Identifying novel antimicrobial peptides from venom gland of spider Pardosa astrigera by deep multi-task learning

Antimicrobial peptides (AMPs) show promises as valuable compounds for developing therapeutic agents to control the worldwide health threat posed by the increasing prevalence of antibiotic-resistant bacteria. Animal venom can be a useful source for screening AMPs due to its various bioactive components. Here, the deep learning model was developed to predict species-specific antimicrobial activity. To overcome the data deficiency, a multi-task learning method was implemented, achieving F1 scores of 0.818, 0.696, 0.814, 0.787, and 0.719 for Bacillus subtilis, Escherichia coli, Pseudomonas aeruginosa, Staphylococcus aureus, and Staphylococcus epidermidis, respectively. Peptides PA-Full and PA-Win were identified from the model using different inputs of full and partial sequences, broadening the application of transcriptome data of the spider Pardosa astrigera. Two peptides exhibited strong antimicrobial activity against all five strains along with cytocompatibility. Our approach enables excavating AMPs with high potency, which can be expanded into the fields of biology to address data insufficiency.

Automated summary

Antimicrobial peptides (AMPs) are promising compounds for developing therapeutic agents against antibiotic-resistant bacteria. Animal venom is a useful source for screening AMPs. A deep learning model was developed to predict species-specific antimicrobial activity, and a multi-task learning method was applied to overcome the lack of data. The model successfully predicted the antimicrobial activity of several bacteria and identified two peptides with strong antimicrobial activity.