Prediction of antimicrobial resistance based on whole-genome sequencing and machine learning.

MOTIVATION: Antimicrobial resistance (AMR) is one of the biggest global problems threatening human and animal health. Rapid and accurate AMR diagnostic methods are thus very urgently needed. However, traditional antimicrobial susceptibility testing (AST) is time-consuming, low throughput, and viable only for cultivable bacteria. Machine learning methods may pave the way for automated AMR prediction based on genomic data of the bacteria. However, comparing different machine learning methods for the prediction of AMR based on different encodings and whole-genome sequencing data without previously known knowledge remains to be done.
RESULTS: In the current study, we evaluated logistic regression (LR), support vector machine (SVM), random forest (RF), and convolutional neural network (CNN) for the prediction of AMR for the antibiotics ciprofloxacin (CIP), cefotaxime (CTX), ceftazidime (CTZ), and gentamicin (GEN). We could demonstrate that these models can effectively predict AMR with label encoding, one-hot encoding, and frequency matrix chaos game representation (FCGR encoding) on whole-genome sequencing data. We trained these models on a large AMR dataset and evaluated them on an independent public data set. Generally, RFs and CNNs perform better than LR and SVM with AUCs up to 0.96. Furthermore, we were able to identify mutations that are associated with AMR for each antibiotic. AVAILABILITY: Source code in data preparation and model training are provided at GitHub website (https://github.com/YunxiaoRen/ML-iAMR). SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

1 Introduction

The rise of antimicrobial resistance (AMR) is one of the greatest threats to global health, food security and societal development. Estimates indicate that the number of yearly deaths will be at 10 million worldwide with a cost of $100 trillion if no steps to tackle AMR are taken by 2050 (Naylor ). Traditional antimicrobial susceptibility testing (AST) is widely used for AMR analysis in clinical practice. However, this approach requires professional facilities and technicians for implementation and is viable only for cultivable bacteria (Boolchandani ). Recently, many studies highlight the potential of machine learning methods in predicting AMR combining sequencing methods and well-known databases with phenotypic information for AMR (Boolchandani ; Liu ; Lv ). For instance, Yang and Kouchaki et al. (2018) analyzed AMR using different machine learning algorithms [e.g. support vector machine (SVM), logistic regression (LR) and random forest (RF)] trained on whole-genome sequencing and achieved high accuracy on AMR prediction. Deep learning algorithms also showed significant potential for predicting new antibiotic drugs, AMR genes and AMR peptides (Arango-Argoty ; Stokes ; Veltri ). However, these studies focused on genome variants (such as single-nucleotide polymorphisms, SNPs) or other features only related to resistant genes identified in previous studies or resistant databases. The potential of machine learning models for predicting AMR without using known resistance mutation databases or annotated genes remains to be clarified.

To use machine learning methods for the classification of AMR, the input sequences (here: genomic sequences) need to be encoded into numerical values. A practical and informative encoding method for the whole-genome sequence is, thus, crucial for downstream analysis. There are various encoding methods for sequences (Spänig and Heider, 2019), e.g. one-hot encoding or label encoding. One-hot encoding, also referred to as sparse encoding, encodes the DNA sequence into a binary matrix, which is then vectorized and used as input for the machine learning models. Label encoding is another simple and straightforward encoding method, where each label is assigned a unique integer.

Thus, in this study, we use label encoding, one-hot encoding and Chaos Game Representation (CGR) to encode the genomic data. CGR is a recurrent iterative function system, which can be used to visualize sequences by building fractals from sequences of symbols, i.e. from an alphabet . Jeffrey (1990) was the first who applied the CGR algorithm to DNA sequences, i.e. n = 4 and , thus the resulting fractals are constructed from squares. Since the development of the CGR and its application in life science, it has been used for the analysis and alignment-free comparison of whole-genome sequences (Joseph and Sasikumar, 2006; Kania and Sarapata, 2021; Lichtblau, 2019). It has been shown that CGR is an excellent representation for genomes and that CGR-driven phylogeny leads to reliable predictions (Deschavanne ). In particular, the comparison between genomes using CGR is straightforward and fast (Hoang ). CGR has been used, for instance, for a fast comparison of SARS-CoV2 strains (Sengupta ). Extensions of CGR include color grids (Deschavanne ) and frequency matrix chaos game representation (FCGR) (Almeida ). Wang used FCGR to calculate the image distance between genomes to generate phylogenetic trees. Rizzo showed that deep neural networks (DNNs) trained on genomes encoded with FCGR yielded very accurate predictions. They used a convolutional neural network (CNN) to divide bacteria into three different phyla, order, family and genus, and showed very high accuracy for the method.

