Modular within and between score for drug response prediction in cancer cell lines

Abstract

Drug response prediction in cancer cell lines is vital to discover new anticancer drugs. However, it's still a challenging task to accurately predict drug responses in cancer cell lines. In this study, we presented a novel computational approach, named as MSDRP (modular within and between score for drug response prediction), to predict drug responses in cell lines. The method is based on a constructed heterogeneous drug–cell line network with multiple information. Compared with other state-of-the-art methods, MSDRP acquired better predictive performance, and identified potential associations between drugs and cell lines, which have been confirmed by the published literature. The source code of MSDRP is freely available at https://github.com/shimingwang1994/MSDRP.git.

---

1 Introduction

Drug development often consumes substantial time and cost.¹ The rapid growth of various high-throughput biological data such as Cancer Cell Line Encyclopedia (CCLE)² and Genomics of Drug Sensitivity in Cancer (GDSC),³ makes computational approaches possible and rational to predict the drug response in cancer cell lines. Hence, some methods were proposed to predict drug sensitivity to cell lines using the drug response data. Geeleher et al. predicted clinical drug response using a ridge regression model and obtained good experimental results on before-treatment baseline tumor gene expression data.⁴ Muhammad et al. employed two types of information (cell line genomic feature and drug chemical properties) to construct the KBMF (kernelized Bayesian matrix factorization) model to predict drug response.⁵ Recently, several methods were proposed to predict drug response in cancer cell lines through combing more information, including compound chemical and therapeutic properties, baseline gene expression levels in cell lines and the cancer genomic alterations.^6–8 These methods usually obtained better predictive ability. Network-based methods are also widely used to predict drug responses. Zhang et al. constructed a dual-layer network (cell line similarity network and drug similarity network) to predict drug responses.⁹ Stanfield et al. used a random walk algorithm to analyze cell line and drug network.¹⁰ GloNetDRP was developed to predict drug response by capturing the global information of drug and cell line in a heterogeneous network.¹¹ HNMDRP¹² applied an information flow-based algorithm¹³ in a heterogeneous network, which consists of the drug–target interaction network, protein–protein interaction(PPI) network, drug–cell line association network, drug similarity network and cell line similarity network, to predict drug response. Machine learning is another effective pattern for predicting drug response.¹⁴ For example, a recursive feature selection tool was employed to extract genomic features from the gene expression profile, and then the support vector machine (SVM) was applied to build a predictor in cell lines.¹⁵ NLLSS was proposed as a semi-supervised learning method to predict drug response.¹⁶ CaDRReS used a recommender system, and reflected its advantages on understanding drug response mechanisms.¹⁷ Xu et al. constructed an autoencoder network based on genetic features, such as mutations and RNA expression, and selected important features using the Boruta algorithm and random forest to predict drug response.¹⁸

Although many methods have been developed to predict drug response in cancer cell lines, there are still some problems to be solved. For example, predicting drug response to cell lines without any known drug response data still deserves attention. In addition, the topological structure feature of the response network is not taken into account sufficiently. Therefore, we proposed a novel method to solve these problems, called MSDRP (modular within and between score for drug response prediction). The proposed method integrates a drug–cell line response network, a drug chemical structure similarity network, the cell lines’ gene expression similarity network and a Gaussian interaction profile kernel similarity network to predict drug response in cancer cell lines (see Fig. 1).

Fig. 1 Framework of MSDRP. (a) Integrate multiple similarity and response information of drugs and cell lines. (b) Construct a drug similarity network, cell line similarity network and drug–cell line response network. (c) Divide drugs (cell lines) into drug (cell line) modules via the investigated cell line (drug). (d) Calculate the modular within-score and modular between-score of the investigated drug and cell line to acquire the association score.

