Deep Neural Network Model for Predicting Gene Activity Using Three-Dimensional Structures of Chemical Compounds

• Activity Number: 241 - SLDS CPapers New
• Type: Contributed
• Date/Time: Monday, July 30, 2018 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Pingzhao Hu* and Md. Mohaiminul Islam and Kevin Jeffers and Andrew M Hogan and Rebecca Davis and Silvia Cardona
• Companies: University of Manitoba and University of Manitoba and University of Manitoba and University of Manitoba and University of Manitoba and University of Manitoba
• Keywords: deep neural network; three-dimensional structure; chemical compounds; drug discovery; gene activity

Abstract

Experimental approaches to drug discovery are time-consuming and expensive. It is well-known that three-dimensional (3D) structures of chemical compounds contain rich information for drug screening. Therefore, it is critical to develop new models to measure compound structure-activity relationships. To solve this issue, we first developed an algorithm to extract compound structure-specific features from atomic coordinates of conformers created on a specific molecular conformation. A denoising stacked autoencoder model was then proposed to generate deep features. The network was built by stacking layers of denoising autoencoders in a convolutional way. Chemogenetic interactions were then predicted using a support vector machine based on the learned high-level feature representations of the 3D structures of the compounds. The models were evaluated using 59 compounds with 6413 conformers and 242 gene products generated by a chemical genomics strategy for mechanism-based profiling of antibacterial compounds. We demonstrated that the proposed model has excellent performance to classify chemogenetic interactions using the structure features extracted from the chemical compounds.