Antibiotic resistance genes (ARG) are certain genes found in bacterial deoxy nucleic acid (DNA) that provide antibiotic resistance. These genes can be acquired either through horizontal gene transfer, in which bacteria trade genetic material with one another, or through mutation. Plasmids, which are compact, circular DNA units that are easily transferred between bacteria, include ARG that can spread quickly throughout a bacterial population [12, 51]. To cure diseases brought on by bacteria resistant to antibiotics, it is crucial to target the genes responsible for antibiotic resistance. To combat AMR, it is crucial to raise public knowledge of the hazards associated with improper usage and excessive use of antibiotics. Also, it is crucial to correctly diagnose the infection to determine the kind of bacteria that caused it and, consequently, apply the right antibiotic treatment [6, 51]. The first step in creating efficient treatments for diseases brought on by resistant bacteria is to pinpoint the genes responsible for AMR. Identification of differentially expressed genes using gene expression data is another crucial component of AMR study; it helps to comprehend the state of the infection and offers more clarity for identifying ARGs.

The complete pipeline of this work is depicted by the block diagrams in Fig. 3. It comprises several quality control methods applied to the data preprocessing, various model stages, and the outcome. The architecture of aiGeneR gene identification model uses an extensive quality control pipeline to preprocess gene expression data, which includes min-max normalization and Log2 transformation while filtering genes according to a stringent p-value threshold of 0.05. Next, it makes use of XGBoost for feature selection and a deep neural network to classify infected data samples. Power analysis, evaluation of sample size effects, generalization abilities, and quantification of memorizing tendencies are some of the factors that are used for evaluating model performance. Additionally, aiGeneR’s biological validation highlights the importance of hub genes and the discovery of antibiotic-resistance genes, emphasizing its applicability in the fields of gene expression analysis and infectious disease investigation.

Fig. 3.

The overall architecture of aiGeneR and other studied models. XGBoost, eXtreme Gradient Boosting.

A large number of samples are needed to train a deep-learning model because a limited training set will result in overfitting. The accuracy curves and loss curves of the training and validation sets provide the most detailed insight into the fitting process. The training and validation set curve trends should be comparable to one another for optimal fit. A reduction in model complexity is required if the accuracy or loss of the training set differs from those of the validation set. These differences indicate overfitting. The performance of the model prediction needs to be enhanced in the absence of underfitting [52].

We construct a basic Multilayer Perceptron (MLP) neural network to perform a binary classification job with prediction probability for DNN. The Keras library, which is based on Tensorflow, is commonly used in Python 3.7 (Python software foundation, Wilmington, DE, USA) [53]. The input dimension of the dataset is 30. One hidden layer comes before one output layer. The accuracy score is the measurement of the model performance. If there has been a significant rise in accuracy ($>$80%) after 20 epochs, the learning process is stopped using the early stopping callback. For aiGeneR 1.0, we construct the architecture with two hidden layers with 12 nodes each and the input layer is of 30 nodes. With this architecture, we can visualize there is a significant improvement in the accuracy ($>$90%) after 17 epochs. We evaluate all the implemented models including the ANN and aiGeneR with Python 3.7 using Jupyter Notebook in Anaconda Navigator 2.3.1.

The dataset for this work is obtained from the National Center for Biotechnology Information (NCBI) and the source (URL) of the dataset is “https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE98505”. The dataset explores the historical function of the synthetic protein MalE-LacZ72–47 in causing cellular stress and its deadly impact on bacteria. The study’s focus on downstream metabolic processes shows that the ROS-dependent component of antibiotic lethality and MalE-LacZ lethality are identical. Growing in M63 medium, E. coli MC4100 cells expressing a MalE-LacZ hybrid protein under a maltose promoter (MM18) were stimulated with 0.2% maltose. To extract RNA, the cells were shaken and incubated at 37 °C for five hours. Samples were taken every hour. Increased susceptibility is seen in oxidative stress-sensitive mutants, suggesting that reactive oxygen species (ROS) cause cell death. The number of samples and genes in this dataset are summarized in Table 1. However, it is found that the dataset taken for our experiment is balanced with both positive and negative samples. The raw data and the processed data are the same since no genes with null values greater than 30% were discovered during the imputation phase.

We found that the dataset (GSE98505) is having null values and the expression value ranges from 0 to 16. We aim to remove the genes which are having more than 30% null values but, there are no such genes identified. To reduce the computational burden, we apply the normalization process to the dataset. The data pre-processing phase includes data imputation, normalization of the raw data, Log2 transform, and p-value measure [14].

In the first step of data processing the duplicate values are removed. Some well-accepted imputation methods for numerical features includes rounded mean. In this case, the approach substitutes null values for that feature’s mean, rounded mean, or median values found across the whole dataset. The rounded mean data imputation technique is used to fill in the null or missing values. The method aids in maintaining the data’s overall distribution by substituting missing values with the rounded mean [54]. The rounded mean imputation technique keeps part of the variable’s statistical characteristics [55].

Data normalization is done in the second step of data preprocessing. Here we deploy the min-max normalization technique. The min-max normalization normalizes the data without disturbing the other data due to variance in their original scale and it reduces all features to a standard and single scale which is best fit for our dataset [56]. However, it is also found that many machine learning algorithms’ convergence rates and performance can be enhanced by normalizing features using min-max normalization [57].

