Evaluating a therapeutic window for precision medicine by integrating genomic profiles and p53 network dynamics

Abstract

The response variation to anti-cancer drugs originates from complex intracellular network dynamics of cancer. Such dynamic networks present challenges to determining optimal drug targets and stratifying cancer patients for precision medicine, although several cancer genome studies provided insights into the molecular characteristics of cancer. Here, we introduce a network dynamics-based approach based on attractor landscape analysis to evaluate the therapeutic window of a drug from cancer signaling networks combined with genomic profiles. This approach allows for effective screening of drug targets to explore potential target combinations for enhancing the therapeutic window of drug responses. We also effectively stratify patients into desired/undesired response groups using critical genomic determinants, which are network-specific origins of variability to drug response, and their dominance relationship. Our methods provide a viable and quantitative framework to connect genotype information to the phenotypes of drug response with regard to network dynamics determining the therapeutic window.

Fig. 1. Network dynamics-based estimation of therapeutic window and application to p53 network.

a Overview of the 6 steps in application of network dynamics-based estimation of therapeutic window for drug-target discovery and stratification of patients. b Dose–response curves (right) were obtained from simulations of dose-dependent perturbations in the p53 regulatory network (left). Red and blue graphs denote dose–response curves of cancer and control network, respectively. Efficacy and toxicity were calculated by maximum effects of dose–response curves in cancer and control network, respectively. Potency was calculated by IC50 values of dose–response curves. Therapeutic window was calculated by comparing drug response curves from cancer and control networks. c Evaluation cell lines as tumor models by comparison of network features according to genomic alterations from CCLE (cell lines) and TCGA (patient samples). Functional genomic alterations were projected onto the nominal p53 network. Node status of the p53 network was determined based on the genomic data, and assigned in a ternary fashion, such that node activity is either constantly activated (black), constantly inactivated (white), or input-dependent (gray) (left). A systematic comparison of the networks in the tumors and cell lines is performed to identify cancer cell lines with the highest network similarity to those of cancer patients by determining network similarity through correlation index with values from –1 to 1 (from red to blue, right). d Selection of cell lines and cancer networks that best match those of patient tumors (left) and genomic alteration profiles (right). Cancer cell lines and patient tumors that have the same node activity profile were matched to an identical single network.

Fig. 2. Comparison of perturbation simulation with database for drug response.

a, b Comparison of predicted drug response with (a) IC50 and AUC from GDSC data and b GR50, GRmax, and GR_AOC from GRbrowser data. Each value from database and simulation of corresponding cancer cell-specific networks was categorized as sensitive or resistant according to a threshold. Prediction rates were calculated for selected drugs (upper panel). Correlation coefficients were calculated between values from database and simulation of corresponding cancer cell-specific networks or 100 networks with random alterations (bottom). c Comparison of predicted drug response between simulation results of the control network and experimental data of MCF10A cells. Each sensitivity metric was normalized to its threshold for comparison. d Comparison of predicted synergistic effects with DREAM challenge data for combination of AKT_1 and BCL2_2 compounds in HCC1187, MCF7, and MDA-MD-436 cell lines. (left). Synergy score over 0 is categorized as sensitive in DREAM challenge data. Combination index was calculated from simulation of corresponding cell line-specific networks, and combination index under 1 is categorized as sensitive. Combination index from simulation of 100 networks with random alterations was mostly 0, indicating no synergistic effect (right). We used experimental values of the AKT inhibitor omipalisib or dactolisib, the BLC2 inhibitor navitoclax, the CYCE inhibitor RO-3306 or purvalanol A, the MDM2 inhibitor CCT007093, the ATM inhibitor KU-55933 or TCS 2312, and the MDM3-p53 interaction inhibitor nutlin-3 depending on data availability.