While most existing studies on CGR encoding focused on CGR for DNA, there also exist a smaller number of studies dealing with other alphabets, e.g. the encoding of protein sequences. Yu used the CGR algorithm for protein classification by separating the amino acids into four groups based on their properties and used multifractal and correlation analysis to construct a phylogenetic tree of Archaea and Eubacteria. In other approaches, the amino acids were retranslated into DNA for CGR (Yang ). Sun used a three-dimensional CGR representation for protein classification, and Löchel used FCGR for resistance prediction in HIV-1 with CNNs.

Thus, in this study, we analyzed the potential of different statistical and machine learning methods, including LR, SVM, RF and CNN with label encoding, one-hot encoding and FCGR encoding for predicting AMR based on whole-genome sequencing of Escherichia coli (E.coli).

2 Materials and methods

The workflow of the study is shown in Figure 1.

Fig. 1.

Workflow of the study. WGS data from Giessen and the public data from Moradigaravand were processed, and single nucleotide polymorphisms (SNPs) were called. The SNP data were encoded by label encoding, one-hot encoding and FCGR encoding for subsequent machine learning. The Giessen dataset was used to train and validate the four machine learning algorithms using cross-validation. The public data were used for the final evaluation of the models. Finally, we analyzed the association of SNPs and SNPs-adjacent genes with AMR using EFS. Created with BioRender.com

2.1 Data collection and sample phenotype

Escherichia coli is an important model organism that can cause severe infections in humans and animals, it also represents a significant resistance gene pool that may be responsible for treatment failure in humans and veterinary medicine (Poirel ).

In our study, we used two datasets, referred to as the Giessen data and the public data. The first dataset (Giessen) was collected as part of our study and contains whole-genome sequencing data (WGS) and corresponding phenotypic information for several antibiotics for, in total, 987 E.coli strains. These isolates were obtained from human and animal clinical samples. Antimicrobial susceptibility testing was performed using the VITEK® 2 system (bioMérieux, Nürtingen, Germany) and interpreted following EUCAST guidelines. DNA isolation and whole-genome sequencing were performed, as described by Falgenhauer .

The latter dataset (public) consists of WGS of 1509 E.coli strains and corresponding phenotypic information (Moradigaravand ). In our study, we focused on the four antibiotics ciprofloxacin (CIP), cefotaxime (CTX), ceftazidime (CTZ) and gentamicin (GEN).

CIP belongs to the class of fluoroquinolones and is widely used to treat various infections, including gastroenteritis, respiratory tract infections or urinary tract infections (Heeb ). CIP is particularly effective against Gram-negative bacteria, such as E.coli. However, due to overuse, resistances evolve rapidly. CTX and CTZ belong to the class of cephalosporins and are also widely used to treat various infections, such as meningitis, pneumonia, urinary tract infections, sepsis and gonorrhea. They are broad-spectrum antibiotics with activity against numerous Gram-positive and Gram-negative bacteria, including E.coli. Nevertheless, resistance is also increasing noticeably (Gums ; Sharma, 2013).

GEN belongs to the aminoglycoside class and is widely used to treat various infections, including meningitis, pneumonia, urinary tract infections and sepsis. It is active against a wide range of bacterial infections, mostly Gram-negative bacteria including E.coli. It binds to the 30S subunit of the bacterial ribosome and negatively affects protein synthesis (Garneau-Tsodikova and Labby, 2016).

We used data of 900 isolates with resistance information for CIP (418 resistant, 482 susceptible), 930 isolates with resistance information for CTX (455 resistant, 475 susceptible), 841 isolated for CTZ (291 resistant, 550 susceptible) and 926 isolates for GEN (216 resistant, 710 susceptible).

