AI4AMP: an Antimicrobial Peptide Predictor Using Physicochemical Property-Based Encoding Method and Deep Learning

Abstract

Antimicrobial peptides (AMPs) are innate immune components that have recently stimulated considerable interest among drug developers due to their potential as antibiotic substitutes. AMPs are notable for their fundamental properties of microbial membrane structural interference and the biomedical applications of killing or suppressing microbes. New AMP candidates must be developed to oppose antibiotic resistance. However, the discovery of novel AMPs through wet-lab screening approaches is inefficient and expensive. The prediction model investigated in this study may help accelerate this process. We collected both the up-to-date AMP data set and unbiased negatives based on which the protein-encoding methods and deep learning model for AMPs were investigated. The external testing results indicated that our trained model achieved 90% precision, outperforming current methods. We implemented our model on a user-friendly web server, AI4AMP, to accurately predict the antimicrobial potential of a given protein sequence and perform proteome screening. IMPORTANCE Antimicrobial peptides (AMPs) are innate immune components that have aroused a great deal of interest among drug developers recently, as they may become a substitute for antibiotics. New candidates need to fight antibiotic resistance, while discovering novel AMPs through wet-lab screening approaches is inefficient and expensive. To accelerate the discovery of new AMPs, we both collected the up-to-date antimicrobial peptide data set and integrated the protein-encoding methods with a deep learning model. The trained model outperforms the current methods and is implemented into a user-friendly web server, AI4AMP, to accurately predict the antimicrobial properties of a given protein sequence and perform proteome screening. Author Video: An author video summary of this article is available.

Keywords: antimicrobial peptide; deep learning; protein-encoding method; real-world data; web service.

FIG 1

Hierarchical clustering plot of physicochemical properties. Six selected physicochemical properties used in our PC6 are marked in blue. Seven physicochemical properties used in the original AC7 method are marked with asterisks.

FIG 2

The Venn diagram presents the overlapping of misplaced AMP prediction, false negative (left) and false positive (right), on an external test by PC6/deep learning and Antimicrobial Peptide Scanner vr.2.

FIG 3

Data processing in this study. (A) Data for model tuning and construction. (B) Data for external testing. (C) Data for the final model.

FIG 4

Algorithm-designed peptide AMP activities [−log(MIC)] and their corresponding AI4AMP scores. Plot areas in high AI4AMP scores (>0.95) were expanded to show the crowd data points. MICs beyond the testing range (>128 or >256 μg/ml) were assigned to 512 [e.g., −log₂(512) = −9].

FIG 5

AI4AMP web server. The left panel shows the user interface (UI) input. (1 and 2) Users may either copy-paste sequences in FASTA format directly in the “input FASTA” form or upload a FASTA file to “Fileupload (*.txt).” Users may submit the query with a valid email address in “Your Email” to trace back to the job result page or stay on the UI for redirecting to the output. (3 and 4) Users may browse the scores and prediction results on the result page (the right panel) and retrieve the output CSV file in the download area. A pie chart summarizes the proportion of predicted AMPs and non-AMPs in the input sequences.

FIG 6

(A) PC6 protein-encoding method. Each input sequence will be transformed into a 200 × 6 matrix, respectively. (B) Deep neural network model. The PC6 encoded data matrix will pass through one convolution layer, one LSTM layer, and one dense layer.