The third step of the data processing includes the Log2 transformation. For gene expression data, the Log2 transformation reduces the dynamic range, makes interpreting fold changes easier, and improves statistical stability and visualization. The fourth and final step of data processing holds the processed data based on a p-value less than 0.05. R statistical software (version 4.2.0, The R Foundation for Statistical Computing, https://www.r-project.org/foundation/) was used to perform all statistical analyses [58]. With a statistical significance criterion of p 0.05 (unless otherwise stated), the Log2 transformation was used to retrieve significantly enriched genes for all database functional analyses.

Artificial neural networks (ANNs) and deep neural networks (DNNs): Because they are capable of accurately capturing the intricate interactions between genes and phenotypes, ANNs and DNNs are frequently utilized in gene expression data processing. These models work especially well for tasks like predicting disease outcomes and classifying gene expression. As our work also focuses on gene network analysis, where the objective is to find interactions between genes, the performance of ANNs and DNNs is found significant [59, 60, 61]. The reason behind choosing ML models like XGBoost, SVM, and RF is, these models can handle high-dimension data, are robust to overfitting, and have the ability of non-linear transformation [24]. In addition to this, XGBoost can be utilized to predict the course of a disease or find biomarkers for particular illnesses (cancer) [18]. SVM is frequently employed in the study of gene expression data because it is capable of revealing intricate connections between genes and phenotypes. Similarly, RF is frequently utilized to predict the course of a disease or find biomarkers for particular diseases using gene expression data [62, 63].

In this study, the proposed aiGeneR is the capsule that binds the DNN algorithm for classification which performs incredibly well on the feature that the XGBoost model has selected. The model accuracy of aiGeneR improved significantly compared with the model run on raw data and the model run with selected features. The general equation for XGBoost feature selection is shown in Eqn. 1,

(1)$$\widehat{a}={\sum }_i{w}_n{f}_n(b)$$

Where the predicted value of input data b is $\widehat{a}$, the total number of distinct trees n in the ensemble is denoted by $\sum$i, and w${}_n$ is the weight given to tree n based on how much it helped to lower the overall loss function. The prediction for tree n on input x is called, $f_n\left(b\right)$, and it is determined by going around the decision tree and giving each leaf node a value depending on the input attributes.

The aiGeneR performs the classification problem by combining the XGBoost feature selection algorithm and DNN architecture. This paradigm gives the genes that are prone to antibiotic resistance, informative for disease prediction, and hub genes, which are in charge of tightly managing a large number of genes through strong cluster correlation. The biological validation section (section VII) explains in various points about this.

The following is the algorithm for XGBoost feature selection and DNN (aiGeneR) which gives the best classification result compared with other ML models.

Step 1: Divide the dataset into train test sets (7:3). Run the XGBoost model with all the features (Baseline model).

Step 2: Repeat each feature’s evaluation using XGBoost to determine its significance. Metrics like feature gain is used to evaluate a feature’s significance.

Step 3: Choose the Top-10, Top-20, and Top-30 features from the XGBoost feature ranking output.

Step 4: To create and train the DNN classifier, import the necessary libraries, such as TensorFlow or Keras.

Step 5: The Top-10, Top-20, and Top-30 features will be the input to the DNN model.

Step 6: Finally make the training and test sets on the input. The testing set will be utilized for evaluation, while the training set will be used to train the DNN classifier.

Algorithm 1 aiGeneR

##Taking the gene expression raw dataset as input

Input: Dataset DS (X = 36, Y = 10576): The set of samples and genes

## The quality control and feature ranking

Output: Normalized DS, p 0.05, Log2 transformation

Feature selection (X = 36, Y = 5730)

Feature Ranking [DS1(X = 36, Y = 10), DS2(X = 36,

Y = 20), DS3(X = 36, Y = 30)]

## Splitting the ranked features into to train-test set

Split the DS to DS${}_{\text{Tr}}$ and DS${}_{\text{Te}}$ as the train and test dataset with a split ratio of 7:3

## Proposed DNN model implementation phase

FOR (ILR = 1$\to$20) do

Weight {W${}_{\text{i}}$ = W${}_1$, W${}_2$, ……, W${}_{12}$}:

FOR (HLR = 1$\to$12) do

FOR (W= W${}_1$$\to$W${}_{12}$) do

FOR (WE${}_{\text{i}}$ = W${}_{10}$$\to$ W${}_{21}$) do

FOR (N${}_{\text{i}}$ = 1 $\to$ 12) do

N${}_1$ = Wi* ILR${}_1$ + Wi* ILR${}_2$ + …….+

Wi* ILR${}_{20}$ + WEi

OP${}_1$ = W${}_{\text{O}}$P${}_{11}$*N${}_1$ + W${}_{\text{O}}$P${}_{12}$*N${}_2$ +

………+ W${}_{\text{O}}$P${}_{22}$*N${}_{12}$ + W${}_{\text{O}}$P${}_{10}$

OP${}_0$ = W${}_{\text{O}}$P${}_{01}$*N${}_1$ + W${}_{\text{O}}$P0${}_2$*N${}_2$ +

………+ W${}_{\text{O}}$P${}_{12}$*N${}_{12}$ + W${}_{\text{O}}$P0${}_{10}$

END

END

END

END

END

The DNN used in aiGeneR is intended to classify E. coli bacterium infection in biological samples. It consists of several artificial neural layers, with two hidden layers positioned in between the input and output layers. The network architecture is specifically designed to handle the input data with 27 features and generate accurate classification results.

Architecture:

(a) Input layer: There are 27 nodes in the input layer, each of which corresponds to a different attribute that was taken from the biological samples. These qualities include the expression value of different genes in the sample.

