PRO-LDM: Protein Sequence Generation with a Conditional Latent Diffusion Model

Abstract

AbstractsProtein design with deep-learning based algorithms represents an emerging and highly promising field in molecular biology. Yet it remains a challenging task due to the complexity of protein sequences. Recent developments in deep generative models such as diffusion model have demonstrated impressive performance in protein structure design, yet left much to be desired in creating new sequences. Here we present PRO-LDM: an efficient protein sequence design framework that utilizes diffusion model in the latent space, utilizing the joint autoencoder as the backbone, to capture latent variable distribution and generate meaningful embeddings of protein sequences. PRO-LDM is capable of (1) learning feature representations in natural proteins both at amino-acid level and sequence level; (2) generating new sequences with natural properties and high diversity; and (3) conditionally designing new protein sequences with designated fitness. Our model presents a feasible pathway a generalizable new tool to extract both physical and chemical properties of amino acids, as well as evolutionary information within the protein sequences, for protein sequence design and functional optimization.