Rational Discovery of Antimicrobial Peptides by Means of Artificial Intelligence

Abstract

Antibiotic resistance is a worldwide public health problem due to the costs and mortality rates it generates. However, the large pharmaceutical industries have stopped searching for new antibiotics because of their low profitability, given the rapid replacement rates imposed by the increasingly observed resistance acquired by microorganisms. Alternatively, antimicrobial peptides (AMPs) have emerged as potent molecules with a much lower rate of resistance generation. The discovery of these peptides is carried out through extensive in vitro screenings of either rational or non-rational libraries. These processes are tedious and expensive and generate only a few AMP candidates, most of which fail to show the required activity and physicochemical properties for practical applications. This work proposes implementing an artificial intelligence algorithm to reduce the required experimentation and increase the efficiency of high-activity AMP discovery. Our deep learning (DL) model, called AMPs-Net, outperforms the state-of-the-art method by 8.8% in average precision. Furthermore, it is highly accurate to predict the antibacterial and antiviral capacity of a large number of AMPs. Our search led to identifying two unreported antimicrobial motifs and two novel antimicrobial peptides related to them. Moreover, by coupling DL with molecular dynamics (MD) simulations, we were able to find a multifunctional peptide with promising therapeutic effects. Our work validates our previously proposed pipeline for a more efficient rational discovery of novel AMPs.

Keywords: antimicrobial; artificial intelligence; graphs; molecular dynamics; peptides.

Figure 1

AMPs rational discovery pipeline. (1) A peptide library was generated by cutting the Escherichia coli genome in silico. (2.1) The improved DL algorithm analyzes the library to select promising candidates with membrane activity. (2.2) Candidates were filtered using physicochemical properties to obtain viable AMPs. (2.3) Molecular Dynamics was implemented to find candidates exhibiting cell-penetrating capability. (3.1) AMPs candidates were evaluated in vitro to obtain the MIC. (3.2) Peptides with additional cell-penetrating activity were evaluated in mammalian cells via confocal microscopy. Created with biorender.com.

Figure 2

Amino acids distribution. Abundance of each amino acid within all sequences of AMPs and Non-AMPs.

Figure 3

AMPs-Net overview. The FASTA sequence of a candidate peptide was transformed into a graph representation and used as input to a Graph Convolutional Neural Network. Based on the message-passing algorithm, a 256-Dimensional updated graph was obtained and averaged over the feature dimensions. The representative vector was concatenated with the physicochemical properties of the peptide, and a linear layer was then used to classify the peptide into an AMP or Non-AMP. Peptides predicted as AMPs were further analyzed by a similar network that outputs the probability of finding them within the four sub-classes of antimicrobial peptides.

Figure 4

Graph Representation of Glycine. In the graph representation of molecules, atoms are represented as nodes and bonds as the edges. Each amino acid atom is represented within the graph by nine physicochemical properties (Table 3). Likewise, bonds between the atoms were described by three properties. The same color implies the same atom and/or bond and, therefore same feature vector. Created with Biorender.com.

Figure 5

Average percentage of each amino acid within all sequences of predicted AMPs.

Figure 6

RD10 peptide interaction with a model lipid bilayer as estimated by Molecular Dynamic (MD) simulations. (a) Penetrating capability of multiple promising candidates. Only the RD10 peptide can penetrate the cellular membrane. (b) PMF profiles of RD10 and three already validated cell-translocating peptides in vitro. (a) Flat bottom. (a1) Peptides position traces within the simulation box for 500 ns. Cellular membrane positioned at 2.4 nm. Only the RD10 sequence is likely to penetrate the cellular membrane; however, the other candidates seem to interact with the headgroups of the phospholipid bilayer strongly and remain therein. (a2) Final position of the CFD peptide. (a3) In its final position, the RD10 peptide has completely penetrated the bilayer and is located deeper within the hydrophobic core. (b) Umbrella Sampling. Free energy profile (PMF) for peptides translocating a simplified eukaryotic membrane. Free energy profile (PMF) for peptides translocating a simplified eukaryotic membrane. (b1) TP2, a cell-penetrating peptide added for comparison. (b2,b3) The antimicrobial peptide Frenatin 2.3 and the antimicrobial, cell-penetrating, and DNA binding peptide Buforin II. Also added for comparison. (b4) RD10, an antimicrobial and cell-penetrating candidate.

Figure 7

Stability of the RD10 peptide inside the cellular membrane. A minor variation on (A) RMSD and (B) Rg indicates that the peptide maintains its 3D conformation along the translocation process. (C) Coulombic (COUL) energies dominate the interactions over Lennard-Jones (LJ) energies. Headgroups (P-HG) and acyl chains (P-AC) play a significant role in peptide-membrane interaction. (D) Density distribution of the peptide along the z-axis of the membrane. The right y-axis presents the density scale for peptides, while the left y-axis represents the membrane components and bulk water. It indicates that the peptide remained mainly within the acyl chains, with a minor fraction interacting with the headgroups.

Figure 8

Cell-Penetration assay. (A) Internalization of bare CNPs into NHA cells (B) and CNPs-RD10 nanobioconjugates (20X magnification, 3 h of exposure). The scale bar corresponds to 100 $\mathsf{\mu }$M. (C) Quantification of colocalization between CNPs and CNPs-RD10 nanobioconjugates by the Pearson correlation coefficient (PCC) and the fraction of cytosol area covered by CNPs and CNPs-RD10 nanobioconjugates. There is a statistically significant difference between both treatments. p ≤ 0.05 (*), p ≤ 0.01 (**). (D) Visual inspection of colocalization studies via confocal imaging. Yellow arrows point to colocalization regions between the green and the red channels, showing CNPs or CNPs-RD10 nanobioconjugates trapped in endosomes. The white arrows indicate non-colocalized regions where CNPs or CNPs-RD10 nanobioconjugates likely escaped endosomes or reached the intracellular space by different internalization mechanisms.