AICpred: Machine Learning-Based Prediction of Potential Anti-Inflammatory Compounds Targeting TLR4-MyD88 Binding Mechanism

Abstract

Toll-like receptor 4 (TLR4) has been implicated in the production of uncontrolled inflammation within the body, known as the cytokine storm. Studies that employ machine learning (ML) in the prediction of potential inhibitors of TLR4 are limited. This study introduces AICpred, a robust, free, user-friendly, and easily accessible machine learning-based web application for predicting inhibitors against TLR4 by targeting the TLR4-myeloid differentiation primary response 88 (MyD88) interaction. MyD88 is a crucial adaptor protein in the TLR4-induced hyper-inflammation pathway. Predictive models were trained using random forest, adaptive boosting (AdaBoost), eXtreme gradient boosting (XGBoost), k-nearest neighbours (KNN), and decision tree models. To handle imbalance within the training data, resampling techniques such as random under-sampling, synthetic minority oversampling technique, and the random selection of 5000 instances of the majority class were employed. A 10-fold cross-validation strategy was used to evaluate model performance based on metrics including accuracy, balanced accuracy, and recall. The XGBoost model demonstrated superior performance with accuracy, balanced accuracy, and recall scores of 0.994, 0.958, and 0.917, respectively, on the test. The AdaBoost and decision tree models also excelled with accuracies ranging from 0.981 to 0.992, balanced accuracies between 0.921 and 0.944, and recall scores between 0.845 and 0.891 on both training and test datasets. The XGBoost model was deployed as AICpred and was used to screen compounds that have been reported to have positive effects on mitigating the hyperinflammation-associated cytokine storm, which is a key factor in COVID-19. The models predicted Baricitinib, Ibrutinib, Nezulcitinib, MCC950, and Acalabrutinib as anti-TLR4 compounds with prediction probability above 0.90. Additionally, compounds known to inhibit TLR4, including TAK-242 (Resatorvid) and benzisothiazole derivative (M62812), were predicted as bioactive agents within the applicability domain with probabilities above 0.80. Computationally inferred compounds using AICpred can be explored as potential starting skeletons for therapeutic agents against hyperinflammation. These predictions must be consolidated with experimental screening to enhance further optimisation of the compounds. AICpred is the first of its kind targeting the inhibition of TLR4-MyD88 binding and is freely available at http://197.255.126.13:8080.

1. Introduction

Toll-like receptor 4 (TLR4) is a key transmembrane receptor that recognises both infectious and non-infectious inflammatory stimuli [1]. TLR4 signalling leads to pro-inflammatory action through the production of pro-inflammatory cytokines, which are small proteins responsible for controlling antibody production, recruiting immune cells, and regulating inflammation for combating infection [1,2]. The primary cytokines involved in signalling between cells of the immune system are interleukins (ILs), tumour necrosis factors (TNFs), interferons (IFNs), and chemokines which can be classified as either pro-inflammatory or anti-inflammatory [3]. During viral infections, antiviral responses in neighbouring cells are activated, and the recruitment of innate and adaptive immune cells is triggered by damage-associated molecular patterns (DAMPs) and pathogen-associated molecular patterns (PAMPs) [4,5,6]. TLR4 recognises invading pathogens via PAMPS and molecules with endogenous origin from damaged tissues via DAMPS to activate protective self-defence mechanisms for the body. Myeloid differentiation primary response 88 (MyD88) is an important adaptor protein in TLR4 signalling [7]. MyD88 interacts with TLR4 through its Toll/interleukin-1 receptor (TIR) domain [8]. The interaction initiates a signalling cascade leading to nuclear factor kappa B (NF-κB) activation and cytokine production [7,9]. NF-κB is a key mediator of pro-inflammatory activity in the body [9].

