Predictive Gene Signatures Determine Tumor Sensitivity to MDM2 Inhibition Free

The p53 protein plays a pivotal role in the tumor suppression and maintenance of genome integrity, acting as a transcription factor that can induce apoptosis, cell-cycle arrest, DNA repair and/or senescence in response to cellular stresses. Although approximately half of all cancers harbor TP53 mutations, wild-type p53 is also frequently inactivated by other mechanisms. The E3 ligase murine double minute 2 (MDM2, also known as HDM2) is frequently overexpressed in various cancer (1, 2), directly inactivating p53 function by binding to its transactivation domain and also inducing its degradation through p53 ubiquitination. Pharmacologic inhibition of the p53–MDM2 interaction has been developed into a therapeutic approach for exerting p53-mediated antitumor effects.

Several preclinical (3–8) and early clinical (9–11) studies of MDM2 inhibitors demonstrated the proof-of-concept of their antitumor effects via p53 activation in both solid and hematologic tumors. There are multiple ongoing clinical studies (12, 13) of MDM2 inhibitors in several cancers, including the novel MDM2 inhibitor DS-3032b, still in an early phase (solid tumors and lymphoma, NCT01877382; myeloma, NCT02579824_DS3032b; leukemia NCT02319369). Clinical responses in these trials have been limited overall, but some patients clearly achieve clinical benefit by monotherapy with MDM2 inhibitors (9–13). Therefore, there is a need for improved MDM2 inhibitors, or efficacious combination therapies, as well as predictive biomarkers for monotherapy with current MDM2 inhibitors.

We report here that DS-3032b exerts highly potent p53-dependent antitumor effects both in vitro and in vivo. In addition, to test the hypothesis that gene signatures can predict sensitivity to MDM2 inhibition, we measured sensitivities to MDM2 inhibition in a broad range of cancer cell lines, patient-derived xenograft models, and primary leukemia specimens and used these to derive predictive signatures from gene expression profiles determined before treatment. We report two different bioinformatic approaches to define predictive gene signatures: one used a panel of 240 molecularly annotated cancer cell lines to find a common signature among various tumor types and the other focused primarily on acute myeloid leukemia (AML) as an example to define tumor-type–specific signatures. Considering that TP53 mutation status is a predictive but not a sufficient biomarker, we also combined the gene signatures with TP53 mutation status, aiming at improved prediction of response.

DS-5272 (8), DS-3032b, and DS-3032a (a salt-free form of DS-3032b) were developed and synthesized by Daiichi Sankyo Co., Ltd. (Fig. 1A). (-)-Nutlin-3a was purchased from Cayman Chemical Company.

SJSA-1, Saos-2, WM-115, and DLD-1 cells were purchased from the ATCC. MOLM-14, OCI-AML3, DOHH-2, and Z-138 cells were purchased from the Leibniz-Institut Deutsche Sammlung von Mikroorganismen und Zellkulturen. The authenticity of the cell lines was confirmed by DNA fingerprinting with the short tandem repeat method, using a PowerPlex 16 HS System (Promega) within 6 months before the experiments. Mycoplasma was tested by PCR method and confirmed the 30 negativity for all the cell lines used (14). Cryopreserved cells, or fresh patient samples, were cultured as described previously (14) and used at passages 4 to 6 for the in vitro and in vivo experiments.

For 48-hour drug treatment experiments, cell lines were harvested at log-phase growth, plated at a density of 1.5 × 10⁵/mL for MOLM-14, OCI-AML3, and Z-138, or 1.0 × 10⁶/mL for primary AML and MCL cells, and treated with DS-3032b (0, 25, 50, 100, 250, 500, 1,000 nmol/L). Annexin V and propidium iodide (PI; purchased from Sigma-Aldrich) binding assays were performed to assess apoptosis as described previously (14). Annexin V- and PI-negative cells were counted as live cells, and the percentage of live cells was calculated by normalizing to untreated controls. Based on the live-cell counts measured at each concentration, IC₅₀ was calculated by using GraphPad Software (ver. 6). Calculations of specific apoptosis and area under the curve (AUC) based on the percentage of live cells are given in Supplementary Information.

