Leukemia Research 76 (2019) 1–10 Contents lists available at ScienceDirect Leukemia Research journal homepage: www.elsevier.com/locate/leukres A precision medicine test predicts clinical response after idarubicin and T cytarabine induction therapy in AML patients David Martínez-Cuadróna,b, Cristina Gilc, Josefina Serranod, Gabriela Rodrígueze, Jaime Pérez-Oteyzaf, Raimundo García-Boyerog, Santiago Jiménez-Bravoh, Susana Vivesi, María Belén Vidrialesj, Esperanza Lavillak, José A. Pérez-Simónl, Mar Tormom, Mercedes Coloradon, Juan Berguao, Juan A. Lópezp, Pilar Herreraq, Pilar Hernández-Campor, Julián Gorrochateguir, Daniel Primor, Jose Luis Rojasr, Jesús Villorias, Federico Moscardóa, Iñaki Troconizt, María Linares Gómezu, Joaquín Martínez-Lópezu, Joan Ballesterosr, ⁎ Miguel Sanza,b, Pau Montesinosa,b, ,1,2, on behalf of the Spanish PETHEMA group a Hospital Universitari i Politècnic La Fe, Valencia, Spain b CIBERONC, Instituto Carlos III, Madrid, Spain c Hospital General Universitario de Alicante, Alicante, Spain d Hospital Universitario Reina Sofía, Córdoba, Spain e Hospital General Universitario Gregorio Marañón, Madrid, Spain f Hospital de Madrid Norte Sanchinarro, Madrid, Spain g Hospital Universitario General de Castellón, Castellón, Spain h Hospital Universitario de Gran Canaria Doctor Negrín, Gran Canaria, Spain i ICO-Hospital Germans Trias i Pujol, Josep Carreras Leukemia Research Institute, Universitat Autònoma de Barcelona, Badalona, Spain j Hospital General Universitario de Salamanca, Salamanca, Spain k Hospital Universitario Lucus Augusti, Lugo, Spain l Hospital Universitario Virgen del Rocío, Sevilla, Spain m Hospital Clínico Universitario, Valencia, Spain n Hospital Universitario Marqués de Valdecilla, Santander, Spain o Hospital San Pedro de Alcántara, Cáceres, Spain p Complejo Hospitalario de Jaén, Jaén, Spain q Hospital Universitario Ramón y Cajal, Madrid, Spain r Vivia Biotech, Tres Cantos, Madrid, Spain s Medicxact, Madrid, Spain t Universidad de Navarra, Pamplona, Spain u Hospital Universitario 12 de Octubre, Madrid, Spain A R T I C LE I N FO A B S T R A C T Keywords: Complete remission (CR) after induction therapy is the first treatment goal in acute myeloid leukemia (AML) Acute myeloid leukemia patients and has prognostic impact. Our purpose is to determine the correlation between the observed CR/CRi Ex vivo assay rate after idarubicin (IDA) and cytarabine (CYT) 3 + 7 induction and the leukemic chemosensitivity measured Pharmacological profile by an ex vivo test of drug activity. Bone marrow samples from adult patients with newly diagnosed AML were Complete remission included in this study. Whole bone marrow samples were incubated for 48 h in well plates containing IDA, CYT, Clinical correlation or their combination. Pharmacological response parameters were estimated using population pharmacodynamic models. Patients attaining a CR/CRi with up to two induction cycles of 3 + 7 were classified as responders and the remaining as resistant. A total of 123 patients fulfilled the inclusion criteria and were evaluable for corre- lation analyses. The strongest clinical predictors were the area under the curve of the concentration response curves of CYT and IDA. The overall accuracy achieved using MaxSpSe criteria to define positivity was 81%, predicting better responder (93%) than non-responder patients (60%). The ex vivo test provides better yet similar ⁎ Corresponding author. E-mail address: montesinos_pau@gva.es (P. Montesinos). 1 Hematology Department (Torre F, planta 7), Hospital Universitario La Fe, Avinguda Fernando Abril Martorell, 106, CP 46026 Valencia, Spain. Tel.1: +34 96 1244000. Ext 411966, Tel.2: +34 96 1245876, Fax: +34 96 1246201, Web: http://www.hematologialafe.es/. 2 CIBERONC, Instituto Carlos III, Madrid, Spain. https://doi.org/10.1016/j.leukres.2018.11.006 Received 6 June 2018; Received in revised form 29 October 2018; Accepted 12 November 2018 Available online 13 November 2018 0145-2126/ © 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/BY-NC-ND/4.0/). D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 information than cytogenetics, but can be provided before treatment representing a valuable in-time addition. After validation in an external cohort, this novel ex vivo test could be useful to select AML patients for 3 + 7 regimen vs. alternative schedules. 1. Introduction to determine the ability of the PharmaFlow PM test to predict the re- sponse to first line induction chemotherapy with CYT and IDA (3 + 7) Several clinical and biological features, like cytogenetic and mole- in young AML patients. cular alterations, may predict the short- and long-term outcomes in patients with AML treated with intensive approaches [1,2]. However, 2. Patients and methods the main prognostic factor after front-line induction treatment is the leukemic cell sensitivity to the chemotherapy itself (i.e, to achieve or 2.1. Patients and study design not a first CR as well as the quality and duration of the CR) [3,4]. In order to predict response to chemotherapy, many individualized sen- This multicenter, non-interventional prospective cohort study was sitivity and resistance assays have been deployed for detecting ex vivo carried out in 43 Spanish institutions of the PETHEMA group. The in- drug-inducible cell death. Nevertheless, as far as we know, none of clusion period lasted five years (2011–2016), enrolling patients aged 18 them have settled into the routine medical practice [5]. years and older with newly diagnosed AML (de novo or secondary to The combination of IDA and CYT (3 + 7 schedule) is widely used as myelodysplastic syndromes, or therapy-related). Patients with acute front-line regimen in younger AML patients, resulting in CR or CR with promyelocytic leukemia and those with AML derived from a chronic incomplete recovery (CRi) rates of 70–80% after one or two cycles [6]. myeloid leukemia were not screened. Patients were eligible for the Among patients who do not achieve CR/CRi after induction therapy, current analysis if they received front-line induction therapy using IDA/ categorized as primary refractory or resistant disease, the prognosis is CYT (3 + 7). All adult patients treated with other front-line regimens dismal [4]. We hypothesize that a reliable and automated ex vivo drug were excluded from the correlation analysis, but their AML samples sensitivity test accurately predicts the antileukemic efficacy of 3 + 7 were used to build pharmacodynamics (PD) population-based models. induction would be valuable to preselect this regimen as the optimal Diagnosis and classification of AML was made according to the WHO upfront chemotherapy schedule. (World Health Classification) criteria [8]. The study was conducted We present here a non-interventional prospective study assessing a according to the Spanish law 14/2007 of biomedical research, and was novel and actionable native environment precision medicine (PM) approved by the Research Ethics Board of each participating institution. method [7] (PharmaFlow platform) that could overcome some of the All patients provided informed consent. previous shortcomings of ex vivo testing. The PharmaFlow platform is a cell-based multi-color screening flow cytometry platform that in- 2.2. Chemotherapy regimen, drugs and evaluation corporates both automated sample preparation and automated flow cytometry, together with proprietary analytical software and a database Induction therapy consisted of up to two cycles of the combination that achieve rapid data acquisition, analysis, and reporting of results. of intravenous (IV) IDA (12 mg/m2/day), from days one to three, and IV This platform has the capacity to evaluate hundreds of drug combina- continuous perfusion of CYT (200 mg/m2/day), from days one to seven. tions that form the basis of treatment protocols in AML, together with A second 3 + 7 induction cycle was administered in patients showing a the fact that these drug combinations are evaluated in whole bone partial remission (PR) after the first cycle. Supportive measures were marrow retaining the biological system as intact as possible. The aim is given according to local policies of each participating institution. Fig. 1. Sequential workflow of experimental (1) and analytical (2 & 3) methods applied in the study. Whole bone marrow samples [A] were incubated preserving the native microenvironment with drugs and drugs mixtures. Automated flow cytometry [B] followed by dot-plots analysis [C] allowed the counting after incubation of Live Pathologic Cells (LPC) at control wells and wells with increasing drugs concentrations. Data was uploaded into the LIMS system. Response vs drug concentration relationships were analyzed through non-linear mixed effect population modelling [E]. Predicted pharmacodynamic profiles were integrated between the 80% confidence interval of the individual estimate of EC50 in order to calculate the area under the curve (AUC) used as a single activity marker. Similarly, a double integration of the two-variables interaction surface function allows the calculation of the volume under the surface (VUS) that is effected by the sign (synergy or antagonism) of the interaction [F]. Correlation of activity markers with clinical output was analyzed by Generalized Additive Models (GAM) [G] and ROC curves [H]. (See details on individual figures). 2 D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 Table 1 Samples were incubated with CYT, IDA and/or CYT + IDA for 48 h. A Patients’ characteristics and clinical response of the 123 patients incubated with more detailed description of the procedure has been published else- CYT and IDA. where [7]. Clinical response (ref. standard) 2.3.2. Modeling of ex vivo activity of CYT, IDA, and their combination CR/CRi PR/Res p-value Evaluation of drug response was done by counting the number of (N = 92) (N = 31) live pathological cells (LPC) remaining after incubation at increasing Age (years) drug concentrations [10]. Dying cells (apoptosis) were excluded using Median (range) 49 (19-71) 55 (22-67) 0.224 a Annexin V-FITC. Pharmacological responses were analyzed using PD population-based models [10] which essentially perform the fitting of b 18–29 [n (%)] 8 (8.7) 2 (6.5) 0.758 30–39 [n (%)] 17 (18.5) 5 (16.1) the dependent variable (natural log of LPC) in a non-linear mixed-ef- 40–49 [n (%)] 23 (25.0) 5 (16.1) 50–59 [n (%)] 22 (23.9) 10 (32.3) fects model to derive typical population values (fixed effects) and the > 60 22 (23.9) 9 (29.0) magnitude of inter-patient and residual variability (random effects). Gender Model development was performed with the first-order conditional Male [n (%)] 42 (45.7) 16 (51.6) 0.565 b estimation method using interaction option with the software Female [n (%)] 50 (54.3) 15 (48.4) NONMEM (v7.2) [11], according to the following equation: ECOG Where LPC0 parameter refers to the number of LPC after incubation 0–1 [n (%)] 69 (87.3) 27 (100.0) 0.062 c in the absence of drug, Emax represents the maximum fractional de- 2–4 [n (%)] 10 (12.7) 0 (0.0) crease in LPC that the drug can elicit, EC50, is the drug concentration FAB subtype exerting half of Emax, and γ is the parameter governing the steepness of b M0 [n (%)] 3 (3.6) 5 (17.9) 0.166 the LPC vs drug concentration (C) curve. For interaction analysis a M1 [n (%)] 19 (22.6) 4 (14.3) Surface Interaction model [12] was used to estimate the degree of sy- M2 [n (%)] 23 (27.4) 9 (32.1) M4 [n (%)] 22 (26.2) 5 (17.9) nergy, referred as α parameter, between both drugs. M5 [n (%)] 16 (19.0) 5 (17.9) Interpatient variability (IPV) associated to all parameters was de- M6 [n (%)] 1 (1.2) 0 (0.0) scribed by means of an exponential model of the components of var- WBC (count × 109L−1) iance. An additive error