The balance between pro-inflammatory cytokines, which are responsible for producing an inflammatory response, and anti-inflammatory cytokines, responsible for controlling inflammation in the body, is complex and important [10]. During viral infections, a variety of immune cells are triggered, which leads to a disruption in the cytokine balance, causing pro-inflammatory cytokines to increase by a considerable amount [11]. While pro-inflammatory cytokines work to ameliorate the further spread of the virus, their uncontrolled production can lead to severe undesirable outcomes, known as cytokine storm (CS) [10]. The cytokine storm (CS) is a critical factor in the poor prognosis of both infectious and non-infectious diseases, contributing to severe symptoms and increased fatality rates [1,12]. Given its involvement in the cytokine storm, inhibiting TLR4 signalling, particularly by targeting the TLR4-MyD88 interaction, emerges as a promising therapeutic strategy for managing inflammatory conditions [13,14].

To identify anti-inflammatory compounds, machine learning (ML) techniques were employed to develop predictive models for inhibitors of the cyclooxygenase-2 (COX-2) enzyme, a key mediator in inflammatory response [15,16]. These models achieved accuracies of 0.746 to 0.754, sensitivity scores of 0.612 to 0.686, and specificity scores of 0.777 to 0.814. In addition, a deep learning (DL)-based model was developed to predict the anti-inflammatory activity of peptides with an area under the curve (AUC) of 0.919 and MCC of 0.735 [17]. In a recent study, deep variational autoencoders and contrast learning, which are advanced computational methods, were used in the identification of anti-inflammatory peptides [18]. Although studies that target TLR4 specifically are limited, a recent study developed ML models using a compiled list of 78 compounds from the ChEMBL database to predict TLR4 inhibitors specifically for the treatment of Mycoplasma pneumonia disease [19,20]. These studies, however, do not specifically target the TLR4-MyD88 binding, which has been shown by multiple studies to be a plausible target for the treatment of the cytokine storm [13,15,17,18]. Hence, there is a pressing need for accessible, user-friendly, and robust applications that specifically target the TLR4-MyD88 interaction, which has been shown to be a plausible drug target for mitigating the cytokine.

In this study, we developed and hosted the first robust ML-based application, AICpred, to predict potential inhibitors against the TLR4-MyD88 binding mechanism. AICpred is user-friendly and easily accessible. We evaluated the potential of AICpred by predicting compounds implicated in inhibiting CS as anti-TLR4. AICpred attained high performance on standard evaluation metrics, which makes the application a potential resource in drug discovery pipelines. The predicted compounds have the potential to be utilised as initial backbones to experimentally evaluate their propensity to inhibit the CS via TLR4-MyD88 inflammatory mechanisms. Although the precise interaction mechanism is unknown, studies have revealed that the COVID-19 spike protein interacts with TLR4 to cause hyperinflammation [21,22]. These observations have been supported by both in silico and in vitro studies [23,24,25]. This gives merit to investigating AICpred predicted compounds as therapeutics for COVID-19-induced cytokine storm.

2. Materials and Methods

2.1. Methods

In this study, five ML algorithms comprising random forest, k-nearest neighbours (KNN), decision tree, eXtreme gradient boosting (XGBoost), and adaptive boosting (AdaBoost) were employed to construct predictive models for identifying potential anti-inflammatory compounds targeting TLR4. Random forest is frequently used for QSAR prediction due to its superior robustness, predictability, and ease of use [26]. The KNN algorithm, known for its simplicity and effectiveness for classification studies, offers the additional advantage of interpretability, allowing for the identification of structural features that may contribute to biological activity [27]. Decision tree algorithms have also been recognised for their ability to interpret based on the clear decision rules that link descriptors with biological activity [28,29]. Ensemble methods such as XGBoost and AdaBoost are particularly well-suited for classification problems, offering enhanced accuracy by minimising bias in imbalanced datasets and rapidly converging to global minima [30,31]. Their superior performance in QSAR prediction has been well-documented [32,33]. XGBoost, characterised by the intelligent aggregation of weak learners (trees) to develop a robust model (Figure 1), has been widely reported for its exceptional performance in QSAR modelling, making it an optimal choice for this study [32,33,34].

Although support vector machines (SVMs) have been successfully used to develop high-performing QSAR models, they were excluded from this study for several reasons [36,37]. While SVM-based models can achieve strong predictive performance across various datasets, their computational complexity—both in terms of time and memory—grows substantially as the dataset size increases, limiting their scalability for large-scale molecular screening [38,39,40]. In contrast, algorithms like random forest, XGBoost, and AdaBoost efficiently handle large datasets and have demonstrated optimal performance when applied to QSAR modelling [41,42,43,44,45].

