Groundbreaking method dmSLAY improves antimicrobial peptide effectiveness and safety

By Tarun Sai LomteReviewed by Susha Cheriyedath, M.Sc.Aug 1 2024

In a recent study published in the journal Nature Biomedical Engineering, researchers developed a method to measure the maintenance or loss of antimicrobial peptide (AMP) activity for thousands of peptide sequence variants.

AMPs are evolutionarily conserved peptides that fight pathogens. Generally, AMPs kill bacteria by disrupting their cell membrane. However, many AMPs exhibit difficulties differentiating between bacterial and mammalian cell membranes. Conventional methods assessing AMP residue flexibility and importance are time-consuming and limited to a few variants.

Further, newer approaches, such as deep mutational scanning, which allows large-scale assessment of millions of mutations, have not been as successful or widely used with AMPs. The study’s authors have developed surface-localized antimicrobial display (SLAY), a novel high-throughput method to simultaneously examine the antibacterial potential of hundreds of thousands of synthetic peptides.

Study: Deep mutational scanning and machine learning for the analysis of antimicrobial-peptide features driving membrane selectivity. Image Credit: Zapp2Photo / Shutterstock

The study and findings

In the present study, researchers developed deep mutational SLAY (dmSLAY) to assess how thousands of amino acid changes to a known AMP, protegrin 1 (PG-1), impact its activity. They designed a deep mutational library of PG-1 variants, encoding 7–12 mutations at each position and 1–9 changes per variant. Overall, the library comprised 7,105 unique sequences. Next, the library was cloned into a surface display plasmid and transformed into Escherichia coli W3110.

SLAY screening of the library identified 1,940 inactive variants, whereas 1,203 variants were predicted to possess antimicrobial activity. Next, 40 variants were synthesized, and their minimal inhibitory concentration (MIC) was measured against E. coli W3110. A receiver operating characteristic curve was used to determine optimal cutoffs for inactive and active variants based on MIC and log2(fold change) scores.

Interestingly, log2(fold change) scores did not correlate with antibacterial potency, although a positive or negative score predicted at 86% accuracy whether a variant lost or retained antimicrobial activity. This included 16 accurate predictions for single-residue variants, indicating that dmSLAY was highly precise in predictions. Further, residue changes in the tail regions of PG-1 retained activity, while most residue changes in the β-sheet regions caused a loss of activity.

Consistently, multiple-residue variants predicted to lose activity had mutations in the β-sheet regions, whereas those retaining activity had changes in the tail region. Next, circular dichroism (CD) spectroscopy was performed to examine secondary structure changes of variants with and without the lipopolysaccharide (LPS), mimicking the bacterial membrane. Without LPS, there was no correlation between MIC and CD spectra.

However, with LPS, variants with spectra closer to that of native PG-1 (PG-1.0) were more potent. Next, the team analyzed the hemolytic activity of PG-1 variants. No correlation was observed between percent hemolysis and CD spectra without LPS. Likewise, in the presence of LPS, hemolytic variants retained secondary structure close to PG-1.0.

Besides, the team multiplied the percent hemolysis and MIC to generate a selectivity score for each variant. Variants with a lower selectivity score were more bacterially selective. Selectivity scores of PG-1 variants were compared with PG-1.0. This comparison identified five variants with enhanced membrane selectivity. Further, propidium iodide (PI) uptake assays were performed to determine whether selective variants had the exact mechanism of membrane lysis as PG-1.0.

E. coli treated with PG-1.0 showed a strong PI uptake. Likewise, selective variants exhibited strong uptake of PI, suggesting that variants probably had a similar lysis mechanism. Furthermore, the selective variants were over 1000-fold more specific to bacterial membranes than the erythrocyte cell membrane. This high specificity highlights their therapeutic potential.

The researchers trained three machine learning (ML) models on a biochemical dataset. Models were used to identify the mutational profiles that decrease mammalian toxicity (hemolysis < 2%), increase bacterial specificity with log(selectivity score) < 0.5, or have a high probability of being potent (MIC ≤ 8 µg/ml) individually. While the models provided information on how the PG-1 sequence impacted specific biochemical features individually, they could not select highly potent variants specific to mammalian or bacterial membranes.

Therefore, the team combined the models and identified over 95,000 sequences passing the three cutoffs. They observed that reduction of hydrophobicity within the β-sheet regions improved bacterial membrane specificity. By contrast, increasing hydrophobicity in the loop and tail regions, particularly with tryptophan, and increasing the overall charge enhanced mammalian membrane specificity.

To validate these findings, thirty-two variants predicted to be bacterially selective were selected. Of these, 75% of variants were active, all with improved selective scores relative to PG-1.0. Notably, ML predictions outperformed dmSLAY in identifying non-toxic and membrane-selective variants while maintaining a similar level of accuracy in predicting antibacterial potency and activity.

Next, the team measured the toxicity of the best ML and dmSLAY variants relative to colistin (a clinical AMP) and PG-1.0. HEK293 cells were treated with each AMP, and their uptake of SYTOX Green, a membrane-impermeable dye, was measured. The results showed that cells treated with bacterially selective variants had considerably less dye uptake than those treated with PG-1.0. Further, those treated with (ML-selected) bsPG-1.2 had less dye uptake than colistin, suggesting relatively less cytotoxicity.

Moreover, PG-1.0 and bacterially selective variants were unaffected by the mobile colistin resistance gene expression. Finally, the researchers investigated the maximum tolerated dose (MTD) of PG-1 and bsPG-1.2 in mice. The MTD of bsPG-1.2 was over five-fold higher than that of PG-1, suggesting bsPG-1.2 was significantly less toxic in mice. Additionally, bsPG-1.2 exhibited antibacterial activity in mice infected with Acinetobacter baumannii AB5075.

Conclusions

Together, dmSLAY offered a rapid, high-throughput method to determine positional residue flexibility and importance and predict how multiple concurrent sequence changes impact antimicrobial activity. In addition, ML allowed for the virtual examination of 5.7 million variants for membrane selectivity, antimicrobial activity, and mammalian toxicity. This strategy could be applied to numerous AMPs to delineate the rules governing membrane selectivity, which could inform the synthetic design of therapeutic peptides. These findings have significant implications for the development of new antimicrobial therapies with reduced toxicity and enhanced selectivity.