Learning from biological context for protein fitness estimation and design

点击翻译按钮获取中文摘要

Proteins, sequences of amino acids, are the fundamental building blocks of life, serving as the workhorses of cellular processes. Their diverse functions, from catalysing chemical reactions to providing structural support, underpin the complex machinery of living organisms. Evolution has conducted a massive experiment over the space of all possible amino acid sequences: those that encode a functional protein survive; those that don't are extinct. By looking at a set of related proteins throughout life, we can begin to understand its evolutionary history, key to unlocking numerous advancements in medicine, biotechnology, and various fields of biology. This project falls within the EPSRC Artificial Intelligence Technologies research area and is co advised by Professor Debbie Marks at Harvard University.Computational biologists have used more and more complex statistical models to analyse protein evolution. Extending the models to the whole proteome has recently been made possible by large protein language models, that aim to uncover the language of life. These models have been shown to recapitulate protein evolution, or phylogeny, even when the set of related homologs is small. The project aims to leverage new methodologies from the fields of in-context learning and non-parametric modelling to protein statistical modelling. In context learning allows models to learn from context, for example by the addition of a few examples. Nonparametric modelling allows the model to learn from explicit data points instead of having to memorize an entire dataset in its parametrised weights. By combining these methods, the aim is to better leverage and retrieve the context provided by protein evolution, to improve performance at same compute cost. These methods will allow to better study unalienable protein sequence such as disordered regions or antibodies, as well as to model insertion and deletion of sequences. The development of such method allows to both quantify pathogenicity of a given protein sequence, to diagnose disease as well as optimize a sequence for its function, with relevance to bioengineering of proteins for chemical processes and drug development.