The predictive models were implemented using the scikit-learn library (version 1.0.2), an open-source machine learning framework for Python (version 3.7.12) library. Bioassay data were obtained from PubChem for the cross-validation and testing of these models [46]. Their performance was evaluated using balanced accuracy, precision, F1 score, recall, Matthew’s correlation coefficient (MCC), and the area under the receiver operating curve (AUROC) as appropriate metrics. Additionally, the models were further validated against known TLR4 inhibitors to ensure their reliability. A summary of the study methodology is shown in Figure 2.

2.2. Dataset Extraction

The dataset of compounds with experimentally measured activity against TLR4 was downloaded from PubChem with Assay ID 861 [46]. This assay is part of broader research comprising four additional assays aimed at identifying inhibitors of TLR4 [51]. The assay identifies compounds that exhibit an inhibitory effect on the toll-like receptor by measuring the inhibition of constitutive TLR4–MYD88 binding [51,52]. The compounds were classified based on an algorithm that calculates a threshold value by summing the percent inhibition of all compounds and adding three times their standard deviation [51]. Any compound exhibiting greater percent inhibition than the threshold was labelled as active. An activity score was subsequently calculated by normalising the percent inhibition of each test compound to the highest observed inhibition value, with negative inhibition values assigned an activity score of zero [51,52]. The dataset consisted of 356 active and 195,623 inactive compounds.

2.3. Descriptors Computation

The Mold2 Descriptor Generator Software (version 2.0.0), designed for efficient and rapid calculation of a broad range of descriptors [53], was used to calculate the molecular descriptors for each compound in the active and inactive sets. Molecular descriptors are mathematical representations of the properties of a molecule, obtained by a well-specified algorithm [54]. As machine learning (ML) models require quantitative data, molecular descriptors provide a means to quantify the physiochemical properties of the compounds [55]. These descriptors serve as the features that are used to train and test the ML models.

A total of 777 2D molecular descriptors (features) were computed for each molecule in the entire dataset. These features were labelled D001 to D777, covering properties such as the counts for functional groups, structural features, and counts for atoms and bonds. For instance, descriptors D012 to D019 were features described by the number of multiple bonds, number of circuit structures, number of rotatable bonds, rotatable bond fractions, number of double bonds, number of aromatic bonds, and the sum of conventional bond orders, respectively. The descriptions of each of the 777 molecular descriptors are provided by the United States Food and Drug Administration [56].

2.4. Data Pre-Processing and Feature Selection

Data preprocessing is a crucial step in ML workflows, involving the cleaning and transformation of data to ensure it is suitable for analysis [57]. A custom Python script was used to remove null values, duplicates, and irrelevant non-compound features from the data. Additionally, the activity column was encoded in a binary format, where active and inactive compounds were assigned values of 1 and 0, respectively. Eighty percent (80%) of the data were allocated for training and validating the models, while the remaining twenty percent (20%) was held out as a test set.

A basic feature selection process was employed in developing the predictive models. To ensure that the ML algorithms perform optimally, low variance features, which are features with little or no useful information, were eliminated using the VarianceThreshold function from scikit-learn’s feature selection package, with a threshold of 0 [58].

The dataset exhibited a high-class imbalance between the active and inactive compounds, which can affect model performance by underrepresenting the minority class [59]. To address this, three resampling techniques were applied to the training data, generating our different models for each of the three resampling techniques in addition to no resampling. The resampling techniques are included herein (1) random undersampling, where data points from the majority class were randomly selected without replacement in an active to inactive ratio of 1:2 [60,61]. This was achieved using the RandomUnderSampler function from Imbalanced-learn (version 0.10.1) with a sampling_strategy of 0.5. (2) Synthetic Minority Oversampling Technique (SMOTE), which creates synthetic instances of the minority class by finding the k-nearest neighbours to the feature vector using mathematical computations [44]. (3) Five thousand inactive compounds randomly sampled for training along with the original number of active compounds.

