11/16(月) RIMD セミナー： Dr. Spyros Lytras & Dr. Liam Brierley

大阪大学 微生物病研究所本館1F 微研ホール（吹田キャンパス）

演者

Spyros Lytras（Antigen Evolution and Design Laboratory, Department of Structural Biology and Chemistry, Institut Pasteur）

演題

Leveraging structural predictions to understand virus glycoprotein evolution on the macro and micro-scale

概要

Viral glycoproteins mediate membrane fusion, dictate host range and facilitate virus immune evasion. The proteins’ important roles paired with the fast evolutionary rate of viruses lead to rapid accumulation of changes in their sequences, hindering homology detection over deep evolutionary time. Recent AI-based methods such as AlphaFold and ESMFold have enabled fast and accurate prediction of protein structures from sequence alone. On the macro-scale, we leverage these structural prediction methods at scale to infer the evolutionary histories of viral glycoproteins based on both sequence but also structural similarity between them. Applied to large virus taxa such as the Flaviviridae and the Coronaviridae, this approach allows us to uncover the wide diversity and complex evolutionary origins of these proteins. On the micro-scale, we can predict co-structures of glycoproteins together with host proteins they are expected to interact with, such as entry receptors. By assessing the prediction confidence of glycoprotein-receptor co-structures for closely related viruses and receptor orthologs we can infer the hosts that each virus strain can infect as well as the interacting residues controlling the ability to infect. These computational predictions are consistent with experimental validation of receptor-based virus infectivity, showcasing the usefulness of structural predictions in assessing the properties of novel viruses. Overall, structural predictions can be instrumental for uncovering both the deep scale evolution of virus glycoproteins, obscured by accumulated diversity, but also the phenotypic effects of even single mutations that disrupt protein-protein structural interactions.

演者
Liam Brierley（MRC-University of Glasgow Centre for Virus Research, University of Glasgow）

演題
Biologically meaningful machine learning to predict host-virus compatibility and zoonotic threats

概要
Virological data about cross-species transmission is challenging to make use of. It's often nonlinear, multimodal and heavily structured.  Machine learning gives us a way to use these diffuse signals to make genotype-to-phenotype predictions, but models must be tailored to the biological nuances. I'll discuss recent computational work from our group that aims to make better inference of which viruses might infect which hosts in future, including generalisable model ensembles to predict zoonotic status from avian influenza virus sequences. I will also introduce graph-based AI methods that can use richly structured networks to enhance training of models to predict viral hosts enhanced by phylogeny and models to predict viral tissue tropisms enhanced by gene expression networks.