Machine learning for drug discovery and beyond

Qian, Wei

Permalink

https://hdl.handle.net/2142/116180 Copy

Description

Title

Machine learning for drug discovery and beyond

Author(s)

Qian, Wei

Issue Date

2022-07-06

Director of Research (if dissertation) or Advisor (if thesis)

Peng, Jian

Doctoral Committee Chair(s)

Peng, Jian

Committee Member(s)

• El-Kebir, Mohammed
• Han, Jiawei
• Wiltschko, Alex

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• cheminformtics
• drug discovery
• computational biology
• bioinformatics
• machine learning
• deep learning
• artificial intelligence

Abstract

The advent of digitized, large-scale, and high-throughput technologies has generated unprecedented data, presenting an excellent opportunity for today's drug discovery program to leverage machine learning (ML). By identifying relevant problems and suitable formulations in ML, we can translate these ever-increasing data into discovering better drugs and shorten the drug development cycles leading to cheaper drug and therapeutic options for previously incurable diseases. In this dissertation, we present four ML methods to tackle different challenges in today's drug discovery pipeline to quickly deliver more viable drug candidates for the clinical trial and eventually improve the quality of life for all humans. First, we introduce a batch equalization method that leverages style-transfer generative adversarial networks to mediate the batch effect commonly found in cellular images such that we can use them more effectively for high-throughput in vitro screening. Second, we describe an energy-inspired SE(3)-equivariant model to efficiently and accurately estimate the distribution of molecular conformations such that we can improve the accuracy for in silico structured-based screening. Third, we propose a 3D full-atom diffusion framework for target-aware molecule generation such that we can explore new chemistry beyond existing screening libraries and propose novel drug candidates for binding targets of challenging diseases. Fourth, we describe a reaction prediction algorithm that brings together rule-based systems (integer linear programming) and data-driven approaches (graph neural network) such that we can use efficiently synthesize drug candidates from the described screening pipeline or generative models. In the end, we use a graph neural network to model odorant molecules (instead of drugs) and find a universal odor space shared by many species. We hypothesize that the biology of metabolism drives such convergent evolution, and our ability to model these volatile organic compounds related to different metabolic processes could have great implications on how we understand animal olfaction and study human health. Put together, this dissertation shows the potential of machine learning to transform drug discovery and human health in the era of big data.

Graduation Semester

2022-08

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Wei Qian

Owning Collections