Intelligent De Novo Design of Novel Antimicrobial Peptides against Antibiotic-Resistant Bacteria Strains

The rise of antibiotic-resistant infections has intensified the need for new antimicrobial agents. AMPs, natural peptides with lower propensity to induce resistance, are promising but traditionally costly and time-consuming to discover. The study proposes an AI-driven, in silico pipeline to accelerate AMP discovery by training a deep generative model on known AMPs to produce novel sequences predicted to have antimicrobial activity. The work leverages advances in AI/ML for drug discovery, including predictive modeling of biological/chemical properties and deep generative methods, to design and prioritize AMP candidates for experimental validation.

Prior computational AMP design includes algorithmic and deep learning approaches. Generative models such as LSTM-based RNNs and variational autoencoders (VAEs) have been used to capture AMP sequence patterns and create novel candidates. Classifiers like AI4AMP and other DNN-based or ensemble predictors have improved AMP activity prediction. GAN-based methods (e.g., DCGAN, BiCGAN in AMPGAN) have shown promise in biomolecule generation, though often with in silico validation. The study builds on WGAN-GP to enhance training stability and mitigate mode collapse, and integrates physicochemical property-aware encoding (PC6) with downstream AI classification to prioritize candidates.

Data collection: 3,195 antibacterial AMPs were compiled from four databases (APD3, LAMP, CAMPR3, DRAMP). Sequences shorter than 10 residues, longer than 30 residues, or containing uncommon amino acids (B, J, O, U, Z, X) were excluded to ensure synthesizability of short peptides. Encoding (PC6): Each peptide was encoded using the PC6 method into a (6,k) matrix capturing six scaled physicochemical properties (scaled to −1 to 1). Sequences shorter than 30 were padded with a zero-vector token "X" to length 30, yielding a real PC6 matrix of shape (1,30,6) per peptide. Model architecture: A DCGAN-based architecture with WGAN-GP loss was used for stability and to reduce mode collapse. Generator: five transposed convolution blocks (first four: ConvTranspose2D + BatchNorm2D + ReLU; final: ConvTranspose2D + tanh) mapping input noise (100×1×1) to (1,30,6). Discriminator: five convolution blocks (first four: Conv2D + leaky ReLU; final: Conv2D) outputting a scalar score. Layer normalization principles of WGAN-GP guided training; Earth-Mover distance with gradient penalty enforced 1-Lipschitz continuity. Training: Generator:Discriminator update ratio 1:5; batch size 128; optimizer Adam with lr=1e-4, β1=0, β2=0.9. Training ran for 60,000 epochs. Every 5,000 epochs, 128 sequences generated from a fixed noise vector were converted back to amino acids using cosine similarity between generated property rows and amino acid property vectors; trailing residues after the first zero-like vector ("X") were removed. Identity to known AMPs was monitored (maximum property-based cosine similarity identity across 3,195 AMPs) and stabilized after ~50,000 epochs. In silico evaluation: To verify non-randomness and AMP-likeness, the team compared amino acid composition and eight physicochemical properties (aliphatic index, aromaticity, Boman index, charge density, net charge, hydrophobic ratio, instability index, isoelectric point) among four groups: real AMPs, GAN-designed peptides, randomly shuffled sequences, and designed helical sequences (generated via modlAMP). t-SNE on the eight properties showed GAN peptides clustering with real AMPs and separate from controls. Candidate selection: From 3,195 GAN peptides, duplicates were removed, yielding 1,970 unique sequences. Filters retained sequences whose eight physicochemical properties lay within mean ± 1 SD of those of real AMPs (computed with modlAMP v4.3.0). Remaining sequences were scored by AI4AMP; those with predicted AMP probability >0.98 were kept. Sequences were categorized by identity to real AMPs: very similar (80–98%), moderately similar (40–60%), dissimilar (<20%). From these, 8 peptides were selected for synthesis: GAN-pep 1–4 (very similar) and GAN-pep 5–8 (moderately similar). Experimental validation: Strains tested included E. coli (SG13009), methicillin-susceptible S. aureus (MSSA; S01-10-0202), methicillin-resistant S. aureus (MRSA; N07-10-0043), carbapenem-susceptible P. aeruginosa (S07-10-0059), and carbapenem-resistant P. aeruginosa (M06-06-0213). Positive control: polyphemusin I; negative control: bovine serum albumin. Assays: disc diffusion (7.8125–500 µg/mL; zones of inhibition) and MIC determination (minimum inhibitory concentration after overnight incubation). All experiments were performed at least in triplicate under aerobic conditions in LB media at 37°C.