While the CIP and CTX data are balanced, the Giessen datasets are imbalanced on the CTZ and GEN data (34% and 23% resistant isolates, respectively). The public dataset is imbalanced for all antibiotics. For CIP, CTX, CTZ and GEN, there are only 267, 115, 73 and 101 resistant samples, representing 18%, 8%, 5% and 7% of all isolates in the public dataset, respectively.

The summary of the datasets is shown in Table 1.

Table 1.

Overview of the datasets

Drug	CIP		CTX		CTZ		GEN
Source	Giessen	Public	Giessen	Public	Giessen	Public	Giessen	Public
Resistant	418	267	455	115	291	73	216	101
Susceptible	482	1229	475	1313	550	1398	710	1398
Total	900	1496	930	1428	841	1471	926	1489

2.2 Variants calling of whole-genome sequencing data

The raw whole-genome sequencing reads were first quality checked and filtered by fastp (Chen ). The filtered reads were then aligned to the E.coli reference genome (E.coli K-12 strain. MG1655) using BWA-mem (Li ). Bcftools (Danecek ) was used for calling variants. Samtools (Li and Durbin, 2009) was used to sort the aligned reads, and vcftools (Danecek ) was used to filter the raw variants. We used default parameters for all tools.

2.3 SNPs pre-processing and encoding

We first extracted reference alleles, variant alleles and their positions, and merged all isolates based on the position of reference alleles. We filtered out the loci without variation (N replaces a locus without variation), and we built the final SNP matrix, where the rows represent the samples and columns are the variant alleles.

To encode the SNPs for subsequent machine learning, we used label encoding, one-hot encoding and FCGR encoding. For the label encoding, the A, G, C, T and N in the SNP matrix were converted to 1, 2, 3, 4 and 0. In one-hot encoding, the DNA sequence is encoded into a binary matrix, which is subsequently vectorized. For the FCGR encoding, we used the R package kaos to transform the sequences into an image-like matrix with a resolution of 200 (Löchel ).

2.4 Machine learning and model evaluation

We used four machine learning methods, including LR, SVM, RF and CNN. For LR, RF and SVM, we used the Scikit-learn python package (Pedregosa ). LR was used with default parameters, except that we used 1000 iterations. RF was used with default parameters and 200 trees. For SVM, we used a linear kernel and default parameters.

We implemented CNNs using the Keras (https://keras.io/) package and TensorFlow (https://tensorflow.org). The CNN architecture is based on eleven hidden layers, including four convolutional layers, two batch normalization layers, two pooling layers, one flattening layer, one fully connected layer and one dropout layer. The structure of the networks for label encoding and one-hot encoding are the same, which differ from FCGR encoding-based CNNs only in the convolutional layers and pooling layers (see Supplementary Fig. S1). For FCGR, we used the Conv2D and MaxPooling2D function, while the CNN for the label encoding used the 1D versions instead.

We used eight filters in the first two convolution layers with a kernel size of three, rectified linear unit activation function and same padding. The last two convolution layers used 16 filters instead. The pool size of all pooling layers is two. We used the softmax activation function in the final fully connected layer and compiled the model with Adam optimization and cross-entropy loss.

2.5 Statistical evaluation

We optimized the machine learning models on the Giessen data using five times 5-fold stratified cross-validation. We applied an up-sampling strategy to balance the samples in the training set. For the final evaluation on the public data, we analyzed the performance on the raw public dataset and on a balanced set using a down-sample strategy.

We evaluated the models using the receiver operating characteristics curve (ROC) and the area under the curve (AUC). We also calculated precision and recall for all models. Statistical comparisons were made by the DeLong test (Demler ).

2.6 Marker genes identification located around SNPs

To identify the SNPs that are associated with resistance, we performed a marker gene identification using the EFS R package (Neumann ). The EFS package aggregates eight feature selection methods for binary classification tasks (Neumann ). We used EFS with default parameters. We then annotated the corresponding genes of SNPs using SnpEff software (Cingolani ).

