Predicting mitochondrial functional state : An imaging based SVM tool

Mitochondria are dynamic organelles, shown to provide signatures of onset of diseases or cellular stresses. In this work we attempt to probe whether specific patterns exist for stress in mitochondria which is assessed though quantitative imaging followed by machine learning. Feature extraction is performed to assess mitochondrial morphology, intensity and intra-mitochondrial structural heterogeneity features from confocal micrographs of mitochondria stained with micro-viscosity sensing and potentiometric dyes. These features examined from the acquired con-focal images of cells with mitochondria in apriori known perturbed (or stress) or unperturbed functional state. Perturbation dependent signatures in these assessed features were studied by employing supervised learning - Support Vector Machine. The classification between perturbed and unperturbed states was performed with an accuracy of nearly 93%. Furthermore, accuracy improvement methods were engineered by optimising the feature space through which an optimal classification accuracy of 98% was achieved and accounted the importance of intra mitochondrial heterogeneity. Overall, the derived features showed the presence of computationally identifiable unique functional state dependent patterns strengthening the way for assembling predictive models for assaying mitochondrial functional states. This could assist in developing less resource-intensive method to classify and study stress conditions. This tool provides a promising application of mathematics in biology that could be applied to characterize stress conditions, drug screening to aid the identification of functional stress state.