DLBCL is a molecularly heterogeneous disease. Many clinical trials have failed to demonstrate improvements to survival by altering first line therapy. This may be due to failure to stratify patients in a way that reflects this molecular heterogeneity. Over the last 5 years, three large DLBCL sequencing studies molecularly subtyped DLBCL using differing sequencing methods and subtyping strategies. They converged on highly similar molecular subtypes. Retrospective application of these molecular subtypes to failed clinical trial data demonstrated the importance of targeting biological agents appropriately. The aim of this thesis was to develop and validate a custom somatic variant calling and annotation pipeline for use on DLBCL FFPE biopsy and plasma ctDNA samples. This was conducted in the context of a prospective observational clinical trial, DIRECT. After establishing the variant calling pipeline, the next aim of the project was to compare genotyping profiles from baseline biopsy and plasma samples. Additionally, I wanted to explore the clinical features associated with plasma genotyping success. I demonstrated that our custom pipeline, AULE (Automated Ultra-sensitive Lymphoma Evaluation) calls somatic variants in DLBCL driver genes at comparable rates in biopsy and paired adequately genotyped plasma samples, as well as the literature. Two different machine learning methods highlighted raised lactate dehydrogenase and total metabolic tumour volume (TMTV) over 100ml as highly predictive of ade- quate genotyping from plasma. This is the first time plasma genotyping adequacy has been linked to clinical features. I also demonstrated that patients with inade- quate plasma genotyping have superior survival. This is likely because patients with adequate plasma genotyping are more likely to have a TMTV greater than 100ml, and TMTV greater than 100ml is associated with inferior survival. The superior survival of the inadequately genotyped group mitigates the reduced ability to detect targetable variants in many of these cases. Genotyping adequacy also affected the ability of the LymphGen molecular classifier to assign samples to molecular subtypes. All inadequately genotyped biopsy and x plasma samples were assigned to LymphGen “Other”. I also showed that adequately genotyped plasma samples assigned to LymphGen “Other” had genotyping profiles with fewer variants than those assigned to true LymphGen subtypes. Combining the genotyping profiles of cases where both biopsy and plasma samples were as- signed to LymphGen “Other” did not improve classifier performance. The inclusion of structural variant data to the input for LymphGen allowed a small number of cases to be reclassified from “Other” to a true molecular subtype. However, a small number of cases lacking structural variants, but with complex genotyping profiles overlapping several molecular subtypes were reassigned to different molecular subtypes on repeated classification. Finally, I demonstrated that BCL2 somatic hypermutation status is highly predictive of BCL2 structural variant status. This could be used instead of fluorescent in situ hybridisation to call BCL2 structural variants in DLBCL. In the experience of the DIRECT trial, genotyping from plasma ctDNA has a failure rate equivalent to the rate that FFPE biopsy blocks contain insufficient material for sequencing. The sequencing of plasma rescues 22% of cases from failed biopsy genotyping.