DeepSynergy: predicting anti-cancer drug synergy with Deep Learning.
Preuer K, Lewis RPI, Hochreiter S, Bender A, Bulusu KC, Klambauer G. · 2018
Read DeepSynergy: predicting anti-cancer drug synergy with Deep Learning. by Preuer K, Lewis RPI, Hochreiter S, Bender A, Bulusu KC, Klambauer G. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5925774.txt.
Introduction
Preuer and co-authors use deep learning to predict anti-cancer drug synergy. Their article appeared in Bioinformatics. Drug pairs can beat single drugs in cancer care. The space of pairs is far too large to test all. Screens now give data to train prediction models. Their model, DeepSynergy, takes chemical and genomic inputs.
DeepSynergy beat four rival machine learning methods, they report. Its error was 7.2 percent below the next best model. Predictions matched measured synergy with strong correlation. As a classifier, it reached high accuracy scores as well. All methods struggled with wholly new drugs or cell lines. Dataset size and diversity likely cause that limit.
Catalog record
| CREATOR | Preuer K, Lewis RPI, Hochreiter S, Bender A, Bulusu KC, Klambauer G. |
|---|---|
| YEAR | 2018 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DeepSynergy: predicting anti-cancer drug synergy with Deep Learning By Kristina Preuer, Richard P I Lewis, Sepp Hochreiter, Andreas Bender, Krishna C Bulusu, Günter Klambauer, Jonathan Wren Published in: Bioinformatics DOI: 10.1093/bioinformatics/btx806 ABSTRACT Abstract Motivation While drug combination therapies are a well-established concept in cancer treatment, identifying novel synergistic combinations is challenging due to the size of combinatorial space. However, computational approaches have emerged as a time- and cost-efficient way to prioritize combinations to test, based on recently available large-scale combination screening data. Recently, Deep Learning has had an impact in many research areas by achieving new state-of-the-art model performance. However, Deep Learning has not yet been applied to drug synergy prediction, which is the approach we present here, termed DeepSynergy. DeepSynergy uses chemical and genomic information as input information, a normalization strategy to account for input data heterogeneity, and conical layers to model drug synergies. Results DeepSynergy was compared to other machine learning methods such as Gradient Boosting Machines, Random Forests, Support Vector Machines and Elastic Nets on the largest publicly available synergy dataset with respect to mean squared error. DeepSynergy significantly outperformed the other methods with an improvement of 7.2% over the second best method at the prediction of novel drug combinations within the space of explored drugs and cell lines. At this task, the mean Pearson correlation coefficient between the measured and the predicted values of DeepSynergy was 0.73. Applying DeepSynergy for classification of these novel drug combinations resulted in a high predictive performance of an AUC of 0.90. Furthermore, we found that all compared methods exhibit low predictive performance when extrapolating to unexplored drugs or cell lines, which we suggest is due to limitations in the size and diversity of the dataset. We envision that DeepSynergy could be a valuable tool for selecting novel synergistic drug combinations. Availability and implementation DeepSynergy is available via www.bioinf.jku.at/software/DeepSynergy. Supplementary information Supplementary data are available at Bioinformatics online. 1 INTRODUCTION Administering drug combinations instead of monotherapy can lead to an increased efficacy compared to single drug treatments (Csermely et al., 2013; Jia et al., 2009). Furthermore, host toxicity and adverse side effects are likely reduced, since doses of drug combinations are typically lower than doses of single agents (Chou, 2006; O’Neil et al., 2016). Drug resistance can be decreased or even overcome through combination therapy (Huang et al., 2016; Kruijtzer et al., 2002; Tooker et al., 2007). Therefore, drug combinations are investigated across various medical areas, such as cancer (Al-Lazikani et al., 2012), viral disease including human immunodeficiency virus (HIV) and hepatitis C virus infections (Clercq, 2007) as well as fungal (Chen et al., 2016; Groll and Tragiannidis, 2009) and bacterial infections (Tamma et al., 2012; Worthington and Melander, 2013). However, drug combination effects can be adverse and even lead to shorter progression free survival of cancer patients (Hecht et al., 2009; Tol et al., 2009). Therefore, finding synergistic drug pairs for a particular cancer type is important for improving efficacy of anticancer treatment. Until