An in vivo study in patient-derived tumor xenograft mice was performed by Champions Oncology Inc. Thirteen tumor models, including melanoma, non–small cell lung carcinoma (NSCLC), colorectal and pancreatic, were selected from the Champions TumorGraft library. Gene expression data and TP53 genotype information were provided by Champions Oncology. Female nu/nu mice (Harlan) between 6 and 9 weeks of age were implanted bilaterally in the flank region with tumor fragments harvested from host animals, each implanted from a specific passage lot. When tumors reached approximately 125 to 250 mm³, animals were matched by tumor volume into treatment and control groups and dosing initiated (day 0). Animals bearing tumors were grouped into treatment and vehicle groups (n = 10, respectively), and DS-3032b was administered at 10 mg/kg/day for 10 days. Tumor dimensions were measured twice weekly by digital caliper. Tumor volume (TV) was calculated using the formula (i): TV = (width)² × (length) × 0.52. Percent tumor growth inhibition (%TGI) values were calculated for each treatment group (T) versus control (C) using tumor measurements of initial (i) and at the end of dosing or the closest tumor measurement (f) by the formula (ii): %TGI = 1 − T_f − T_i / C_f − C. All experimental procedures were performed in accordance with the in-house guideline of the Institutional Animal Care and Use Committee of Champions Oncology, Inc.

For all the animal studies in the present study, the study protocols were approved by the Institutional Animal Care and Use Committee of Champions Oncology Inc. (patient-derived xenograft mice) and Covance Laboratories Inc. (SJSA-1 and MOLM-13 xenograft mice). Studies with primary human leukemia samples were approved by the institutional review board at MD Anderson Cancer Center. Forty-one peripheral blood or bone marrow samples from 38 patients with AML, newly diagnosed, or relapsed/refractory after chemotherapy were collected.

A differential expression analysis of the mRNA datasets derived from sensitive and resistant cell lines was performed. Genes were ranked by P value according to Student two class t test in each cell line in OncoPanel. Genes differentially upregulated in sensitive cell lines, defined by GI₅₀ values of DS-5272 (<1.5 μmol/L), contributed to the sensitivity signature. The signature contained the top 1% of ranked genes (n = 175). The robustness of the gene signatures was inspected in several ways. First, heat maps of the entire 1% gene lists were generated. Visual inspection of the heat map was performed to characterize the extent and consistency of the differential expression across cell lines and down the gene list. Second, the Q value was calculated as (P value/P value rank) × number of genes measured. As a guideline, all signatures were required to have ≥10 genes with Q value ≤ 0.1 to support further analysis. Lastly, the sensitivity signature was compared against other drug signatures by comparing the total number of genes at the Q value threshold 0.05 and the Q value of the least significant gene in the top 1% signature.

The expression values of the genes within the signature were Z score normalized. Once normalized, the average Z score of the sensitivity signature genes was calculated on an individual cell line, murine or human sample basis to generate a sensitivity score, respectively. The same calculation was used in GEP data derived from Oncomine and primary AML samples. Data from the microarray studies have been submitted to the Gene Expression Omnibus. The accession number is GSE110087.

