Comparative analysis of machine learning algorithms on the microbial strain-specific AMP prediction

Abstract

The evolution of drug-resistant pathogenic microbial species is a major global health concern. Naturally occurring, antimicrobial peptides (AMPs) are considered promising candidates to address antibiotic resistance problems. A variety of computational methods have been developed to accurately predict AMPs. The majority of such methods are not microbial strain specific (MSS): they can predict whether a given peptide is active against some microbe, but cannot accurately calculate whether such peptide would be active against a particular MS. Due to insufficient data on most MS, only a few MSS predictive models have been developed so far. To overcome this problem, we developed a novel approach that allows to improve MSS predictive models (MSSPM), based on properties, computed for AMP sequences and characteristics of genomes, computed for target MS. New models can perform predictions of AMPs for MS that do not have data on peptides tested on them. We tested various types of feature engineering as well as different machine learning (ML) algorithms to compare the predictive abilities of resulting models. Among the ML algorithms, Random Forest and AdaBoost performed best. By using genome characteristics as additional features, the performance for all models increased relative to models relying on AMP sequence-based properties only. Our novel MSS AMP predictor is freely accessible as part of DBAASP database resource at http://dbaasp.org/prediction/genome.

Keywords: AMP prediction; antimicrobial peptides; machine learning.

Figure 1

Flowchart of MSS prediction algorithms.

Figure 2

Balance accuracies for the models from three subgroups of the MSSPM G2 group: (a) G25, (b) G28 and (c) G29 (abbreviations are described in Section 3.2). Test sets SQTS_ij were created based on data corresponding to a particular pair (i,j): i-th strain and j-th GF. The first index takes on the following values: i = 1,…8, with i = 1 corresponding to Escherichia coli ATCC 25922, i = 2—Pseudomonas aeruginosa ATCC 27853, i = 3—to Klebsiella pneumoniae ATCC 700603, i = 4—to Salmonella typhimurium ATCC 14028, i = 5—to Acinetobacter baumannii ATCC 19606, i = 6—to Staphylococcus aureus ATCC 25923, i = 7—to Enterococcus faecalis ATCC 29212, and i = 8—to Bacillus subtilis ATCC 6633. The second index takes on the values j = 5, 8, 9, where j = 5 corresponds to mono+di nucleotide compositions and j = 8, 9 correspond to SF: j = 8 corresponds to genome similarity index dDDH and j = 9 corresponds to index that relies on similarity between gyrB genes; training sets SQTS_cj = SQTS_1j ∪ SQTS_2j ∪,…,∪ SQTS_8j. BAC was evaluated using 10-fold cross-validation.

Figure 3

Balance accuracies for the models from three subgroups of the MSSPM G3 group: (a) G35, (b) G38 and (c) G39 (abbreviations are described in Section 3.2). Test sets SQTS_ij were created based on data corresponding to a particular pair (i,j): i-th strain and j-th GF. The first index takes on the following values: i = 1,…,8, with i = 1 corresponding to Escherichia coli ATCC 25922, i = 2—Pseudomonas aeruginosa ATCC 27853, i = 3—to Klebsiella pneumoniae ATCC 700603, i = 4—to Salmonella typhimurium ATCC 14028, i = 5—to Acinetobacter baumannii ATCC 19606, i = 6—to Staphylococcus aureus ATCC 25923, i = 7—to Enterococcus faecalis ATCC 29212, and i = 8—to Bacillus subtilis ATCC 6633. The second index takes on the values j = 5, 8, 9, where j = 5 corresponds to mono+di nucleotide compositions and j = 8, 9 correspond to SF: j = 8 corresponds to genome similarity index dDDH and j = 9 corresponds to index that relies on similarity between gyrB genes; training sets ^i-SQTS_cj = SQTS_cj–SQTS_ij. BAC was evaluated using 10-fold cross-validation.

Figure 4

Average accuracies of the best predictive models proposed on applying the Y-scrambling test with different shuffling percentages of the true activity.