2.5. Model Training

Model training was carried out on the Kaggle platform, leveraging its computational resources, robust environment, and support for ML tasks, using Scikit-Learn, Pandas, NumPy, and other Python libraries. For each algorithm employed in this study, four different models were trained on the original unbalanced data and the three data resampling techniques described above. The random forest algorithm was employed, which constructs multiple decision trees trained on a random subset of features and samples. The final prediction is determined by majority voting of the individual tree predictions [62]. Decision trees consist of a hierarchical structure with root nodes, internal nodes, and leaf nodes, utilising a sigmoid activation function [28]. Boosting algorithms such as AdaBoost and XGBoost were also applied, where they use a combination of weak learners to create a stronger, more robust model (Figure 1) [63]. For the K-nearest neighbours (KNN) algorithm, a non-parametric model, no explicit architecture or activation function is required [64]. Hyperparameters for the KNN model, such as the number of neighbours (set to 10) and the number of jobs (set to −1), were optimised.

2.6. Model Evaluation and Validation

The developed models were evaluated using a 10-fold cross-validation strategy and further validated on the held-out test set. The evaluation metrics used were balanced accuracy, precision, F1 score, recall, Matthew’s correlation coefficient (MCC), and the area under the receiver operating curve (AUROC). The formulae and interpretation of these performance metrics are shown in

Table 1. Performance metrics for model evaluation. The accuracy, balanced accuracy, recall, precision, F1 score, and MCC were used to evaluate the developed ML models at both the cross-validation and testing stages.

Metrics	Mathematical Definition	Interpretation
Accuracy	${\displaystyle \frac{TP+TN}{TP+FP+FN+TN}}$	1—Perfect 0—Poor
Recall	${\displaystyle \frac{TP}{TP+FN}}$	1—Perfect 0—Poor
Precision	${\displaystyle \frac{TP}{TP+FP}}$	1—Perfect 0—Poor
F1 score	$2\times {\displaystyle \frac{recall\times precison}{recall+precision}}$	1—Perfect 0—Poor
Balanced accuracy	${\displaystyle \frac{Sensitivity+Specificity}{2}}$	1—Perfect 0—Poor
MCC	${\displaystyle \frac{\left(TP\times TN-FP\times FN\right)}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}}$	+1—Perfect −1—Poor

. A confusion matrix based on the true positive (TP), true negative (TN), false positive (FP), and false negative (FN) parameters was used at each model evaluation stage [65].

Accuracy represents the ratio of correctly predicted instances to total instances but may not be appropriate for imbalanced data without the application of data sampling techniques to introduce balance, making it a viable metric for model comparison [44,66]. In addition to accuracy, balanced accuracy assesses the true measure of accuracy for each class in imbalanced data [67]. Precision determines the fraction of correctly predicted positives, while recall measures the fraction of actual positives correctly predicted [68]. F1 score is a result of a concordant mean of precision and recall [69]. MCC only presents a high score when the model prediction produces good results in all four categories of the confusion matrix [70]. The AUROC score evaluates a model’s ability to distinguish between classes using error costs [71]. The models were further validated with known (experimentally determined) inhibitors of TLR4. A total of five known inhibitors of TLR4, including Resatorvid, M62812, ZINC25778142, (+)-Naltrexone, and (+)-Naloxone, were evaluated via the trained model to further assess their performance [47,48,49,50].

Table 2. Experimentally determined Toll-like receptor 4 (TLR4) inhibitors and their half-maximal inhibitory concentration (IC₅₀) values retrieved from the literature.

Known Inhibitor	IC50	Reference
Resatorvid/TAK-242	11 to 33 nM	[47]
M62812	7 µM	[48]
ZINC25778142	16.6 µM	[49]
(+)-Naloxone	105.5 µM	[50]
(+)-Naltrexone	94.4 µM	[50]

states these inhibitors and their half-maximal inhibitory concentrations (IC₅₀).

