a confounder to control and as a mechanistic signal that helps interpret how cfDNA is generated and processed.