AMPGAN v2: Machine Learning-Guided Design of Antimicrobial Peptides

Abstract

Antibiotic resistance is a critical public health problem. Each year ∼2.8 million resistant infections lead to more than 35 000 deaths in the U.S. alone. Antimicrobial peptides (AMPs) show promise in treating resistant infections. However, applications of known AMPs have encountered issues in development, production, and shelf-life. To drive the development of AMP-based treatments, it is necessary to create design approaches with higher precision and selectivity toward resistant targets. Previously, we developed AMPGAN and obtained proof-of-concept evidence for the generative approach to design AMPs with experimental validation. Building on the success of AMPGAN, we present AMPGAN v2, a bidirectional conditional generative adversarial network (BiCGAN)-based approach for rational AMP design. AMPGAN v2 uses generator-discriminator dynamics to learn data-driven priors and controls generation using conditioning variables. The bidirectional component, implemented using a learned encoder to map data samples into the latent space of the generator, aids iterative manipulation of candidate peptides. These elements allow AMPGAN v2 to generate candidates that are novel, diverse, and tailored for specific applications, making it an efficient AMP design tool.

Deep generative models for peptide design

“…Therefore, WGANs are widely used in various kinds of data generation tasks. 88,98,99 Regarding peptide generation, several attempts have been made to utilize GANs and their variants to generate antimicrobial, 39,40,49,50 anticancer, 48 and immunogenic 51 peptides. For example, upon training on 16K AMPs and 5K non-AMPs acquired from several datasets, PepGAN 50 designed a mixed loss function to utilize AMP label information so that the discriminator was encouraged to distinguish real AMP sequences.…”

“…Wang et al 41 Tran et al 53 Capecchi et al 42 Das et al 45 Tucs et al 50 Ferrell et al 39 Oort et al 40 ToxinPred's dataset Gupta et al 118 Toxicity 1805 toxic peptides Das et al…”

“…Through experimental validation, the authors further showed that PepGAN was able to generate an AMP twice as strong as the conventional antibiotic ampicillin. In addition, AMPGAN 39 and the follow-up model AMPGAN v2 40 adopted a bidirectional CGAN framework to generate peptides with desired targets, mechanisms, and minimum inhibitory concentration (MIC) values. By introducing an encoder neural network to map peptide sequence to latent representations, the learnt latent space can be more structural and consequently leads to improved data modeling.…”

Machine Learning in Antibacterial Drug Design

“…The group also highlighted three important peptide descriptors as essential for antibacterial activity, namely, charge, polarity, and pseudo-amino acid composition ( Wani et al, 2021 ). The field of in silico tools for designing antibacterial peptides using machine learning is also gaining traction, and targeted tools such as AMPGAN v2 are being developed ( Van Oort et al, 2021 ). AMPGAN v2 is a bidirectional conditional generative adversarial network (BiCGAN) that targets de novo generation of antibacterial peptides.…”

Emerging Computational Approaches for Antimicrobial Peptide Discovery

“…On the contrary, those attribute-controlled models based on recurrent neural networks, variational autoencoders, adversarial autoencoders, generative adversarial networks may encourage novelty of designed sequences. That is the case of the specific bidirectional conditional generative adversarial network developed in AMPGAN v2 [ 94 ] that learns data driven priors through generator-discriminator dynamics and controls generation using conditioning variables. Thus, a learned encoder mapping data samples into the latent space of the generator implements the bidirectional component that aids iterative manipulation of novel, diverse, and application-tailored candidate peptides.…”