In the context of this project, our main contribution consists of a systematic study of open challenges in the evaluation of structural variations callers. Through this work, we identified several often overlooked pitfalls that can bias SV benchmarking and we proposed field-standard best practices that can aid researchers in evaluating their software tools. In addition, the new benchmarking dataset we helped create is expected to become a cornerstone for assessing SV callers' performance. This will help the bioinformatics community to develop next-generation tools capable of accurately characterizing structural variations, including those that remain insufficiently captured today. Finally, the somatic structural variation caller we are developing achieves very high accuracy, outperforming the top state-of-the-art competitor. Improving the characterization of somatic structural variations in both easy and complex regions of the human genome will shed more light on how these rearrangements contribute to cancer development.
Our contributions in the area of pangenomics were designed specifically to address gaps in the state-of-the-art. Our goal was to provide the scientific community with new methods, data structures, and tools that enable analyses that were previously not possible. First, we developed the first pangenome-based approach for the detection and differential quantification of alternative splicing events. This work represents an initial step toward the adoption of spliced pangenomes, enabling more direct investigations of how genetic variation influences splicing patterns. Second, we introduced the first data structure for indexing pangenome graphs that can scale to large graphs without pruning complex regions and support queries of arbitrary length. Third, we proposed an approach to augment a pangenome graph directly from raw sequencing reads. Unlike state-of-the-art methods that rely on high-quality de novo assemblies, our approach avoids the costly and computationally intensive assembly step and the need for high-coverage data coming from multiple sequencing technologies. Finally, our work on pangenome graph personalization provides a cleaner combinatorial problem formulation, making it easier to analyze theoretically and opening new avenues for further theoretical investigation.