structure was used for the residual variability. Median (range) 22.1 (0-288.4) 22.3 (1-157) 0.461 a Population PD models were built with BM samples from 473 patients b 0–10 [n (%)] 31 (33.7) 12 (38.7) 0.528 that were incubated with CYT, 456 with IDA and 443 with CYT + IDA. 10–50 [n (%)] 37 (40.2) 14 (45.2) Bayesian estimation methods were then used to retrieve individual > 50 [n (%)] 24 (26.1) 5 (16.1) patient parameters based on their available exposure-response mea- Cytogenetic risk profile surements in conjunction with the PD population parameters. b Favorable [n (%)] 11 (13.1) 0 (0.0) < 0.001 Intermediate [n (%)] 68 (81.0) 11 (40.7) Evaluation of the population PD models was done by the simulation- Adverse [n (%)] 5 (6.0) 16 (59.3) based procedure visual predictive check [13]. Five hundred experi- mental scenarios equal to the original ones were simulated using the FLT3-ITD status Wild type [n (%)] 72 (79.1) 28 (90.3) 0.161 b selected models and the corresponding parameters. In each simulated Mutant [n (%)] 19 (20.9) 3 (9.7) set and for each concentration lever the 2.5, 50, and 97.5th percentiles NPM1 status of the LPC distribution were calculated, then the 95% confidence in- Wild type [n (%)] 47 (57.3) 22 (78.6) 0.045 b tervals for the above mentioned percentiles were computed and re- Mutant [n (%)] 35 (42.7) 6 (21.4) presented graphically together with the 2.5, 50, and 97.5th percentiles obtained from the raw data. Patients with missing data have not been included in the denominators of the relative frequencies. a 2.3.3. Probability of clinical outcome modeling clinical correlation Mann-Whitney test. b Pearson’s chi-square test. Individual response profiles normalized with respect to LPC0, were c Fisher’s exact test. integrated between the concentration points corresponding to the 20th and 80th percentiles of the distribution of estimated individual EC50 Response to induction chemotherapy was assessed according to the values, to obtain the values of the areas under the curves (AUCs) that revised Cheson criteria [9]. Patients were considered as responders if were used as a descriptor of the ex vivo drug effect (i.e., the higher the they achieved CR or CRi within the first two identical 3 + 7 induction AUC, the lower the cytotoxic effect (efficacy or potency) of the drug. cycles. Patients dying during induction before response assessment The individual AUC values were correlated to the actual patient´s were considered as non-evaluable. The remaining patients were clas- response after induction therapy (non-responder [PR or resistant dis- sified as resistant. ease] vs. responder [CR or CRi]). The probability of being non-re- sponder was modeled using binary logistic generalized additive models (GAM) based on the binomial distribution that included either one bi- 2.3. Vivia’s PharmaFlow PM test or two univariate smooth functions of the AUCs of CYT and IDA. Additionally, univariate smooth functions of the LPC0, α, and the pre- A representative workflow of the PharmaFlow PM test is shown in post incubation difference of the percentage of LPC in control wells (to Fig. 1, collecting the experimental and analytical methods applied in detect for any possible effect of spontaneous cell death) were included this study. as well, but discarded afterwards because they were not related to the clinical response. Also, the predictive ability of relevant patients’ 2.3.1. Native environment whole bone marrow sample characteristics (age and sex, presenting leukocyte count, performance Ex vivo drug sensitivity analysis was made using the PharmaFlow status, mutations in the NPM1 or FLT3 genes, and cytogenetic risk platform (previously termed ExviTech®) [7] maintaining the bone group) on top of pharmacodynamics data was explored by introducing marrow (BM) microenvironment. A minimum BM sample volume be- them as parametric model terms in auxiliary GAMs. P-spline bases were tween one and two ml was collected by aspiration at AML diagnosis, used as smoothers for univariate smooth functions; tensor products of before starting induction chemotherapy, and was processed by an au- univariate P-spline smooths were used for constructing bi-variate tomated method in Vivia Biotech laboratories 24 h after extraction. smooth functions. All smoothing bases had dimension three. 3 D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 Fig. 2. Visual predictive check of the population pharmacokinetic models of cytarabine and idarubicin. Open circles are the observed data points, the solid and dashed red lines are, respectively, the median and the 5-95th percentiles of the observed distribution of ln(cells), and the semitransparent red and blue bands represent, respectively, the simulation-based 95% confidence intervals for the median and 5-95th percentiles of the estimated population distribution of ln(cells). Coefficients of the smooth functions were estimated using penalized 2.4. Data collection and study endpoints iteratively re-weighted least squares. Minima of the scaled Akaike in- formation criteria were used to find the optimal values of the smoothing Demographic data (gender, age) were prospectively collected since parameters. diagnosis, as well as the following parameters: WBC in PB, ECOG per- formance status, type of AML (FAB classification, de novo vs. secondary AML), karyotype [14], FLT3 and NPM1 mutation status, hematological 4 D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 Table 2 survival (OS) probability was also calculated according to the observed Estimation of the ex vivo population pharmacodynamic parameters. and predicted response after induction. Parameters typical and random (variability and residual error percentage) Since the prediction of outcome to front-line induction therapy is are shown together with the corresponding relative standard error calculated likely to be most beneficial to elderly