For random forest methods with cross-validation in primary AML samples, gene expression profiling of the 41 primary AML samples was used. Eleven each among 33 TP53 wild-type samples, or 14 each among 41 (including 8 TP53-mutant samples) were selected as sensitive or resistant to DS-3032 based on AUC based on the percentage of live cells. Classification was performed using the random forest method to identify a predictive algorithm with a particular gene set. Forests were created with 10,000 decision trees. Performance was assessed by AUC of a receiver-operating characteristic (ROC) curve, from internal out-of-bag (OOB) testing results. Each tree uses a different random bootstrap sample of two thirds of cases from the original data. A test set classification is obtained for each case in about one third of the trees. By default, the random forests method performs a binary classification (in the present study, sensitive vs. resistant to DS-3032b). However, it also reports a probability for being sensitive to DS-3032b. The threshold for predicting sensitive or resistant was set as 0.5 probability. Overall accuracy was defined as a ratio of the cases correctly predicted as sensitive or resistant to the total cases. Positive predictive value was defined as a ratio of the cases correctly predicted as sensitive to the total cases predicted as sensitive. Negative predictive value was defined as a ratio of the cases correctly predicted as resistant to the cases predicted as resistant. Sensitivity was defined as a ratio of the cases correctly predicted as sensitive to the total sensitive cases.

The chemical structure of DS-3032b is shown in Fig. 1A. To examine the in vitro inhibitory effects of DS-3032b on MDM2–p53 interaction, we utilized homogeneous time-resolved fluorescence (HTRF). DS-3032b inhibited the MDM2–p53 interaction with IC₅₀ values of 5.57 nmol/L (Supplementary Fig. S1A). Consistently, nanomolar concentrations (30–1,000 nmol/L) of DS-3032b increased p53 levels in a dose-dependent manner by 24 hours, along with induction of the p53 targets MDM2 and p21, in various cancer cell lines (SJSA-1, Colo829, MOLM-13, DOHH-2, NCI-H929, and WM-115) harboring wild-type TP53 (Fig. 1B). Treatment with DS-3032b also induced PARP cleavage, indicating apoptosis induction (Fig. 1B). Dose-dependent increase in expressions of the p53 target genes CDKN1A (p21) and BBC3 (PUMA) was also confirmed in SJSA-1 cells (Supplementary Fig. S1B). These results proved that DS-3032b is a potent inhibitor of MDM2 and indeed activates p53 function.

We next treated six human cancer cell lines derived from various tumor types (MOLM-13, AML; DOHH-2, non-Hodgkin B-cell lymphoma; SJSA-1, osteosarcoma; WM-115, melanoma; DLD-1, colon adenocarcinoma; Saos-2, osteosarcoma) with DS-3032b or nutlin-3a, and determined their GI₅₀ values (Supplementary Table S1). DLD-1 and Saos-2 cells with TP53 mutations were resistant to either of the drugs, while the other four cell lines with wild-type TP53 were more sensitive with nanomolar values of GI₅₀. Compared with GI₅₀ values of nutlin-3a, DS-3032b was approximately 10-fold more potent. We then used a TP53 wild-type AML and mantle cell lymphoma (MCL) cell lines (MOLM-14, OCI-AML3, and Z-138) and knocked down TP53 expression using short hairpin RNA to test for p53 dependency of DS-3032b. As expected, TP53 knockeddown cells are significantly resistant to DS-3032b– and nutlin-3a–induced apoptosis (Fig. 1C; Supplementary Fig. S1C), confirming the p53-dependency of apoptosis induction of these two inhibitors. Of note, we also treated circulating MCL cells obtained from 5 patients. All samples had wild-type TP53 as determined by Sanger sequencing. DS-3032b induced apoptosis in a dose-dependent manner in all of the samples (Supplementary Fig. S1D).

We further tested the in vivo activity of DS-3032b using tumor xenograft models with TP53 wild-type MOLM-13 and SJSA-1 cells. In both models, DS-3032b inhibited tumor growth in a dose-dependent manner, especially with oral administration at 25 and 50 mg/kg/day (Fig. 1D and E; Supplementary Fig. S1E and Supplementary Table S2), indicating the potential of clinical use of this agent. Daily and intermittent dosing (every 3 or 7 days; Fig. 1E) both showed in vivo efficacy, suggesting relatively long-lasting efficacy of DS-3032b in vivo. We also confirmed induction of p53, MDM2, and p21 protein in tumor cells 6 and 24 hours after in vivo treatment with DS-3032b (Fig. 1F).

