During the last decades the amount of chemical and biological data is rapidly growing up due to a progression in the screening campaign and especially in High-Throughput Screening (HTS) techniques. As results, the several web sites, including ChEMBL and PubChem, have been established that aggregate data from published screening and medicinal chemistry efforts. This comprehensive mapping of chemical biology space enables the development of large-scale quantitative structure-activity relationships (QSAR) modeling. It was previously shown that aggregated results from the different related biological targets could improve the quality of QSAR models. Thus, the application of these models lead to acceleration of the translational science and drug discovery process. The two most successful approaches in this field are Protocemometrics (PCM) and Multitask Deep Learning (MDL). In this study, we have developed and rigorously validated QSAR models based on both PCM and MDL techniques together with “classical” QSAR methods. To compare the approaches we collected and generated the following data sets: i) 659 data sets related to the different targets (receptors or enzymes) were created based on ChEMBL data; ii) 47 data sets related to different toxicity assays were created using Tox21 data. For the PCM approach the QSAR models were based on the following descriptor sets: i) chemical based descriptors: QNA (Quantitative Neighbourhoods of Atoms) and RDkit; ii) target based descriptors: PROFEAT. Random Forest, Generalized Boosted Machine and Deep Learning were used to construct the PCM model and “classical” QSAR models. For MDL and “classical” approaches only chemical based descriptors were used (QNA and RDkit). The models obtained were compared using external validation sets. In our study, MDL-based models showed the better performance compared to PCM and “classical” QSAR approaches. The obtained MDL models were incorporated into publicly available NCATS web services at https://tripod.nih.gov/predx.

Large-scale QSAR modeling: Proteochemometrics vs. multitask deep learning

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.