patients, we performed sensitivity as the ratio between the standard error provided by NONMEM and the es- analyses of the clinical correlation by re-running the GAMs within the timate. Estimates of inter-patient variability (IPV) are expressed as coeffi- cohort aged 60 years or more (n = 31). cient of variation (%). Single Drugs 2.5. Statistical analyses Parameter (units) Cytarabine Idarubicin The probability of response modeling was performed with the mgcv LPC0 (cells) 7530 (4.2) 7270 (4.8) package (v1.8–23) run in the R environment (v3.4.3) for statistical EMAX (unitless) 1 (-) 1 (-) computing [15]. The empirical ROC curves were calculated for the EC50 (μM) 6.94 (13.3) 0.087 (9·2) probabilities of being non-responder from each GAM. The AUCs of ROC γ(unitless) 0.684 (-) 1.14 (-) Residual Error (log(μM)) 0.231 (2.9) 0.237 (3.2) curves were computed using the trapezoidal rule. In addition, three cut- Inter-patient variability (IPV) points to define positivity and derive classification probabilities (sen- LPC0 89.7 (2.4) 92.7 (2.6) sitivity and specificity) were established for each ROC curve. One used EMAX N/D N/D a geometric criterion, by selecting the closest point to the (1,0) co- EC50 229 (4.4) 159 (4.5) ordinate (left upper corner of the [sensitivity,1−specificity] plane), γ N/D N/D Residual Error 50.5 (3.8) 45.8 (4) another was set by maximizing both sensitivity and specificity (MaxSpSe) and the other by minimizing a misclassification cost term Drug Combination (mMCT) [16], assigning a greater cost to false positives than to false Parameter (units) Cytarabine + Idarubicin negatives (prioritizing specificity over sensitivity). The OS was described with the Kaplan-Meier method and compared α (unitless) 1.1 (13) IPV α [CV(%)] 176 (4·8) between the patients predicted to be non-responder and responder as Residual Error [log(μM)] 0.299 (4.2) per the aforementioned three different cut-points using simple Cox re- IPV Residual Error 58.3 (4.4) gression. 3. Results 3.1. Patient characteristics Overall, 954 BM samples from patients with AML suspicion were received at the laboratory. Of them, 316 (33%) were not evaluable because of the following laboratory technical issues: 1) low sample cellularity (187 patients), 2) low cell viability (below 60%) in control wells after incubation (67), 3) insufficient sample volume (< 500 μL) (38), and 4) other reasons such as clotted sample (24). Other 26 pa- tients (3%) did not fulfill the diagnosis criteria. Among the 612 ana- lyzed samples, 139 where used only for assay adjustment and did not contain necessary data for the final model. Overall, 473 patients sam- ples (50%) were used to build the PD models, and a complete data set was monitored in 237 of them (50%). Among the monitored patients, 114 were not evaluable for the correlation analyses due to: 1) induction death (20 patients), 2) not first line of treatment (11), and 3) other induction schedule (83). Finally, 123 monitored patients (52%) fulfilled the inclusion criteria defined in the study and were evaluable for the correlation analyses (see CONSORT diagram in supplementary Fig. 1). The main patient and disease characteristics of these 123 patients re- garding the clinical response are displayed in Table 1. In summary, median age was 50 years (range, 19 to 71), 109 patients (89%) were diagnosed with de novo AML, and 21 patients (17%) were categorized Fig. 3. Regression hyperplane of the predicted probability of resistance as having high-risk cytogenetics. Only the cytogenetic risk group and, over the AUCs of cytarabine and idarubicin. The AUCs are a summary of marginally, the presence of mutations in the NPM1 gene were sig- pharmacodynamic parameters such that the higher the AUC the lower the cy- nificantly associated with clinical response to induction and the result totoxic effect (efficacy or potency) of the drug. The regression hyperplane has of the PM test. Post-remission therapy consisted of allogeneic stem cell been obtained using bi-dimensional smooth functions in a binary logistic GAM. transplant (SCT) in 33 patients (27%), and chemotherapy with or AUC: area under the curve, GAM: generalized additive model. without autologous SCT in 66 patients (54%). response, number of 3 + 7 induction cycles, date of response, date of 3.2. Ex vivo PharmaFlow test characterization of CYT-IDA combination last follow-up, and post-remission therapy. All data collection forms and clinical records were monitored. Visual predictive checks graphs were generated for the single drugs The primary end-point was evaluate the predictive capacity of the PD models (Fig. 2). Most of the observations were contained within the ex vivo results. First, the CR/CRi rate observed in patients treated with simulation-based 95% confidence intervals of the 2.5–97.5 th popula- up to two induction cycles of 3 + 7 was recorded and monitored. This tion percentiles proving good predictability of the selected models. was correlated with the ex vivo drug sensitivity analyses performed in Pharmacodynamic population parameters as well as variability and the same cohort of patients. As a secondary end-point, the overall error values are shown in Table 2. The typical parameter values for 5 D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 Fig. 4. Empirical and smoothed (binormal) ROC curves of the probability of resistance obtained in the binary logistic GAM. A. - Open circles are the pairs of sensitivity and 1−specificity values at the estimated discrete individual values of the probability of resistance (used as a marker to classify the patients as responder or resistant), the solid large circles represent the pairs of sensitivity and 1−specificity values at the selected cut-points that were obtained with each of the three criteria specified in the text: ‘MaxSpSe’ selects the point that maximizes both, the sensitivity and the specificity; ‘Geometric’ selects the closest point to the (1,0) coordinate (left upper corner of the [sensitivity,1−specificity] plane); and ‘mMCT’ selects the point that minimizes a misclassification cost term that assigned a greater cost to false positives than to false negatives (prioritizes specificity over sensitivity). AUC: area under the curve, CI: confidence interval, FPF: false positive fraction, NPV: negative predictive value, PPV: positive predictive value, Se: sensitivity, Sp: specificity, TPF: true positive fraction. B.