(b) Hidden Layer: This deep neural network has two hidden layers, each with 12 nodes. These hidden layers act as processing units in between, converting the incoming data into a feature space that is more abstract and representative. A rectified linear unit (ReLU) function serves as the activation function for each node in the hidden layers, which each apply a weighted sum of inputs from the layer before. This non-linearity makes it possible to identify intricate linkages in the data.

(c) Output layer: There is just one node in the output layer. The anticipated chance that the input sample is contaminated with E. coli is represented by the output node’s activation value in this binary classification problem. Typically, a sigmoid activation function is used to compress this number into the range [0, 1], with values closer to 1 denoting a higher likelihood of infection.

A labelled dataset of E. coli-infected and non-infected samples is utilized to train the DNN. Through the use of an optimization technique called Adam, the network learns to modify the weights and biases attached to each link between nodes in the layers. Utilizing a loss function that measures the discrepancy between expected and real labels, the network’s performance is assessed. Binary cross-entropy is a typical loss function for binary classification applications. To achieve optimum performance, hyperparameters like the learning rate are set at 0.0001 and batch size is 42. In order to prevent over fitting, a 3-fold cross-validation is also used during training. This deep neural network architecture in aiGeneR 1.0, which includes 27 input nodes and two hidden layers, was created especially for classifying E. coli infections in biological samples.

In the above algorithm, ILR contains the input layer nodes and HLR contains the hidden layer nodes. W and WE${}_{\text{i}}$ are the weights for the input and hidden layer respectively. OP${}_1$ and OP${}_0$ are the two nodes of the output layer. The algorithm is based on a deep network having one input layer with 27 nodes, two hidden layers with 12 nodes each, and the output layer where the classification results were obtained.

In this section, we discussed the working procedure of the DNN classification model. The deployment of the proposed model is done with the architectural modification of the baseline DNN model. We focus on the model evaluation techniques, evaluation metrics used, and baseline model of DNN for our work. The implemented DNN model has having input layer, two hidden layers, and one output layer. The DNN model was trained for 20 epochs, with 2 samples in each batch. To prevent overfitting, an early stopping mechanism was also implemented. The early halting mechanism, which reduced the learning rate to 0.001 of the previous learning rates, was activated specifically if the accuracy in the validation set did not increase by 0.0001 within 17 epochs. The Top-10, Top-20, and Top-30 features chosen by the XGBoost feature selection model is used to determine the number and dimensions of aiGeneR model’s input nodes.

The proposed aiGeneR model consists of four major steps namely quality control, effective feature selection, classification, and biological interpretation as shown in Fig. 3. The main functionality of this model is to extract significant features, observe the model performance, and reduce the computation burden. However, the computational time is much less if the learning model operates with selected features [66].

The different steps of this deployed model are, step-1 includes the used dataset, step-2 holds the data preprocessing and feature selection used for data preparation, and step-3 is meant for the classification of infected samples. The last section of our proposed model (step 4) represents the hub gene identification and biological validation. The basic operation of the model starts with the data pre-processing and feature selection process as used by our group previously [67]. Here we evaluate the XGBoost feature selection model to find the most significant features from the dataset. The evaluation is based on training the XGBoost model on our dataset using the labels as the target variable and the gene expression levels as features. According to how much each feature (gene) contributes to the prediction, XGBoost automatically gives importance scores for every feature (gene) during the training phase. The advantages of the XGBoost feature selection technique help to find significant features which helps to increase model accuracy. The ability of the XGBoost feature selection technique to deal with missing values, outliers, and non-linear data makes it more popular, which is shown in this section [68].

Features selection and optimization are crucial processes in the analysis of gene expression data. The selection of the most pertinent features becomes essential for correct insights and model performance because many genes may influence outcomes [24]. Finding a selection of genes that cause the observed changes is the goal of this technological study.

The genes are selected by deploying the XGBoost feature selection model. The top-ranked genes selected by the XGBoost model are then used by the different classifiers proposed in this work. The XGBoost feature selection model is implemented on the 5571 genes selected after data preprocessing. The XGBoost feature selection model selects and ranks 479 genes as shown in Fig. 4.

Fig. 4.

The genes selected by the XGBoost feature selection model with their importance score and Gene number. XGBoost, eXtreme Gradient Boosting.

In Fig. 4, a few Top-ranked genes which are having feature importance scores of more than 0.01 are marked with different color (blue, orange, and dark green) than other selected genes. The highest feature importance score obtained is 0.24 and the lowest is 0.00014. We then take the Top-10, Top-20, and Top-30 ranked genes and form three different datasets, and applied the classification model to these datasets. The Top 30 genes based on their feature importance score are shown in Fig. 5.

Fig. 5.

Top-30 ranked genes with their rank value and gene number.

Classification is just one of the many machine-learning tasks that can be performed with ANNs. An artificial neural network collects input data for a classification problem and outputs a categorical result. The classification performance of the learning model highly depends on the model tuning. Model tuning is a crucial phase in the ML process since it can enhance the model’s functionality and increase its predictive power [70]. A few key parameters for the deployed ML models in this work are discussed below.

