Abstract 2719: Molecularly-informed prediction of treatment efficacy in GDSC cell line data using computational reasoning.

Abstract The Digital Drug Assignment (DDA) system is a knowledge-graph-based computational method that automates reasoning at the patient level and scores molecularly targeted agents (MTAs) based on the full tumor genomic data. This approach was predictive of relative benefit of the agents as used in the SHIVA01 trial (DOI: 10.1038/s41698-021-00191-2). Here, we evaluated the predictive power of DDA on a larger scale by analyzing tumor genomic and drug sensitivity data from the GDSC database. Our study was based on data from 659 cell lines derived from a broad spectrum of solid tumors. Corresponding drug sensitivity data were available for 34,713 treatment datapoints involving 87 types of MTAs. All tumor genomic profiles were processed using DDA, which scores MTAs and stratifies them according to predicted efficacy. Consequently, the same MTA can receive different drug scores across tumors, depending on their individual molecular profiles. Treatments were then ranked for each tumor based on their DDA scores, resulting in 72 treatment groups defined by ranking positions (i.e., drugs ranked at the same position across tumors formed one treatment group). Sensitivity to treatment was determined using Z-scores of IC50 values, with treatments showing negative Z-scores classified as sensitive. Among the top-ranked MTAs, 54% of treatments were sensitive, with sensitivity gradually decreasing across lower-ranking groups and reaching 0% in the bottom group. A linear trend across DDA score rank groups from top to bottom was confirmed by the Cochran-Armitage trend test (Z = -10.42, p = 2.08e-25), indicating a very strong negative trend across ordered groups from the top towards the bottom. Increased confidence in the benchmark drug response classification was achieved by excluding treatments around the median with progressively larger absolute IC50 Z-score thresholds. With IC50 Z-score exclusion thresholds of absolute 1, 2, 2.5, and 3, the sensitivity of cell lines to top-ranked treatments was 59%, 67%, 74%, and 83%, respectively, while remaining 0% in the bottom groups in all cases. Thus, increasing confidence in drug response correlated with higher predictive accuracy. Although group sizes decreased with stricter thresholds - reducing statistical power - the results remained highly significant throughout, the Cochran-Armitage trend test Z-values gradually increased from -8.33 to -4.68 (all p < 0.001). These results demonstrate the predictive power of DDA-score-based treatment ranking across solid tumors and MTAs, with top-ranked drugs having the greatest efficacy. The findings strongly support the notion that there is a correlation between aggregated scientific evidence and drug sensitivity. DDA can potentially address challenges with complex molecular profiles in routine clinical settings and clinical trial design. Citation Format: Robert Doczi, Akos Takacs, Anna Dirner, Dora Lakatos, Barbara Vodicska, Dora Gorog-Tihanyi, Reka Szalkai-Denes, Eniko Kispeter, Andras Makkos, Aniko Gorbe, Peter Ferdinandy, William T. Beck, Christophe Le Tourneau, Petak Istvan. Molecularly-informed prediction of treatment efficacy in GDSC cell line data using computational reasoning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2719.