3 Results

3.1 Performance of different machine learning methods for predicting AMR on Giessen data

We used the filtered SNPs matrix encoded by label encoding, one-hot encoding and FCGR encoding from the Giessen dataset to train the four machine learning methods LR, RF, SVM and CNN. The performance of the four machine learning models was evaluated using five times 5-fold cross-validation. The ROC curves and AUC values of the different machine learning models range from 0.69 to 0.96, demonstrating that all models can effectively predict AMR compared with random null models (Fig. 2). We observed that the mean AUC of the RFs was higher than for LR, SVM and CNN classifiers for all antibiotics with both encoding methods (Fig. 2). In particular, RFs were significantly better than LR (P = 0.03), SVMs (P = 0.01) and CNNs (P = 0.02) for CIP with label encoding (Supplementary Fig. S2). RFs were also better than the other three classifiers for GEN with label encoding and FCGR encoding (P < 0.05). For CTZ, RFs significantly outperformed SVMs with all encoding methods (P < 0.05) (Supplementary Fig. S2). For CTX, RFs are significantly better than LR and SVM with label encoding and one-hot encoding (P < 0.05), while there are no significant differences if the FCGR encoding is used (Supplementary Fig. S2).

Fig. 2.

ROC curves for the models with label encoding, one-hot encoding and FCGR encoding on the Giessen data. First row: ROC curves for CIP with label encoding (A), one-hot encoding (B) and FCGR encoding (C), respectively. Second row: ROC curves for CTX with label encoding (D), one-hot encoding (E) and FCGR encoding (F), respectively. Third row: ROC curves for CTZ with label encoding (G), one-hot encoding (H) and FCGR encoding (I), respectively. Fourth row: ROC curves for GEN with label encoding (J), one-hot encoding (K) and FCGR encoding (L), respectively

Moreover, all models show high precision and recall using label (Table 2), one-hot (Table 3) and FCGR encoding (Table 4) for CIP. For CTZ and GEN, the models show high recall but lower precision, which may be related to the imbalanced resistant and susceptible isolates. In sum, RF, CNN, LR and SVM can predict AMR for CIP, CTZ, GEN and CTX with three encoding methods in E.coli.

Table 2.

Results of the four machine learning models with label encoding on the Giessen data

Classifiers/drug	Precision	Precision	Precision	Precision	Recall	Recall	Recall	Recall
	CIP	CTX	CTZ	GEN	CIP	CTX	CTZ	GEN
CNN	0.88 ± 0.04	0.75 ± 0.04	0.81 ± 0.02	0.76 ± 0.03	0.87 ± 0.01	0.65 ± 0.10	0.89 ± 0.03	0.91 ± 0.02
LR	0.88 ± 0.05	0.71 ± 0.04	0.81 ± 0.03	0.77 ± 0.02	0.90 ± 0.03	0.69 ± 0.08	0.92 ± 0.05	0.96 ± 0.03
RF	0.92 ± 0.04	0.75 ± 0.03	0.84 ± 0.03	0.79 ± 0.02	0.89 ± 0.03	0.73 ± 0.07	0.90 ± 0.06	0.97 ± 0.03
SVM	0.85 ± 0.03	0.69 ± 0.02	0.78 ± 0.03	0.75 ± 0.02	0.89 ± 0.04	0.73 ± 0.03	0.89 ± 0.03	0.96 ± 0.03

Table 3.

Results of the four machine learning models with one-hot encoding on the Giessen data

Classifiers/drug	Precision	Precision	Precision	Precision	Recall	Recall	Recall	Recall
	CIP	CTX	CTZ	GEN	CIP	CTX	CTZ	GEN
CNN	0.87 ± 0.05	0.75 ± 0.00	0.84 ± 0.01	0.80 ± 0.00	0.90 ± 0.01	0.71 ± 0.03	0.84 ± 0.03	0.87 ± 0.05
LR	0.89 ± 0.05	0.71 ± 0.04	0.80 ± 0.03	0.78 ± 0.02	0.89 ± 0.03	0.73 ± 0.08	0.89 ± 0.05	0.95 ± 0.02
RF	0.92 ± 0.05	0.75 ± 0.01	0.82 ± 0.02	0.80 ± 0.03	0.90 ± 0.02	0.73 ± 0.07	0.90 ± 0.07	0.97 ± 0.03
SVM	0.86 ± 0.05	0.68 ± 0.03	0.77 ± 0.03	0.76 ± 0.03	0.89 ± 0.03	0.69 ± 0.06	0.89 ± 0.06	0.95 ± 0.04

Table 4.

