Prediction model for drug response of acute myeloid leukemia patients

Abstract

Despite some encouraging successes, predicting the therapy response of acute myeloid leukemia (AML) patients remains highly challenging due to tumor heterogeneity. Here we aim to develop and validate MDREAM, a robust ensemble-based prediction model for drug response in AML based on an integration of omics data, including mutations and gene expression, and large-scale drug testing. Briefly, MDREAM is first trained in the BeatAML cohort (n = 278), and then validated in the BeatAML (n = 183) and two external cohorts, including a Swedish AML cohort (n = 45) and a relapsed/refractory acute leukemia cohort (n = 12). The final prediction is based on 122 ensemble models, each corresponding to a drug. A confidence score metric is used to convey the uncertainty of predictions; among predictions with a confidence score >0.75, the validated proportion of good responders is 77%. The Spearman correlations between the predicted and the observed drug response are 0.68 (95% CI: [0.64, 0.68]) in the BeatAML validation set, -0.49 (95% CI: [-0.53, -0.44]) in the Swedish cohort and 0.59 (95% CI: [0.51, 0.67]) in the relapsed/refractory cohort. A web-based implementation of MDREAM is publicly available at https://www.meb.ki.se/shiny/truvu/MDREAM/ .

Fig. 1. Overview of MDREAM.

a Omics data of the BeatAML cohort are used to train and validate the performance of the model. We further use the Clinseq cohort and the LeeAML cohort as two external validation datasets. b The ensemble-based prediction model uses the stacking approach including two layers: (1) the first layer contains 122 base models for prediction (D₁, D₂,..., D_n), each is built from the data of a single drug; (2) the second layer consists of 122 ensemble models (E₁, E₂,..., E_n) corresponding to 122 drugs. Each ensemble model is built based on the prediction output of the first layer. An advantage of ensemble learning is that it can exploit the correlation of drug sensitivity between drugs. c Comparison of performance between the base models (x-axis) and the ensemble models (y-axis) in the BeatAML validation set using the correlation between the predicted and observed response of individual drugs. Orange points above the diagonal line indicate higher correlation—better performance—achieved by the ensemble models. d Prediction results with information on confidence score for a new patient sample from the external validation dataset. This plot shows the results of the patient from the ClinSeq cohort with a median Pearson correlation between predicted AUC and observed drug sensitivity. The drugs with low AUC and high confidence (confidence score >0.75) might be of interest for further investigation in the treatment of that patient. e, f Waterfall plots of the predicted AUC and observed drug sensitivities of the patient with median Pearson correlation from the BeatAML validation set (e) and the Clinseq dataset (f). Note that the negative correlation in f panel is as expected since the AUC and drug sensitivity score (DSS) have opposite directions: lower AUC and higher DSS correspond to better drug response.

Fig. 2. Performance of MDREAM in BeatAML, Clinseq, and LeeAML.

a Boxplots show the correlation of individual drugs from the tenfold cross-validation in the training set. Each boxplot displays the interquartile range (IQR) between the 25th percentile (the lower boundary) and the 75th percentile (the upper boundary). The center line of the box presents the median, and the whiskers are within the 1.5 IQR value. The points (blue or red) in each boxplot represent the correlation results of venetoclax and trametinib. b Prediction performances of the BeatAML testing set (x-axis) versus the Clinseq dataset (y-axis) across 76 overlapping drugs using the correlation between predicted values and observed values. The drugs at the bottom right indicate the drugs with good predictions in both cohorts. c–e Predicted AUCs vs observed AUCs of all data points (c) and individual drugs, including Trametinib (d) and Venetoclax (e) from the BeatAML testing set. f–k Similar scatter plots for predicted AUCs vs observed DSSs of the Clinseq dataset and LeeAML dataset, respectively.

Fig. 3. Confidence score for drug response prediction.

a Relationship between AUC and IC₅₀ for the drug dovitinib in the BeatAML dataset; this is used to determine the threshold of good response based on the AUC (T_AUC). b. The proportion of validated good responders (observed $\mathrm{AUC}\le T_{\mathrm{AUC}}^D$ in the BeatAML testing set) increases with higher confidence scores. Each boxplot displays the interquartile range (IQR) between the 25th percentile (the lower boundary) and the 75th percentile (the upper boundary). The center line of the box presents the median, and the whiskers are within the 1.5 IQR value. $T_{\mathrm{AUC}}^D$ is the threshold of AUC for good drug responses in drug D. The x-axis represents four categories of confidence scores.

Fig. 4. AUC versus z-score for different AML molecular subtypes.

Each point represents samples of an individual subtype group. The y-axis is the median value of drug response (AUC) of each subtype group. The z-score (x-axis) is a statistic of network enrichment analysis (NEA), which is used to assess the interaction between the subtype-specific genes and the drug-target genes in the functional gene network. A higher z-score indicates a stronger interaction than expected in a random permutation of the network. The blue line is the linear regression line. The yellow rectangle suggests some interesting subtypes, which respond to the drug potentially via the functional interaction between the subtype-specific and the drug-target genes.