Deep learning approaches for pre-clinical drug discovery

Thesis

Deep learning methods have experienced a revolution, driven by their successful application in fields such as computer vision and natural language processing. In this thesis, we describe several novel methodologies leveraging deep learning for applications to pre-clinical drug discovery.

First, we propose a generative approach to the design of molecular linkers which incorporates basic 3D information. In large-scale tests, we find that our method substantially outperforms a database-based approach, the previous de facto approach for this problem. Through a series of case studies, we demonstrate the application of our approach to scaffold hopping, fragment linking and PROTAC design. We then extend this framework to incorporate physically meaningful 3D structural information, providing a richer prior for the generative process, and also apply our method to molecular elaboration tasks, such as R-group design.

We then turn our attention to predictive modelling, in particular structure-based virtual screening. We find that the advances in convolutional neural networks (CNNs) for general computer vision tasks are applicable to structure based virtual screening. In addition, we propose two techniques to incorporate domain-specific knowledge into this framework. First, we show that limitations in docking necessitate the use of multi-pose scoring and demonstrate the benefits of an average scoring policy. Second, we propose a transfer learning approach to construct protein family specific models, utilising knowledge of the differences between protein families.

Finally, we investigate how a generative approach can be used to improve the training and benchmarks sets employed in structure-based virtual screening. We propose a deep learning method that generates decoys to a user’s preferred specification in order to control decoy bias or construct sets with a defined bias. We show that our approach significantly reduces the bias contained in such sets. We validate that our generated molecules are more challenging for docking-based approaches to separate from bioactive compounds than previous decoys. In addition, we show that CNN-based structure-based virtual screening methods can be trained on such compounds.

Authors: Imrie, FM

• DOI: 10.5287/ora-kqypkmj0m
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: virtual screening; machine learning; drug discovery; compound design; deep learning
• Subjects: Machine learning; Computer aided drug design; Computational chemistry