High-throughput screening of a large number of molecules is a widely used approach to identify lead compounds exerting a beneficial effect on a given phenotype. In the context of cancer, libraries of chemical entities have been tested in this way against panels of cell lines grown in different conditions and with heterogeneous genomic backgrounds [1]. Following the pioneering work of the “NCI-60”, a collection of 59 human cancer cell lines developed by the National Cancer Institute for in vitro drug screening [2], recent hallmark studies have shown that screening very large cell line collections can recapitulate known and identify novel molecular genomic determinants of drug sensitivity [1], [3]–[5].

In these studies, using systematic statistical inference and regression methods, determinant such as oncogenic lesions, high or low levels of basal gene expression and other genotypic traits have been associated to profiles of increased sensitivity/resistance to specific compounds. For instance, by applying a multivariate analysis of variance [6] and the ‘Elastic Net’ regression framework [7] established drug-genotype associations have been confirmed and complemented with markers of tissue-specificity and novel connections, e.g. the EWS-FLI1 translocation in Ewing’s sarcoma and sensitivity to PARP inhibitors, have been identified and further experimentally validated. Results of these studies have been made publicly available, providing unique resources that support the discovery of new predictive biomarkers for personalised cancer therapy.

Increasing further the size of the considered cell-line/compound panels would be very beneficial, as it provides the basis to improve the accuracy and predictive power of the inferred associations. However, this requires larger infrastructures and the cost grows with the screening size. In addition, due to various technical and logistical reasons in a high-throughput screen [7], the resulting compound-by-cell line matrix of drug efficacy (typically summarised in their IC₅₀, the half maximal (50%) inhibitory concentration of a substance with respect to cell viability) is often not complete. Although many steps are automated, filling experimentally each gap could be expensive and laborious [6]. Hence, an accurate tool to impute missing IC₅₀s and estimate them for novel cell lines would be of great value for drug screening design.

Furthermore, a robust prediction tool for in silico identification of potentially effective drugs for treating a specific cancer could be used for drug repositioning [8], [9]. An approach of this kind is represented by the COMPARE algorithm [10], [11] that uses drug response profiles of the NCI-60 screening, through a ‘guilt-by-association’ paradigm. Following this principle, drugs eliciting a similar drug-response profile across the cell lines in the NCI-60 panel are hypothesized to share a common mode of action (MoA), thus enabling MoA discovery for novel drugs (if their tumour-suppression profile is similar to that of a known and well characterized drug) as well as the discovery of novel or secondary effects for established drugs.

Ultimately, in silico methods to accurately predict the effectiveness of drugs based on the molecular making of tumours (i.e. genome, transcriptome) would be a major milestone towards personalized therapies for cancer patients based on molecular biomarkers [12].

We therefore investigated whether it is possible to build machine learning models (for details see “Materials and Methods” section, “Machine learning” subsection) that can predict drug sensitivity using cell line screening experimental data, where cell lines are treated with variable concentration of a given drug and the resulting dose-response curve summarized by an IC₅₀. We focused on the most comprehensive cancer drug screening dataset available to date, from the “Genomics of Drug Sensitivity in Cancer” (GDSC) project [3]. For each drug, a neural network model was trained to predict its IC₅₀ profile across the panel of cell lines based on the genomic background of each cell, as characterised by microsatellite instability status (1 =  unstable or 0 =  stable), somatic coding variants in the coding sequence of 77 cancer genes (1 =  any change in protein sequence and 0 =  wild type) and copy number alterations denoting gene amplification and deletion of those cancer genes (1 =  amplification/more than 7 copy numbers, 0 =  wild type/between 1 or 7 copy numbers, and –1 =  deletion/no copy number). However, the predictive power of these initial models was limited, especially for those drugs without a well-known oncogene-to-drug response dependency.

We reasoned that cancer cell sensitivity to drug molecules is driven by features from both cells and drugs. Whereas cell features are ultimately connected to the inner workings of the cell, drug features include physicochemical properties that are correlated with the ability of the molecule to cross the cell membrane (e.g. lipophilicity) or its selectivity to intracellular targets (e.g. fingerprints encoding the chemical structure).