2.7. Applicability Domain Analysis

The applicability domain for predictive models refers to the physiochemical, structural, or biological information on which the training set of the model has been generated and within which the model’s predictions for new compounds are considered relevant and reliable [72]. The applicability domain is necessary for defining the boundary in which a given model’s prediction can be regarded to be reliable [73]. In this study, the applicability domain of the trained models was assessed by generating a plot of standardised descriptor values for each compound in both the training and test sets. The descriptors for each compound were standardised as part of this analysis [74]:

$${\mathit{S}}_{\mathit{k}\mathit{i}}={\displaystyle \frac{|{\mathit{X}}_{\mathit{k}\mathit{i}}-{\overline{\mathit{X}}}_{\mathit{i}}|}{{\mathsf{\sigma }}_{{\mathit{X}}_{\mathit{i}}}}}$$

(1)

where

• $\mathit{k}=1,2,3$… $\mathit{n}$ number of compounds;
• $\mathit{i}=1,2,3$… $\mathit{n}$ number of molecular descriptors;
• ${\mathit{S}}_{\mathit{k}\mathit{i}}=$ Standardized descriptor $\mathit{i}$ for compound $\mathit{k}$ from the training or test set;
• ${\mathit{X}}_{\mathit{k}\mathit{i}}=$ Actual descriptor $\mathit{i}$ for the compound $\mathit{k}$ from the training or test set;
• ${\overline{\mathit{X}}}_{\mathit{i}}=$ Mean value for the descriptor ${\mathit{X}}_{\mathit{i}}$ from the training compounds only;
• ${\mathsf{\sigma }}_{{\mathit{X}}_{\mathit{i}}}=$ Standard deviation of the descriptor ${\mathit{X}}_{\mathit{i}}$ from training compounds only.

2.8. Web Server Development

An interactive web application was developed utilising the best-performing models with widely adopted tools and libraries. The backend was implemented using Flask [75], while the front end was designed with Hypertext Markup Language (HTML), Cascading Style Sheets (CSS), and JavaScript [76].

2.9. Screening of COVID-19-Induced CS Inhibitors

The deployed models were utilised to screen inhibitors that have been involved in clinical trials, clinical observations, in vitro assays, and animal experimentation to significantly alleviate the cytokine storm associated with COVID-19. These include inhibitors of protein targets such as JAK, Nod-like receptor family pyrin domain-containing 3 (NLRP3), and Bruton’s Tyrosine Kinase (BTK), whose inhibition has been shown to demonstrate anti-inflammatory effects in COVID-19 cases (

Table 3. Nod-like receptor family pyrin domain-containing 3 (NLRP3), Janus Kinase (JAK), and Bruton’s Tyrosine Kinase (BTK) inhibitors screened using the deployed XGBoost model.

Inhibitor	Target	PubChem CID	Reference
Baricitinib	JAK1 and JAK2	44205240	[80]
Nezulcitinib	All JAK isoforms	146421275	[81]
Ibrutinib	BTK	24821094	[82]
Acalabrutinib	BTK	71226662	[83]
MCC950	NLRP3	9910393	[84]

) [77,78,79]. The Simplified Molecular Input Line Entry System (SMILES) of each of the compounds was obtained from PubChem and used as inputs for the deployed models [46].

3. Results

3.1. Data Pre-Processing

Following data cleaning, the dataset consisted of 194,888 inactive compounds and 356 active compounds, all with 777 features (descriptors). To facilitate model development and evaluation, the dataset was partitioned into training and test sets. The training set contained 155,917 inactive compounds and 278 active compounds, while the test set consisted of 38,971 inactive compounds and 78 active compounds. Dimensionality reduction was applied, resulting in a final training set of 156,195 rows and 645 columns and a testing set of 39,049 rows and 645 columns.

The three data resampling iterations were then performed to balance the training data shown in Figure 3. No resampling was performed on the heldout test set. Random undersampling resulted in a balanced training set of 556 inactive compounds and 278 active compounds. The application of SMOTE increased the number of active compounds to 77,958, while the number of inactive compounds remained at 155,917. In the fourth iteration, a subset of 5000 inactive compounds was randomly selected and combined with the 356 active compounds. Here, 4000 inactive compounds and 284 active compounds were used for training, while 1000 inactive compounds and 72 active compounds were used for testing.