- Confusion matrix for MaxSpSe cutoff. maximal fractional effect (Emax) were set to 1 for both drugs, and was (29.4%) than the GAM over the AUC values (40.8% in the subset of 111 limited to the range 0-3. The typical value for the alpha parameter of patients in whom the cytogenetic risk was informed) (Fig. 6). Fur- the interaction model was 1.1 (Table 2), indicating slight synergistic thermore, the model with the pharmacodynamics data explained sig- interaction between IDA and CYT in the ex vivo combination experi- nificantly more variability than the cytogenetic information alone ments. (p = 0.001, details available or request). The sensitivity analyses showed that the predictive ability of the PM 3.3. Clinical responses among AML patients treated with CYT-IDA test remained intact within the cohort aged ≥60 years (Supplemental Fig. 3), although in this case most of the discriminative information was CR/CRi was achieved after one (88, 96%) or two (4, 4%) identical provided by CYT data; IDA AUC values were in general higher in older induction cycles in 92 out of 123 patients (75%) included in the cor- patients. relation study. 3.5. OS according to the ex vivo activity and observed clinical response 3.4. Correlation between ex vivo activity and clinical response to CYT-IDA The OS was significantly shorter in patients predicted to be non- Fig. 3 depicts the predicted surface fitted by the GAM representing responders than in patients predicted to be responders regardless of the the probability of being non-responder for the observed range of the cut-point used to classify them. The median OS among patients pre- individual AUC values. The model presented used a bivariate smooth dicted to be non-responders ranged from 344 to 589 days (Fig. 5). It function of CYT and IDA; the models that used univariate smooths (see was not reached in patients predicted to be responders. The hazard Supplementary Fig. 2) achieved worse fit. Higher CYT and IDA AUC ratios (HR) of death (patients predicted to be non-responders vs. re- values were associated with greater probability of being non-responder, sponders) ranged from 2.46 (1.38–4.36) to 3.44 (1.88–6.28). The values albeit the relationship was non-monotonical. Sensitivity/specificity for the groups defined by actual clinical response were similar (median values ranged from 81% / 82%, to 61% / 95%, based on the cut-point OS among non-responders: 279 days; HR [resistant vs. CR/CRi]: 3.17). selected (Fig. 4A). The positive/negative predictive values (PPV/NPV) ranged from 60% / 93% to 79% / 88%. Fig. 4B shows the confusion 4. Discussion matrix obtained using the MaxSpSe cutoff point. In the auxiliary models, only the cytogenetic risk group provided This study shows good correlation between the hematological re- independent additional predictive information on top of CYT and IDA sponse to IDA/CYT 3 + 7 induction and the observed leukemic che- pharmacodynamics, the latter remaining significantly associated with mosensitivity measured by a novel approach to ex vivo testing of drug response no matter of patients’ clinical characteristics. A simple logistic activity. The statistical model that was built using the PharmaFlow PM model of the probability of being non-responder over the cytogenetic test showed a good correlation (81%) with hematological responses risk group (favorable/intermediate/adverse) explained less variability observed in patients. A 93% of patients predicted sensitive indeed 6 D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 upfront clinical setting. Several studies have analyzed the correlation between ex vivo drug testing and clinical outcomes in adult AML patients by different methods [17–23], but none was adopted in the routine clinical practice. Our novel approach was designed to overcome some of the ex vivo testing limitations of the last 30 years. Thus, the PharmaFlow PM test has several relevant aspects differing from previously reported assays: 1) contrarily to others using ficol isolated leukemic cells, we used whole BM samples diluted to keep, at least in part, the native environment. We have shown this prevents major artifacts on drug activity [7]. As the tumor microenvironment has been implicated in the drug-resistance mechanisms [24–27], we can hypothesize that using whole BM samples could lead to more reliable results; 2) we employed PD models to analyze the data albeit the correlations were performed in relatively small number of patients. This method increases accuracy by fitting dose-response population data in one single step [28]; and 3) we used a proprietary automated flow cytometry platform (PharmaFlow) that could lower experimental errors, providing accurate data from patient samples. Our study methodology has several limitations that deserve some comments: 1) the study was designed to create a predictive model for the clinical response in terms of CR/CRi vs. PR/resistance, using the classical criteria, whilst no systematic/centralized minimal residual disease assessment was performed. We can speculate that a more pre- cise response assessment by flow cytometry could improve the accuracy of our observations; 2) the PharmaFlow PM test analyzed the main baseline leukemic population, not facing up the presence of phenoty- pically different subclones before and after induction; 3) the test was not predicting the actual