2 Methods

MSDRP includes three parts (see Fig. 1). Firstly, we collect drug's chemical structure information, drug–target interaction information, cell lines’ gene expression information and cell lines’ drug response data to calculate multiple similarities about drugs and cell lines (Fig. 1a). Then, a heterogeneous drug–cell line network is constructed, including the drug similarity network, cell line similarity network and drug–cell line response network (see Fig. 1b). Secondly, drugs (cell lines) are divided into several modules according to cell lines (drugs) associated with them (see Fig. 1c). Finally, the modular within-score and modular between-score of drugs and cell lines are calculated to obtain the association score of drugs and cell lines (see Fig. 1d).

2.1 Construction of the heterogeneous network

2.1.1 Drug–cell line response network. We download drugs’ log-normalized IC50 values in cancer cell lines¹⁹ from GDSC (Downloaded in April 2018, Release 7.0). Staunton et al.'s method²⁰ is used to classify cell lines into two classes: sensitive or resistant, according to their IC50 values with drugs. Then, a drug–cell line response network with nd drugs and nc cell lines is constructed (an edge represents that a cell line is sensitive to a certain drug, and a node represents a drug or a cell line), and its adjacency matrix is A ∈ R^nd×nc.

2.1.2 Drug chemical structure similarity. The drugs’ chemical structure is downloaded from PubChem,²¹ and the 1-D and 2-D structure properties are extracted by PaDEL.²² Then, we obtain the chemical structure feature to characterize drugs. Similarly to existing studies,^7–9,11,12 the Pearson correlation coefficient²³ (Pcc) value of two drugs is calculated to construct the chemical structure similarity matrix CTS.

2.1.3 Cell lines’ gene expression similarity. We download the gene expression profile for cancer cell lines from GDSC and normalize the data. Similar to previous studies,^9,12 the Pcc value of each cell line pair is also calculated to construct the gene expression similarity matrix GES.

2.1.4 Gaussian interaction profile kernel similarity for drugs. Based on the fact that similar cell lines and similar drugs exhibit similar drug responses,^8,9 the topological structure information in the drug–cell line response network can be considered to calculate the Gaussian interaction profile kernel similarity matrix for drugs. Inspired by previous methods,^24–27 the similarity between drug d_i and drug d_j is defined as below:

GSD[d_i,d_j] = exp(−α‖A(d_i) − A(d_j)‖²) (1)

where vector A(d_i) and A(d_j) represent the ith and jth row of matrix A respectively, and α is a parameter to control the kernel bandwidth defined as below:

Previous methods^24–27 set the bandwidth parameter α′ as 1. Here, we modify α′ by the number of cell lines associated with d_i and d_j simultaneously, as below:

An example demonstrates the rationality of the modified bandwidth parameter visually (see Fig. 2). There is only one cell line associated with d₁ and d₂ simultaneously, two cell lines associated with d₁ and d₃, and two cell lines associated with d₂ and d₃ (see Fig. 2a). So the similarity between d₁ and d₂ should be lower than the other two drug pairs. However, the similarities calculated by the previous bandwidth parameter are the same, which are 0.4724 (see Fig. 2b). Our modification reflects the difference among the three drug pairs (see Fig. 2c). GSD[d₁,d₃] and GSD[d₂,d₃] have the same value of 0.7788, and exceed GSD[d₁,d₂] of 0.6873, which is more rational.

Fig. 2 An example of the rationality of our modified bandwidth parameter. (a) A drug–cell line response network with three drugs, four cell lines and eight edges. (b) Similarity of drugs calculated by setting the bandwidth parameter as 1. (c) Similarity of drugs calculated with our modified bandwidth parameter.

In addition, we collect the drugs’ target information from GDSC and construct the drug–target interaction network. We also consider the target information of drugs, so the steps above are carried out again in this network and get the Gaussian interaction profile kernel similarity matrix GST for drugs.

2.1.5 Gaussian interaction profile kernel similarity for cell lines. Similar to Section 2.1.4, vector A(c_i) and A(c_j) are the ith and jth column of matrix A respectively, the Gaussian interaction profile kernel similarity between cell line c_i and cell line c_j is defined as below:

GSC[c_i,c_j] = exp(−β‖A(c_i) − A(c_j)‖²) (4)

Finally, Gaussian interaction profile kernel similarity matrix GSC can be acquired.