The dataset we have taken has having small sample size, a short validation set would not give a reliable indication of the model’s performance. K-fold cross-validation is one way to handle such a situation [71, 72]. Except for the class distribution of the dataset being kept throughout the splits, the splitting technique is similar to the repeated K-fold cross-validation. In other words, each fold will have an identical distribution of samples across classes as the original dataset. So, for classification tasks with unbalanced class distributions, stratified K-fold cross-validation will be more appropriate [52, 72]. In our implementation phase, we take the k value as 3 for all the models. The deployed XGBoost, SVM, and RF classification model has followed the K-fold cross-validation from the train-test split. Based on the validation accuracy, precision, recall, f-score, false positive rate (FPR), and false negative rate (FNR), the XGBoost, SVM, and RF model performance is evaluated.

The deep network-based classification model used in this study was tested using the same methodology as the classification models mentioned in Section 4. Training and validation accuracy curves and loss curves were initially plotted to pre-screen experimental configurations with good performance to choose the best set of hyper-parameters for the model. The best parameter combination was then chosen by repeating the trial settings with good performance 3CV 10 times and using the average AUROC as an evaluation indicator. The performance of the ANN and DNN models is measured based on the validation accuracy (ACC), precision (PRE), sensitivity (SEN), specificity (SPE), f-score (F1), FPR, and FNR (Appendix C).

Our proposed model, aiGeneR, is quantified in this section, along with a thorough examination of its correctness. For its remarkable predictive abilities in a variety of tasks, from classification to regression, the aiGRNER 1.0 algorithm, a variation of the XGBoost method with the DNN classification algorithm, has drawn a lot of attention. Our goal is to thoroughly evaluate the accuracy of aiGeneR and learn more about its performance traits using various datasets.

The model metrics for different learning models with raw datasets (without feature selection) are shown in Table 4, and Fig. 6 shows the performance of these learning model metrics. With an impressive classification accuracy of 75%, the non-linear aiGeneR model outperforms the linear SVM. The measures show that the proposed aiGeneR model exceeds the other model in terms of classification accuracy which is more than 20% than XGB+ANN, XGB+XGB, and XGB+SVM classification models. It is observed that the XGB+RF classification model resulted in poor accuracy of only 37% and 0% specificity which indicates a large number of false positives and an inability to correctly detect negative examples.

Fig. 6.

Classification model metrics for all the models on raw data.

Across three different feature sets, the aiGeneR model showed promise in classification tasks as shown in Table 5. The model produced relatively high accuracy and precision while maintaining a reasonable balance between recall and precision when tested using the Top-10 attributes. When the aiGeneR model was tested using the Top-20 features, its performance significantly improved. A stronger overall ability to predict and a more balanced trade-off between precision and recall are shown by improvements in accuracy, recall, and F1 score. The experiment is carried out based on the protocol discussed in the experimental protocol (EP) in section IV.

A stronger overall ability to predict and a more balanced trade-off between precision and recall are shown by improvements in accuracy, recall, and F1 score. The perfect specificity and low false positive rate demonstrate that the model has sustained superior performance in correctly classifying negative samples. However, for the Top-30 feature, the performance of aiGeneR is unchanged. The proposed aiGeneR (XGBoost feature selection and DNN classification) model successfully used feature information to generate precise predictions for the classification problem. It is crucial to note that the model’s performance was considerably impacted by the choice of the most significant features, highlighting the significance of feature engineering and selection in machine learning pipelines.

To choose the most insightful features from the dataset, we used three separate datasets based on feature selection (ranking). The datasets are Top-10, Top-20, and Top-30. The five different classification algorithms aiGeneR, ANN, XGB, SVM, and RF were also coupled with these features to create a comparative classification model. Performance measures like accuracy, precision, recall, F1 score, specificity, false positive rate, and false negative rate are taken into consideration for the deployed model’s potential testing on the identification of infected and non-infected samples. Figs. 7,8,9 show the model metrics on Top-10, Top-20, and Top-30 genes respectively, and Fig. 10 summarizes the performance of all these model metrics in terms of classification accuracy.

Fig. 7.

Classification metrics of all the models (Top-10 features).

Fig. 8.

Classification metrics of all the models (Top-20 features).

Fig. 9.

Classification metrics of all the models (Top-30 features).

Fig. 10.

All Models accuracy on Top-10, Top-20, and Top-30 features.

Using XGBoost feature selection techniques, we compared how well machine learning models performed at classification tasks during the experiment phase. The outcomes revealed that the adoption of the feature selection technique significantly affects the model’s classification performance. When compared to how these models perform on raw data, it is also seen that classification models applied to the Top-20 features yield the best classification accuracy. Additionally, models with fewer features lighten the computational load and offer the best classification accuracy. In addition to this, high accuracy and precision were continuously attained by aiGeneR, making it an excellent contender for classification tasks for our defined objective. This observation is obtained with model evaluation metrics in experimental protocol (EP) section (section IV).

The observation in figure (Fig. 10) clearly shows that the aiGeneR model acquires a higher classification accuracy of a minimum of 10.08% (for all the 30 ranked feature datasets) and a maximum of 14.9% (for the Top-10 ranked feature dataset) in comparison to the other proposed models. However, it is also seen that all the proposed models perform well on the selected feature set of Top-20 and Top-30 as compared with the Top-10 feature set (Appendix Table 11). It can be concluded from our hypothesis that; the proper selection of significant features boosts the performance of the classification model in terms of accuracy. A minimum of 20 features is needed by aiGeneR to attain the best model accuracy.