3.2. Model Development and Evaluation

The performance of the models was assessed using a 10-fold cross-validation based on accuracy, balanced accuracy, recall, precision, F1 score, and MCC (Figure 4). The models were then tested on held-out data to further assess their performance on unseen data (

Table 4. Performance of the best models for each algorithm during a 10-fold cross-validation (CV) and on the external test data. All best-performing models for each algorithm were developed on the random 5000 data sampling technique.

Model	Process	Accuracy	Balanced Accuracy	Precision	Recall	F1 Score	AUROC	MCC
Random Forest	CV	0.972	0.787	1.000	0.573	0.723	0.971	0.743
	Test	0.968	0.770	0.975	0.542	0.696	0.987	0.714
Decision Trees	CV	0.981	0.939	0.839	0.891	0.863	0.939	0.854
	Test	0.981	0.938	0.842	0.889	0.865	0.938	0.855
AdaBoost	CV	0.987	0.921	0.958	0.845	0.897	0.983	0.893
	Test	0.992	0.944	0.985	0.889	0.934	0.998	0.931
XGBoost	CV	0.994	0.958	1.000	0.915	0.955	0.998	0.954
	Test	0.994	0.958	1.000	0.917	0.957	0.999	0.955
KNN	CV	0.942	0.604	0.685	0.215	0.323	0.757	0.360
	Test	0.936	0.611	0.548	0.236	0.330	0.778	0.332

). The AUROC was also calculated for each model based on their true positive and false negative rates of prediction (Figure 5).

XGBoost models had the highest performances (Table 4), with evaluation metrics above 0.85 except the model trained with data from the random undersampled iteration (Table S1). In contrast, KNN showed the weakest performance with balanced accuracies as low as 0.51, precision scores as low as 0.01, and F1 scores as low as 0.02 (Table S2). Among the four different training and testing iterations, random undersampling iterations produced the poorest results with precisions as low as 0.03 for random forest (Table S3), 0.05 for decision tree (Table S4), and 0.01 for KNN. Performance of the AdaBoost models was comparable to the XGBoost models (Table S5).

3.3. Validation with Known Inhibitors of TLR4

The eight top-performing models were further validated using experimentally determined inhibitors of TLR4. Each of the models predicted at least 1 inhibitor as active except the AdaBoost model trained on unsampled data. The XGBoost model trained on the random 5000 data resampling technique was able to predict all five known inhibitors as active with prediction probabilities ranging from 0.83 to 0.997 (

Table 5. Model validation with known inhibitors. The activity of five known inhibitors was predicted using the best-performing XGBoost model, which was trained on the random 5000 data resampling technique presented in this study. The prediction probability of the model is also presented.

Known Inhibitor	IC50	Reference	XGBoost Prediction Probability
Resatorvid/TAK-242	11–33 nM	[47]	0.830
M62812	1–3 µg/mL	[48]	0.983
ZINC25778142	16.6 µM	[49]	0.997
(+)-Naloxone	105.5 µM	[50]	0.996
(+)-Naltrexone	94.4 µM	[50]	0.996

). Due to its performance across metrics and validation with known inhibitors, this XGBoost model was selected for deployment as a web server.

3.4. Results of Applicability Domain Analysis

According to the descriptor standardisation approach, 1.16% of the training data were classified as outliers while 0.09% of the external test data fell outside the model’s domain of applicability (Figure 6). These outliers represent compounds with molecular descriptors significantly deviating from the majority of the data, indicating that their predictions may be less reliable. The low percentage of test set outliers suggests that the model’s applicability domain is well-defined and capable of handling most of the new compounds with confidence.

3.5. Model Deployment

The models have been made accessible via a web server application named AICpred, which can be freely accessed at http://197.255.126.13:8080 (accessed on 24 December 2024) (Figure 7). Users can utilise this web server platform to predict potential inhibitors of TLR4 by submitting a compound in SMILES format or a text file containing the compound ID and its corresponding SMILES.

3.6. Evaluating COVID-19-Induced CS Inhibitors

The AICpred web server was employed to screen known inhibitors of JAK, NLRP3, and BTK. With the exception of Nezulcitinib, all inhibitors were within the domain of applicability of both models and were predicted as active against TLR4 with prediction probabilities ranging from 0.992 to 0.996 (

Table 6. Activity prediction of JAK, NLRP3, and BTK inhibitors using the AICpred web server. The applicability of the models to these inhibitors is also highlighted.