2.1.6 Integrated similarity for drugs and cell lines. To construct an accurate similarity network for drugs and cell lines, multiple similarity information is integrated by combining the Gaussian interaction profile kernel similarity with drugs’ chemical structure feature similarity and cell lines’ gene expression similarity, respectively. The integrated drug similarity matrix SD and cell line similarity matrix SC is defined as below:

2.2 Module partition of drugs and cell lines

For an investigated pair between drug d_i and cell line c_j, all drugs are divided into drug modules via c_j, and all cell lines are divided into cell line modules via d_i (see Fig. 1c).

Step 1, all drugs are divided into md modules ND^j₁, ND^j₂,…,ND^j_mdvia c_j, where md is the number of drugs associated with c_j. In order to keep balance among all modules, we control the size of each module not exceeding ⌈nd/md⌉. Firstly, md empty modules are initialized, and each drug associated with c_j is assigned into each module and marked as the core of it. Then, for the drug unassigned, the most similar core with it can be found and the drug is added into the core's module. When all drugs are assigned, check the size of all modules. If the size of the module ND^j_s exceeds ⌈nd/md⌉, the similarity between the core and other drugs in the module will be sorted in descending order, and the last |ND^j_s| − ⌈nd/md⌉ drugs will be removed from ND^j_s. Repeat this process until the size of each module is not more than ⌈nd/md⌉.

Step 2, all cell lines are divided into mc modules NCⁱ₁, NCⁱ₂,…,NCⁱ_mcvia d_i in a similar way to step 1.

2.3 Drug response prediction using the modular within-score and modular between-score

The method WBSMDA²⁶ mentioned the conception of the within-score and between-score. For the investigated drug d_i and investigated cell line c_j, the within-score (between-score) of d_i is the highest similarity with d_i among the group of drugs with (without) known association with c_j, and the within-score (between-score) of c_j is the highest similarity with c_j among the group of cell lines with (without) known association with d_i. MSDRP adds the modular information into within-score and between-score (see Fig. 1d). According to Section 2.2, the core in one drug module is associated with cell line c_j, and other drugs in this module are not associated with c_j. Hence, we can calculate the within-score and between-score of drug d_i in each drug module, which is from the local perspective. Then, the within-score and between-score in each drug module can be integrated by the weight of drug modules to obtain the modular within-score and modular between-score of d_i, which is from the global perspective. It is analogous to calculate the modular within-score and modular between-score of c_j. The addition of modular information makes the use of similarity information more comprehensive.

MW_d(d_i,c_j) is defined as the modular within-score of d_i as below:

where ND^j_p is a drug module and ND^j_p(core) is the core of it. weightd(p) is the weight of ND^j_p as below:

MB_d(d_i,c_j) is defined as the modular between-score of d_i as below:

Similarly, MW_c(d_i,c_j) is defined as the modular within-score of c_j as below:

where NCⁱ_q is a cell line module and NCⁱ_q(core) is the core of it. weightc(q) is the weight of NCⁱ_q as below:

Similarly, MB_c(d_i,c_j) is defined as the modular between-score of c_j as below:

To calculate the association score between d_i and c_j, the modular within-score and modular between-score of d_i and c_j are integrated comprehensively as below:

Besides, for a new drug d which does not have any known associations with any cell lines, MSDRP integrates the modular within-score and modular between-score of drugs as below:

Meanwhile, for a new cell line c which does not have any known associations with any drugs, MSDRP integrates the modular within-score and modular between-score of the cell lines as below:

3 Results

3.1 Performance evaluation

To evaluate the predictive performance of MSDRP, Leave-one-out cross validation (LOOCV) is implemented. Each sensitive drug–cell line pair is left out in turn as the test data and the other drug–cell line pairs are used as training data for model learning. The test association with a higher rank represents a greater possibility of correct predictions. The receiver operating characteristic (ROC) curve is drawn and the areas under the ROC curve (AUROC) are computed to show the predictive ability of MSDRP. Meanwhile, because the number of resistant response is much more than sensitive response in the drug–cell line response network (see Table 1), we adopt the precision–recall (PR) curve and the areas under the PR curve (AUPR) to evaluate the predictive performance of our model on this imbalanced data. Hence, the results are evaluated by using the ROC curve (or AUROC) and PR curve (or AUPR) simultaneously in the following sections.