Results of the four machine learning models with FCGR encoding on the Giessen data

Classifiers/drug	Precision	Precision	Precision	Precision	Recall	Recall	Recall	Recall
	CIP	CTX	CTZ	GEN	CIP	CTX	CTZ	GEN
CNN	0.87 ± 0.04	0.74 ± 0.04	0.81 ± 0.03	0.75 ± 0.02	0.91 ± 0.03	0.84 ± 0.04	0.87 ± 0.06	0.96 ± 0.01
LR	0.79 ± 0.08	0.70 ± 0.04	0.73 ± 0.05	0.69 ± 0.04	0.85 ± 0.04	0.79 ± 0.05	0.85 ± 0.04	0.86 ± 0.02
RF	0.91 ± 0.03	0.74 ± 0.01	0.82 ± 0.02	0.80 ± 0.02	0.87 ± 0.03	0.72 ± 0.07	0.90 ± 0.07	0.98 ± 0.01
SVM	0.81 ± 0.03	0.72 ± 0.03	0.73 ± 0.01	0.69 ± 0.02	0.88 ± 0.03	0.81 ± 0.05	0.87 ± 0.03	0.92 ± 0.03

3.2 Evaluation of the models on public data

We performed a further evaluation of our models using the public data of E.coli of Moradigaravand . The public data are highly imbalanced and thus performance metrics are difficult to interpret. Thus, to evaluate the performance of the models, we performed a down-sampling to balance the public data. For completeness, results for the imbalanced set are shown in Supplementary Tables S1–S3.

The resulting ROC curves clearly show that the machine learning models generalize well and can predict AMR (Fig. 3). The AUCs of RFs are higher compared with those from LR, SVM and CNN with three encoding approaches, except for CTZ and GEN with FCGR encoding. Consistent with the results from the Giessen data, all classifiers have high precision and recall for three encoding methods (Tables 5–7).

Fig. 3.

ROC curves for the models with label, one-hot and FCGR encoding on the public data. First row: ROC curves for CIP with label encoding (A), one-hot encoding (B) and FCGR encoding (C), respectively. Second row: ROC curves for CTX with label encoding (D), one-hot encoding (E) and FCGR encoding (F), respectively. Third row: ROC curves for CTZ with label encoding (G), one-hot encoding (H) and FCGR encoding (I), respectively. Fourth row: ROC curves for GEN with label encoding (J), one-hot encoding (K) and FCGR encoding (L), respectively

Table 5.

Evaluation of the machine learning models with label encoding on the public data

Classifiers/drug	Precision	Precision	Precision	Precision	Recall	Recall	Recall	Recall
	CIP	CTX	CTZ	GEN	CIP	CTX	CTZ	GEN
CNN	0.94	0.71	0.79	0.84	0.88	0.88	0.81	0.70
LR	0.93	0.76	0.80	0.82	0.90	0.84	0.75	0.62
RF	0.95	0.75	0.81	0.83	0.90	0.85	0.77	0.61
SVM	0.94	0.71	0.75	0.77	0.87	0.84	0.74	0.60

Note: Precision and recall are calculated based on balanced data using down-sampling.

Table 7.

Evaluation of the machine learning models with FCGR encoding on the public data

Classifiers/drug	Precision	Precision	Precision	Precision	Recall	Recall	Recall	Recall
	CIP	CTX	CTZ	GEN	CIP	CTX	CTZ	GEN
CNN	0.84	0.71	0.72	0.74	0.93	0.89	0.86	0.71
LR	0.85	0.77	0.79	0.80	0.89	0.87	0.86	0.74
RF	0.92	0.77	0.83	0.83	0.88	0.89	0.78	0.59
SVM	0.88	0.78	0.77	0.75	0.90	0.86	0.86	0.74

Note: Precision and recall are calculated based on balanced data using down-sampling.

3.3 Marker genes associated with antibiotic resistance

We performed an SNP association study on the Giessen and public data using the EFS R package with default parameters. In this analysis, we did not include the known resistance genes. Thus, we aimed at identifying secondary mutations that contribute to the resistance directly or indirectly, e.g. compensatory mutations. This data-driven approach does not need AMR expert knowledge and can also be used and predict resistance even without knowing the resistance genes but by identification of the secondary mutations. EFS provided a ranking of the SNPs for each antibiotic. The ten most important SNPs for each antibiotic are shown in Figure 4. These SNPs are part of 19 different genomic regions. We then annotated and analyzed the corresponding genes of these regions (Table 8).