We also treated a panel of 240 cell lines from various cancers (OncoPanel, Eurofins Panlabs) with DS-3032a (a free form of DS-3032b), DS-5272 (another prototypic MDM2 inhibitor developed previously; ref. 8), and nutlin-3a. As expected, GI₅₀ profile of DS-3032b is highly correlated with that of DS-5272 (r = 0.98, P = 2.2e−16; Supplementary Fig. S2A; Supplementary File S1).

We have shown in clinical studies of the MDM2 inhibitor RG7112 that therapeutic response to RG7112 as a single drug varies from patient to patient in cancer patients harboring wild-type TP53 tumors (10). We established first that sensitivity to MDM2 inhibition is in part dependent on baseline MDM2 levels (4) and confirmed this finding in the AML clinical trial of RG7112, where both baseline and induced MDM2 levels were correlated with response (10). The heterogeneous response in p53 wild-type cell lines had also been observed in cell lines from a single disease entity (5, 15).

Indeed, the data from OncoPanel cell lines treated with the three MDM2 inhibitors showed that TP53 wild-type cell lines have a large range of GI₅₀ values (e.g., varying from <0.1 to >10 μmol/L for both DS-5272 and DS-3032b; Supplementary Fig. S2B; Supplementary Table S3 and Supplementary File S1). Cell lines that were TP53 wild-type but were resistant to all the three inhibitors included HeLa, C4I, and C32 cells, which may be related to human papilloma virus infection (HeLa and C4I) and ARF gene deletion (C32) that inactivate p53. However, the 24% of TP53 wild-type cell lines that were deemed resistant did not all have known causes of resistance (Fig. 2A). Also, our in vitro data using cell lines or primary MCL cells with wild-type TP53 (Supplementary Table S1 and Supplementary Fig. S1D) showed that sensitivity varied over a certain range (GI₅₀ < 100 to > 250 nmol/L).

To confirm the variance of sensitivities to DS-3032b of wild-type TP53 tumor cells through ex vivo experiments, we collected 41 primary AML specimens obtained from 38 patients, which were all clinically annotated (Supplementary Table S4). Eight (19.5%) had TP53 point mutations, detected by next-generation sequencing and Sanger sequencing. Samples with TP53 wild-type were significantly more sensitive to DS-3032b than those with TP53 mutations (Fig. 2B). Because exact GI₅₀ values could not be calculated in highly sensitive (<25 nmol/L) or highly resistant (>1,000 nmol/L) samples, we also used the AUC based on percentages of live cell number (normalized to untreated controls) at each concentration (Figs. 2C; Supplementary Fig. S2C and Supplementary File S2). We then focused on the outer one third of sensitive or resistant samples, and assessed how efficiently TP53 mutation status alone could predict the sensitivity to DS-3032b (Fig. 2C; Supplementary Table S5). Reflecting that some of the wild-type TP53 samples were as resistant or even more resistant than TP53-mutated samples, the positive predictive value (PPV; 62%) and overall accuracy (68%) were relatively low, compared with high sensitivity of 93% (Table 1). This indicates that predictive biomarkers with higher PPV than that of TP53 mutation status alone are desired for more effectively predicting tumor sensitivity to MDM2 inhibition.

TP53 mutation status can also be a continuous variable, based on the allele frequency of mutant TP53. Of note, AUC values based on live-cell numbers (Fig. 2D) and apoptosis (Fig. 2E) are both correlated with allele frequencies of TP53 mutations in a positive and negative manner, respectively. This indicates that in cases with mutated TP53, the allele frequency of TP53 mutations could potentially serve to predict sensitivity to MDM2 inhibition.