Table 1 Global characteristic of the drug–cell line response network

Number of drugs	Dimension of drug chemical structure feature	Number of cell lines	Dimension of cell lines’ gene expression profile	Number of drug–cell line sensitive associations	Number of resistant drug–cell line pairs
251	1444	990	17737	23774	178269

---

Max degree of drugs	Min degree of drugs	Average degree of drugs	Max degree of cell lines	Min degree of cell lines	Average degree of cell lines
146	34	94.72	168	1	24.01

---

3.2 Performance analysis of MSDRP

The global characteristic of the drug–cell line response network is shown in Table 1. The degree of a drug (cell line) is the number of its associations with cell lines (drugs). The drug–target interaction network mentioned in Section 2.1.4 contains 250 drugs, 239 targets and 456 interactions. As a result, MSDRP acquires an AUROC value of 0.8446 and AUPR value of 0.4007 on LOOCV. In addition, we set the bandwidth parameter as 1 to repeat MSDRP and obtain an AUROC value of 0.7529 and AUPR value of 0.3197 (see Fig. 3). These results demonstrate the effectiveness of our modified bandwidth parameters.

Fig. 3 ROC curves and PR curves of MSDRP with different bandwidth parameters. (a) ROC curves. (b) PR curves.

3.3 Comparison with existing methods

We compare MSDRP with five methods: Xu's method,¹⁸ WBSMDA,²⁶ Zhang's method,⁹ HNMDRP¹² and CaDRReS.¹⁷ The ROC curves and PR curves of these methods are shown in Fig. 4. Our model has the highest AUROC value and AUPR value. AUROC values of each drug obtained by these methods are calculated and shown in a box-and-whisker plot, where the P-values of paired t-tests between each method and MSDRP are marked (see Fig. 5a). Our model obtains statistically significantly better results than the other five methods according to all P-values not more than 1.2974 × 10⁻⁵. For AUPR of all drugs, we also provide a box-and-whisker plot (see Fig. 5b) and five P-values of the paired t-test are not more than 5.2771 × 10⁻⁶. These results illustrate that our model possesses excellent ability for predicting drug–cell line associations and potential confidence to detect new associations.

Fig. 4 ROC curves and PR curves on LOOCV of different methods. (a) ROC curves. (b) PR curves.

Fig. 5 The box-and-whisker plots of AUROC and AUPR values of drugs by different methods with the P-value of the paired t-test. (a) AUROC values. (b) AUPR values.

3.4 Tissue specific cell line type

Since one cell line belongs to one certain tissue type, we detect differences of drug responses in diverse tissue types. All cell lines are divided into 19 tissue types and the percentage of each tissue type is counted (see Fig. 6a). The major tissue types are lung_NSCLC, urogenital_system, leukemia and aero_dig_tract, which take up 11.21% (111), 10.4% (103), 8.3% (82) and 7.98% (79) of 990 cell lines, respectively. Comparison of AUROC values on the four major tissue types is exhibited (see Fig. 6b). Our model acquires the highest AUROC values no matter in which tissue type, which are 0.829, 0.8314, 0.8035 and 0.8325 from left to right. The AUPR values of the major tissue types are shown in Fig. 6c, where MSDRP gets 0.2587, 0.3286, 0.5677 and 0.2991 from left to right, and still exceeds the other five methods. Meanwhile, predictive performance on the other fifteen tissue types is also detected (see Fig. S1, ESI†) and our model still has an advantage, which illustrates the high efficiency of our model when aiming at a specific tissue type.

Fig. 6 Distribution of tissue types and predictive performance in the four major tissue types. (a) Distribution of nineteen tissue types. (b) Comparison of AUROC values between our method and other methods in the four major tissue types. (c) Comparison of AUPR values between our method and other methods in the four major tissue types.