Fig. 4.

EFS analysis for each antibiotic for both datasets. The left four figures are the identified ten most important SNPs for CIP (A), CTX (C), CTZ (E) and GEN (G) from the Giessen dataset. The right figures are the corresponding SNPs from the public dataset

Table 8.

SNPs and corresponding genes associated with AMR

SNP Position	Gene location	SNP annotation	Gene	Gene biotype	Drug
18169	17489 → 18655	Synonymous	nhaA	CDS	CTX, CTZ, GEN
898919	898518 → 899645	Synonymous	rlmC	CDS	CIP, CTX
2008324	2008277 → 2009482	Synonymous	yedE	CDS	CTX, CTZ, GEN
2017588	2016554 → 2017927	synonymous	fliI	CDS	CIP, CTX, CTZ, GEN
2148909	2147674 → 2149026	Synonymous	yegD	CDS	GEN
2655873	2655075 → 2656358	Synonymous	pepB	CDS	CIP, CTX, GEN
3099618	3098558 → 3099565	Upstream gene	yggM	CDS	CIP, CTX, CTZ, GEN
3644715	3643140 → 3645182	Synonymous	prlC	CDS	CTX, CTZ, GEN
4101302	4100810 → 4101430	Missense	sodA	CDS	CIP, CTZ
4127700	4127286 → 4127894	Synonymous	yiiX	CDS	CTX, CTZ, GEN
4172893	4172057 → 4173085	Missense	murB	CDS	CIP, CTX, GEN
4230581	4230354 → 4231226	Synonymous	rluF	CDS	CTZ, GEN
4441487	4439872 → 4441215	Upstream gene	ytfL	CDS	CIP, CTX, CTZ, GEN
4453756	4453583 → 4454578	Synonymous	yjfF	CDS	CIP, CTX, CTZ, GEN
4466572	4466299 → 4467246	Synonymous	treR	CDS	CIP, CTX, CTZ, GEN
4477553	4477307 → 4478311	Missense	argI	CDS	CIP, CTX, CTZ
4483166	4480982 → 4483837	Synonymous	valS	CDS	GEN
4605418	4604875 → 4605663	Synonymous	fhuF	CDS	CIP, CTX, CTZ, GEN
4627668	4627315 → 4628547	Synonymous	nadR	CDS	CIP

Note: The first column shows the positions of the identified SNPs for the four antibiotics. The second column and third column show the gene location and SNP annotation. The fourth column and fifth column show the genes annotated from SNPs and gene biotype. The final column is the antibiotics that are associated with the SNPs.

Some of these genes are well-known genes conferring antibiotic resistance, such as marA. marA is a gene related to multiple drug resistance (Abdolmaleki ). In comparison, the other genes have not been well studied so far. For instance, the gene nhaA (associated with CTX, CTZ and GEN resistance) displays a Na+/H+ antiport activity in E.coli that can regulate the permeability, which may further affect drug resistance (Padan ). The gene rlmC encodes a 23S RNA methyltransferase that methylates the 23S rRNA, of antibiotic binding sites and is related to antibiotic resistance (Pletnev ; Stojković ). It has been reported that the gene fliI encodes a virulence factor, and some studies focused on the correlation between antimicrobial resistance and bacterial virulence (Beceiro ; Deng ). The gene pepB encodes the peptidase B, which is related to the production of bacteriocins, narrow-spectrum antimicrobial peptides produced by bacteria (Suzuki ; Telhig ). MurB is the key biosynthetic enzyme involved in the synthesis of peptidoglycan, the key component of the cell wall (Nasiri ; Walsh and Wencewicz, 2014). In sum, the marker genes and SNPs identified by EFS can be used as a reference for further AMR studies.

