AbstractMultiscale computational models of heart are being extensively investigated for improved assessment of drug-induced Torsades de Pointes (TdP) risk, a fatal side effect of many drugs. Model-derived metrics (features) such as action potential duration, net charge carried by ionic currents (qNet) and others have been proposed in the past as potential candidates for classifying TdP risk. However, the criteria for selection of new risk metrics are still poorly justified, and they are often trained/tested only on small datasets. Moreover, classifiers built on derived features have thus far not consistently provided increased prediction accuracies compared to classifiers based on in vitro measurements of drug effects on ion channels (direct features). In this paper, we analyze a large population of virtual drugs to examine systematically the sensitivity of several model-derived features. The influence of different ion channels in regulation of the model-derived features is identified using global sensitivity analysis (GSA). Specifically, the analysis points to key differences in the input parameters that affect several model-derived features and the generation of early afterdepolarizations (EAD), thus opposing the idea that these features and sensitivity to EAD might be strongly correlated. We also demonstrate that previously proposed model-derived features could be well fitted by a linear combination of direct features. This well explains the observed comparable performances of classifiers built on direct features and model-derived features. Combining GSA and simple probability analysis, we also show that the odds of any linear metric constructed from direct features to perform as well as qNet is very low. Nevertheless, despite high predictive power of qNet to separate drugs into correct categories of TdP risk, the GSA results suggest that the actual mechanistic interpretation for qNet’s improved performance deserves further investigation. In conclusion, analyses like ours can provide more robust feature selection/construction. Improved experimental designs with increased focus on the critical model parameters indicated by GSA can potentially reduce the uncertainties of key model components and result in increased confidence of TdP risk predicted by in silico models.Author SummaryBiophysical models often have extremely involved intrinsic structure. In the majority of research, either complex methods of non-linear dynamics and empirical analysis are employed to explore the underlying structure of cell processes such as transmembrane ionic currents. Global sensitivity analysis (GSA) could be considered as a brute force alternative to study the model relationships between physical processes, discovering the mechanisms responsible for phenomena of interest. As we demonstrated here, GSA application could be extended to explore the structure of features derived from outputs of biophysical models and used in statistical models to build regressions or classifiers. In particular, GSA seems to be valuable to formalize the methods of feature selection/construction that are used for classification of drugs with respect to their cardiotoxicity.
An Approach for Optimal Feature Selection in Machine Learning using Global Sensitivity Analysis