Among the 240 OncoPanel cell lines, we defined 62-sensitive and 164-resistant cell lines, based on GI₅₀ values of DS-5272 (cutoff values of <1.5 and >10 μmol/L, respectively; Supplementary Fig. S2B and Supplementary Table S6). This categorization was almost identical to that for DS-3032b. Specifically, only three (786-o, HS294T, and RPMI8226) among 240 cell lines on the panel were categorized discordantly. As to 786-o and HS294T, the dose–response curves are essentially similar (Supplementary Fig. S2D), and we therefore concluded that the sensitivities of these lines to DS-3032b and DS-5272 are basically similar. The reason for discordant data for RPMI8226 cells remains unclear.

Using publicly available gene expression profiling data for all the cell lines, we then selected 175 genes ranked at the top 1% of the genes of higher expression level with lowest P value in the 62 sensitive cell lines than resistant cell lines (Supplementary File S3). We thus defined the average of Z-scores of the 175 genes as a “sensitivity score.” The sensitivity score similarly predicted responses to three different MDM2 inhibitors (DS-5272, nutlin3a, and DS-3032a; free form of DS-3032b; Supplementary Fig. S2E), indicating its robustness as predictive gene signature for different MDM2 inhibitors. Of note, we also defined a “resistance score” by selecting the top 1% of genes of higher expression level with lowest P value in resistant lines (Supplementary File S3); however, it did not predict the sensitivity/resistance well (worse than predictive performance by TP53 mutation status alone); thus, we decided to further develop the sensitivity score.

A combination of TP53 genotype and gene signature showed further enrichment of sensitive cell lines compared with using either of them alone (Supplementary Fig. S3A). Specifically, by applying the sensitivity signature, 41 of 55 p53 wild-type cell lines had high-sensitivity scores and the rest of 14 lines had low scores. Thirty-nine of 41 (95%) cell lines with high scores were sensitive, while 11 of 14 (79%) cell lines with low scores are resistant to DS-5272, suggesting potentially high predictive performance of this signature.

To validate the 175-gene signature, we used patient-derived tumor xenografts. In vivo antitumor activities of DS-3032b were evaluated in 13 total tumor models, derived from melanoma, NSCLC, colorectal, and pancreatic cancers (Supplementary Table S7A). A positive correlation was observed between the sensitivity scores and TGI values (Fig. 3A, Spearman r = 0.67, P = 0.031). The prediction accuracy, sensitivity, and PPV were 85%, 88%, and 88%, respectively (Supplementary Table S7B).

Performing gene expression profiling in addition to ex vivo cell viability assays of AML samples described above (Supplementary Table S4), we tested whether the sensitivity score of each sample predicts the sensitivity to DS-3032b. First, we focused only on the 33 cases with wild-type TP53. As the outer third of sensitive or resistant samples, eleven p53 wild-type samples each were selected as sensitive or resistant to DS-3032 based on AUC of %live cell number, and the 175-gene signature was tested. The prediction accuracy, sensitivity and PPV were 73%, 91% and 67%, respectively (Fig. 3B; Supplementary Fig. S3B). Next, we also tested the 175-gene signature in all 41 samples including additional 8 cases with TP53 mutations. Focusing on the outer third (fourteen each) sensitive or resistant samples in the 41 samples, the prediction accuracy, sensitivity and positive predictive values (PPV) were 79%, 93% and 72% (Fig. 3C; Supplementary Fig. S3C) respectively. Compared with results for TP53 mutation status alone described above (68%, 93%, and 62%, respectively; Supplementary Table S5), the 175-gene signature improved PPV (Table 1). Using the same dataset of our primary AML samples, we also tested a recently reported 4-gene predictive signature (16), which showed relatively low predictive performance (prediction accuracy, sensitivity, and PPV were 68%, 86%, and 63%, respectively; Supplementary Table S8 and Table 1), compared with the 175-gene signature (Supplementary Fig. S3D).

