Quantitative scoring of differential drug sensitivity for individually optimized anticancer therapies

Abstract

We developed a systematic algorithmic solution for quantitative drug sensitivity scoring (DSS), based on continuous modeling and integration of multiple dose-response relationships in high-throughput compound testing studies. Mathematical model estimation and continuous interpolation makes the scoring approach robust against sources of technical variability and widely applicable to various experimental settings, both in cancer cell line models and primary patient-derived cells. Here, we demonstrate its improved performance over other response parameters especially in a leukemia patient case study, where differential DSS between patient and control cells enabled identification of both cancer-selective drugs and drug-sensitive patient sub-groups, as well as dynamic monitoring of the response patterns and oncogenic driver signals during cancer progression and relapse in individual patient cells ex vivo. An open-source and easily extendable implementation of the DSS calculation is made freely available to support its tailored application to translating drug sensitivity testing results into clinically actionable treatment options.

Figure 1. Implementation of the drug sensitivity scoring (DSS) pipeline in the AML samples.

(a) Each compound was tested in a dose response series in 10-fold dilutions at 5 different concentrations (typically 1–10,000 nM). The response readout (CellTiter-Glo reagent) was normalized using positive and negative controls on each dose plate to provide the response measure (relative inhibition %). (b) Dose-response parameters estimated through logistic function model include IC₅₀ (half-maximal inhibitory concentration), slope of the curve at IC₅₀, and the bottom and top asymptotes of the curve (R_min and R_max). (c) Schematic illustration of the differential DSS calculation (dDSS, the grey area). The two dose-response curves show clearly differential activity patterns, yet their relative IC₅₀ is equal, showing an example in which IC₅₀ is not informative enough for detecting selective responses in patient samples. Inset: analytic calculation of the DSS statistic as an integral over the dose range where the drug response exceeds a given minimum activity level A_min. (d) Waterfall plots of the individual dDSS profiles enable identification of cancer-selective drugs for a given patient sample. (e) Heatmap plots of the dDSS profiles over all the samples enable identification of drug-sensitive patient sub-groups. dDSS for the control samples reflect the variability among the control sample responses. (f) Network maps of the kinases the particular sample is addicted to enable identification of oncogenic driver signals.

Figure 2. Predictive accuracy of the response scores against a visual evaluation.

(a) Average drug-response profiles in five activity classes. The error bars indicate standard error of the mean (SEM). A subset of 795 dose-response curves was visually classified into either inactive (612), low active (70), semi active (65), active (30) or very active (18) classes by an experienced drug screener (T.P.), who was blind to the response parameters during the visual evaluation. The reproducibility of the expert-assigned classifications was confirmed by repeating the visual classification six months later, showing high reproducibility (97.5% of the curves were assigned to the same class by the expert across the five activity classes). (b) Predictive accuracy of each response score was evaluated using the receiver operator characteristic (ROC) analysis, where the dose-response curves were ordered according to the increasing value of the response score (see Methods). The area under the ROC curve (AUROC) is listed for each response score when distinguishing between 612 inactive and 183 active dose-response curves (Supplementary Table 2 details the AUROC for each activity class separately). The statistical significance of observed AUROC differences between the scores (table in the inset) was calculated using the DeLong's test.

Figure 3. Unsupervised clustering of the compounds based on their drug response profiles.

(a) Clustering dendrogram of the compound screening panel. The DSS drug response profiles over all the AML patient samples, relative to the control samples, were clustered using the Ward's hierarchical clustering algorithm and Spearman's rank-based correlation coefficient (see Supplementary Fig. 6 for dendrograms from the other response scores). The primary mechanism of action (MoA) classification of the compounds is illustrated in color coding (Supplementary Table 1). (b) Comparison of the response scores in terms how accurately their compound clustering reflects the established MoA classes in terms of the adjusted Rand index (see Supplementary Fig. 2 for other evaluation indices). The empirical statistical significance of the relative differences in the cluster evaluation indices was assessed with respect to permutation-based random null distribution (see Methods).

Figure 4. Distributions of the drug response scores in the CCLE in vitro cancer cell models.

Distribution of the three scores in response to (a) PLX4720 treatment in melanoma cell lines; and (b) PD−0325901 treatment in hematopoietic and lymphoid cells. The boxes depict the median and the interquartile range of the response score, and the p-values the difference in the treatment sensitivity between the BRAF-V600E or RAS-mutated and the wild type cells, respectively (Wilcoxon rank-sum test). (c) Individual breast cancer cell responses to lapatinib treatment. The sub-group of highly responsive samples (dotted box) was identified automatically using the observed skewness value γ and its significance level (D'Agostino test15).

Figure 5. Distributions of DSS₃ responses across the primary cancer samples ex vivo.

The AML patient and control sample responses to (a) ruxolitinib and (b) entinostat. The sub-group of highly responsive samples (dotted box) was identified automatically using the observed skewness values γ and their significance levels (D'Agostino test15). Tables below list the molecular profiles (significant AML mutations and recurrent gene fusions), disease stages (D, diagnosis; D*, secondary AML diagnosis; R, relapsed and/or refractory), and French–American–British (FAB) classification of the patients to illustrate the lack of correlation between functional drug sensitivity and somatic mutation profiles in this limited cohort. Examples of drug-response curves behind some of the individual response values are shown in Supplementary Fig. 7.

Figure 6. Monitoring of treatment response using drug sensitivity and target addiction profiling.

(a) Correlation of DSS₃ response profiles in an individual patient (252) before and after dasatinib treatment (252_1 vs 252_5). (b) Correlation of target addiction profiles estimated with the kinase inhibition sensitivity score (KISS, see Methods). (c) Network view of the kinase addiction changes before and after treatment (left and right panels, respectively). The kinase addiction sub-networks show connections among the initially most active and selective kinase targets (KISS > 5 in the sample 252_1). Node coloring indicates the degree of kinase addiction (KISS, Eq. (2)), and edges connect kinases with similar inhibitor selectivity profiles (Spearman's rank-based correlation > 0.5) based on a biochemical screen of kinase inhibitor specificities. Non-expressed kinase targets were excluded from the networks. Dynamic changes in the kinase addiction maps during the all the serially sampled phases of the disease progression in this patient are shown in Supplementary Fig. 8.