overall response (it was not aimed to predict induction death); 4) the test was not evaluable in a sizable proportion of patients mainly due to low cellularity or low viability samples; 5) al- though the incubation time was relatively short, additional transpor- tation and processing time could lead, in some patients, to a non-af- fordable delay to start induction chemotherapy while receiving the test report; and 6) our results are not yet validated in an independent co- hort. Finally, we should speculate about some potential causes about why the PharmaFlow could not predict some selected cases: 1) the test analyzed the main leukemic population, not facing up the presence or marginal resistant subclones, 2) the test was not reproducing the leu- kemic stem-cell niche who could have an in vivo protective effect against chemotherapy agents, 3) the test was not taking into account some pharmacodynamic and pharmacokinetic parameters which may influence the actual response to chemotherapy (e.g., drug-drug inter- actions during induction, pharmacogenetic polymorphisms); and 4) technical issues/limitations of the test. To identify the ex vivo variables associated to hematological re- sponse, sophisticated regression analyses were performed. The AUC of the concentration-response curves yielded a useful summary of the pharmacodynamic parameters for the purposes of predicting the clin- ical response. Both drugs provided meaningful and independent in- formation, as can be observed in predicted surface of being non-re- Fig. 5. Kaplan-Meier plots of overall survival. The three panels deptict the sponder shown in Fig. 3; When IDA is inactive at maximal IDA AUC, the pairs of survival functions for patients classified as responder (solid black lines) AUC of CYT leads the prediction with a classical dose-response curve and resistant (solid red lines) according to the cut-points of the estimated (red). Conversely, When CYT is inactive IDA leads the prediction with a probability of resistance that were obtained with each of the three criteria classical dose response curve (blue). When one of the drugs is very specified in the text (panel A: ‘MaxSpSe’, panel B: ‘Geometric’, and panel C: active (low AUC) the other shows a more limited effect, still consistent ‘mMCT’). The dashed lines represent the survival functions of clinical re- with higher AUC corresponding to higher probability of resistance. This sponders (black lines) and resistant patients (red lines). The hazard ratios of death were obtained from a Cox regression model that used the patients who behavior is coherent and consistent with the expectations; higher AUC were predicted to be responder as the reference category (patients predicted to of either drug, i.e. lower activity, implies higher probability of re- be resistant over patients predicted to be responder). sistance. CI: confidence interval, HR: hazard ratio Notably, the observed interaction between both drugs was not highly synergistic, and did not independently correlate with hemato- achieved CR/CRi, potentially allowing for better selection of the front- logical response. Contrarily to other reports [29], our data suggest that line chemotherapy schedule. However, a validation study in an external the effectiveness of the CYT-IDA combination may rely in their com- cohort must be performed before using the PharmaFlow PM test in the plementary activities, and marginally on their synergism. Diagnostic accuracy of a predictive test is usually assessed by the 7 D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 Fig. 6. Comparison between clinical correlation of cytogenetics (left) and PM Test (right), on a cohort of 111 patients sharing both results. ROC curves (top) and confusion matrices for MaxSpSe cutoff (down). Deviance explained is 29.4% for cytogenetics, and 40.9% for PM test. maximum sensitivity and specificity values achieved in ROC curves The information about the cytogenetic risk was less predictive than (Fig. 4A, “maxSpSe”). However, we considered an alternative approach the pharmacologic parameters; Deviance explained 29.4% for cytoge- to maximize the ability to identify patients as non-responder (i.e, PPV) netics, and 40.8% for PM test. Nonetheless, cytogenetics still proved to or responder (i.e, NPV), while prioritizing specificity over sensitivity be of prognostic relevance in these patients. Fig. 6 compares clinical (“mMCT”) to account for the high clinical cost of false positives (de- correlation of cytogenetics vs PM Test, on a cohort of 111 patients nying potentially effective 3 + 7 schedule). Although the geometric sharing both results. In both approaches prediction of sensitive patients cutoff point balanced both aspects, the final selected cutoff point in is better than resistant patients (NPV vs PPV). While PM Test has better ROC curve was the MaxSpSe to construct the confusion matrix and prediction of sensitive patients (NPV = 93% vs 88%), cytogenetics derive classification probabilities, achieving high values for both spe- shows a 20% improvement in the prediction of resistant patients. cificity and sensitivity, as well as good PPV and NPV (Fig. 4B). The However, results from cytogenetic risk are available typically in 10–14 accuracy achieved by the PharmaFlow PM test using the aforemen- days and thus after patient treatment, while results from this novel ex tioned MaxSpSe criteria to define positivity was 81%, predicting better vivo PM Test are available in 48–72 h prior to patient treatment. It responder (93%, NPV) than non-responder patients (60%, PPV). We can should be noted that the model showed an acceptable performance speculate that, in order to improve first CR/CRi rates, patients predicted among patients older than 60 years old (which are at higher risk of as non-responder to CYT-IDA could be proposed for alternative induc- treatment related death), and retained