Indeed, extensive work has been done on Quantitative Structure-Activity Relationship (QSAR) approaches to predicting whole-cell activity of molecules based of their chemical properties [13]–[16], including applications to predicting anti-cancer activity in drugs [17], [18]. However, such QSAR approaches exclusively based on chemical features cannot distinguish between resistant and sensitive cell lines. For instance, building a model without any information of the cell lines, the model will be not capable of predicting cell line A to be more resistant than cell line B to drug C, which is the main aim of integrating chemical and genomic features in our models.

We therefore extended our machine learning models to include as input chemical features from the drugs, besides the molecular characterization of the cell lines (see Fig 1). This integrative approach not only integrates two complementary streams of information, but also allows the model to be trained with much larger amounts of data, which is often a key factor to improve predictive performance (see Fig 2). Consequently, data was pre-processed to include 689 chemical descriptors of the drugs and 138 genomic features for differentiating the cell lines, resulting in an input space of 827 features.

Chemical descriptors were generated with PaDEL software [19] from simplified molecular-input line entry system (SMILES) structures. Descriptors include physicochemical features such as weight, lipophilicity, rule of five, and additionally fingerprints of the drugs (for details see “Materials and Methods” section, “Features” subsection, and http://padel.nus.edu.sg/software/padeldescriptor/).

For building our model, we used GDSC screening data from 608 genomically characterised cell lines and 111 drugs for which chemical information were available (see Fig 2 and Methods for details). The published version of this matrix holds 38,930 IC₅₀ values (∼58% of the total, due to technical and logistic reasons).

We performed an 8-fold cross-validation, where the test set of each fold was not used for training so as to measure the predictive power of the resulting models across all drugs rather than for each drug separately. Neural networks were able to impute missing log(IC₅₀) values on the test sets with an averaged Pearson correlation coefficient (R_p), coefficient of determination (R²) and root mean square error (RMSE) (Text S1) of 0.85, 0.72 and 0.83 across all 111 drugs, respectively (Fig 3A). Alternatively, random forests achieved comparable performances (R_p of 0.85, R² of 0.72 and RMSE of 0.84; full details in supplementary materials). Furthermore, we conducted a blind test using 13,565 new experimental IC₅₀ values only received after training our models in order to verify our cross-validation results (drug-to-cell line matrix updated by ∼18%, with these newly generated IC₅₀s exclusively used as the blind test set). The results on the blind test were almost as good as in the cross-validation, obtaining an R_p of 0.79, R² of 0.64 and an RMSE of 0.97 (Fig S1, Text S2). The accuracy of the predictions encouraged us to train the networks with fewer IC₅₀ values. Remarkably, the predictive power of the models did not fall appreciably off in quality, even if the amount of training data was reduced to 20% of the total (Fig 3B).

Using an analysis of variance (ANOVA) to identify drug-to-oncogene associations, we investigated how well the IC₅₀ values predicted for the test set using our model recapitulate associations manifested in the experimental data, for instance, whether a given mutation is causing sensitivity or resistance against a drug [3]. Using only predicted IC₅₀ values, we correctly captured 79% (168/213) of the significant observations with the same t-test tendency (positive or negative effect on drug sensitivity) identified with the experimental IC₅₀s. When only considering significant associations from our model (p-value adjusted with Benjamini-Hochberg, FDR = 0.2), we correctly predicted 28% (59/213) of all experimentally identified associations. Where we failed to detect an association the ANOVA effect size is often small, or the experimental correlation is associated with a mutation either not or infrequently represented within the subset of cell lines with predicted IC₅₀ values. Notably, as example of the utility of this approach, using only predicted IC₅₀ values we identified known drug-to-oncogene associations such as sensitivity of BRAF-mutated cells lines to MEK1/2-inhibitors (Fig 4B) [20]. The range of predicted IC₅₀ values for a drug are typically narrower than for the observed values and is likely because currently available genomic dataset are in sufficient to explain the observed range of drug responses across the cell lines.

In addition, we assessed the predictive power of our model for unknown cell lines. Therefore, we applied a more stringent 8-fold cross-validation, where a cell line was either included in the train or test set. These models achieved an R_p of 0.82, R² of 0.68 and an RMSE of 0.89 (Fig S2) demonstrating the accuracy of our model to predict IC₅₀ values for completely new cell lines. In an additional simulation, we left out all cancer cell lines from a specific tissue, e.g. we removed all lung cancer cell lines (106 out of 608 cell lines) and still obtained an R_p of 0.79, R² of 0.61 and RMSE of 0.99 (Fig S3).

