Deep learning enables the design of functionalde novoantimicrobial proteins

Abstract

Protein sequences are highly dimensional and present one of the main problems for the optimization and study of sequence-structure relations. The intrinsic degeneration of protein sequences is hard to follow, but the continued discovery of new protein structures has shown that there is convergence in terms of the possible folds that proteins can adopt, such that proteins with sequence identities lower than 30% may still fold into similar structures. Given that proteins share a set of conserved structural motifs, machine-learning algorithms can play an essential role in the study of sequence-structure relations. Deep-learning neural networks are becoming an important tool in the development of new techniques, such as protein modeling and design, and they continue to gain power as new algorithms are developed and as increasing amounts of data are released every day. Here, we trained a deep-learning model based on previous recurrent neural networks to design analog protein structures using representations learning based on the evolutionary and structural information of proteins. We test the capabilities of this model by creating de novo variants of an antifungal peptide, with sequence identities of 50% or lower relative to the wild-type (WT) peptide. We show by in silico approximations, such as molecular dynamics, that the new variants and the WT peptide can successfully bind to a chitin surface with comparable relative binding energies. These results are supported by in vitro assays, where the de novo designed peptides showed antifungal activity that equaled or exceeded the WT peptide.

Deep generative models for peptide design

“…(3) Incorporation of peptide structure information into deep generative models. While peptide structures can provide mechanistic information to better guide the model to generate peptides with desired functions, the majority of deep generative models (except Caceres-Delpiano et al 54 and Rossetto et al 48 ) in peptide design have only used sequence information. We hypothesize that the dynamic and flexible nature of peptide structures make them difficult to be inputted into deep generative models.…”

“…For instance, Rossetto et al 48 used a 4D tensor Schemes commonly used to represent peptides in the generation task itself include direct sequence representation 41,42,45,46,[49][50][51][52][53] and learned embeddings. 38,39,44,45,54,55 A natural way to encode peptides is through their primary structure (i.e., amino acid sequence). A peptide of length L can be represented by a string of characters or integers of length L, 39,45,46,49,50 or a L Â n matrix such that each amino acid has a unique n-dimensional vector.…”

“…In addition to these approaches, Capecchi et al 42 utilized a transfer learning strategy to fine tune the RNN model trained on DBAASP 73 to two smaller non-hemolytic AMP datasets in order to address the scarcity of non-hemolytic AMP data. Taking a different approach, Caceres-Delpiano et al 54 adopted a multitask training strategy to simultaneously perform language modeling, secondary structure, contact map, and structure similarity prediction. After model training, novel peptide sequences were generated by mutating the target sequence and after meeting certain structure similarity and energy criteria based on the trained multitask model.…”

“…After model training, novel peptide sequences were generated by mutating the target sequence and after meeting certain structure similarity and energy criteria based on the trained multitask model. Through wet-lab validation, Nagarajan, 52 Capecchi et al , 42 and Caceres-Delpiano et al 54 managed to identify two, eight, and two novel AMPs, respectively. Grisoni 46 also used transfer learning to fine tune the RNNs trained on 10K α -helical cationic amphipathic peptide sequences to 26 known anticancer peptides (ACPs).…”

Using AlphaFold for Rapid and Accurate Fixed Backbone Protein Design

“…A variety of neural network architectures have been used including variational autoencoders (Greener et al, 2018;Hawkins-Hooker et al, 2021), deep exploration networks (Linder et al, 2020), graph neural networks (Strokach et al, 2020), recurrent neural networks (Alley et al, 2019) and autoregressive models (Shin et al, 2021;Trinquier et al, 2021). Ultimately the hope is that faster and more accurate protein design with deep learning will lead to the design of functional proteins (Tischer et al, 2020;Caceres-Delpiano et al, 2020).…”