A Systematic Implementation of Machine Learning Algorithms for Multifaceted Antimicrobial Screening of Lead Compounds ^†

Abstract

This study employed machine learning algorithms to identify lead compounds that inhibit the antibiotic targets DNA gyrase and Dihydrofolate reductase in Escherichia coli, and identified new, multifaceted antimicrobial compounds. This study used three separate datasets: (1) 326 Escherichia coli DNA gyrase inhibitors and 132 non-inhibitors, (2) 346 Escherichia coli Dihydrofolate reductase inhibitors and 176 non-inhibitors, and (3) 18387 non-specific drug-like chemicals. All datasets were then processed using ECFP-4 fingerprints and split into train, test, and validation datasets according to a 70–15–15 train-test-validation split. We explored the potential of six different classification algorithms, all optimized with Bayesian optimization. Our results indicate that the Gradient Boosting Classifier (GBC) performed the best at identifying a compound’s efficacy towards DNA gyrase with an accuracy, precision, recall, F1-score, and AUC of 0.91, 0.92, 0.86, 0.88, and 0.933, respectively. The Random Forest Classifier (RFC) performed optimally for identifying a compound’s effectiveness towards Dihydrofolate reductase with an accuracy, precision, recall, F1-score, and AUC of 0.86, 0.83, 0.85, 0.84, and 0.944, respectively. As a result, the GBC and RFC were used to search for compounds that inhibited both DNA gyrase and Dihydrofolate reductase. Out of 18387 compounds, we identified five novel compounds that have a predicted probability greater than 95% to inhibit both DNA gyrase and Dihydrofolate reductase, suggesting a high antimicrobial potential. The models evaluated in this study, particularly the GBC and RFC models, hold tremendous promise in computationally screening large libraries of compounds for antimicrobial potential.

1. Introduction

1.1. Background on Antibiotics and Antibiotic Resistance

Amidst the recent explosion of antibiotic use in both humans and agriculture, antibiotic resistance in bacterial strains has begun to spike. This has led to the advent of “superbugs”, bacteria that have developed resistance to multiple antibiotics [1]. As a result, there have been numerous research efforts in recent years aiming to identify new antibiotics.

1.2. Recent Advances in Computational Drug Discovery: Applications to Antimicrobial Compounds

Recent advances in machine learning and computational biology have demonstrated the potential to accelerate computational drug discovery by filtering the chemical space for target molecules [2,3]. Previous researchers have demonstrated the efficacy of random forest classification models when predicting for life-extending chemicals [4]. Furthermore, most recently, researchers have demonstrated the potential of deep learning models, particularly in identifying novel antibiotics that are successful in vivo against a wide range of bacterial infections, indicating the potential for computational methods to revolutionize antibiotic bacteria [5].

1.3. DNA Gyrase and Dihydrofolate Reductase as Antimicrobial Targets

Many current antimicrobial compounds operate by inhibiting the function of key proteins that are vital to bacterial function [6]. This study focused on two such proteins proposed by prior literature as antimicrobial targets: DNA gyrase and Dihydrofolate reductase. DNA gyrase functions as topoisomerase in bacteria, aiding in the process of ATP-dependent negative supercoiling of DNA in bacteria [7]. Previous successful antibiotic classes such as Coumarins and Quinolones have modulated the function of DNA gyrase, leading to antimicrobial function via the breakdown of bacterial function [8]. Similarly, Dihydrofolate reductase has also been a popular target for antimicrobial agents due to its crucial role in nucleotide synthesis [9].

1.4. Purpose

In an effort to speed up antibiotic discovery, this study demonstrated the promise of machine learning classification models in multifaceted antimicrobial compound screening and identification.

2. Materials and Methods

2.1. Datasets and Dataset Preprocessing

The breakdown of the three datasets used in this research are displayed in Table 1. The datasets consisting of Escherichia coli DNA gyrase and Dihydrofolate reductase inhibitors were sourced from ChEMBL [10]. The dataset consisting of 18,387 non-specific drug-like chemicals was sourced from Zinc 15 [11].

Table 1. Dataset Breakdown.

Data Type	DNA Gyrase	Dihydrofolate Reductase	Unspecific
Inhibitor	326	346	0
Non-inhibitor	132	176	0
Unspecific	0	0	18,387

All compounds in datasets were characterized using ECFP-4 fingerprints. All datasets were then split into train, test, and validation datasets according to a 70–15–15 train–test–validation split.

2.2. Machine Learning Models

This study employed six classification models in total: logistic regressions (LR), support vector machines (SVM), random forests (RFC), k-nearest neighbors (K-NN), AdaBoost (ADA), and Gradient Boosting (GBC). All models were evaluated using accuracy, prevision, recall, F1-score, and area under curve (AUC).

2.3. Bayesian Optimization

This study also implemented Bayesian optimization in order to optimize all six classification models. For each of the classification algorithms, Bayesian optimization was run to optimize the parameters. The models were optimized using the validation dataset in order to minimize overfitting when evaluating model metrics with the test dataset.

3. Results

3.1. Machine Learning Model Evaluation

3.1.1. DNA gyrase Machine Learning Model Evaluation

All six optimized machine learning models were trained on DNA gyrase inhibitors and evaluated using accuracy, prevision, recall, F1-score, and AUC. The gradient-boosting algorithm performed the best with an accuracy, precision, recall, F1-score, and AUC of 0.91, 0.92, 0.86, 0.88, and 0.933, respectively (Table 2, Figure 1 and Figure 2).

Table 2. DNA gyrase Machine Learning Model Accuracy Metrics.

