Novel Antimicrobial Peptides Designed Using a Recurrent Neural Network Reduce Mortality in Experimental Sepsis

Abstract

The search and development of new antibiotics is relevant due to widespread antibiotic resistance. One of the promising strategies is the de novo design of novel antimicrobial peptides. The amino acid sequences of 198 novel peptides were obtained using a generative long short-term memory recurrent neural network (LSTM RNN). To assess their antimicrobial effect, we synthesized five out of 198 generated peptides. The PEP-38 and PEP-137 peptides were active in vitro against carbapenem-resistant isolates of Klebsiella aerogenes and K. pneumoniae. PEP-137 was also active against Pseudomonas aeruginosa. The remaining three peptides (PEP-36, PEP-136 and PEP-174) showed no antibacterial effect. Then the effect of PEP-38 and PEP-137 (a single intraperitoneal administration of a 100 μg dose 30 min after infection) on animal survival in an experimental murine model of K. pneumoniae-induced sepsis was investigated. As a control, two groups of mice were used: one received sterile saline, and the other received inactive in vitro PEP-36 (a single 100 μg dose). The PEP-36 peptide was shown to provide the highest survival rate (66.7%). PEP-137 showed a survival rate of 50%. PEP-38 was found to be ineffective. The data obtained can be used to develop new antibacterial peptide drugs to combat antibiotic resistance.

Keywords: LSTM; RNN; antibiotic resistance; antimicrobial peptide; carbapenem-resistance; neural network; peptide design; sepsis.

Figure 1

Comparison of peptide characteristics between the training data (Training, orange), the generated sequences (Sampled, blue), the pseudo-random sequences with the same amino acid distribution as in the training set (Ran, purple), and the manually created hypothetical amphipathic helices (Hel, green). The horizontal dashed lines represent the mean (violin plots) and median (box plots) values; the whiskers extend to the outermost non-outlier data points. Graphs from left to right: Eisenberg hydrophobicity, Eisenberg hydrophobic moment, and sequence length. The figure was generated using the modlAMP’s GlobalAnalysis.plot_summary method in Python [16].

Figure 2

Comparison of probability of survival (%) in each group: control (sterile saline), PEP-36, PEP-38 and PEP-137. The peptides were injected once in a dose of 100 μg 30 min after infection of mice with 6.75 × 10⁸ CFU suspension of a carbapenem-resistant isolate of K. pneumoniae. *—significant differences from the control group using the Kaplan-Meier method and Log-rank (Mantel-Cox) test.

Figure 3

Visual comparison of the structures of LL-37 (yellow; PDB ID: 2K6O) and synthesized peptides: PEP-137—blue, PEP-38—red, PEP-36—green (by AlphaFold).

Figure 4

Initial and final snapshot of peptide-membrane system before and after 225 ns of MD simulation. The membrane is shown in lines (cyan) and the peptides are shown in helical structure (colored according to the type of amino acids). Phosphate atoms of the membrane are shown in gray color.

Figure 5

Top view of the peptide-membrane system (left) at 225 ns of MD simulation showing the unravelling of the helical structure. Phosphate atoms of the membrane are shown in gray color. Side view (right) showing the penetration of peptide at 225 ns.