We further tested the predictive performance when the 175-gene signature was applied after TP53 mutation status was incorporated as a first predictive factor for resistance. Specifically, we predicted samples with TP53 mutations as resistant, and sensitivities to DS-3032b of the rest of samples with wild-type TP53 was predicted using the 175-gene signature (Fig. 3D, top). The accuracy, sensitivity, and PPV for predicting the top 14 sensitive or resistant samples were 82%, 86%, 81%, respectively (Fig. 3D, bottom, and Table 1), showing potentially more balanced performance with higher PPV compared with using TP53 mutation status or the gene signature alone.

Of note, the top-ranked genes in the signature were well-known p53-inducible genes (Supplementary Table S9), suggesting that the gene signature may reflect high basal activity of p53, which may lower the threshold for inducing p53-mediated apoptosis by MDM2 inhibition. With this notion, supporting the superiority of our gene signature to TP53 mutation status alone, the 175-gene signature applied to genomic data of more than 15,000 clinical samples obtained from the OncoMine database revealed that only a subset of tumors relatively enriched with wild-type TP53 (melanoma, lymphoma, myeloma, renal cell cancer, and leukemia) contains high-sensitivity scores (Fig. 3E; Supplementary Table S10). On the other hand, cervical cancer has low p53 mutation frequency and a low sensitivity score, probably owing to p53 inactivation via HPV infection in cervical cancer.

Whereas a common gene signature regardless of tumor types is useful in terms of its broad application, tumor-type–specific gene signatures perhaps need to be explored to achieve a higher performance for predicting drug sensitivity in a particular tumor type. Thus, we examined another bioinformatics approach by utilizing only the ex vivo dataset of primary AML samples described above.

We first hypothesized that genes with high expression variance have the highest predictive value for determining drug sensitivity. Using a fairly routine bioinformatics method for classifying gene signatures (17), we first selected the 1,500 top-ranked genes based on expression variance among 33 primary AML samples with wild-type TP53 (Supplementary File S2). We used these genes’ expression values to predict the results of ex vivo treatment with DS-3032b with the random forest method (18), ranking every gene based on the significance for the prediction (referred to as “variable importance”; Fig. 4A, top). Cross-validation was performed by holding out one third of the total samples for testing, and the learners (decision tree) ensembles were trained on only the remaining two thirds of the data. The overall accuracies, sensitivity, and PPV of the identified algorithm with the 1,500 genes for predicting the top 11 sensitive or resistant samples among 33 were 68%, 73%, and 67%, respectively (the confusion matrix is shown in Fig. 4A, bottom), showing similar accuracy and PPV to the 175-gene signature described above. Of note, only 8 of 175 genes in our cell line–based signature (Supplementary File S1) and none of 13 genes (19) or 4 genes (16), previously reported gene signatures, were ranked in these 1,500 genes. These results indicate that genes expressed with high variance in a particular tumor type, regardless of the relationship to p53, also participate in determining the drug sensitivity to a certain extent, perhaps in different way(s) from the other gene sets enriched with p53 transcriptional targets. In contrast to the unbiased approach based on gene expression variance above, we next selected 32 genes (Supplementary Table S11) based on previous reports suggesting the potential predictive values of the genes. Those included 13 genes previously published in predictive gene signatures for the sensitivity to MDM2 inhibitors (16, 19), 22 genes that were upregulated or downregulated by an MDM2 inhibitor (RG7112) determined in a phase I clinical trial (10), TRIAP1 and E2A (TCF3), which are reported to be transcriptional cofactors of p53 (20), and Drosha, which is reported to be downstream of MDM2 (21). Various cell-intrinsic mechanisms of p53 inactivation are known (22), and most of them (e.g., MDM2 overexpression, ARF deletion, NPM1 mutation, FLT3-ITD, PML-RARalpha, activation of PI3K or MAPK pathways) result in hyperactivation of MDM2, which inhibits p53-mediated signaling. XPO1, a nuclear exporter of p53, contributes another mechanism of p53 inactivation. Therefore, we included these two key genes in the list. We then applied these 32 genes in the random forest method, which showed the accuracy of 59.1%, sensitivity of 73%, and PPV of 57% (Fig. 4B). Next, we combined the 1,500 genes selected by the unbiased approach above and the referenced 32 genes. The random forest method created the algorithm with accuracy of 77%, sensitivity of 82%, and PPV of 75% (Fig. 4C; Table 1), indicating that combining the two gene selection methods improved the predictive performance.

