DEEP LEARNING-BASED COMPOUND ACTIVITY INFERENCE USING PHENOTYPIC PROFILES and CHEMICAL STRUCURES

Predicting compound activity using phenotypic profiles and chemical structures can reduce the cost of developing new drugs. A dataset containing information on the binary activities of 16170 molecules for 100 different target proteins was used to evaluate the models. We found that deep learning models outperform classical machine learning methods on all quality metrics for each modality. The feasibility of a combined model was also investigated, showing significant improvements in prediction accuracy and reliability. Our results indicate the significant potential of applying deep learning in conjunction with chemical structures and phenotypic profiles of molecules to enhance activity prediction. Further research in this direction could lead to the development of even more effective methods for predicting molecule activity, thereby accelerating the drug discovery process.