Cancer Drug Response Profile scan (CDRscan): A Deep Learning Model That Predicts Drug Effectiveness from Cancer Genomic Signature

Abstract

In the era of precision medicine, cancer therapy can be tailored to an individual patient based on the genomic profile of a tumour. Despite the ever-increasing abundance of cancer genomic data, linking mutation profiles to drug efficacy remains a challenge. Herein, we report Cancer Drug Response profile scan (CDRscan) a novel deep learning model that predicts anticancer drug responsiveness based on a large-scale drug screening assay data encompassing genomic profiles of 787 human cancer cell lines and structural profiles of 244 drugs. CDRscan employs a two-step convolution architecture, where the genomic mutational fingerprints of cell lines and the molecular fingerprints of drugs are processed individually, then merged by ‘virtual docking’, an in silico modelling of drug treatment. Analysis of the goodness-of-fit between observed and predicted drug response revealed a high prediction accuracy of CDRscan (R2 > 0.84; AUROC > 0.98). We applied CDRscan to 1,487 approved drugs and identified 14 oncology and 23 non-oncology drugs having new potential cancer indications. This, to our knowledge, is the first-time application of a deep learning model in predicting the feasibility of drug repurposing. By further clinical validation, CDRscan is expected to allow selection of the most effective anticancer drugs for the genomic profile of the individual patient.

Quantitative Structure-Mutation-Activity Relationship Tests (QSMART) Model for Protein Kinase Inhibitor Response Prediction

“…102 The resulting ROC curves for 23 cancer types and the overall curve are shown in Fig 2c. 103 The overall AUC is 0.981 and comparable to a recent deep neural network-based 104 study [33] (AUC > 0.98). AUC for each cancer-centric model is available in S1 Table. 105…”

“…Furthermore, based on the feature sets used to validate the contribution of previous 201 components in the framework, neural networks also outperformed random forests, SVM, 202 and Lasso regression (S2 Table-S6 Table). A previous study [33] also showed that neural 203 networks outperformed random forests and SVM (R 2 = 0.843, 0.698, and 0.562 for 204 DNN, random forests, and SVM, respectively) in drug response prediction. Interestingly, 205 neural network was only slightly better than Lasso in overall performance when 206 randomly selected features were used as inputs (R 2 = 0.125 vs. 0.116, p-value = 0.015, 207…”

“…models. The overall residual plot (Fig 2b) shows that there is no specific U shape, 95 inverted U shape, or funnel shape, which means our prediction models need no more 96 higher-order features to capture the variation of drug responses (S1 Fig shows residual threshold used in a previous study [33] (IC 50 = -2), we set multiple thresholds (-4, -3, -2, 101 -1, and 0) and averaged the results to avoid overestimating the prediction performance. 102 The resulting ROC curves for 23 cancer types and the overall curve are shown in Fig 2c.…”

“…13 Ding et al [36] DNN (EN, SVM) MUT, EXP, CNV 25-fold CV AUC = 0.87 2018.03.08 He et al [30] Kernel (EN, Ridge, RF) EXP 3-fold CV Precision = ∼0.35 2018.06. 11 Chang et al [33] CNN (RF, SVM) SNP 5% leave-out R 2 = 0.843 2018.07.01 Cichonska et al [31] Kernel SNP, MET, EXP, CNV 10-fold CV r = 0.858 2018.09.14 Le et al [19] Network (Kernel) MUT, EXP 5-fold CV r = 0.804 2018.09.14 Juan-Blanco et al [20] Network MUT, EXP LOOCV AUC = ∼0.72 2018.10.10 Yang et al [21] Network + SVM (Kernel) MUT, MET, CNV, PPI 5-fold CV AUC = 0.788 2018.12.07 Liu et al [22] Network EXP 10-fold CV r = 0.73 2019.01. 22 Wei et al [23] Network EXP LOOCV r = 0.63 2019.01.…”

“…To compare our prediction performances with those in a previous study [33], we used 406 the same methods to generate (1) fingerprints, (2) extended fingerprints, and (3) 407 graph-only fingerprints by PaDEL-descriptor (version 2.21) [41] for each drug. In total, 408 there are 3,072 binary descriptors as drug features in the compared training sets.…”