De Novo Protein Design using Generative Machine Learning
Year: 2024
Type: dissertation
Abstract: In this thesis, methods are developed for computationally designing novel protein \nsequences and structures using deep generative machine learning algorithms. It is \ndivided into six chapters. \nChapter 1 provides an introduction to contemporary protein design, generative \nmachine learning, and the nascent field emerging from the intersection of the two. \nThis particularly focuses on the challenge of de novo protein design and the \ndemonstrated uses of generative deep learning methods. \nChapter 2 describes the first study performed for this thesis. It explores the use \nof variational autoencoders for protein sequence design with two separate in silico \ndesign tasks, one functional and one structural, respectively. Being one of the \nearliest works on this topic, it provided evidence that generative approaches had \nmerit as a focus for further design investigations. \nChapter 3 encompasses an analysis of the performance of state-of-the-art \ndeep protein structure prediction algorithms, primarily the AlphaFold method, on \npreviously de novo designed protein sequences. As expected, it finds AlphaFold is \nable to accurately and confidently predict the structures of these proteins, \nsupporting its use as a tool in the development of future design methods. \nChapter 4 explores the development of protein sequence language models \ntrained on synthetic sequences to avoid the detriments of training with natural \nsequences. It also describes the evaluation of generated sequences with \nstate-of-the-art structure predictor AlphaFold. \nChapter 5 describes a technique for fixed-backbone protein design using \ngreedy sequence optimization of AlphaFold structure predictions that leverages the \nmodels developed in the previous chapter. Initial in vitro validation of a small \nnumber of designed sequences provides optimistic signs of success. \nThe final chapter highlights the key contributions of this thesis to the field of \ncomputational protein design and concludes with implications for future design \nmethod development.
Author: Moffat, Lewis Iain
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7055885338