3.5 Drug response prediction in the missing data

In the response network consisting of 251 drugs and 990 cell lines, 46447 drug–cell line pairs don’t have log-normalized IC50 values. In other words, 18.69% of 2512990 drug–cell line possible combinations are missing data. Nevertheless, MSDRP has the ability of predicting potential associations in the missing data. Here, we implement drug response prediction in the missing data for the major four tissue types and find literature evidence for the top 20 potential drug–cell line associations in each tissue type manually. When predicting for tissue types of Lung_NSCLC, urogenital_system, leukemia and aero_dig_tract, six, five, four and four association pairs in the top 20 were confirmed sensitive, respectively (see Table S1, ESI†). The drugs involved are used to treat non-small cell lung cancer (NSCLC), endometrium adenocarcinoma (EA), ovarian cancer (OC), endometrial cancer (EC), acute myeloid leukaemia (AML), head-and-neck cancer (HNC) and head and neck squamous cell carcinoma (HNSCC). We also provide the rank of these associations in the other five methods and our method acquires higher rank than them. Here, we introduce the association between drug palbociclib and cell line NCI-H292 in detail. Cell line NCI-H292 belongs to tissue type Lung_NSCLC, which has the highest proportion among all tissue types. We sorted the association scores of all pairs belonging to Lung_NSCLC, and this pair ranked 1st. Gopalan et al. gave palbociclib orally to 19 previously-treated patients with recurrent or metastatic NSCLC at 125 mg daily on days 1–21 of a 28 day cycle.²⁸ Experiments showed that palbociclib therapy alone was well-tolerated, and 50% of evaluable patients kept a stable disease (SD), which suggested the induction of cellular senescence. These results indicate that MSDRP has advantageous performance in predicting unknown drug response and can provide convenience for prediction in networks with missing data.

3.6 Prediction of cell lines without any known drug response data

GDSC (Release 7.0) provides 1001 cell lines with whole exome sequencing, including 990 cell lines with drug response data and 11 cell lines without any drug response data. For the 11 cell lines, which are D-245MG, SNU-283, NCI-H660, U-CH2, MHH-CALL-4, KO52, SC-1, RERF-LC-FM, NCI-H740, NCCIT and BONNA-12, MSDRP can calculate the prediction scores with all drugs and a higher score means a potential sensitive association. After sorting the predictive score of one certain cell line, the top 20 pairs are counted to find literature evidence manually. All evidence with the detailed descriptions is provided (see Table S2, ESI†). The drugs involved are used to treat glioblastoma multiforme (GBM), glioma, glioblastoma, colorectal cancer (CRC), colorectal adenocarcinoma (CRAC), colon cancer (CC), castration-resistant prostate cancers (CRPC), prostate cancer (PC), osteosarcoma, B-Cell Chronic Lymphocytic Leukemia (CLL), AML, diffuse large B-cell lymphoma (DLBCL), lung_small_cell_carcinoma (SCLC), testicular germ cell tumor (TGCT), and hairy cell leukemia (HCL). Here, we introduce the association between drug cetuximab and cell line NCI-H660 in detail. We sorted the association scores between NCI-H660 and all drugs, and cetuximab ranked 1st. NCI-H660 belongs to cell type prostate. The EGF receptor (EGFR) might be a valid therapeutic target of metastatic castration-resistant prostate cancers (mCRPC), because of its overexpression in mCRPC. Cathomas et al. found that cetuximab could improve the outcome of patients with mCRPC by inhibiting EGFR.²⁹ In conclusion, these results demonstrate that MSDRP is an efficient approach to predict the drug response for new cell lines and can provide valuable applications in future research studies.