We then tested the algorithm of combined genes on all 41 samples including ones with TP53 mutations. Specifically, another set of 1,500 genes was selected according to expression variance among all the 41 AML samples, combined with the same 32 referenced genes, then the random forest method was performed. The accuracy, sensitivity, and PPV for predicting the top 14 sensitive or resistant samples were 68%, 72%, 67%, respectively (Fig. 4D and Table 1), showing reduced predictive performance than when focusing on only TP53 wild-type samples.

Finally, we combined this algorithm with TP53 mutation status. Specifically, we predicted samples with TP53 mutations as resistant, and then applied the random forest method using 1,532 genes to the remaining wild-type TP53 samples (Fig. 4E, upper). The accuracy, sensitivity and PPV for predicting the top 14 sensitive or resistant samples were 75%, 68%, 82%, respectively (Fig. 4E and Table 1), showing better accuracy and PPV than the random forest method alone. This indicated that combining the random forest method with the prediction by TP53 mutation status provides a better algorithm compared with each alone. Of note, when we followed the same flow chart and ran the random forest method with the 50 top-ranked genes in the 1,532 genes (Fig. 4F), the high accuracy, sensitivity and PPV were conserved (96%, 93%, 100%; Fig. 4F). Although re-selecting the 50 genes defined by the first learning iteration has a possibility of overfitting, it suggests that narrowing down gene numbers based on the rank defined by the random forest method could be an option for defining a more feasible scale of gene set for further clinical implementation.

Our study demonstrated that DS-3032b is a highly potent MDM2 inhibitor and induces p53-mediated apoptogenic effects in various cancer cells. In addition, to overcome insufficient prediction of tumor cell sensitivities to MDM2 inhibition by TP53 mutation status alone, two different approaches for defining genetic biomarkers were developed: (i) a 175-gene signature defined by a cell line–based approach; and (ii) an AML-specific 1,532-gene signature defined by the random forest method. We propose that a combination of the classical factor TP53 mutation status and the gene signatures provides the best algorithms for predicting tumor sensitivities to MDM2 inhibitors (Fig. 5A).

Our 175-gene signature showed a robust predictive performance of the sensitivities to DS-3032b in validation datasets of human ex vivo specimens and patient-derived xenograft murine models. Importantly, although the 175 genes in the present study included all of the 4 genes, which were previously reported as predictive gene signature for other MDM2 inhibitors (16), our signature of 175 genes had higher predictive power. Of note, although another 13-gene signature has also been reported (19), it was found later not to be predictive due to the revised annotation of TP53 mutations in the cell lines (23). These findings collectively suggest that incorporating a broader spectrum of p53 transcriptional targets into the gene signature is superior to focusing on fewer genes. A concern for further developing this signature is that it is derived from cell lines of various cancers, not a single disease entity. Considering that a variety of p53 target genes are differently regulated depending on tissue types (24, 25), the cell line panel–based identification of key predictive genes may not be the best approach.