Fig. 3. Target screening and patient stratification using drug response categorization.

a Categorization of for according to “selective control” group as S_1–4 by efficacy using dose–response curves obtained from simulation of inhibition of a target node and its outgoing links (left). Categorization of responses according to “optimal control” group as O_1–3 by toxicity and potency by comparing to dose–response curves obtained from simulation of cancer and control networks with inhibition of a target node or link (right). b All the drug responses from 17 cancer networks with 480 perturbations were plotted based on efficacy, potency, and toxicity. Colors denote 12 drug response categories (S_n,O_m). c Distribution of the 480 perturbations in cancer networks with regard to the response categorization. d The example triangle map representing all the response categories of 480 drug perturbations to screen for optimal targets for a single cancer-cell network. Each square on the triangle map represents the (S_n,O_m) classification obtained from the dose-dependent simulation for corresponding drug or drugs. The NA denotes the case of a link combination in which the two links were from the same node or the case of a node-link mixed combination in which the link from a node were combined with that node. The graph shown in the upper left represents the dose–response curve for corresponding drug or drugs. The graph in the lower left represents the dose–response landscape for corresponding drug combinations. e Illustrative workflow to identify critical determinants and their dominance relationship for patient stratification. For each cancer network, we analyze dose–response curves of all the subnetworks, including the control network, in the two-dimensional efficacy–potency plot (step 1). The critical determinant is the common and minimum genomic alteration(s) among the networks exhibit the same drug response with the original cancer network. By determining the effect of combinations of the genomic alterations to the drug response curve, we establish the dominance relationships among them (step 2). We can predict the response of cancer networks and stratify them using critical determinants and their dominance relationship (step 3).

Fig. 4. Drug-target screening by cancer cell network responses to identify targets with therapeutic windows.

Analysis of the 480 single and combination perturbations in NT_8 (A, B) and NT_9 (C, D). a, c Triangle maps of the (S_n,O_m) classification for NT_8 and NT_9. Capital letters in a and c correspond to graphs in b, d, respectively. b, d Dose–response curves and dose–response landscapes for the selected single (A, B, C, D, H, I) and combination (E, F, G, J) perturbations. Red and blue graphs denote dose–response curves of cancer and control network, respectively.

Fig. 5. Prediction of drug response and stratification of patients based on critical determinants.

a The selected 3 single and 3 combination perturbations from the cell line-specific networks. b Critical determinants and their dominance relationship obtained from the cell line-specific networks. Upward pointing triangle indicates an activating genetic alteration; downward pointing indicates inactivating genetic alteration. Red indicates D response; blue indicates U response. Critical determinants at higher levels (Lv) are dominant over those at lower levels in determining the drug response. c Prediction of drug responses in the patient-specific networks using statistically identified markers or the dominance relationship between critical determinants (CD). d Stratification of patient-specific networks to the response of AKT inhibition based on unbiased genetic alterations (left) or on the dominance relationship between critical determinants obtained from patient-specific networks (right). The first column shows D or U responses of patient-specific networks. The second column shows group for each critical determinant. The remaining columns show genetic alterations in patient-specific networks (black: constantly activated alteration, white: constantly inactivated alteration, gray: no alteration). Among the genomic alterations, critical determinants for each patient are highlighted in colors (light blue: constantly inactivated alteration as an U response marker, dark blue: constantly activated alteration as U response marker, red: constantly activated alteration as D response marker, pink: constantly inactivated alteration as D response marker). The cluster 1–15 are notated by the critical determinants in Supplementary Fig. 5 while the cluster 16 has no critical determinant.

Fig. 6. Overview of network dynamics-based evaluation of the therapeutic window and stratification of patients for advancing precision medicine.

The patients with different cancer types are clustered into groups, G1, 2, and 3, according to their genomic profiles. The current basket trial analyzes relationships between the patient group’s genomic profile and given drug responses (left). On the other hand, our in silico basket trial analyzes the dynamics of patient group-specific networks, NT_1, 2, and 3 representing G1, 2, and 3, respectively, through virtual experiments (right). The two approaches lead to different drug response predictions for the new group, G4 (bottom).