Prediction of therapeutic response and cancer outcomes in solid tumours via in silico clinical trials.

Abstract

e13603 Background: Oncology therapeutic development has high failure rates with substantial costs. Advances in technology, including molecular characterisations of cancer and computational power, provide the opportunity to better model therapeutic response for existing and emerging therapies. Methods: Concr adapted Bayesian statistical principles used by astrophysicists to effectively integrate multi-dimensional data (genomic & transcriptomic) for accurate modelling of cancer biology. Our proprietary computational framework was applied to predict clinical response and survival through construction of digital twins and in silico simulations of clinical trials. Predicted log odds ratios (LOR) were generated for drug response and compared to published data from 8 suitable historical clinical trials. Three distinct models underpin this framework: Drug Efficacy, Treatment Response & Overall Survival (OS). Three publicly available datasets were used for training: Cancer Therapeutic Response Portal ( in vitro dose-response data), Cancer Cell Line Encyclopaedia and TCGA. Results: For all 8 clinical studies, the digital twin model accurately simulated both trial arms, compared drug efficacy across arms and predicted which treatment was most active. Unblinded evaluation: Our model accurately predicted gemcitabine would have greater clinical benefit than 5-FU (LOR -0.10, P <0.0001) in metastatic pancreatic cancer. In advanced breast cancer, the model predicted docetaxel had a higher response rate than doxorubicin, and carboplatin response rates were similar to docetaxel in BRCA wild-type patients. In platinum-sensitive recurrent ovarian cancer, the model predicted cisplatin-based combination had a higher response rate than paclitaxel. Blinded evaluation: Using data for paclitaxel, the model correctly predicted that nab-paclitaxel+gemcitabine response rates were higher than gemcitabine in metastatic pancreatic cancer (predicted LOR -0.090, p = <0.001). In all 3 early breast cancer studies, predictions mirrored clinical trial results, adjuvant ECF > CMF, adjuvant TC > AC and adjuvant tamoxifen was superior to capecitabine. The underlying pan-tumour Overall Survival model used a Random Survival Forest approach, trained on 23 cancers with predictive accuracy of 0.78 AUC ROC and C-index 0.71. To enable cohort enrichment for drug response, the model segmented patient cohorts into treatment responder and non-responder. Using this approach, the log odds increased for drug response rate for 16 of the 19 different cancers and 14 of the different 17 drugs. Conclusions: Our digital twin model represents transformative technology for clinical trial simulation with strong potential for accelerating and de-risking clinical drug development, and utility for synthetic control arms.