Therefore, an AML-specific gene signature was developed for correcting the potential weakness of the 175-gene signature. However, this approach is also challenging due to the limited availability of independent datasets for validation. As a partial solution, we utilized the random forest method (18), an ensemble classifier that leverages the predictive power of multiple learners through an ensemble approach, for two reasons: (i) it is capable of handling a large number of features (i.e., >1,500 genes analyzed in an algorithm with 10,000 trees in the present study), (ii) it provides intrinsic cross-validation during the testing and assessment procedure. Thus, the process of 3-fold cross-validation, intrinsic in the random forest procedure, maximized the generalizability of the defined signature. However, we recognize the limitation of the present study because we lack another independent AML dataset for complete validation of the developed gene signature. A matched clinical cohort from recent MDM2 inhibitor trials would be optimal, but only one such database exists (from the completed phase I study of RG7112; ref. 10), which is not publicly available. However, the ongoing clinical trial of DS-3032b in AML (NCT02319369) will definitely provide such a dataset in near future. One could also argue the feasibility of 1,532-gene expression profiling for clinical implementation. In this regard, we could narrow down the number of selected genes because every gene is ranked according to its predictive significance. For example, the 50 top-ranked genes used in Fig. 4F will be also tested in the trial dataset (Fig. 5A).

The cell line–based 175-gene signature and the AML-specific 1,532-gene signature share only a limited number of genes (26 genes), for at least two reasons: (i) the signatures were developed through entirely different algorithms (one selected highly expressed genes in sensitive cell lines, while the other used random forest machine learning), (ii) cell types used for gene expression profiling were different (AML specific or not). Considering that only about 60 genes were found to be common p53 target genes in different tissue types in the meta-analysis of 16 genome-wide datasets (24, 25), the 26-gene overlap in our signatures is not surprising, and may rather well reflect AML-specific characteristics. Therefore, even though validation is more advanced for the 175-gene signature, we believe that it will be important to develop these two signatures in parallel. Specifically, we plan to validate the two signatures in the ongoing clinical study of DS-3032b in AML patients (NCT02319369; Fig. 5A). Furthermore, because the tumor-type–specific approach can be applied to relatively small datasets (e.g., 33 samples as in the present study), ongoing clinical studies of DS-3032b and other MDM2 inhibitors for any cancers could additionally define improved tumor-type–specific gene signatures (Fig. 5B). Moreover, a similar tumor-type–specific approach may be potentially applied to other molecularly targeted agents, if a referenced gene set related to the agent's mechanism of action (corresponding to the 32 referenced genes for MDM2 inhibition) can be identified.

J. Ishizawa has ownership interest (including patents) in Daiichi Sankyo Co., Ltd. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J. Ishizawa, K. Nakamaru, T. Seki, K. Kojima, K.P. Patel, A. Tse, R.E. Davis, A. Rao, M. Andreeff

Development of methodology: J. Ishizawa, K. Nakamaru, T. Seki, K. Tazaki, K. Kojima, A. Rao

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Ishizawa, T. Seki, K. Kojima, D. Chachad, R. Zhao, L. Heese, C. DiNardo

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J. Ishizawa, K. Nakamaru, T. Seki, K. Tazaki, K. Kojima, R. Zhao, L. Heese, W. Ma, M.C.J. Ma, C. DiNardo, K.P. Patel, A. Tse, R.E. Davis, A. Rao, M. Andreeff

Writing, review, and/or revision of the manuscript: J. Ishizawa, K. Nakamaru, K. Tazaki, K. Kojima, K.P. Patel, A. Tse, R.E. Davis, A. Rao, M. Andreeff

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): D. Chachad, C. DiNardo, S. Pierce, A. Rao

Study supervision: K. Nakamaru, T. Seki, M. Andreeff

This work was supported in part by grants from the Japan Society for the Promotion of Science Postdoctoral Fellowship for Research Abroad Award (to J. Ishizawa); the Ministry of Education, Culture, Sports, Science and Technology in Japan (26461425), the Princess Takamatsu Cancer Research Fund (14-24610), the Osaka Cancer Research Foundation (to K. Kojima); and the National Institutes of Health USA (CA49639, 100632, CA136411, and CA16672) and the Paul and Mary Haas Chair in Genetics (to M. Andreeff). We thank Jairo A. Matthews for clinical data collection. We thank Dr. N. Hail and Mr. D.R. Norwood for editorial assistance. We also thank Kenji Watanabe for technical support.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.