The false positive rate (FPR) and false negative rate (FNR) for all the implemented models on the Top-ranked feature dataset are shown in Fig. 11. The minimum FNR and FPR for each feature set is denoted as A1, A2, A3 and B1, B2, B3 respectively. The FPR and FNR values are the same for SVM in Top-10 features set is denoted by C1 and for ANN model on Top-20, and Top-30 feature set is denoted by C2 and C3. The average FPR for the Top-10, Top-20, and Top-30 feature datasets is 0.98, 0.52, 0.74 and the average FNR is 1.39, 0.72, and 0.69 respectively. It is observed that the FPR and FNR are reduced with Top-20 and Top-30 ranked features (genes) as compared to Top-10 ranked features.

Fig. 11.

False positive rate and False negative rate of all the studied models for Top-10, Top-20, and Top-30 ranked features.

The XGBoost feature selection algorithm applied to the raw data selects 471 (four hundred seventy-one) initial features as shown in Fig. 4. The selection is based on feature ranking which uses the Gini index for ranking the selected genes and can be visualize in Fig. 5. In this work, we take the Top-30 ranked genes for the analysis of the performance of the proposed models. We carefully searched for the presence of the AMR genes in the dataset, and it was found that there is a single AMR gene present in the dataset, and that gene is selected and ranked among the Top-30 genes by the XGBoost model. The selected Top-30 ranked genes and their feature importance number (the position of genes in the dataset), and gene symbol are shown in Table 6 and the characteristics of these (aiGener-identified) genes are shown in Appendix Table 12 (Ref. [57, 58, 59, 60, 61, 62, 67, 68, 69, 70]) (Appendix F).

In addition to the above, the Top-30 ranked genes selected by the XGBoost feature selection model with their rank, feature importance number (F#), gene id, and gene name (gene symbol). This gene ranking table presents a prioritized list of the genes in the dataset based on their feature importance ratings. The genes that are ranked 1, 3, 8, 9, 22, 28, and 30 are highly correlated with other genes based on the number of genes connected to them.

Escherichia coli (E. coli) often carries the tetM gene. Tetracycline, a popular antibiotic used to treat various genes (gene-id) there are only 15 genes which are having their gene symbols. The tetracycline resistance genes are a family of genes that includes the tetM gene. ‘tetM’, a ribosome protection protein, is a protein that is produced by the tetM gene. It works by attaching to the ribosome and blocking the antibiotic tetracycline from attaching to the ribosomal target site [75]. The tetM gene is identified by our proposed model and ranked in 15th place as shown in Table 6. Due to the limited gene expression data availability for E. coli, the presence of ARG is very less. In this work, we deployed XGBoost feature selection method for its simplicity and significant performance over gene expression data. Several feature selection methods like PCA, LDA, t-SNE, PCA Polling can be tested on this data and comparison of classification performance may include in future work.

The Receiver Operating Characteristic (ROC) curve is a crucial indicator of a classification model’s efficacy. We examine the performance analysis of our proposed aiGeneR along with ANN, XGBoost, SVM, and RF with a value of p 0.05. The K-3 cross-validation is used to figure out how the accuracy of each of these models varies as the amount of training data changes. The dataset employed in this work is non-linear and complicated, which makes conditionality problematic. These problems are essentially handled by the quality control process used in this study.

More importantly, the feature selection technique which provides the most significant features helps to improve the performance of the aiGeneR model. Fig. 12 shows the ROC performance of the five classification models (aiGeneR, ANN, XGBoost, SVM, and RF). Our proposed model aiGeneR has accomplished a remarkable milestone with a robust area under the curve (AUC) value of 98.4%. However, the ROC value of RF is lowest compared with all other classification models. In the analysis process of gene expression data despite the challenges of the implemented complex non-linear dataset aiGeneR achieves the best AUC value.

Fig. 12.

Receiver operating characteristic of all the classification models.

This study also comprehended the implemented model’s performance on all possible train-test split and the comparison of classification accuracy on test data. The size of training data has an impact on the learning model and makes the model generalized well to unseen data [34]. We evaluate our proposed model aiGeneR along with four other classifiers used in this study on a used dataset with different train-test splits. It is observed that aiGeneR requires very minimal cases for generalization whereas other models require a greater number of cases. The detailed discussion on the effect of data size on our proposed model is discussed in this section. All the possible train-test splits on the used dataset and the comparison of classification accuracy on test data are shown in Fig. 13.

Fig. 13.

Visualization of classification accuracy achieved in different train-test splits of all the studied learning models.

The goal is to track how the train-test split ratio influences the performance of the model as per the EP (section IV) effect of data size. In the case of the aiGeneR classification model, as the percentage of training data rises, accuracy progressively rises. When using a 70:30 train-to-test split ratio, the model obtains the best accuracy of 93%. The XGBoost-based ANN, XGBoost, and SVM classification model achieves the best classification accuracy with the 70:30 train-test split and the RF classification model reaches the maximum accuracy with a 60:40 train-test split. Our observation on this analysis concludes, that all the studied learning model archives an optimum accuracy with a 70:30 train-test split.

Model generalization is the capacity of the model to function effectively on novel, untested data, suggesting its resilience and applicability for practical applications. The least number of unseen instances and minimum amount of data needed for the generalization of the proposed learning models are shown in Table 7. The least number of new instances and the minimum amount of data needed for model generalization are shown in the table for each machine learning model. We evaluate the model generalization on the Top-30 selected features with 36 samples (cases). A minimum of 40 data points is needed for generalization in both DNN and ANN models. Additionally, to validate the model’s performance on fresh data, at least 16 previously unreported cases are required. To achieve generalization, the XGBoost model needs a larger dataset with at least 70 data points and minimum of 25 instances is required for verifying unrecognized circumstances. Similarly, to achieve generalization, SVM and DT require 60 data points with 22 unseen instances.

