RP31 - Streamlined uncertainty aware data analysis for molecular tumor boards

In the first two cohorts, we have developed methods for dealing with genomic haplotypes in precision medicine with the goal for enabling uncertainty aware predictions for molecular tumor boards. These are complemented by approaches to enhance the transparency of latent variables by extensive visualizations as well as LLM based chat interfaces, ensuring interpretability of the results and a comprehensive assessment of the entailed evidence. In cohort 1, we developed a method for determining the HLA type [1]. In cohort 2 we are developing a method for determining the combined impact of genomic variants on proteins or neoantigens. Both build on our genomic variant calling model Varlociraptor [2] which comprehensively captures uncertainty in the processes underlying DNA/RNA sequencing and their analysis for determining existence and category of individual genomic variants under consideration of the experimental scenario. All these methods provide accurate and informative results, however, to some extend, at the expense of computational speed. In the third cohort, we will focus on improving the algorithmic and computational efficiency of these approaches. We will redesign intermediate data representation to minimize IO stress, leverage SIMD/SIMT and general parallelization, to maximize use of computational resources, and work algorithmically to improve the theoretical computational complexity. Moreover, we aim to generalize the previously developed approaches towards further applications.

[1] H. Uzuner, A. Paschen, D. Schadendorf, and J. Köster, “Orthanq: transparent and uncertainty-aware haplotype quantification with application in HLA-typing” BMC Bioinformatics, vol. 25, no. 1, p. 240, Jul. 2024, doi: 10.1186/s12859-024-05832-4.

[2] J. Köster, L. J. Dijkstra, T. Marschall, and A. Schönhuth, “Varlociraptor: enhancing sensitivity and controlling false discovery rate in somatic indel discovery” Genome Biology, vol. 21, no. 1, p. 98, Apr. 2020, doi: 10.1186/s13059-020-01993-6.