its independent predictive value tion schedules or front-line clinical trials. among the adverse cytogenetic subgroup of patients (which are fre- The estimated OS was significantly better in patients predicted to be quently offered non-intensive options because of high resistance rate). responder. This is not surprising, because PharmaFlow PM test accu- We can hypothesize that the PM test could be valuable for selecting rately predicted the response, and achieving a first CR leads to im- candidates to IDA and CYT 3 + 7 among these critical subsets. proved OS after 1st line in AML [30]. Interestingly, the OS probability In conclusion, our novel approach to ex vivo testing using the was similarly discriminated by the test result (responder vs. non-re- PharmaFlow PM platform provided drug sensitivity parameters that sponder) compared with the actual clinical response (CR/CRi vs. re- were integrated in a flexible generalized additive logistic regression sistance). model with an outstanding predictive accuracy for hematological 8 D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 response after front-line induction with CYT-IDA 3 + 7. After validation relapsed acute myeloid leukemia, Blood 126 (2015) 319–327. in an external cohort, our diagnostic tool could be useful to select AML [5] D. Schrag, H.S. Garewal, H.J. Burstein, D.J. Samson, D.D. Von Hoff, M.R. Somerfield, American Society of Clinical Oncology Technology Assessment: patients for 3 + 7 regimen vs. alternative schedules. The PETHEMA chemotherapy sensitivity and resistance assays, J. Clin. Oncol. 22 (2004) AML group is launching a clinical trial to confirm the predictive value 3631–3638. of this new PM Test. [6] M. Low, D. Lee, J. Coutsouvelis, S. Patil, S. Opat, P. Walker, A. Schwarer, H. Salem, S. Avery, A. Spencer, A. Wei, High-dose cytarabine (24 g/m2) in combination with idarubicin (HiDAC-3) results in high first-cycle response with limited gastro- Acknowledgements intestinal toxicity in adult acute myeloid leukaemia, Intern. Med. J. 43 (2013) 294–297. [7] T.A. Bennett, P. Montesinos, F. Moscardo, D. Martinez-Cuadron, J. Martinez, This study was supported by grants from the Spanish Ministry of J. Sierra, R. Garcia, J.P. de Oteyza, P. Fernandez, J. Serrano, A. Fernandez, Economy, Industry and Competitiveness, Enisa (Empresa Nacional de P. Herrera, A. Gonzalez, C. Bethancourt, G. Rodriguez-Macias, A. Alonso, J.A. Vera, Innovación, S.A.) and Soprea (Sociedad para la Promoción y B. Navas, E. Lavilla, J.A. Lopez, S. Jimenez, A. Simiele, B. Vidriales, B.J. Gonzalez, C. Burgaleta, J.A. Hernandez Rivas, R.C. Mascunano, G. Bautista, J.A. Perez Simon, Reconversión Económica de Andalucía, S.A.U.). L. Fuente Ade, C. Rayon, I.F. Troconiz, A. Janda, A.G. Bosanquet, P. Hernandez- DMC., JML, JB and PM designed the study, supervised research, and Campo, D. Primo, R. Lopez, B. Liebana, J.L. Rojas, J. Gorrochategui, M.A. Sanz, wrote the paper; CG, JS, GR, JPO, RGB, SJB, SV, MBV, EL, JAPS, MT, J. Ballesteros, Pharmacological profiles of acute myeloid leukemia treatments in MC, JB, JAL, PH and MLG contributed samples/patient data and pro- patient samples by automated flow cytometry: a bridge to individualized medicine, Clin. Lymphoma Myeloma Leuk. 14 (2014) 305–318. vided critical input; PHC and DP supervised research, analyzed the data [8] J.W. Vardiman, N.L. Harris, R.D. Brunning, The World Health Organization (WHO) and provided critical input; JG, JLR, JV and IT performed the statistical classification of the myeloid neoplasms, Blood. 100 (2002) 2292–2302. data analysis, supervised research and provided critical input; FM and [9] B.D. Cheson, J.M. Bennett, K.J. Kopecky, T. Buchner, C.L. Willman, E.H. Estey, C.A. Schiffer, H. Doehner, M.S. Tallman, T.A. Lister, F. Lo-Coco, R. Willemze, MS design the study and supervised research. All authors contributed to A. Biondi, W. Hiddemann, R.A. Larson, B. Lowenberg, M.A. Sanz, D.R. Head, the preparation of the manuscript and approved the submission in its R. Ohno, C.D. Bloomfield, Revised recommendations of the international working current form. group for diagnosis, standardization of response criteria, treatment outcomes, and reporting standards for therapeutic trials in acute myeloid leukemia, J. Clin. Oncol. We would like to thank the following investigators and institutions 21 (2013) 4642–4649. from the Spanish PETHEMA group for their support and collaboration: [10] R.N. Upton, D.R. Mould, Basic concepts in population modeling, simulation, and Dr. Carmen Burgaleta (Hospital Príncipe de Asturias, Alcalá de Henares, model-based drug development: part 3-introduction to pharmacodynamic modeling methods, CPT Pharmacometrics Syst. Pharmacol. 3 (2014) e88. Madrid), Dr. Mª Ángeles Fernández (Hospital Xeral Cies, Vigo), Dr. Juan [11] S.L. Beal, L.B. Sheiner, A.J. Boeckmann, R.J. Bauer, NONMEM Users Guides. Icon Antonio Vera (Hospital Virgen de la Macarena, Seville), Dr. Arancha Development Solutions, Ellicot City, Maryland, 1989–2001. Alonso (Hospital Universitario Quirón, Madrid), Dr. Jordi Sierra [12] W.R. Greco, G. Bravo, J.C. Parsons, The search for synergy: a critical review from a response surface perspective, Pharmacol. Rev. 47 (1995) 331–385. (Hospital de la Santa Creu y Sant Pau, Barcelona), Dr. Ataulfo González [13] M. Bergstrand, A.C. Hooker, J.E. Wallin, M.O. Karlsson, Prediction-corrected visual (Hospital Clínico San Carlos, Madrid), Dr. Begoña Navas (Hospital predictive checks for diagnosing nonlinear mixed-effects models, AAPS J. 13 (2011) Universitario Moncloa, Madrid), Dr Esperanza Lavilla Rubira (Hospital 143–151. Lucus Augusti, Lugo), Dr. Adriana Simiele (Hospital Povisa, [14] D. Grimwade, R.K. Hills, A.V. Moorman, H. Walker, S. Chatters, A.H. Goldstone, K. Wheatley, C.J. Harrison, A.K. Burnett, Refinement of cytogenetic classification in Pontevedra), Dr. Concepción Bethancourt (Hospital Regional