Therapeutic enzyme engineering using a generative neural network

Abstract

Enhancing the potency of mRNA therapeutics is an important objective for treating rare diseases, since it may enable lower and less-frequent dosing. Enzyme engineering can increase potency of mRNA therapeutics by improving the expression, half-life, and catalytic efficiency of the mRNA-encoded enzymes. However, sequence space is incomprehensibly vast, and methods to map sequence to function (computationally or experimentally) are inaccurate or time-/labor-intensive. Here, we present a novel, broadly applicable engineering method that combines deep latent variable modelling of sequence co-evolution with automated protein library design and construction to rapidly identify metabolic enzyme variants that are both more thermally stable and more catalytically active. We apply this approach to improve the potency of ornithine transcarbamylase (OTC), a urea cycle enzyme for which loss of catalytic activity causes a rare but serious metabolic disease.

Prediction of designer-recombinases for DNA editing with generative deep learning

“…Changing the encoding to amino acid properties or a learned representation for protein sequences (Alley et al 2019; Rao et al 2019; Wittmann et al 2021) would give RecGen additional information that could help to predict more reliably. Another way to adapt the network would be to change the fully connected neural network layers to convolutional or recurrent network layers, which could further improve the performance (Hawkins-Hooker et al 2021; Giessel et al 2022). While improvements to the algorithm are important, the data used to train the model is likely to be even more critical.…”

“…Thanks to this approach, it is possible to sample from this data distribution to generate new sequences. The algorithms that are most commonly used for protein sequence generation are Generative Adversarial Networks (GANs; (Goodfellow et al 2014; Gupta and Zou 2018; Repecka et al 2021) and Variational Autoencoders (VAEs; (Kingma and Welling 2013; Riesselman et al 2018; Costello and Martin 2019; Davidsen et al 2019; Das et al 2021; Hawkins-Hooker et al 2021; Giessel et al 2022)). In addition to these two, there are also deep learning algorithms for natural language processing that have been adapted for generative modeling.…”

Designing human Sphingosine-1-phosphate lyases using a temporal Dirichlet variational autoencoder

“…Experimental results showed that TDVAE generated a large ensemble of HsS1PL variants with preserved functional features, including the presence of the key catalytic lysine residue. Surprisingly, obtaining these variants did not require training on large sequence datasets or multiple sequence alignment information [21], which are usually difficult to obtain when sequences are highly divergent as S1PL. We then further validated our results by predicting the structure of a subset of variants and performing molecular dynamics simulations to assess enzyme structural stability and integrity; here we found HsS1PL variants to maintain favorable inter-chain contacts to form stable, compact and largely invariant homodimeric complexes.…”

“…However, current methods are usually limited by the number of homolog sequences and the corresponding quality of multiple sequence alignments, along with the availability of known tertiary structures. Recently, instead, deep generative learning has proven to be a viable solution to generate new, unobserved functional proteins [21], either by learning evolutionary constraints from highly curated multiple sequence alignments [22] or directly from protein sequences [23, 24]. However, these methods require a large number of sequences to be effectively trained and extensive computational resources.…”

Learning Epistasis and Residue Coevolution Patterns: Current Trends and Future Perspectives for Advancing Enzyme Engineering

“…The authors pointed out that higher-order coevolutionary effects have the potential to provide a deeper understanding of the structure− function relationship in novel ways. 125 In 2022, Notin et al used ESM-1v, 126 a protein language model, and MSA transformer 127 in the framework of tranception, an approach based on autoregressive transformers and inference-time retrieval for fitness prediction of various proteins and enzymes. 128 They pointed out that one limitation of the approach is that it neglects potentially important epistatic effects.…”