AMPlify: attentive deep learning model for discovery of novel antimicrobial peptides effective against WHO priority pathogens

Abstract

Background: Antibiotic resistance is a growing global health concern prompting researchers to seek alternatives to conventional antibiotics. Antimicrobial peptides (AMPs) are attracting attention again as therapeutic agents with promising utility in this domain, and using in silico methods to discover novel AMPs is a strategy that is gaining interest. Such methods can sift through large volumes of candidate sequences and reduce lab screening costs.

Results: Here we introduce AMPlify, an attentive deep learning model for AMP prediction, and demonstrate its utility in prioritizing peptide sequences derived from the Rana [Lithobates] catesbeiana (bullfrog) genome. We tested the bioactivity of our predicted peptides against a panel of bacterial species, including representatives from the World Health Organization's priority pathogens list. Four of our novel AMPs were active against multiple species of bacteria, including a multi-drug resistant isolate of carbapenemase-producing Escherichia coli.

Conclusions: We demonstrate the utility of deep learning based tools like AMPlify in our fight against antibiotic resistance. We expect such tools to play a significant role in discovering novel candidates of peptide-based alternatives to classical antibiotics.

Keywords: Antimicrobial peptide; Attention mechanism; Deep learning.

Fig. 1

Model architecture of AMPlify. Residues of a peptide sequence are one-hot encoded and passed to three hidden layers in order: the bidirectional long short-term memory (Bi-LSTM) layer, the multi-head scaled dot-product attention (MHSDPA) layer and the context attention (CA) layer. The output layer generates the probability that the input sequence is an AMP

Fig. 2

Performance comparison of different AMP prediction tools based on the test sequence similarities to their corresponding training sets. F1 scores of AMP prediction tools were calculated on test subsets based on similarities to sequences in the training sets. All the AMP/non-AMP test subsets were derived from the AMPlify test data, with subsets containing 10 or fewer sequences removed. The size of the round makers indicates the number of sequences remaining in the test subset given the similarity threshold

Fig. 3

Visualization of AMPlify model performance and the AMP discovery pipeline application results. a Receiver operating characteristic (ROC) curves of AMPlify and comparators are plotted, with round dots marking the performance at the threshold of 0.5. The iAMP-2L online server only output labels of AMP/non-AMP without the corresponding probabilities, so it appears as a single point on the plot. b AMPlify prediction scores against peptide lengths of 101 sequences analyzed by AMPlify. The grey dotted line represents the score threshold of 0.5 used to distinguish AMPs from non-AMPs. Inset shows amplified view of the upper left region of the plot to enhance visualization of the majority of the selected sequences