Novel antimicrobial peptide discovery using machine learning and biophysical selection of minimal bacteriocin domains

Abstract

17Bacteriocins are ribosomally produced antimicrobial peptides that represent an untapped 18 source of promising antibiotic alternatives. However, inherent challenges in isolation and 19 identification of natural bacteriocins in substantial yield have limited their potential use as 20 viable antimicrobial compounds. In this study, we have developed an overall pipeline for 21 bacteriocin-derived compound design and testing that combines sequence-free prediction of 22 bacteriocins using a machine-learning algorithm and a simple biophysical trait filter to generate 23 minimal 20 amino acid peptide candidates that can be readily synthesized and evaluated for 24 activity. We generated 28,895 total 20-mer peptides and scored them for charge, a-helicity, and 25 hydrophobic moment, allowing us to identify putative peptide sequences with the highest 26 potential for interaction and activity against bacterial membranes. Of those, we selected sixteen 27 sequences for synthesis and further study, and evaluated their antimicrobial, cytotoxicity, and 28 hemolytic activities. We show that bacteriocin-based peptides with the overall highest scores for 29 our biophysical parameters exhibited significant antimicrobial activity against E. coli and P. 30 aeruginosa. Our combined method incorporates machine learning and biophysical-based 31 minimal region determination, to create an original approach to rapidly discover novel 32 bacteriocin candidates amenable to rapid synthesis and evaluation for therapeutic use.

Exploring Amyloidogenicity of Peptides From Ribosomal S1 Protein to Develop Novel AMPs

“…Despite the wide variety of approaches to assessing antibacterial activity, it is difficult to create a universal template that could distinguish between antimicrobial and non-antimicrobial peptides, which is a significant limitation in the development of new AMPs. Obviously, there is a need for laboratory testing of the effectiveness of predicted AMPs, including for further refinement and improvement of the results of the predictive programs ( Fields et al, 2020 ). We have proposed a new mechanism of AMP action, a mechanism of directed co-aggregation, which is based on the interaction of a peptide capable of forming fibrils with a target protein.…”

Machine learning enables accurate and rapid prediction of active molecules against breast cancer cells

“…All quantitative compound-cell associations (cell-based assays, assay type: F) for available BC cell lines and normal BC cell lines were collected from ChEMBL (Mendez, et al, 2019) (downloaded in March 2021) after the exclusion of metastatic cell lines. Each BC cell dataset was then processed using the following steps: (i) compounds with biological activity reported as IC 50 , EC 50 , or GI 50 were kept, whereas molecules that had no bioactivity record were removed; (ii) the units of bioactivity (i.e., g/mL, M, nM) were converted into the standard unit in μM; (iii) for a molecule with multiple bioactivity values, the final bioactivity value was obtained by averaging the available bioactivity records; (iv) according to previous studies (Fields, et al, 2020; Ye, et al, 2021), compounds with bioactivity values (e.g., IC 50 , EC 50 , GI 50 ) ≤10 μM were considered as active and vice versa; molecules whose label could not be unequivocally assigned (e.g., activity <100 μM or activity >1 μM) were excluded from the dataset; (v) all molecules were processed by removing salt and optimized based on the MMFF94X force field using MOE software (version 2018) with the default parameters. Finally, 14 cell lines with the number of active molecules (actives) and inactive molecules (inactives) >50 were retained.…”

“…For example, in 2020, Stokes et al first reported directed message passing neural network models using a collection of 2,335 compounds for those that inhibited the growth of Escherichia coli (phenotype screening data) and then identified the lead compound halicin with broad-spectrum antibacterial activity (Stokes, et al, 2020). Other machine learning-based models have been established to identify new agents against Methicillin-Resistant Staphylococcus aureus (Wang, et al, 2014), Mycobacterium tuberculosis (Ye, et al, 2021), Pseudomonas aeruginosa (Fields, et al, 2020), Plasmodium falciparum (Ashdown, et al, 2020), and Schistosoma (Zheng, et al, 2021). In the field of anticancer drug design and discovery, phenotypical whole cell-based screening methods have substantially advanced our ability to identify new anticancer drugs.…”

Mining and unearthing hidden biosynthetic potential

“…Using this method, a number of synthetic-bioinformatic natural products (syn-BNPs) were discovered, including the peptide humimycin (1) with potent antimethicillin-resistant Staphylococcus aureus (MRSA) activity 22 , the antibiotic paenimucillins (e.g., paenimucillin A, 2) 23,24 , and an antifungal peptide 23 . Similar methodologies were applied to identify novel RiPPs with antibacterial activity 25 and a new class of RiPPs, the pyritides 26 . Based on the architecture of a BGC in Micromonospora rosaria, the corresponding natural products were predicted to undergo a formal, enzymatic [4 + 2]cycloaddition with subsequent elimination of the leader peptide and water to produce a pyridine-based macrocycle (pyritide A2, 3) 26 .…”

“…1). Later on, class-independent RiPP genome mining tools utilizing alternative probes such as the RiPP recognition elements (RRE) were established 16,25,[55][56][57][58] . Whereas potential producers of RiPPs can thus be identified through comparative genome mining, additional methods are required to actually find the corresponding metabolite.…”