Decentralizing Clinical Tumor Exome Analysis: On-Premises GPU-Based Secondary Analysis in a Clinical Molecular Laboratory

Background/Objectives: Clinical tumor sequencing has become central to precision oncology, yet secondary genomic analysis remains heavily dependent on cloud or centralized computational infrastructure, creating barriers related to cost, turnaround time, data governance, and accessibility. We demonstrate that secondary analysis of clinical tumor exomes can be performed entirely within a clinical molecular laboratory using workstation-class GPU hardware, eliminating dependence on external computational infrastructure. Methods: Sequence data generated off site were analyzed on premises with a variant-calling workflow run on benchtop GPU workstations. Accuracy was assessed against the Genome in a Bottle germline and Sequencing Quality Control Phase 2 (SEQC2) HCC1395 somatic reference materials and recall of the variants listed on the clinical molecular reports for 172 institutional colorectal tumor exomes. Results: We analyzed all 172 exomes from reads to called variants on device, at a median of 25.1 min of on-device compute across the 170 exomes with complete timing. On the germline reference material, an F-measure of 0.9963 was obtained genome-wide, and 0.9945 restricted to the exome target. Against those reports, we recovered all 153 reported variants across the four driver genes (KRAS, NRAS, PIK3CA, and BRAF). Coding-variant recall on this development cohort was 95.74% (1168/1220) requiring exact reference and alternate allele agreement, and 98.44% (1201/1220) when 33 indels matching on coding or protein position alone were also counted. Call sets were identical across triplicate runs of five exomes on both devices. Conclusions: Secondary analysis of solid tumor exomes was completed on workstation-class hardware in a colorectal cancer cohort, showing that clinical molecular laboratories can independently perform this step without reliance on external computing resources. The architecture, in which each site analyzes its own reads, can serve as a node in a federated genomic system.