Our results show that by using genomic features from the cell lines and chemical information from drugs, it is possible to build in silico multi-drug models to impute missing IC₅₀ values with non-parametric machine learning algorithms such as neural networks and random forests. As output for our method, we chose to explore IC50 values as generated by Garnett et al. [3], which enables us to compare our results to them, however other metrics (such as a capped IC50 or area under the curve), might provide additional insight and potentially lead to more robust models.

The Pearson correlation (Fig. 2A) and coefficient of determination (Fig. 2B) of the multi-drug model are significantly better than the single-drug models, while the RMSE error is similar (Fig 2C). This means that the error (on average) of predicting a given IC50 value is the same in the multi-drug and single-drug models (RMSE) and, since some drugs are active at different concentration ranges, the model is able to cover a much larger dynamic range with a similar precision. The coefficient of determination balances these two terms, and thus a broader range with the same RMSE increases R². Thanks to the use of chemical descriptors, multi-drug models are trained with a volume of data that is two orders of magnitude bigger than the data to train each single-drug model. This larger dataset weights the difficulty in training heterogeneous response values across drugs.

In several instances, the use of multi-drug models permitted the in silico identification of genomic events associated with altered drug sensitivity, which is only possible when genomic properties are considered.

Although our models did not capture all known gene to drug associations, we anticipate that as larger drug sensitivity and genomic datasets become available in coming years the predictive power of these models will increase. We believe that the predictive power of our models is due to the large number of cell lines and broad range of drugs in the GDSC panel that samples intensively the chemical space of common cancer drugs (chemotherapeutic and kinase inhibitors). It remains to be determined how these models will predict completely unknown families of therapeutic agents.

The predictive ability of our methods for individual values is still limited and could be further improved by extending the set of input features with additional layers of molecular characterization of the cell lines, such as basal transcriptional profiles and phosphoproteomic data. These data types have been used to predict drug responses in various contexts [21]–[24]. Another valuable extension could be the inclusion of gene expression data following drug treatment, a powerful in silico resource for predicting treatment outcomes and elucidating compound mode of action [25], [26], as well as a promising gateway to the identification of new drug repositioning opportunities [27]. Additionally, epigenetics data could enhance the prediction capabilities of future methods [28].

Our method uses purely experimental data, but additional predictive power can be expected from including knowledge of the underlying network [29]. It has been shown that the prediction of drug response and mode of action by transcriptional profiling is significantly enhanced when paired with known a priori gene and protein networks [30], [31] and drug similarities have been inferred based on the corresponding in silico predicted impinged pathway [32]. Prior knowledge could also increase the interpretability of the results. Known regulatory relationships between genes and transcriptional data [33] and protein networks [34] can be used to identify deregulated pathways, and be further linked to the genomic alterations that drive them [35], highlighting subnetworks of importance for drug response.

Incorporation of these additional features will require a scheme to prioritize the input features based on their impact on the final trained model. Associations between features and outcomes could be explicitly unveiled by integrating in our models feature selections criteria and dimensionality reduction techniques.

In terms of predictive models, we have used standard machine learning methods (neural networks and random forests), given their flexibility and robustness as predictive models. A fertile ground for further research is investigating the application of other modeling techniques, including linear regression methods (e.g. LASSO, ElasticNets).

Our results also show that one can estimate the accuracy of prediction for different degrees of sparseness in the data, which may have utility when designing experiments where coverage has to be balanced with accuracy. Furthermore, because models are able to predict IC₅₀ on cell lines not screened yet, predictions from these models can be used to decide whether it is worthwhile expanding the panel of cell lines, or rather focus on a few selected ones.

The implications of our results go beyond their utility to optimise the experimental design of drug screenings. Once a model is built, it could be used to systematically test the potential effect of novel drugs in silico, based on their chemical features and similarity. These predictions can help to evaluate the potential activity of new drugs, e.g. from large chemical libraries, to be screened. Furthermore, predictions on clinically approved drugs is expected to reveal candidates for drug repurposing and potentially identify specific disease sub-types that would be most responsive [8]. Although cell lines are not an exact replica of real tumours, comprehensive predictive models such as ours together with expanded genomic and epigenomic datasets may be a good proxy to facilitate the development new therapeutic strategies tailored to individual patients [12].