题目：Latent generative landscapes as maps of functional diversity in protein sequence space
Literature source: Nature Communications | (2023) 14:2222
Code: https://github.com/morcoslab/LGLVAE/
Introduction: Variational autoencoders are unsupervised learning models with generative capabilities that, when applied to protein data, classify sequences through phylogenetics and generate de novo sequences that preserve statistical properties of protein composition. While previous studies have focused on clustering and generating features, here the authors evaluate latent manifolds that embed sequence information. To investigate the properties of the latent manifold, they utilized direct coupled analysis and a Potts-Hamiltonian model to construct a latent generative landscape. Here we show how this landscape captures the phylogenetic grouping, function, and fitness properties of multiple systems, including globulins, β-lactamases, ion channels, and transcription factors. The authors provide support for how landscapes can help understand the effects of sequence variability observed in experimental data and provide insights into directed and natural protein evolution. We propose that combining the generative properties and functional prediction capabilities of variational autoencoders, together with coevolutionary analysis, may benefit applications in protein engineering and design.
main content:
-------------------------------------------
Welcome to like, collect and forward!
See you next time!