Machine Learning Model Predicts Antimicrobial Peptide Synergy With High Test Accuracy

A supervised machine learning model developed by researchers can accurately predict synergistic therapeutic interactions between antimicrobial peptides and conventional antimicrobial agents, offering a computational pathway to identify potent combination treatments without relying on exhaustive benchtop trials. Addressing infectious diseases remains a vital objective for healthcare systems challenged by opportunistic pathogens and drug-resistant bacterial strains that emerge when traditional agents are administered at suboptimal concentrations. Combining antimicrobial peptides, which are cationic and amphiphilic host defense molecules with low tendencies to provoke resistance, alongside standard antimicrobial drugs represents an appealing strategy to broaden pathogen targets and lower requisite dosages. However, assessing these dual regimens conventionally requires laborious experimental procedures such as checkerboard assays to compute the fractional inhibitory concentration index. To circumvent the extensive time, material consumption, and financial expense inherent to combinatorial laboratory testing, investigators formulated a supervised predictive framework to classify pairwise interactions as either synergistic or lacking significant interaction directly from physicochemical, drug-specific, and microbial features.

Strict curation standards were applied to peptide structures to eliminate confounding structural variations, excluding sequences bearing intrachain bonds, coordination bonds, or nonstandard amino acids, while specifying an unmodified hydrogen atom at the amino terminus and an amide group at the carboxyl terminus. From an initial pool of 286 qualifying antimicrobial peptides, sequences displaying 90 percent or greater sequence similarity were removed, yielding 67 unique peptides evaluated in combination with characterized antimicrobial drugs. In total, 407 curated experimental observations were assembled, comprising 199 confirmed instances of synergy and 208 records showing no substantial interaction. Across this dataset, 101 unique fractional inhibitory concentration index values were documented, ranging from a minimum of 0.01 to a maximum of 1.98, with an overall mean of 0.63. Combinations involving Pseudomonas aeruginosa constituted the largest single microbial cohort, representing 32.2 percent of all evaluated interactions.

To configure the inputs for predictive learning, the investigators partitioned candidate predictors into three functional categories: structural characteristics of antimicrobial peptides, molecular properties of paired antimicrobial agents, and target microbial parameters. Numeric peptide features included sequence length, molecular weight, normalized hydrophobicity, net charge, isoelectric point, penetration depth, tilt angle, disordered conformation propensity, linear moment, amphiphilicity index, average hydrophilicity, and the proportion of hydrophilic residues. Antimicrobial agents were parameterized through molecular weight, physiological charge, octanol-water partition coefficient, aqueous solubility, acid dissociation constant, mechanism of action, and drug class. Microbial characteristics encompassed target species designations, Gram staining status, and the individual minimum inhibitory concentrations of both the peptide and the companion drug when administered alone. Binary outcomes were established according to fractional inhibitory concentration thresholds, where values less than or equal to 0.5 indicated synergism and values spanning 0.5 to 4 indicated no significant interaction, while true antagonistic interactions exceeding 4 were entirely absent from the assembled data.

Before undertaking algorithm development, the research team conducted a correlation matrix evaluation to examine collinearity among numerical predictors and prune redundant variables that could introduce computational noise or inflate model variance. Strong linear correlations emerged between normalized hydrophobicity and the amphiphilicity index, exhibiting a correlation coefficient of 0.85, as well as between net peptide charge and average hydrophilicity at 0.71. Sequence length and peptide molecular weight demonstrated near-perfect correlation at 0.95, while normalized hydrophobicity was inversely correlated with disordered conformation propensity at negative 0.98. In addition, the physiological charge of the antimicrobial agent correlated strongly with its acid dissociation constant at 0.75. Given that literature evidence highlights the primary importance of peptide hydrophobicity, net charge, and drug lipophilicity over disordered conformation or charge states, the team selectively eliminated the amphiphilicity index, average hydrophilicity, sequence length, disordered conformation propensity, and drug charge. Eliminating features possessing correlation coefficients greater than 0.5 or lower than negative 0.5 removed collinear relationships without stripping essential chemical variance.

Data preprocessing protocols were implemented to rescale input distributions, encode categorical attributes, and balance class frequencies prior to formal model training. The researchers compared four distinct numeric scaling techniques across multiple learning architectures, testing min-max scaling, Z-score standardization, maximum absolute scaling, and robust data scaling. Robust scaling demonstrated superior classification accuracy across all baseline architectures, increasing the validation accuracy of the boosting framework from 70.53 percent in its raw form to 75.38 percent after normalization. Nominal predictors, including target species identities and antimicrobial drug classifications, were converted into orthogonal binary vectors using one-hot encoding. To resolve the modest class imbalance between the 199 synergistic and 208 non-interacting samples, the Synthetic Minority Oversampling Technique generated nine synthetic observations to equalize both cohorts at 208 instances. A subsequent one-way analysis of variance yielded a p-value exceeding 0.995, confirming that the oversampling procedure introduced no statistically significant distortion to downstream classifier performance.