We executed a power analysis to establish the minimal sample size required for precisely and accurately calculating a population proportion. The tests were carried out using the technique mentioned in [65, 77, 78]. The sample size calculation formula, denoted by the symbol Sn, is as follows,

(2)$$S_n=[{(z*)}^2\times \left(\frac{\tilde{p}(1-\tilde{p})}{Mo{E}^2}\right)]$$

Here, MoE stands for the margin of error, $\tilde{p}$ is the estimated proportion of the feature in the population and z* is the Z-score associated with the appropriate confidence level. Half the breadth of the confidence interval was used to calculate the MoE${}^2$. We settled on a proportion of 0.5 and a confidence interval of 95% for our experiment. To implement the power analysis, we use MedCalc [76, 79] and the obtained result is shown in Appendix Fig. 18.

As can be observed from Appendix Fig. 18 (Appendix D), the study has a sample size of more than what is necessary to meet the desired level of statistical power and classify accurately. The minimal sample size for the used dataset is also less than the amount of data accessible. However, to increase the classification model’s accuracy, statistical power, and precision, data augmentation may be used.

The scientific validation of our proposed work uses the “Microarray transcriptomic profiling of patients with sepsis due to faecal peritonitis and pneumonia to identify shared and distinct aspects of the transcriptomic response” (E-MAT-5274) dataset which is available in ArrayExpress [73]. The characteristics of the dataset are described in Table 8. We evaluate our proposed model with the E-MAT-5274 dataset keeping all the model configurations and parameters as per our proposed pipeline. It can be observed from Table 9 that, the trend in the Top-20 and Top-30 selected feature groups achieves the same level of classification accuracy as our proposed model with the E. coli dataset.

This experiment indicated our proposed model may be used as a benchmark model for infected sample classification and informative gene identification using the gene expression datasets. The accuracy of classification achieved by aiGeneR on this dataset is still the greatest and has not altered, demonstrating the potential for the generalization of our approach. This indicates the validity of our claim that aiGeneR is a generalized model that can access various gene expression datasets to identify the most important genes.

This section explores the critical function of functional association and gene network analysis in biological validation. By highlighting the potential roles of important genes in particular pathways and processes and revealing coordinated patterns of gene expression, these approaches make it easier to evaluate high-dimensional gene expression data. The key to demonstrating the applicability and precision of these analytical methods is the coupling of computational predictions with experimental confirmation.

7.2.1 Gene Network

A database of observed and anticipated protein-protein interactions is called STRING. Protein-protein interaction networks are mathematical representations of the physical contacts between proteins in the cell [80]. The interactions come from computational prediction, knowledge transfer across species, and interactions gathered from other (primary) databases; they comprise direct (physical) and indirect (functional) correlations. This analysis section provides some summary network information, including the number of nodes and edges. The average node degree is the average number of interactions a protein has in the network. Higher numbers of edges reflect a dense gene cluster and a gene having maximum numbers of edges will be treated as the hub gene. Gene-network study provides a clear view of the identification of significant genes and pathways, discovers the functional association, prediction of gene function, and identification of hub genes. Disease biomarker and drug target identification is also the key contribution of gene-network analysis [81, 82].

The proposed learning model is tested on the Top-30 (thirty) ranked genes and found there are only 24 (twenty-four) gene names (gene symbol) available in the used dataset. Using these 24 genes the gene network is being constructed with the help of STRING and it is found that out of 24 genes, 15 genes are available in the STRING dataset. While comparing the Top-30 and Top-20 feature datasets we found that out of these 15 genes present in the STRING dataset, 11 are also present in the Top-20 feature dataset.

The strain used by our suggested model for the genes chosen is Escherichia coli K12 MG1655. We increase the number of genes in our experiment to build networks which will make it easier to comprehend how genes interact with one another. We, therefore, take into account an additional 60 genes that belong to the same strain as our observed genes (model-predicted genes). Finally, the gene network we tested included 75 genes from the K12 MG1655 strain out of which 15 genes are identified by our suggested model.

We searched for the connections and functional associations between our researched gene sets and other genes in E. coli to further confirm the filtered gene set. Utilizing the stringApp of Cytoscape [83], which maps the genes to the STRING database of interacting proteins [80], identified 15 significant genes (colored red), and 60 other genes were linked to the protein-protein interaction (PPI) network as shown in Fig. 14. STRING involves functional relationships from selected pathways, computational text mining, and prediction techniques as well as tangible connections from experimental data [84].

Fig. 14.

Gene correlation network of Top-30 ranked genes from aiGeneR with 60 other genes in the K12 MG1655 strain of E. coli.

The number of nodes is the same as the number of genes (75) and the expected edges is 156 but, the network constructed in STRING shows the number of edges is 360 which is a sign that the obtained genes create a significantly more interacting network than excepted. The Genes identified by our model (Top-30 gene group), especially paaZ, polB, trpC, trpB, adk, paaX, and trpE shows the maximum number of connected genes and gene cluster to them as shown in Table 6. The genes selected by aiGeneR are given additional properties to serve as hub genes according to the interaction edge we discovered in our gene network and the deep connections among the genes. The tetM an ARG identified by our proposed model is resistant to tetracycline. Both Gram-positive and Gram-negative bacteria can exhibit tetracycline resistance, which is mediated by the genes tetM and other related genes. Through horizontal gene transfer processes like conjugation, transformation, and transduction, this resistance can spread between bacteria [85]. The higher classification performance of aiGeneR with gene network analysis gives us a thorough understanding of the hub genes and the most important genes present in the dataset.

