RamaNet: Computational de novo helical protein backbone design using a long short-term memory generative adversarial neural network

Abstract

<ns4:p>The ability to perform <ns4:italic>de novo</ns4:italic> protein design will allow researchers to expand the variety of available proteins. By designing synthetic structures computationally, they can utilise more structures than those available in the Protein Data Bank, design structures that are not found in nature, or direct the design of proteins to acquire a specific desired structure. While some researchers attempt to design proteins from first physical and thermodynamic principals, we decided to attempt to test whether it is possible to perform <ns4:italic>de novo</ns4:italic> helical protein design ofjust the backbone statistically using machine learning by building a model that uses a long short-term memory (LSTM) generative adversarial network (GAN) architecture. The LSTM-based GAN model used only theφandψangles of each residue from an augmented dataset of only helical protein structures. Though the network’s generated backbone structures were not perfect, they were idealised and evaluated post generation where the non-ideal structures were filtered out and the adequate structures kept. The results were successful in developing a logical, rigid, compact,helical protein backbone topology. This paper is a proof of concept that shows it is possible to generate a novel helical backbone topology using an LSTM-GAN architecture using only theφandψangles as features. The next step is to attempt to use these backbone topologies and sequence design them to form complete protein structures.</ns4:p>

Relevant Applications of Generative Adversarial Networks in Drug Design and Discovery: Molecular De Novo Design, Dimensionality Reduction, and De Novo Peptide and Protein Design

“…There are several studies for de novo peptide and protein design in drug design and discovery using GAN-based approaches, including the LSTM-GAN (Long Short-Term Memory Generative Adversarial Network) structure in peptide design [91], the gcWGAN (Guided Conditional Wasserstein Generative Adversarial Network) structure in peptide folding [92], the DCGAN (Deep Convolutional Generative Adversarial Network) structure in protein backbone design [93], the DCGAN structure in target-specific compounds for cannabinoid receptors [94], the GANDALF (Generative Adversarial Network Drug-tArget Ligand Fructifier) structure in peptide design [95], and the Feedback-GAN structure in antimicrobial peptides [96].…”

“…For example, Sabban and Markovsky [91] suggested that the LSTM-GAN structure, a combination method involving the GAN architecture and long short-term memory units, was able to generate novel helical protein backbone topologies with preferred features in the context of de novo protein design. The LSTM-GAN structure employed a long short-term memory unit in the generative network module and another one in the discriminative network module, where the long short-term memory unit is often utilized in the field of natural language processing [97].…”

Variational Autoencoders for Protein Structure Prediction

“…While it remains more challenging to directly generate a distribution of tertiary structures via generative models, notable efforts have been made [11,22,23,27]. Work in [11] does not provide an end-to-end NN-based framework for PSP, but composes a neural energy function with a Langevin dynamics-based simulator and is able to predict more than one type of structure for a given protein sequence.…”

“…Work in [11] does not provide an end-to-end NN-based framework for PSP, but composes a neural energy function with a Langevin dynamics-based simulator and is able to predict more than one type of structure for a given protein sequence. A body of recent works [22,23,27] represent the state-ofthe-art on generative NN-based frameworks for problems related to PSP. However, these methods, based on the generative adversarial network (GAN) framework, generate structural fragments of an a-priori determined length (32, 64, or 128) that do not in themselves constitute tertiary structures of a given amino-acid sequence.…”

DenseCPD: Improving the Accuracy of Neural-Network-Based Computational Protein Sequence Design with DenseNet

“…In addition to these two generative models, a number of groups have also developed generative models for various purposes in protein design and engineering. [44][45][46][47] Recently, Zhang and coworkers proposed a convolutional neural network for protein sequence design and reached a state-of-the-art accuracy of 42.2% on a test set that shared less than 30% sequence identity with the training set. 48 In this study, we aim to further improve the accuracy of CPD using deep-learning methods.…”