4 Discussion

This study analyzed four different machine learning methods (RFs, LR, SVMs and CNNs) for predicting four antibiotic resistances in E.coli based on whole-genome sequence data with three different encoding schemes, namely, label encoding, one-hot encoding and FCGR encoding. Moreover, our goal was to identify mutations (secondary mutations) contributing to resistance beyond known resistance genes. Thus, we used a reference genome for E.coli without known resistance genes. Our study confirmed that label encoding, one-hot encoding and FCGR encoding could encode genomic data for preparing the input data for subsequent machine learning and deep learning methods. Our results show that the four machine learning methods can effectively predict AMR without the need for a database of known resistance genes or SNPs, which is an essential prerequisite for AMR prediction in less well-studied pathogens and drugs. Furthermore, we provide potential genes and SNPs associated with AMR based that can be used as a reference for the subsequent experiments.

Previous studies reported different SNPs in the bacterial genome associated with multiple drug resistance (Brimacombe ; Figueroa ; Shi ; Su ; Yang ). However, these studies mainly focused on partial SNPs based on available AMR databases (Yang ). Machine learning based on the complete set of SNPs from whole-genome sequencing gives further insights and can be used to identify novel biological mechanisms of resistance.

Encoding the genomic features into a readable format for ma-chine learning and deep learning is an essential step. Label encoding, one-hot encoding and CGR encoding can convert SNPs into machine-recognizable formats very efficiently. Our study used the three approaches to encode SNPs and yield excellent predictions for both encoding methods. Many studies indicated that CNNs outperform other machine learning algorithms in image classification, which was the rationale for incorporating FCGR as an encoding scheme.

We compared four machine learning methods, including RFs, LR, SVMs and CNNs. Overall, the four machine learning methods showed good performance in predicting the four antibiotic resistances of E.coli. We also demonstrated that our models generalize well on unseen data, as proven by validating the results based on an independent public dataset. We were also able to identify SNPs associated with resistance. However, the marker genes located around the SNPs associated with AMR need experimental validation.

Although we only focused on four antibiotics in this study, our method can easily be applied to other antibiotics and can also be extended to other resistance-related SNPs of other pathogens, also from species other than bacteria. Furthermore, our approach can also be applied to other biomedical areas, e.g. for cancer resistance prediction. More importantly, our method may have huge potential in systems medicine, to improve the diagnosis, targeted therapy and disease prevention.

There are also some limitations in our study. For example, we only used SNP data in our models that have been called based on a single reference genome. This, however, spares many genomic regions that might be important resistance factors. This is especially true for diverse species like E.coli. One approach to mitigate this issue would be the selection of more suitable or multiple reference genomes. Another option potentially leading to a more holistic set of potential SNPs would be to use an artificial pseudo-pan-genome incorporating many genomes of a particular species as a reference within the SNP detection workflow. However, other features, e.g. transcriptomics or proteomics data, might be important for AMR as well (Moradigaravand ). Moreover, several other important drugs have not been taken into account yet. However, they may be analyzed with the same methodology when enough data are available.

5 Conclusion

We investigated four machine learning methods for predicting AMR to four different drugs in E.coli from whole-genome sequence data with label encoding, one-hot encoding and FCGR encoding. Our results demonstrated that all methods perform very well also for unseen data. Overall, our study provides a new machine learning-driven approach for resistance prediction and thus, may improve treatment of patients in the future.

We evaluated the performance based on cross-validation on our own data and tested the model performance on public data. Moreover, we identified potential SNPs and corresponding genes that are associated with AMR.

We could demonstrate that label encoding, one-hot encoding and FCGR encoding can be used for whole-genome sequence analyses. Moreover, we provide a comprehensive evaluation of different machine learning algorithms for AMR prediction in E.coli. The results of the study give a rich reference resource for further research on both experimental and computational aspects of AMR.

Click here for additional data file.

Table 6.

Evaluation of the machine learning models with one-hot encoding on the public data

Classifiers/drug	Precision	Precision	Precision	Precision	Recall	Recall	Recall	Recall
	CIP	CTX	CTZ	GEN	CIP	CTX	CTZ	GEN
CNN	0.95	0.83	0.84	0.80	0.90	0.83	0.78	0.62
LR	0.90	0.80	0.76	0.81	0.90	0.85	0.78	0.63
RF	0.90	0.78	0.73	0.81	0.90	0.86	0.78	0.63
SVM	0.89	0.78	0.75	0.73	0.88	0.83	0.77	0.55

Note: Precision and recall are calculated based on balanced data using down-sampling.