7.2.2 The Pathway Analysis

The bar graph in Fig. 15 depicts the findings of a pathway analysis, which revealed significant metabolic processes active in E. coli. Among the identified pathways, the cellular aromatic compound metabolic process, organic cyclic compound metabolic process, and small molecule metabolic process are especially important. These findings are consistent with previous E. coli research that has demonstrated the importance of these pathways in the bacterium’s metabolism [86, 87].

Fig. 15.

Pathways association of selected Top-30 genes.

The analysis report also includes a few genes like PAAZ, PAAI, YFER, and UXUB that are connected to multiple pathways. These genes carry out novel metabolic processes in E. coli, including the hydrolysis of phenylacetyl-CoA and other aromatic molecules [88], which may be essential for E. coli to adapt to diverse environmental circumstances and use various carbon sources. However, certain genes listed in the table, such as POLB and ADK, have well-established roles in DNA replication, repair, and nucleotide metabolism, respectively. TRPB and TRPC, which encode enzymes involved in trypTophan biosynthesis, are also members of the well-studied trp operon in E. coli.

While these genes may not be associated with any new pathways, their presence in multiple pathways highlights their importance in E. coli metabolic processes. These findings provide a comprehensive overview of the metabolic network of E. coli and shed light on the interconnectedness of various pathways and the roles of specific genes within them. Further research into the functional significance of these pathways and genes will help us understand the physiology of E. coli and advance our understanding of microbial metabolism.

These pathways and genes selected by aiGeneR may also have implications for the pathogenesis of E. coli-caused urinary tract infections (UTIs), which are the most common cause of UTIs in humans. Some E. coli metabolism pathways and genes, such as those involved in iron acquisition, adhesion, toxin production, or biofilm formation, may contribute to virulence and survival in the urinary tract environment [89]. The genes identified by aiGeneR and the pathway analysis provide a detailed understanding of how these pathways and genes affect E. coli’s ability to cause UTIs could lead to new prevention and treatment strategies, especially in light of rising antibiotic resistance [89].

7.2.3 Differentially Expressed Genes

The genes displaying significant expression differences between the sick and healthy samples were found using DE analysis. To detect the Differentially Expressed Genes (DEGs), filtering criteria of padj(FDR) less than 0.05 (p 0.05) and Log2Fold-change $>$0.2 was applied. As the dataset has some limitations there is a very small number of significant genes present. Hence, we keep the Log2Fold-change value more than 0.2 to find out the significant genes in the dataset taken for analysis as shown in Fig. 16. The genes named z2263 (1759349), c5398 (1760188), yqek (1760655), c0161 (1767264), c1153 (1768175), c3811 (1762223), yddk (1764611) are positively expressed whereas cusF (1762115) in negatively expressed. The genes with color red are the significant genes (differentially expressed) and genes with color gray are non-significant. However, a few other genes which are positively expressed are missing names in the database.

Fig. 16.

The differentially expressed genes in the dataset after quality control (with p-value 0.05).

The core of our study involved applying hybrid ML models to classify E. coli infection cases and identify the relevant antibiotic resistance genes (ARGs). Deep network models were combined to create these machine-learning models. Therefore, it’s critical to compare our approach to earlier AI models. Considering this, we decided to compare our suggested models with earlier ML models (in AMR and other disease analyses) to directly address the benchmarking efforts.

There is an absence of research that combines ML and gene expression data to identify ARGs. Gene sequence information is used in the majority of research to classify resistance. Here, we evaluated two distinct gene expression and sequencing datasets that were utilized for cancer classification and AMR analysis in our benchmarking section. We chose cancer as the subject of our model benchmarking because machine learning has been used in numerous studies that use gene expression data.

For an accurate AMR analysis, data pre-processing, including cleaning, normalizing, and feature engineering, is essential. Several techniques in aiGeneR quality control pipeline, including min-max normalization, Log2 transform, a p-value criterion of less than 0.05, XGBoost feature selection, and deep neural networks, were used to find significant genes. Metrics like accuracy, precision, recall, and F1 score were used to assess the classification model’s performance on infected E. coli samples. The model achieved an F1 score of 93%, accuracy of 93%, precision of 100%, and recall of 87%. Additionally, the model’s adaptability to changes in the input data, generalizability to new data, and congruence with biological observations were all assessed. It is found that the model is reliable, generalizable, and consistent, according to the findings of these assessments.

Using gene expression data, our proposed aiGeneR model delivers hub genes and ARGs. The maximum classification accuracy is attained by the innovative, non-linear aiGeneR. Furthermore, the efficient feature selection used in our suggested pipeline plays a crucial role in improving classification accuracy. With various gene expression datasets, our suggested aiGeneR has demonstrated its generalizability while maintaining a high level of classification accuracy. The classification performance is enhanced by the significant genes that are identified by aiGeneR. It has also been noted that our approach achieves the maximum classification accuracy with just 20 genes. One of the most crucial features of our aiGeneR pipeline is its capacity to recognize hub genes, and the network analysis of the aiGeneR chosen has already demonstrated this assertion. Additionally, we assert that the aiGeneR identified genes are strongly linked to UTI, as revealed by the pathway analysis of these genes.