4 Conclusions

In this study, we propose a novel method MSDRP to predict drug responses in cancer cell lines and contributions are made at three aspects. Firstly, the innovation on the bandwidth parameter extracts topological information in the drug–cell line response network to the maximum extent, which is beneficial for integrating various similarity networks. Secondly, the addition of modular information into within-score and between-score provides more sufficient similarity information for the investigated drug or cell line than calculating within-score or between-score among the whole response network. Finally, predicting potential drug sensitive associations for new cancer cell lines is a significant development for future biological research. The experiment results confirm these improvements. We hope that MSDRP can provide potential value in future clinical research.

Conflicts of interest

There are no conflicts to declare.

Acknowledgements

This work has been supported by the Natural Science Foundation of Heilongjiang Province (Grant No. F2016016), and National Key Research and Development Program of China (Grant No. 2016YFC0901905).

References

1. A. Gottlieb, G. Y. Stein, E. Ruppin and R. Sharan, PREDICT: a method for inferring novel drug indications with application to personalized medicine, Mol. Syst. Biol., 2011, 7, 496.
2. J. Barretina, G. Caponigro, N. Stransky, K. Venkatesan, A. A. Margolin and S. Kim, et al., The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity (vol 483, pg 603, 2012), Nature, 2012, 492, 290.
3. W. J. Yang, J. Soares, P. Greninger, E. J. Edelman, H. Lightfoot and S. Forbes, et al., Genomics of Drug Sensitivity in Cancer (GDSC): a resource for therapeutic biomarker discovery in cancer cells, Nucleic Acids Res., 2013, 41, D955–D961.
4. P. Geeleher, N. J. Cox and R. S. Huang, Clinical drug response can be predicted using baseline gene expression levels and in vitro drug sensitivity in cell lines, Genome Biol., 2014, 15, R47.
5. A. U. D. Muhammad, G. Elisabeth, G. N. Mehmet, L. Tuomo, K. Olli and W. Krister, et al., Integrative and personalized QSAR analysis in cancer by kernelized Bayesian matrix factorization, J. Chem. Inf. Model., 2014, 54, 2347.
6. S. Gupta, K. Chaudhary, R. Kumar, A. Gautam, J. S. Nanda and S. K. Dhanda, et al., Prioritization of anticancer drugs against a cancer using genomic features of cancer cells: a step towards personalized medicine, Sci. Rep., 2016, 6, 23857.
7. Y. C. Wang, J. W. Fang and S. L. Chen, Inferences of drug responses in cancer cells from cancer genomic features and compound chemical and therapeutic properties, Sci. Rep., 2016, 6, 32679.
8. L. Wang, X. Li, L. Zhang and Q. Gao, Improved anticancer drug response prediction in cell lines using matrix factorization with similarity regularization, BMC Cancer, 2017, 17, 513.
9. N. Q. Zhang, H. Y. Wang, Y. Fang, J. Wang, X. Q. Zheng and X. S. Liu, Predicting Anticancer Drug Responses Using a Dual-Layer Integrated Cell Line-Drug Network Model, PLoS Comput. Biol., 2015, 11, e1004498.
10. Z. Stanfield, M. Coşkun and M. Koyutürk, Drug Response Prediction as a Link Prediction Problem, Sci. Rep., 2017, 7, 40321.
11. D. H. Le and V. H. Pham, Drug Response Prediction by Globally Capturing Drug and Cell Line Information in a Heterogeneous Network, J. Mol. Biol., 2018, 430, 2993–3004.
12. Z. Fei, M. Wang, J. Xi, J. Yang and L. Ao, A novel heterogeneous network-based method for drug response prediction in cancer cell lines, Sci. Rep., 2018, 8, 3355.
13. W. Wenhui, Y. Sen, Z. Xiang and L. Jing, Drug repositioning by integrating target information through a heterogeneous network model, Bioinformatics, 2014, 30, 2923–2930.
