Deep neural networks (DNNs) have become a very popular machine learning technique with wide applications in the computer vision and speech recognition fields. DNNs rose to prominence in the chemoinformatics field after they were applied in the winning submission of Merck’s Kaggle competition in 2012. Different deep learning methodologies have since been applied to predict chemically relevant endpoints such as activity, drug induced liver injury, and solubility. DNNs have shown promise even when the datasets are not very large, a commonly stated requirement for the applicability of deep learning techniques. In this work, we have explored the use of DNNs to support phenotypic assay campaigns to predict both activity in a specific screen and to suggest which target the compound is hitting. Data related to the biological pathway targeted in the screen was extracted from public databases. To generate a proof-of-concept case phenotypic screening data was extracted from PubChem. Based on the assay description, biological pathway models were obtained from WikiPathways and refined using literature sources. Bioactivity data against targets found in the model were extracted from ChEMBL. Additionally, DNNs were compared to commonly used machine learning techniques. We expect this technique to provide useful guidance during phenotypic screening campaigns.

Using deep neural networks with heterogeneous chemical data to support phenotypic assay campaigns

254th National Meeting (2017)

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