Abstract Introduction: Molecular characterization is rapidly becoming a standard of care in clinical oncology practice. Simultaneously, growing volumes of peer-reviewed literature reporting the relationships between genomic aberrations, tumor responsiveness to treatments, and associated patient outcomes has significantly increased the challenge facing clinicians to comprehensively identify and prioritize personalized treatments. This challenge is further compounded by a) the disease-specific nature of most predictive biomarkers, b) discordant published observations, c) the varying levels of clinical validity associated with most biomarker:drug response relationships, and d) the time constraints that all treating oncologists face. To overcome such challenges, we developed an automated computational system to assess the predicted effects of published genomic aberrations on drug responsiveness in a patient/disease-specific manner. Here we describe how computational selection and assessment of peer-reviewed published evidence relevant to a given cancer patient compares to that of a human expert with access to the same body of data. Methods: We developed a computational treatment decision support system (TDSS) for the analysis of cancer patients with a set of known mutations. The system holds a drug-response database (DRDB) containing expert curated information on three levels of clinical validity: 1) clinically endorsed (>100 FDA-approved facts), 2) clinically observed (>5,000 patients), and 3) translationally observed (∼4,000 model systems). Given a patient mutation profile, the TDSS retrieves related DRDB entries and weighs them according to validity level, relationship to patient disease (using the Medical Subject Headings ontology, MeSH), and RECIST response criteria. The TDSS then summarizes whether a given genotype is likely to confer response or resistance to available cancer drugs. To compare the performance of this system to that of a human molecular diagnostics expert, we created 48 virtual patients, each from one of five highly frequent cancer indications (lung, CRC, GIST, melanoma, breast) and with one or two randomly drawn predictive mutations. They were both processed by the TDSS and surveyed by an expert with electronically access to all DRDB data. The task was to recommend up to three treatments with highest likelihood of response, and identify up to three with highest risk of resistance for each case. Results: The TDSS required a few seconds of computation, compared to more than five hours of work by the expert. Both obtained a similar number of actionable predictions, for essentially the same 75% of the cases. For each case, the three treatments selected for response / resistance were compared and critically re-assessed. We found the error rates to be higher for the molecular diagnostics expert (9.6% of treatments for response, 2.1% for resistance) than for the TDSS (3.0% response, 1.4% resistance). Frequent causes of human error include a) overlooking evidence and b) failure to consider the numbers of patient cases observed. Conclusions: Computational aggregation and rules-based analysis of the clinical effects of tumor mutations on cancer drug responsiveness can greatly aid the evidence-based prioritization of treatment options for cancer patients with superior accuracy and turn-around-times. While a large-scale study is planned to compare TDSS performance against multiple experts in a real-life clinical setting, our preliminary insights suggest that the necessity for such computational systems will parallel both the emergence of clinico-molecular knowledge in this domain and the adoption rates for clinical use of genomic technologies. Citation Format: Alexander Zien, Francesca Diella, Anja Doerks, Theodoros Soldatos, Markus Hartenfeller, David B. Jackson. Automated retrieval and assessment of biomarker-related evidence for cancer treatment decision support. [abstract]. In: Proceedings of the AACR Precision Medicine Series: Drug Sensitivity and Resistance: Improving Cancer Therapy; Jun 18-21, 2014; Orlando, FL. Philadelphia (PA): AACR; Clin Cancer Res 2015;21(4 Suppl): Abstract nr A41.
Read full abstract