recently, effective drug combinations were proposed based on clinical experience. However, ‘trial-and-error’ is time-, labor- and cost-intensive. Furthermore, it might expose patients to unnecessary or even harmful treatment (Day and Siu, 2016; Pang et al., 2014). Another strategy to identify synergistic drug pairs without harming patients is high-throughput screening (HTS). A large number of measurements can be produced in reasonable time at low costs (Bajorath, 2002; Bleicher et al., 2003; White, 2000). HTS has also been applied to test for synergistic drug combinations (He et al., 2016). In these screens different concentrations of two drugs are applied to a cancer cell line. Despite the importance of cancer cell lines in biomedical research, their ability to accurately represent the in vivo state is often questioned. The reason is that even if there is a high genomic correlation between the original tumour and the derived cancer cell line, it is still far from perfect (Ferreira et al., 2013). Furthermore, testing the complete combinatorial space with HTS is unfeasible (Goswami et al., 2015; Morris et al., 2016). Computational methods such as machine learning models offer the possibility to efficiently explore the large synergistic space. Accurate predictive models can be generated by leveraging the available HTS synergy data. Reliable predictions provide guidance for in vitro and in vivo research. Furthermore, methods developed to utilize genomic information for their predictions offer the opportunity to apply them also in an in vivo setting. Therefore, these predictive models are a big step towards precision medicine (Bulusu et al., 2016). Anti-cancer synergy prediction has already been tackled with a wide variety of approaches. Methods range from systems biology methods (Feala et al., 2010), kinetic models (Sun et al., 2016), mixed integer linear programming methods based on the diseased gene set (Pang et al., 2014), computational methods based on gene expression profiles after treatment with single drugs and dose response curves (Goswami et al., 2015; Yang et al., 2015), to machine learning approaches including Random Forests and Naive Bayes methods (Li et al., 2015; Wildenhain et al., 2015). However, these methods are restricted to certain pathways, targets or cell lines, or require transcriptomic data of cell lines under compound treatment. In contrast, our approach only requires a single transcriptomic characterization of the cell line without compound treatment, and the chemical structure of the two drugs. Two crowdsourced challenges were focused on computational methods for drug combinations. In 2012, a challenge considered only a single cell line (OCI-LY3) and 14 different drugs (Bansal et al., 2014). Recently, a second challenge was launched to determine the state-of-the-art of prediction of synergy scores (Yu, 2016). The dataset consisted of 3800 synergy score samples of pre-specified drug combinations. Although the research questions addressed in these challenges are undoubtedly of high importance, the competition was grounded on datasets with limited size and did not evaluate the performance of methods for novel drug combinations. Due to the fast development of high-throughput methods, the amount of available synergy data points has tremendously increased. Publicly available databases such as ASDCD (Chen et al., 2014) containing antifungal combinations and DCDB (Liu et al., 2010, 2014) with approved and investigational combinations, play a key role in providing good quality training data for developing computational predictive methods. A review of these resources has been discussed in Bulusu et al. (2016). Recently, a large HTS synergy study (O’Neil et al., 2016) with more than 20 000 synergy measurements was performed, which offers the possibility to evaluate computational methods for predicting novel drug combinations. The dataset covers 38 drugs and 39 cancer cell lines. Therefore, the performed HTS covered 83% of the possible two drug combinations. Both experimental and approved drugs were tested. The used cancer cell lines originated from seven different tissue types. Previous methods were developed and optimized for small datasets. However, methods developed for limited data might not be appropriate anymore. Predictive performance can be improved, since more data is available and methods which learn from tens of thousands of data points can be used. Deep Neural Networks (DNNs), which strongly profit from large datasets, have impacted many scientific disciplines and achieved new state-of-the-art performance (LeCun et al., 2015). Deep Learning has set new records in image (Farabet et al., 2013; Krizhevsky et al., 2012)