Inhibitor	XGBoost Prediction (Prediction Probability)	Within Applicability Domain
Baricitinib	Active (0.996)	Yes
Nezulcitinib	Active (0.994)	No
Ibrutinib	Active (0.995)	Yes
Acalabrutinib	Active (0.993)	Yes
MCC950	Active (0.996)	Yes

). The details of the results are shown in Table 6.

4. Discussion

The role of the cytokine storm in disrupting the immune balance and its clinical implications in sepsis management and multiorgan failure is well established [85]. To address this challenge, we have developed AICpred, a machine learning-based web application designed to facilitate the identification of potential compounds targeting the TLR4-MyD88 binding. Traditional ML algorithms have been reported to perform better on the majority class than the minority class [86]. To mitigate the issue of class imbalance in our dataset, resampling techniques were employed in optimising model performance [87]. Furthermore, dimensionality reduction was applied to eliminate redundant and irrelevant data, which improved computational efficiency, increased learning accuracy, and provided deeper insights into the models [88].

Among the algorithms, boosting algorithms (XGBoost and AdaBoost) and tree-based algorithms (decision trees and random forests) produced the best-performing models. All XGBoost models, with the exception of the model trained with the random undersampling iteration, achieved values above 0.85 for all evaluation metrics. Boosting algorithms are a type of ensemble learning technique that enhance the performance of individual base learners by fusing them into a composite whole [89]. Thus, making them fast and capable of handling possible overfitting and missing data [90]. Notably, XGBoost achieved excellent generalisability and accuracy, even in the presence of imbalanced data, with accuracies reaching 0.99 and balanced accuracies above 0.75, outperforming other models in robustness and prediction accuracy [91,92,93].

In this study, the resampling technique that produced one of the best-performing models for each algorithm (except KNN) was the random selection of 5000 inactive compounds versus the original number of actives (356). In contrast, random undersampling performed the weakest among all iterations, likely due to the loss of important information from the majority class [94]. SMOTE, a popular technique for handling class imbalance in many studies, was applied and performed well with XGBoost models, yielding an accuracy and balanced accuracy both above 0.90 [44,95,96,97].

The performance of our models was compared with previous studies. For example, ML models have been trained on 1565 inhibitors and 1671 non-inhibitors to predict inhibitors against the signal transducer and activator of transcription 3 (STAT3) to alleviate the cytokine storm [98]. The study was based on a 2-D descriptor with the best random forest model yielding accuracies of 0.763 and 0.763. In contrast, the random forest models in our study achieved training and validation accuracies of 0.971 and 0.968, respectively. In addition, the XGBoost models in this study outperformed others by over 25% [98]. In another study, a random forest model has also been developed for the prediction of plant-derived compounds for anti-COVID-19 therapy [99]. Although comparable in terms of accuracy, their random forest model slightly outperformed the model presented in this study, recording a 46% higher recall and a 29% higher F1 score.

To further ensure the reliability of our models, we analysed the applicability domain, which defines the range within which model predictions can be considered reliable [73]. Coincidentally, the best models from each algorithm were trained on the data generated from the random selection of 5000 inactive compounds in addition to the original number of active compounds. The applicability domain contains 99.91% of the external test set, thus, within three standard deviations from the mean of the training data. This structural similarity between the training and test data contributed to the models’ strong performance and reliability [74,100].

The best-performing models for each algorithm were further validated with experimentally determined inhibitors of TLR4 (Table 5). These inhibitors have been shown in vivo or in vitro to inhibit the lipopolysaccharide (LPS)-induced production of pro-inflammatory cytokines. LPS, a vital component of the outer membrane of Gram-negative bacteria, is a pathogen-associated molecular pattern (PAMP) that stimulates TLR4, leading to the production of pro-inflammatory cytokines such as tumour necrosis factor ⍺ (TNF-⍺) and interleukin 6 (IL-6) [101]. The XGBoost model trained on the random 5000 data resampling technique successfully predicted all five experimentally determined inhibitors as active with prediction probabilities ranging from 0.83 to 0.997.