Model	Accuracy	Precision	Recall	F1-Score	AUC
Logistic Regression	0.88	0.88	0.82	0.84	0.919
Support Vector Machine	0.86	0.92	0.74	0.78	0.921
Random Forest	0.87	0.92	0.76	0.80	0.898
K-Nearest Neighbor	0.78	0.74	0.67	0.69	0.754
AdaBoost	0.83	0.79	0.75	0.77	0.920
Gradient Boosting	0.91	0.92	0.86	0.88	0.933

Figure 1. DNA gyrase Machine Learning Model Confusion Matrices.

Figure 2. DNA gyrase Machine Learning Model receiver operating characteristic (ROC) curves.

3.1.2. Dihydrofolate reductase Machine Learning Model Evaluation

All six optimized machine learning models were trained on Dihydrofolate reductase inhibitors and evaluated using accuracy, prevision, recall, F1-score, and AUC. The random forest algorithm performed the best, with an accuracy, precision, recall, F1-score, and AUC of 0.91, 0.92, 0.86, 0.88, and 0.933, respectively (Table 3, Figure 3 and Figure 4).

Table 3. Dihydrofolate reductase Machine Learning Model Accuracy Metrics.

Model	Accuracy	Precision	Recall	F1-Score	AUC
Logistic Regression	0.85	0.82	0.82	0.82	0.949
Support Vector Machine	0.85	0.81	0.83	0.82	0.929
Random Forest	0.86	0.83	0.85	0.84	0.944
K-Nearest Neighbor	0.83	0.81	0.86	0.82	0.926
AdaBoost	0.82	0.78	0.80	0.79	0.866
Gradient Boosting	0.85	0.82	0.82	0.82	0.889

Figure 3. Dihydrofolate reductase Machine Learning Model Confusion Matrices.

Figure 4. Dihydrofolate reductase Machine Learning Model ROC curves.

3.2. Identification and Analysis of Novel Antimicrobial Ligands

By implementing the best performing model for identifying DNA gyrase inhibitors (gradient boosting) and the best performing model for identifying Dihydrofolate reductase (random forest), this study used each model to identify novel compounds that are predicted to inhibit both DNA gyrase and Dihydrofolate reductase. Using both models, five compounds were identifed that had an average predicted probability greater than 0.97. The best performing compound is CN(Cc1cnc2nc(N)nc(N)c2n1)c1ccc(C(=O)N[C@@H](CCC(=O)NO)C(=O)O)cc1, with a predicted probability of 0.9988515310206159 to inhibit DNA gyrase, a predicted probability of 0.9897304236200257 to inhibit Dihydrofolate reductase, and an averaged predicted probability of 0.9942909773203208 (Table 4).

Table 4. Novel Antimicrobial Compounds and Probabilistic Analyses.

Compound	Predicted Probability: DNA Gyrase	Predicted Probability: Dihydrofolate Reductase	Predicted Probability: Average
CN(Cc1cnc2nc(N)nc(N)c2n1)c1ccc(C(=O)N[C@@H](CCC(=O)NO)C(=O)O)cc1	0.9988515310206159	0.9897304236200257	0.9942909773203208
CN(Cc1cnc2nc(N)nc(N)c2n1)c1ccc(C(=O)N[C@H](CCC(=O)O)C(=O)N[C@@H](CCC(=O)O)C(=O)O)cc1	0.9910340430619817	0.9974326059050064	0.994233324483494
CN(Cc1cnc2nc(N)nc(N)c2n1)c1ccc(C(=O)N[C@@H](CCC(=O)N[C@@H](CCC(=O)O)C(=O)O)C(=O)O)cc1	0.9995824679977368	0.9858793324775353	0.9927309002376361
CN(Cc1cnc2nc(N)nc(N)c2n1)c1ccc(C(=O)N[C@H](CCC(=O)N[C@@H](CC(=O)O)C(=O)O)C(=O)O)cc1	0.9908400010423145	0.9691912708600771	0.9800156359511958
CN(Cc1cnc2nc(N)nc(N)c2n1)c1ccc(C(=O)N[C@@H](CCC(=O)N2CCC[C@H]2C(=O)NO)C(=O)O)cc1	0.9993063830032061	0.959349593495935	0.9793279882495705

4. Discussion

Of the six models trained on DNA gyrase, gradient boosting was the most accurate and had the highest F1-score. Unlike the some of the other algorithms, the Bayesian optimization of gradient-boosting models drastically changes the model performance and metrics (i.e., number of trees, learning rate, and maximum depth), greatly enhancing its performance. The use of gradient boosting is further enhanced as the DNA gyrase dataset consists of few outliers, and overall computation time was not a major constraint [12].

For the Dihydrofolate reductase models, the random forest model nominally outperformed other models. With the optimal models from both DNA gyrase and Dihydrofolate reductase, the compounds identified had a probability of 0.97 to inhibit both DNA gyrase and Dihydrofolate reductase.

Since the precision in both models are either similar or higher than respective recall metrics, the false positivity rate will be comparable to the accuracy metrics of the models, ensuring that the chosen compounds maintain a high probability of being effective.

5. Conclusions

This study evaluated the efficacy of machine learning models at identifying novel antimicrobial compounds. The machine learning models evaluated in this study, particularly the gradient-boosting and random forest models, performed very well and hold tremendous potential in computationally screening large libraries of compounds for antimicrobial potential. Furthermore, the compounds identified in this study hold promise as potential, novel antimicrobial compounds. Future investigations should explore alternative classification approaches to antimicrobial compound screening. The compounds identified in this study should also be researched further in vivo to identify additional antimicrobial potential.