An important computational tool in drug design is target prediction where either for a given chemical structure the interacting biomolecules (e.g. proteins) must be identified. Chemical structures interact with different biomolecules if they have similar 3D structure. Thus, the outputs of the prediction are highly interdependent from each other. Furthermore, we have partially labelled molecules since not all training molecules are measured of being active on each biomolecule. The Merck Kaggle challenge on chemical compound activity was won by Hinton?s group with deep networks. This indicates the high potential of deep learning in drug design and attracted the attention of big pharma. However, the unrealistically small scale of the Kaggle dataset does not allow to assess the value of deep learning in drug target prediction if applied to in-house data of pharmaceutical companies. Even a publicly available drug activity data base like ChEMBL is magnitudes larger than the Kaggle dataset. ChEMBL has 13 M compound descriptors, 1.3 M compounds, and 5 k drug targets, compared to the Kaggle dataset with 11 k descriptors, 164 k compounds, and 15 drug targets. On the ChEMBL database, we compared the performance of deep learning to seven target prediction methods, including two commercial predictors, three predictors deployed by pharma, and machine learning methods that we could scale to this dataset. Deep learning outperformed all other methods with respect to the area under ROC curve and was significantly better than all commercial products. Deep learning surpassed the threshold to make virtual compound screening possible and has the potential to become a standard tool in industrial drug design.
Sprache der Kurzfassung:
Neural Information Processing Systems Foundation (NIPS 2014)
Workshop on Machine Learning in Computational Biology