The escalating global health crisis of antibiotic resistance demands innovative approaches to antimicrobial drug discovery. Antimicrobial peptides (AMPs), naturally occurring peptides with inherent resistance-breaking properties, offer a promising avenue for next-generation antibiotics. Traditional AMP discovery methods, however, are time-consuming and costly. This study addresses this challenge by employing a deep learning-based approach to design novel AMPs in silico. Artificial intelligence (AI) and machine learning (ML) are rapidly transforming drug development, enabling predictions of diverse biological and chemical properties, including MIC, and the construction of biomolecule classifiers for identifying AMPs. Existing in silico methods for AMP design encompass various computational algorithms and deep neural networks (DNNs), including generative long short-term memory models, variational autoencoders (VAEs), and generative adversarial networks (GANs). This research builds upon these advancements, focusing on the application of a WGAN-GP, a modified GAN architecture designed to overcome limitations of the original GAN, such as mode collapse, to generate novel AMP candidates. The improved stability and performance of WGAN-GP make it particularly suitable for the complex task of peptide sequence generation.

The study extensively reviewed existing literature on AMP design and AI/ML applications in drug discovery. Several studies have successfully employed various machine learning techniques such as knowledge graph embedding, convolutional graph networks, and DNNs to predict biological properties, design molecules and even develop novel drug candidates. Existing AMP design methods include algorithms like Joker for inserting patterns into sequences and deep learning models like those utilizing recurrent neural networks and variational autoencoders for generating peptides. The authors specifically noted the AI4AMP model, a CNN for predicting AMP activity which is used in this study to evaluate generated sequences. Further, prior work exploring the use of GANs, particularly DCGAN, for DNA and protein design, informed the choice of WGAN-GP in this research. The review emphasized the limitations of previous approaches, particularly the issue of mode collapse in GANs, which the current study aims to address.

This study employed a WGAN-GP, a type of GAN, trained on a dataset of 3195 AMPs collected from four AMP databases. Peptides shorter than 10 amino acids or containing uncommon amino acids were excluded, and only those shorter than 30 amino acids were selected to manage synthesis challenges. The PC6 protein-encoding method was used to transform each peptide sequence into a numerical matrix representing six physicochemical properties, enabling the GAN to learn from the data. The WGAN-GP model consisted of a generator and a discriminator. The generator produced new AMP candidates from random noise input, while the discriminator learned to distinguish between real and generated AMP sequences. The WGAN-GP loss function, which includes a gradient penalty, was used to improve the stability and reduce mode collapse during training. The training process used a batch size of 128 and the Adam algorithm as optimizer, with a learning rate of 1 × 10⁻⁴ and a training step ratio of 1:5 (generator:discriminator). Every 5000 epochs, the generator's output was evaluated by assessing the identity scores—cosine similarity based on physicochemical properties—between generated peptides and known AMPs in the training dataset. Training proceeded for 60,000 epochs. Following training, 1970 unique GAN-designed peptides remained. Eight candidates (GAN-pep 1-8) were selected based on their physicochemical properties (being within one standard deviation of known AMPs' properties) and high probability of antimicrobial activity (predicted by AI4AMP to have >0.98 probability). These were then synthesized and subjected to in-vitro testing.

The in silico analysis of generated peptides demonstrated that the WGAN-GP model successfully captured the patterns of real AMPs, replicating their amino acid distribution and physicochemical properties. The generated peptides clustered closely with real AMPs in t-SNE plots, distinct from random or helical sequences. In vitro experiments, using disc diffusion assays and MIC determinations, revealed that seven out of eight synthesized GAN-designed peptides exhibited antibacterial activity. GAN-pep 3 and GAN-pep 8 displayed particularly promising results, exhibiting broad-spectrum activity against both Gram-negative and Gram-positive bacteria, including antibiotic-resistant strains like methicillin-resistant Staphylococcus aureus (MRSA) and carbapenem-resistant Pseudomonas aeruginosa (CRPA). Specifically, GAN-pep 3 demonstrated consistently low MICs against all tested bacteria strains, outperforming the positive control (polyphemusin I) in several instances. Sequence alignment revealed that GAN-pep 3 exhibited approximately 60% similarity to a Cecropin A-melittin hybrid protein, suggesting a potential structural basis for its antimicrobial activity. Notably, GAN-pep 8 did not show significant similarity to any existing AMPs in the database (BLASTP, E > 1). The study provided a detailed MIC table quantifying the effectiveness of these peptides against various strains of bacteria, with GAN-pep 3 consistently displaying low MIC values, indicating potent antibacterial activity.

The results of this study demonstrate the effectiveness of the proposed AI-guided approach for designing novel AMPs with potent activity against both Gram-negative and Gram-positive bacteria, including antibiotic-resistant strains. The successful generation of peptides with distinct properties, some similar to known AMPs and some showing novel sequences, highlights the ability of the WGAN-GP model to capture the essential features of antimicrobial activity. The low MICs observed for GAN-pep 3 suggest its potential as a highly effective antimicrobial agent. This research contributes to the development of efficient and cost-effective strategies for identifying novel antibiotics. The approach is not limited to AMPs; it could be applied to design peptides with different biological functions, accelerating drug discovery for various therapeutic applications.

This research successfully employed a WGAN-GP to generate novel AMPs with high in-vitro antibacterial activity, addressing the critical need for new antibiotics to combat antibiotic resistance. GAN-pep 3 emerged as a particularly promising lead compound due to its low MICs across multiple bacterial species and strains. The methodology described provides a scalable platform for future drug design, applicable beyond AMPs to a variety of therapeutic peptides. Future work should focus on expanding the predictive models to include factors such as hemolysis to further refine the design process and ensure the safety of the generated peptides.

The study's limitations include the limited number of peptides selected for in-vitro testing. A larger scale synthesis and testing of generated peptides could provide a more comprehensive evaluation of the model's potential. Further, the current in-vitro assessment does not include tests for hemolytic activity, which is a crucial aspect for assessing the safety profile of potential drug candidates. While the AI4AMP model was used to pre-select candidates, further validation of the generated peptide's activity in diverse contexts and against a broader range of bacterial strains is needed. In addition, while the paper discusses similarity to some known AMPs for GAN-pep 3, further investigation into the mechanism of action is warranted for both GAN-pep 3 and GAN-pep 8 to fully understand their activity.