Deep learning-enabled design of synthetic orthologs of a signaling protein

Abstract

Evolution-based deep generative models represent an exciting direction in understanding and designing proteins. An open question is whether such models can represent the constraints underlying specialized functions that are necessary for organismal fitness in specific biological contexts. Here, we examine the ability of three different models to produce synthetic versions of SH3 domains that can support function in a yeast stress signaling pathway. Using a select-seq assay, we show that one form of a variational autoencoder (VAE) recapitulates the functional characteristics of natural SH3 domains and classifies fungal SH3 homologs hierarchically by function and phylogeny. Locality in the latent space of the model predicts and extends the function of natural orthologs and exposes amino acid constraints distributed near and far from the SH3 ligand-binding site. The ability of deep generative models to specify orthologous function in vivo opens new avenues for probing and engineering protein function in specific cellular environments.

ProT-VAE: Protein Transformer Variational AutoEncoder for Functional Protein Design

“…SH3 domains have evolved to perform a variety of functions within various organisms by evolving differential binding specificities, resulting in a number of distinct paralogs (i.e., homologous proteins performing different functions within the same species) within the SH3 family. Recent work by Lian et al trained VAEs over an MSA of ~--5,300 SH3 homologs to develop a deep generative model for synthetic SH3 design [20]. The VAE learned an unsupervised three-dimensional latent space embedding in which the natural sequences demonstrated an emergent hierarchical clustering by phylogeny and function.…”

“…Based on these desired criteria, a number of deep learning architectures have been employed for data-driven protein design. Two approaches in particular have received substantial attention: variational autoencoders (VAEs) [15, 16, 20, 21] and transformer-based models [24–31, 35]. VAEs are deep generative models comprising two consecutive neural networks [36, 37]: an encoder compresses the high-dimensional sequence data into a low-dimensional latent space that is then passed to a decoder whose task is to reconstruct the input sequences as the output of the network.…”

“…The loss function can also be augmented with a semi-supervised prediction head that regresses the assayed function of (some fraction of) the training sequences that have been subjected to experimental measurements [39–41]. This functionality enables iterative semi-supervised re-training of the network after conducting experimental synthesis and testing of novel synthetic sequences to update the model with data from non-natural regions of sequence space and also sculpt and induce functional gradients in the learned latent space [20]. The latent space serves as an information bottleneck that furnishes a projection of the sequences onto a low-dimensional manifold spanned by a small number of correlated patterns of amino acid mutations (i.e., “design rules”) that underpin the sequences in the training data.…”

“…The latent space serves as an information bottleneck that furnishes a projection of the sequences onto a low-dimensional manifold spanned by a small number of correlated patterns of amino acid mutations (i.e., “design rules”) that underpin the sequences in the training data. Annotation of the latent space by phylogenetic or phenotypic properties frequently reveal dimensions within the latent space to be correlated with phylogenetic evolution and/or protein function [15, 20]. The interpretable low-dimensional embedding is extremely valuable in both understanding and interpreting the distribution of sequences in the training data and, crucially, in providing a smooth, continuous, low-dimensional embedding to guide interpolative or extrapolative generative decoding of novel sequences designed to optimize desired protein functions [15, 20].…”

“…Annotation of the latent space by phylogenetic or phenotypic properties frequently reveal dimensions within the latent space to be correlated with phylogenetic evolution and/or protein function [15, 20]. The interpretable low-dimensional embedding is extremely valuable in both understanding and interpreting the distribution of sequences in the training data and, crucially, in providing a smooth, continuous, low-dimensional embedding to guide interpolative or extrapolative generative decoding of novel sequences designed to optimize desired protein functions [15, 20]. Despite these attractive properties, VAE models frequently necessitate that sequences be provided as fixed length vectors, which requires sequences to be arranged within multiple sequence alignments (MSAs) [42].…”