The present invention provides methods for the prediction of in vivo cell-specific toxicity of a compound that combines high-throughput imaging of cultured cells, quantitative phenotypic profiling, and machine learning methods. More particularly, the invention provides a method for the prediction of in vivo renal proximal tubular-, bronchial-epithelial-, and alveolar-cell-specific toxicities of asoluble or particulate compound that comprises contacting cultured human kidney and pulmonary cells with the compound at a range of concentrations, then labelling the cells with DNA, yH2AX and actin markers and obtaining textural features, spatial correlation features, ratios of the markers, intensity features, cell count and morphology, estimating dose response curves and performing automatic classification of the compound using a random-forest algorithm.