Model training and evaluation were carried out using an Intel Core i9-10940X central processing unit and 64 gigabytes of random-access memory, with implementations based on Python and scikit-learn libraries. Seventeen distinct supervised learning classifiers were systematically benchmarked using a 3-repeated 5-fold cross-validation scheme over a 75 percent training split. The benchmark encompassed logistic regression, linear discriminant analysis, Gaussian process classification, k-nearest neighbors, decision tree variants, support vector machines, multilayer perceptrons, and multiple ensemble architectures including random forests, AdaBoost, and gradient boosting algorithms. The baseline Light Gradient Boosted Machine Classifier attained the highest average cross-validation accuracy of 75.75 percent, accompanied by an area under the receiver operating characteristic curve of 0.82, a precision of 77.40 percent, a recall of 75.65 percent, an F1 score of 75.30 percent, and an elapsed training time of 1.68 seconds. Classifiers such as Gaussian processes, bagging classifiers, histogram gradient boosting, random forests, and multilayer perceptrons also demonstrated statistically significant performance differences compared to lower-performing baselines like Gaussian Naive Bayes, which scored only 61.80 percent.

Following baseline selection, the Light Gradient Boosted Machine Classifier underwent automated hyperparameter tuning using a genetic algorithm-driven Tree-based Pipeline Optimization Tool. The optimization procedure operated across 150 generations with a population size of 100 and an offspring size of 25, raising the cross-validation accuracy of the classifier from 72.43 percent to 75.38 percent. When evaluated on the independent 25 percent holdout test dataset containing all available features, the resulting optimized classifier attained a top test accuracy of 76.92 percent and an area under the curve of 80.71 percent. Across holdout classifications, the model yielded identical precision, recall, and F1 scores of 78.18 percent. Analysis of the test confusion matrix revealed that out of 55 instances lacking interaction, the model correctly identified 43 true negatives while misclassifying 12 as false negatives, translating to a specificity of 78.2 percent. For the 49 synergistic pairs, the model correctly classified 37 true positives and misclassified 12 as false positives, achieving a sensitivity of 75.5 percent.

To evaluate the impact of dimensional reduction, the optimized model was re-evaluated on the holdout test data using the pruned feature space that excluded the seven highly correlated predictors. This streamlined model delivered an area under the curve of 0.811 and an overall test accuracy of 75.96 percent, reflecting a minor accuracy reduction of only 0.96 percent despite eliminating nearly a third of the continuous attributes. Confusion matrix analysis under this reduced configuration revealed that synergistic classification accuracy improved from 75.5 percent to 79.6 percent, while non-interaction accuracy decreased from 78.2 percent to 72.7 percent. This indicates that stripping redundant variables enhances the model’s capacity to resolve genuine synergistic interactions while dramatically lowering training complexity and computational overhead. The findings confirm that supervised algorithms can establish robust functional relationships between uncorrelated molecular predictors and antibacterial synergy without requiring dense, multi-collinear descriptor sets.

Information gain metrics were computed to determine which biological and chemical factors contributed most significantly to algorithmic decision-making. In the full-feature model, antimicrobial peptide characteristics accounted for 50 percent of total feature importance, paired antimicrobial properties accounted for 40 percent, and shared interaction variables contributed the remaining 10 percent. The target microbial species emerged as the single most influential predictor, followed by the individual minimum inhibitory concentrations of the antimicrobial peptide and the companion drug when applied independently. In contrast, the isoelectric point of the peptide, the physiological charge of the drug, and pathogen Gram classification displayed the lowest predictive contribution. In the pruned model, peptide features represented 46 percent of total importance, drug characteristics represented 42 percent, and shared variables comprised 12 percent, with peptide minimum inhibitory concentration becoming the primary weighted factor, microbial species ranking second, and peptide tilt angle ascending to the third most critical feature.

Model generalizability was further scrutinized using an external evaluation dataset and leave-one-out validation strategies across distinct biological groupings. On the external test cohort, the trained model achieved an overall accuracy of 71.74 percent and an area under the curve of 0.76. Although Gram-negative pathogens constituted 75.4 percent of the foundational observations, external testing revealed stable predictive fidelity across structural divisions, registering 73.42 percent accuracy for Gram-negative species and 70.73 percent for Gram-positive organisms. Exhaustive leave-one-out iterations produced mean accuracies of 78.70 percent across unique microbial species and 72.86 percent across distinct antimicrobial compounds. The authors noted clear operational boundaries, emphasizing that the absence of an antagonism class with indices exceeding 4, the modest total sample volume of 407 entries, and the exclusive focus on planktonic bacterial cultures rather than resilient biofilm architectures represent key limitations requiring expansion in subsequent computational investigations.

Subject of Research: Supervised machine learning prediction of synergistic interactions between antimicrobial peptides and antimicrobial agents

Article Title: Prediction of the synergistic effect of antimicrobial peptides and antimicrobial agents via supervised machine learning

Article References: Olcay, B., Ozdemir, G. D., Ozdemir, M. A., Ercan, U. K., Guren, O., & Karaman, O. (2024). Prediction of the synergistic effect of antimicrobial peptides and antimicrobial agents via supervised machine learning. BMC Biomedical Engineering, 6(1), Article 1. https://doi.org/10.1186/s42490-024-00075-z

Image Credits: AI Generated

DOI: 10.1186/s42490-024-00075-z

Keywords: antimicrobial peptides, synergistic effect, fractional inhibitory concentration, supervised machine learning, LightGBM, oLGBMC, checkerboard assay, minimum inhibitory concentration, host defense peptides, bioinformatics, feature importance, Gram-negative bacteria, Gram-positive bacteria, Pseudomonas aeruginosa, TPOT optimization, robust normalization, cross-validation