A new approach for prediction of tumor sensitivity to targeted drugs based on functional data

BMC Bioinformatics A new approach for prediction of tumor sensitivity to targeted drugs based on functional data Noah Berlow 0 Lara E Davis 2 Emma L Cantor 2 Bernard Sguin 1 Charles Keller 2 Ranadip Pal 0 0 Department of Electrical and Computer Engineering, Texas Tech University , Lubbock, TX , USA 1 Flint Animal Cancer Center, Colorado State University , Fort Collins, CO , USA 2 Department of Pediatrics, Pape Family Pediatric Research Institute, Oregon Health & Science University , Portland, OR , USA Background: The success of targeted anti-cancer drugs are frequently hindered by the lack of knowledge of the individual pathway of the patient and the extreme data requirements on the estimation of the personalized genetic network of the patient's tumor. The prediction of tumor sensitivity to targeted drugs remains a major challenge in the design of optimal therapeutic strategies. The current sensitivity prediction approaches are primarily based on genetic characterizations of the tumor sample. We propose a novel sensitivity prediction approach based on functional perturbation data that incorporates the drug protein interaction information and sensitivities to a training set of drugs with known targets. Results: We illustrate the high prediction accuracy of our framework on synthetic data generated from the Kyoto Encyclopedia of Genes and Genomes (KEGG) and an experimental dataset of four canine osteosarcoma tumor cultures following application of 60 targeted small-molecule drugs. We achieve a low leave one out cross validation error of < 10% for the canine osteosarcoma tumor cultures using a drug screen consisting of 60 targeted drugs. Conclusions: The proposed framework provides a unique input-output based methodology to model a cancer pathway and predict the effectiveness of targeted anti-cancer drugs. This framework can be developed as a viable approach for personalized cancer therapy. - In the last decade, a number of drugs targeting specific biologically relevant kinases have been developed that are becoming common in cancer research as a basis for personalized therapy. The idea of treating cancer through inhibition of a specific tyrosine kinase was proven by the discovery that patients with Chronic Myeloid Leukemia can be successfully treated by inhibiting the tyrosine kinase BCR-ABL with the kinase inhibitor Imatinib Mesylate [1]. However, the success rate of any one specific targeted drug for other forms of cancer, such as sarcoma, is limited as the tumors exhibit a wide variety of signaling pathways and are not uniformly dependent on the activity of a specific kinase [2-6]. The numerous aberrations in molecular pathways that can produce cancer is one cause to necessitate the use of drug combinations for treatment of individual cancers. Combination therapy design requires a framework for inference of the individual tumor pathways, prediction of tumor sensitivity to targeted drug(s) and algorithms for selection of the drug combinations under different constraints. The current state of the art in predicting sensitivity to drugs is primarily based on assays measuring gene expression, protein abundance and genetic mutations of tumors; these methods often have low accuracy due to the breadth of available expression data coupled with the absence of information on the functional importance of many genetic mutations. A commonly used method for predicting the success of targeted drugs for a tumor sample is based on the genetic aberrations in the tumor (e.g. mutation, amplification). However, the accuracy of prediction of drug sensitivity based on mutation knowledge is limited in many forms of tumors as some of the mutations (or low frequency polymorphisms) may not be functionally important or tumors can develop without the known genetic mutations. Statistical tests have been used in [7] to show that genetic mutations can be predictive of the drug sensitivity in non-small cell lung cancers but the classification rates of these predictors based on individual mutations for the aberrant samples are still low. For specific diseases, some mutations have been able to predict the patients that will not respond to particular therapies: for instance [8] reports a success rate of 87% in predicting non-responders to anti-EGFR monoclonal antibodies using the mutational status of KRAS, BRAF, PIK3CA and PTEN. The prediction of tumor sensitivity to drugs has also been approached as a classification problem using gene expression profiles. In [9], gene expression profiles are used to predict the binarized efficacy of a drug over a cell line with the accuracy of the designed classifiers ranging from 64% to 92%. In [10], a co-expression extrapolation (COXEN) approach is used to predict the binarized drug sensitivity in data points outside the training set with an accuracy of around 75%. In [11], a Random Forest based ensemble approach was used for prediction of drug sensitivity and achieved an R2 value of 0.39 between the predicted IC50s and experimental IC50s. Supervised machine learning approaches using genomic signatures achieved a specificity and sensitivity of higher than 70% for prediction of drug response in [12]. Tumor sensitivity prediction has also been considered as (a) a drug-induced topology alteration [13] using phosphoproteomic signals and prior biological knowledge of a generic pathway and (b) a molecular tumor profile based prediction [7,14]. Most interestingly, in the recent cancer cell line encyclopedia (CCLE) study [15], the authors characterize a large set of cell lines (> 900) with numerous associated data measurement sets: gene and protein expression profiles, mutation profiles, methylation data along with the response of around 500 of these cells lines across 24 anticancer drugs. One of the goals of the study was to enable predictive modeling of cancer drug sensitivity. For generating predictive models, the authors considered regression based analysis across input features of gene and protein expression profiles, mutation profiles and methylation data. The performance (as measured by Pearson correlation coefficient between predicted and observed sensitivity values) of the predictive models using 10 fold cross validation ranged between 0.1 to 0.8. In particular, the correlation coefficient for prediction of sensitivity using genomic signatures for the drug Erlotinib across > 450 cell lines was < 0.35. Erlotinib is a commonly used tryosine kinase inhibitor selected primarily as an EGFR inhibitor. However, studies have shown [16] that these targeted drugs often have numerous side targets that can play significant roles in the effectiveness of the inhibitor drugs. The target inhibition profiles of drugs and sensitivity of trainings set of drugs can provide significant information for enhanced prediction of anti-cancer drug sensitivity as we have recently shown [17]. By incorporating the drugtarget interaction data and sensitivities of training drugs with genom (...truncated)