Four different models, including RF, DNN, DT, and srst2 [90], are implemented in [91]. The performance of DT was found to have a high classification accuracy of 91% when the models were evaluated based on classification accuracy. In this work, gradient boosting tree classifier is implemented with 0.1 learning rate, 300, 600, and 5000 boosting stages, deviance loss, and an 8:2 train-to-test split. Similar genetic characteristics that cause AMR are found in [92] by employing the SVM. Two SVM ensembles were created for each antibiotic case using the same feature matrix and AMR phenotypes: one with 500 SVMs trained on 80% of genomes with all features, and another with 500 SVMs trained on 80% of genomes with 50% of features, aiming to enhance SVM accuracy with high-dimensional biological data. It has been shown that the SVM model’s gene identification accuracy was 90%. However, most models that employ gene expression data choose feature selection techniques. The research published in [18, 30], and [93] used a variety of ML models to identify genes and categorize cancers. SVM, XGBoost, Neural networks, RF, and DT are the ML models used in this work. The XGBoost model in [18] achieves the best classification accuracy of 96.38%, the XGBoost model in [93] achieves the highest classification accuracy of 80%, and the SVM model in [30] achieves the highest classification accuracy of 96.38%. All these ML models are implemented on gene expression data. The work considered for this is shown in (Table 10, Ref. [18, 30, 91, 92, 93, 94, 95, 96]). A DeepPurpose DL model, which makes use of gene expression data, was deployed in [94] for the detection of Target genes and drug-resistant melanoma. The affinity score provided by the Deep Purpose (which is calculated based on the targeted genes and their potential drugs) is used as the performance measure. The model metrics are not provided in the publication; instead, the authors simply provide the number of genes that the model has identified. In [95], an experiment was conducted to predict antibiotic resistance using SVM and gene expression data. The model accuracy that was attained was 86%. Drug resistance and biomarkers in colon cancer identification are conducted by [96]. It obtains an AUC value of 0.6590 using gene expression data and elastic net regression.

It is not feasible to perform benchmarking specifically on ARG identification and classification of infected E. coli samples using gene expression data. As a result, we choose to compare our suggested model with the work in oncology. The model we propose is concentrated on E. coli infectious sample classification and ARG identification. The classification accuracy of the proposed aiGeneR is 93% with an AUC value of 98.4%, which is the highest of any model currently in use for AMR analysis of gene expression data. The generalizability of our model may be demonstrated by the classification accuracy and AUC of aiGeneR and its validation on the E-MAT-5274 gene expression dataset (section VII).

Access to diverse and extensive datasets that contain details on infections, drugs, and resistance mechanisms is necessary for AMR studies. Due to the restricted availability of such data, obtaining it might be difficult, particularly for rare or newly discovered resistance patterns. AMR data is intrinsically complex since it takes into account several variables, including bacterial strains, and environmental circumstances. It is a big problem to integrate and analyze these complicated datasets.

For an accurate AMR analysis, data pre-processing, including cleaning, normalizing, and feature engineering, is essential. Several techniques in the aiGeneR quality control pipeline, including min-max normalization, Log2 transform, a p-value criterion of less than 0.05, XGBoost feature selection, and deep neural networks, were used to find significant genes. Metrics like accuracy, precision, recall, and F1 score were used to assess the classification model’s performance on infected E. coli samples. The model achieved an F1 score of 93%, accuracy of 93%, precision of 100%, recall of 87%, and. Additionally, the model’s adaptability to changes in the input data, generalizability to new data, and congruence with biological observations were all assessed. It is found that the model is reliable, generalizable, and consistent, according to the findings of these assessments.

The aiGeneR learning model revealed that the genes paaI, trpC, polB, pspB, trpB, adk, paaZ, and tetM were significant. Expertise in microbiology, genetics, bioinformatics, and machine learning is frequently needed for effective AMR investigation. To address the complexities of AMR, multidisciplinary collaboration is required.

The application of ML models and neural networks for ARG detection and classification is the primary concern of this work. The work demonstrates a significant improvement in the identification of informative genes, the discovery of ARGs, and the classification of non-linear gene expression data sources, making the suggested aiGeneR a benchmark in the field of ARG identification. In comparison to previous studies on gene expression datasets for ARG detection, the aiGeneR model performs remarkably well. Additionally, the system’s robustness and domain adaptability are demonstrated by cross-validation, biological validation, and unseen implementations, as well as through how effectively it operates in domains other than the specific one on which it was trained.

This pilot study concerning the discovery of ARGs using gene expression data is highly motivated. This study can be expanded upon with data augmentation, perhaps leading to improved model performance. However, physicians do not recommend this strategy (the augmentation of medical data) because it is medically erroneous [97, 98]. If the model has been trained using synthetic data, we may get better model metrics. There are a few biases in our model that could be eliminated with more research, including (i) a smaller number of studies, (ii) the use of data augmentation, (iii) comparisons with other ML and DL models, (iv) no comments on the clinical validation, and (v) a description of benchmarking studies [99, 100, 101, 102, 103, 104].

Future work on enhancing ARG identification will focus on creating fresh datasets and investigating cutting-edge architectural concepts like Synthetic Minority Over-sampling Technique (SMOTE). We aim to assess the performance of these new models and conduct a variability analysis by contrasting them to our current aiGeneR models, such as the combination of ML with exhaustive feature space with DL.

Additionally, to improve the performance of the classification model, we intend to create a new quality control pipeline for the non-linear gene data. We want to work on analyzing research and ranking them according to their bias. Design systems can also be pruned to lower the size of the training models, and artificial intelligence designs are subject to bias.