acute myeloid leukemia: determination of prognostic significance of rare recurring Universitario de Málaga Carlos Haya, Málaga), Dr. Guiomar Bautista chromosomal abnormalities among 5876 younger adult patients treated in the (Hospital Universitario Puerta de Hierro, Majadahonda, Madrid), Dr. United Kingdom Medical Research Council trials, Blood 116 (2010) 354–365. [15] S.N. Wood, Generalized Additive Models. An Introduction with R. Chapman & Hall, Jose Ángel Hernández Rivas (Hospital Infanta Leonor, Madrid), Dr. CRC, Boca Raton, Florida, 2006. Bernando J. González (Hospital Universitario de Canarias, Santa Cruz [16] M. Greiner, Two-graph receiver operating characteristic (TG-ROC): update version de Tenerife, Canarias), Dr. Mª Asunción Mora Casado (Hospital Infanta supports optimisation of cut-off values that minimise overall misclassification costs, J. Immunol. Methods 191 (1996) 93–94. Sofía, Madrid), Dr. Adolfo de la Fuente (MD Anderson Cancer Center, [17] A.G. Bosanquet, P. Nygren, L.M. Weisenthal, Individualized tumor response testing Madrid), Dr. Carlos Javier Cerveró (Hospital Virgen de la Luz, Cuenca), in leukemia and lymphoma, in: G.J. Kaspers, B. Coiffier, M.C. Heinrich, E.H. Estey Dr. Consolación Rayón (Hospital Universitario Central de Asturias, (Eds.), Innovative Leukemia and Lymphoma Therapy, Informa Healthcare, New York, 2008, pp. 23–44. Oviedo), Dr. Teresa Olavé (Hospital Clínico Universitario Lozano Blesa, [18] K.R. Jun, S. Jang, H.S. Chi, K.H. Lee, J.H. Lee, S.J. Choi, J.J. Seo, H.N. Moon, Zaragoza), Dr. Aurelio López (Hospital Arnau de Vilanova, Valencia), H.J. Im, C.J. Park, Relationship between in vitro chemosensitivity assessed with Dr. Olga Salamero (Hospital Universitari Vall d'Hebron), Dr. Edelmira MTT assay and clinical outcomes in 103 patients with acute leukemia, Korean J. Lab. Med. 27 (2007) 89–95. Martí (Hospital de Manises, Valencia), Dr. Juan Muñoz (Hospital Puerta [19] J.M. Norgaard, S.T. Langkjer, T. Palshof, B. Pedersen, P. Hokland, Pretreatment del Mar, Cadiz), Dr. Guillermo Debén Ariznavarreta (Hospital leukaemia cell drug resistance is correlated to clinical outcome in acute myeloid Universitario de A Coruña, A Coruña), Dr. Fernando Ramos (Hospital leukaemia, Eur. J. Haematol. 66 (2001) (2001) 160–167. Universitario de León, León), Dr. María Lourdes Amador (Hospital de [20] T. Pemovska, M. Kontro, B. Yadav, H. Edgren, S. Eldfors, A. Szwajda, H. Almusa, M.M. Bespalov, P. Ellonen, E. Elonen, B.T. Gjertsen, R. Karjalainen, E. Kulesskiy, Montecelo), Dr. Manuel Mateo Pérez (Complejo Hospitalario S. Lagstrom, A. Lehto, M. Lepisto, T. Lundan, M.M. Majumder, J.M. Marti, Universitario de Santiago, Santiago de Compostela, A Coruña) and Dr. P. Mattila, A. Murumagi, S. Mustjoki, A. Palva, A. Parsons, T. Pirttinen, M.E. Ramet, Ana Kerguelen (Hospital Universitario la Paz, Madrid). We also thank M. Suvela, L. Turunen, I. Vastrik, M. Wolf, J. Knowles, T. Aittokallio, C.A. Heckman, K. Porkka, O. Kallioniemi, K. Wennerberg, Individualized systems medicine strategy the patients for providing their samples. to tailor treatments for patients with chemorefractory acute myeloid leukemia, Cancer Discov. 3 (2013) 1416–1429. Appendix A. Supplementary data [21] W.E. Pierceall, S.M. Kornblau, N.E. Carlson, X. Huang, N. Blake, R. Lena, M. Elashoff, M. Konopleva, M.H. Cardone, M. Andreeff, BH3 profiling discriminates response to cytarabine-based treatment of acute myelogenous leukemia, Mol. Supplementary material related to this article can be found, in the Cancer Ther. 12 (2013) 2940–2949. online version, at doi:https://doi.org/10.1016/j.leukres.2018.11.006. [22] P. Staib, E. Staltmeier, K. Neurohr, O. Cornely, M. Reiser, T. Schinkothe, Prediction of individual response to chemotherapy in patients with acute myeloid leukaemia using the chemosensitivity index Ci, Br. J. Haematol. 128 (2005) 783–791. References [23] S. Yamada, T. Hongo, S. Okada, C. Watanabe, Y. Fujii, T. Ohzeki, Clinical relevance of in vitro chemoresistance in childhood acute myeloid leukemia, Leukemia. 15 (2001) 1892–1897. [1] E.H. Estey, Acute myeloid leukemia: 2013 update on risk-stratification and man- [24] Z.W. Li, W.S. Dalton, Tumor microenvironment and drug resistance in hematologic agement, Am. J. Hematol. 88 (2013) 318–327. malignancies, Blood Rev. 20 (2006) 333–342. [2] M.A. Sanders, P.J. Valk, The evolving molecular genetic landscape in acute myeloid [25] E.A. Sison, P. Brown, The bone marrow microenvironment and leukemia: biology leukaemia, Curr. Opin. Hematol. 20 (2013) 79–85. and therapeutic targeting, Expert Rev. Hematol. 4 (2011) 271–283. [3] J.P. Sasine, G.J. Schiller, Emerging strategies for high-risk and relapsed/refractory [26] Y. Tabe, M. Konopleva, Role of microenvironment in resistance to therapy in AML, acute myeloid leukemia: novel agents and approaches currently in clinical trials, Curr. Hematol. Malig. Rep. 10 (2015) 96–103. Blood Rev. 29 (2015) 1–9. [27] H. Zahreddine, K.L. Borden, Mechanisms and insights into drug resistance in cancer, [4] F. Thol, R.F. Schlenk, M. Heuser, A. Ganser, How I treat refractory and early Front. Pharmacol. 4 (2013) 28. 9 D. Martínez-Cuadrón et al. Leukemia Research 76 (2019) 1–10 [28] A. Quartino, M.O. Karlsson, A. Freijs, N. Jonsson, P. Nygren, J. Kristensen, E.J. Feldman, Phase 2 trial of CPX-351, a fixed 5:1 molar ratio of cytarabine/dau- E. Lindhagen, R. Larsson, Modeling of in vitro drug activity and prediction of norubicin, vs cytarabine/daunorubicin in older adults with untreated AML, Blood clinical outcome in acute myeloid leukemia, J. Clin. Pharmacol. 47 (2007) 123 (2014) 3239–3246. 1014–1021. [30] D.K. Webb, Management of relapsed acute myeloid leukaemia, Br. J. Haematol. 106 [29] J.E. Lancet, J.E. Cortes, D.E. Hogge, M.S. Tallman, T.J. Kovacsovics, L.E. Damon, (1999) 851–859. R. Komrokji, S.R. Solomon, J.E. Kolitz, M. Cooper, A.M. Yeager, A.C. Louie, 10