14. M. P. Menden, F. Iorio, M. Garnett, U. McDermott, C. H. Benes and P. J. Ballester, et al., Machine Learning Prediction of Cancer Cell Sensitivity to Drugs Based on Genomic and Chemical Properties, PLoS One, 2013, 8, e61318.
15. Z. Dong, N. Zhang, C. Li, H. Wang, Y. Fang and J. Wang, et al., Anticancer drug sensitivity prediction in cell lines from baseline gene expression through recursive feature selection, BMC Cancer, 2015, 15, 489.
16. X. Chen, B. Ren, M. Chen, Q. Wang, L. Zhang and G. Yan, NLLSS: Predicting Synergistic Drug Combinations Based on Semi-supervised Learning, PLoS Comput. Biol., 2016, 12, e1004975.
17. S. Chayaporn, B. Denis and N. Niranjan, Predicting Cancer Drug Response Using a Recommender System, Bioinformatics, 2018, 34(22), 3907–3914.
18. X. Lu, H. Gu, Y. Wang, J. Wang and P. Qin, Autoencoder Based Feature Selection Method for Classification of Anticancer Drug Response, Front. Genet., 2019, 10, 233.
19. J. L. Sebaugh, Guidelines for accurate EC50/IC50 estimation, Pharm. Stat., 2011, 10, 128–134.
20. J. E. Staunton, D. K. Slonim, H. A. Coller, P. Tamayo, M. J. Angelo and J. Park, et al., Chemosensitivity prediction by transcriptional profiling, Proc. Natl. Acad. Sci. U. S. A., 2001, 98, 10787–10792.
21. E. E. Bolton, Y. Wang, P. A. Thiessen and S. H. Bryant, Chapter 12 – PubChem: Integrated Platform of Small Molecules and Biological Activities, Elsevier Science & Technology, 2008.
22. Y. Chun Wei, PaDEL-descriptor: an open source software to calculate molecular descriptors and fingerprints, J. Comput. Chem., 2011, 32, 1466–1474.
23. P. Ahlgren, J. Bo and R. Rousseau, Requirements for a cocitation similarity measure, with special reference to Pearson's correlation coefficient, J. Assoc. Inf. Sci. Technol., 2014, 54, 550–560.
24. X. Chen, Z. C. Jiang, D. Xie, D. S. Huang, Q. Zhao and G. Y. Yan, et al., A novel computational model based on super-disease and miRNA for potential miRNA-disease association prediction, Mol. BioSyst., 2017, 13, 1202.
25. X. Chen, Y. A. Huang, Z. H. You, G. Y. Yan and X. S. Wang, A novel approach based on KATZ measure to predict associations of human microbiota with non-infectious diseases, Bioinformatics, 2016, 33, 733–739.
26. X. Chen, C. C. Yan and X. Zhang, WBSMDA Within and Between Score for MiRNA-Disease Association prediction, Sci. Rep., 2016, 6, 21106.
27. X. Chen, D. Xie, L. Wang, Q. Zhao and H. Liu, BNPMDA: Bipartite network projection for MiRNA–Disease association prediction, Bioinformatics, 2018, 34, 3178–3186.
28. P. K. Gopalan, M. C. Pinder, A. Chiappori, A. M. Ivey, A. Gordillo Villegas and F. J. Kaye, A phase II clinical trial of the CDK 4/6 inhibitor palbociclib (PD 0332991) in previously treated, advanced non-small cell lung cancer (NSCLC) patients with inactivated CDKN2A, American Society of Clinical Oncology, 2014.
29. R. Cathomas, C. Rothermundt, D. Klingbiel, L. Bubendorf, R. Jaggi and D. C. Betticher, et al., Efficacy of cetuximab in metastatic castration-resistant prostate cancer might depend on EGFR and PTEN expression: results from a phase II trial (SAKK 08/07), Clin. Cancer Res., 2012, 18, 6049–6057.

---

Footnote

† Electronic supplementary information (ESI) available: Fig. S1. Predictive performance in the other fifteen tissue types. (a) Comparison of AUROC values between our method and other methods in the other fifteen tissue types. (b) Comparison of AUPR values between our method and other methods in the other fifteen tissue types. Table S1. Literature evidence of the top 20 potential drug-cell line responses for major tissue types in the missing data. Table S2. Literature evidence of potential drug responses for new cell lines in the top 20. See DOI: 10.1039/c9mo00162j

---