### Introduction

### Materials and Methods

### 1. Study selection

### 2. Data processing

_{50}or area under curve (referred to as ActArea in CCLE) values were classified as sensitive and those above median IC

_{50}or area under curve values were classified as resistance. Detailed step-by-step normalization methods and procedures have been documented previously [9]. Expression values from each data set were normalized and log-transformed. Raw data from Affymetrix platforms, if available, were preprocessed using Robust Multi-array Average (RMA) [10]. Otherwise, we used pre-processed data as provided by the original authors. To generate gene level summarization, we utilized an interquartile range (IQR) method. This allowed us to designate the probe set ID with the largest IQR of expression values out of all multiple probe set IDs as the representative of the gene. Missing expression values are designated using nearest neighbor imputation (R package impute) [11]. To achieve correct batch effect and cross-study normalization, ComBat, an empirical Bayes method, was applied [12].

### 3. Development of the algorithm

*f(s)*of

*n*functions of a single parameter variable s in

*n*−dimensional space. Given a finite

*n*−dimensional random vector

*X=(X*, the projection index is defined as:

_{1}, X_{2}, ..., X_{n})*f(v)=E(X|v*. The PDS of sample

_{f}(X)=s)*i*is defined as the distance along the curve between principal curve

*f*and a reference end point, defined as the centroid of control set of samples (i.e., sensitive cells). Every sample is analyzed in relation to this principal curve and PDSs are assigned using the normalized projection distance for each sample’s pathway. Pathway information used to design the PDS matrix was obtained from three curated pathway databases (the Kyoto Encyclopedia of Genes and Genomes, the BioCarta and the National Cancer Institute–Nature Pathway Interaction Database). Then we used regularized regression on these PDS matrices to fit the model. Regularization techniques have been described in detail in previous reviews [15,16]. The elastic net is a regularized regression method that linearly combines the penalties of the lasso and ridge regression methods and is defined as

_{i}### 4. Evaluation strategies

*p*is the predicted probability and

_{i}*o*is the actual outcome of the event and

_{i}*n*is the sample size.

*BS*is essentially the mean squared error of the probability forecast of a dichotomous event. Hence, a small BS corresponds to a good calibration of predictions. F1 score is a weighted mean of precision and recall, ranging from 0 (worst value) to 1 (best value). ACC is defined as (TP+TN)/(TP+TN+FP+FN). MCC, considered as a balanced measure, is a geometric mean corrected for chance agreement ((TP×TN)–(FP×FN)/square root ((TP+FP)×(TP+FN)×(TN+FP)×(TN+FN))). Prediction performances of all parameters except BS are directly proportional, ranging from 0 to 1. Higher BS denotes worse performance. MCC ranges from –1 (completely incorrect) to 1 (completely correct). All statistical analyses were performed using R ver. 3.2.3 (R Foundation for Statistical Computing Platform, Vienna, Austria).

### Results

_{50}or area under curve values) or resistant (above the median IC

_{50}or area under curve values). After classification, the total of 239, 143, 244 samples were obtained in the drug-sensitive (S) group and 240, 143, 244 in the resistant (R) groups of CCLE-PTX, CGP-PTX, CGPDTX, respectively. In leave-one-out cross-study validation for CCLE and CGP, our algorithm showed a high discrimination ability with an overall AUROC of 0.703 (95% CI, 0.674 to 0.731), an AUPRC of 0.712, a BS of 0.218, a sensitivity of 61.7% and a specificity of 67.1% (Tables 4, 5, Fig. 5A). Considering the accuracies previously reported on consistency between CGP and CCLE were close to random level at 0.5, our model’s performance was remarkable. Next, we tested whether this ITR-based model could predict ATR. Surprisingly, our ITR-based model had a good prediction performance on ATR (overall AUROC, 0.688 [95% CI, 0.539 to 0.837]; AUPRC, 0.735; BS, 0.226; sensitivity, 68.0%; and specificity, 64.0%), suggesting high transferability between ITR and ATR (Table 5, Fig. 5B).

### Discussion

_{50}between CCLE and CGP (Pearson rho of 0.18 in IC

_{50}for PTX with SVM classifier) [2]. Another study reported a similar finding, citing PTX among the major cause for drug/cell line inconsistency (Spearman’s rank correlation coefficient 0.1-0.2) [3]. Recently, Dong et al. [22] applied linear (SVM) and non-linear (random forest) modeling and addressed that, although SVM achieves better performance (0.55) for PTX than RF predicting model (0.482), the values were not much higher than random prediction. Our internal cross-study validation using CCLE and CGP cohorts shows the model with pathway mapping approach is highly consistent between CCLE and CGP (overall cross-study AUC, 0.703), compared to the almost random levels reported in previous studies. Applying this model on ATR cohorts, we further tested whether a model built from intrinsic resistance dataset can predict AR. Surprisingly, prediction parameters were as high, suggesting the model’s generalizability over intrinsic and acquired resistant cell lines (overall AUC, 0.688).