- In silico properties: GAN-designed peptides matched real AMPs in amino acid composition and distributions of eight key physicochemical properties; t-SNE showed GAN peptides clustering near real AMPs and distinct from random/helical controls. - Disc diffusion: Multiple GAN peptides inhibited growth of E. coli, MSSA, MRSA, carbapenem-susceptible and -resistant P. aeruginosa. GAN-pep 3 and GAN-pep 8 showed the broadest inhibitory spectra across all tested bacteria. - Sequence similarity: GAN-pep 3 showed ~60% identity to a Cecropin A–melittin hybrid (ABB29918.1). GAN-pep 8 had no significant BLASTP hits (E < 1), suggesting novelty. - MICs (µg/mL) summary (selected): • E. coli (SG13009): Polyphemusin I 0.7; GAN-pep 2 = 2; GAN-pep 3 = 2; GAN-pep 4 = 2; GAN-pep 5 = 22.5; GAN-pep 8 = 15; GAN-pep 1,6,7 >50. • S. aureus MSSA (S01-10-0202): GAN-pep 3 = 6; GAN-pep 8 = 15; others >50 (polyphemusin I >50). • S. aureus MRSA (N07-10-0043): GAN-pep 3 = 45; GAN-pep 8 = 45; others >50 (polyphemusin I >50). • P. aeruginosa carbapenem-susceptible (S07-10-0059): GAN-pep 3 = 3; GAN-pep 2 = 50; GAN-pep 4 = 50; GAN-pep 8 >50; others >50. • P. aeruginosa carbapenem-resistant (M06-06-0213): GAN-pep 3 = 3; GAN-pep 2 = 5; GAN-pep 4 = 35; GAN-pep 8 >50; others >50. - Overall: 7/8 GAN-designed peptides showed antimicrobial activity against at least one strain. GAN-pep 3 was most potent and broad-spectrum, with low MICs across Gram-negative and Gram-positive strains, including resistant MRSA and CRPA.

The study addresses the challenge of discovering new antimicrobials by demonstrating that a WGAN-GP trained on physicochemical encodings of known AMPs can learn AMP-relevant sequence and property distributions, generating candidates with similar profiles to real AMPs. The pipeline, combining GAN generation, physicochemical filtering, and AI4AMP activity prediction, effectively prioritized a small set for synthesis, of which most were active. The strong performance of GAN-pep 3 (low MICs across E. coli, MSSA, MRSA, and both carbapenem-susceptible and -resistant P. aeruginosa) highlights the feasibility of AI-guided de novo peptide design to counter antibiotic resistance. The findings support that property-aware generative models can propose novel, sometimes dissimilar sequences (e.g., GAN-pep 8) with practical antibacterial activity, thereby expanding the chemical space beyond known AMPs.

An AI-guided de novo design approach using WGAN-GP and PC6 encoding efficiently generated short AMP candidates. After in silico screening and wet-lab validation, 7 of 8 synthesized peptides exhibited antibacterial activity. GAN-pep 3 and GAN-pep 8 showed broad-spectrum effects, including against MRSA and CRPA, with GAN-pep 3 demonstrating the most favorable MIC profile. This framework can accelerate AMP discovery and is adaptable to designing peptides with other functions (antiviral, antifungal, anticancer). Future work should integrate additional predictive models (e.g., hemolysis, Gram specificity, species-specific MIC) and expand experimental validation to safety and efficacy studies to advance candidates toward therapeutics.

- Limited experimental scope: only eight GAN-designed peptides were synthesized and tested, restricting generalizability of success rates. - Assessed in vitro only: no in vivo efficacy, stability, or pharmacokinetic/toxicity data. - Safety not evaluated: hemolysis and cytotoxicity assays were not performed, acknowledged as a necessary next step. - Narrow organism panel: five strains across three species; broader species/strain diversity would strengthen conclusions. - Sequence novelty/variability: selection biased toward sequences within physicochemical ranges of known AMPs and with high AI4AMP scores; fully dissimilar candidates were not synthesized, potentially limiting exploration of novel sequence space.