The deployed models were further utilised to predict the activity of inhibitors targeting JAK, NLRP3, and BTK (Figure 6). The inhibition of these targets has been shown to have anti-inflammatory effects on the COVID-19-induced cytokine storm [77,78,79]. Interestingly, TLR4 has been reported to trigger the activation of the NF-κB pathway, while other studies also indicate that BTK and its upstream activator, haematopoietic cell kinase (HCK), are involved in toll-like receptor signalling [9,82]. The XGBoost model predicted all these inhibitors as active with prediction probabilities ranging from 0.992 to 0.996. These findings suggest the potential utility of the developed models in the prediction of anti-inflammatory compounds for study on the possible alleviation of the COVID-19-induced cytokine storm. However, experimental and clinical validation is required to gain deeper insights into the biology of hyperinflammation and immune response.

The study focused on tree-based, ensemble, and non-parametric supervised methods, namely random forest, decision tree, AdaBoost, XGBoost, and KNN, due to their established performance in QSAR modelling. Future work will incorporate deep learning methods as they demonstrate superior performance compared to traditional ML techniques in applications including drug-induced liver injury prediction and the effect of endocrine-disrupting chemicals on human health [102,103]. QSAR deep learning methods have achieved high accuracy (>0.90) in qualitative predictions and excellent quantitative performance on large datasets (coefficient of determination (R²) = 0.80, predictive square correlation coefficient (Q²) = 0.86) [102]. These approaches show great promise in predicting drug-target interactions and identifying potential inhibitors for drug discovery, utilising architectures such as CNN-RNN hybrids and graph representation learning [104,105]. Incorporating deep learning methods in future studies is a plausible next step in addition to ensemble methods that combine different models to enhance prediction accuracy [106].

By integrating the best-performing model into the AICpred web application, researchers can utilise these models for drug repurposing or screening of large compound libraries for the identification of novel therapeutic agents targeting diseases where TLR4 plays a major role in the mechanisms. However, this study is limited by the highly imbalanced datasets used for the training. Hence, highlighting the need for more high-throughput screening studies to unravel inhibitors of TLR4. Future studies may benefit from larger and more balanced datasets. Additionally, computational predictions of AICpred must be corroborated with in vitro and in vivo studies to gain insights into their mechanism and effects on mitigating hyper-inflammation. The study did not incorporate information about the mechanism of action (MOA) and binding modes but rather used the available bioactive datasets [107], which would have been essential for post-computational insights [107,108]. Future studies may benefit from the incorporation of MOA and binding mode information in training and testing datasets to enhance model predictive efficiency.

5. Conclusions

The binding of TLR4 to key receptors in the immune pathways elicits inflammation-associated mechanisms. In this study, we developed and deployed an XGBoost model as a web-based application for the discovery of potential anti-inflammatory compounds as plausible inhibitors of the TLR4-MyD88 signalling pathway. Robust ML models comprising random forest, adaptive boosting (AdaBoost), eXtreme gradient boosting (XGBoost), k-nearest neighbours, and decision trees were trained and evaluated using a 10-fold cross-validation strategy. Among the models, XGBoost exhibited the best performance, achieving accuracy, balanced accuracy, precision, and recall scores of 0.994, 0.958, 1.000, and 0.917, respectively. The predictions via the XGBoost model were further validated on experimentally determined TLR4 inhibitors and compounds associated with inflammation-mediated CS. Baricitinib, Ibrutinib, Nezulcitinib, MCC950, and Acalabrutinib were predicted as inhibitory compounds against TLR4-MyD88 binding with probabilities above 0.90. Similarly, TAK-242 (Resatorvid) and the benzisothiazole derivative (M62812), known inhibitors of TLR4, were predicted as bioactive agents within the applicability domain with prediction probabilities above 0.80. The XGBoost model has been integrated into a web application, known as AICpred, which is accessible at http://197.255.126.13:8080 (accessed on 24 December 2024). The platform enables the screening of large compound libraries to aid in identifying potential inhibitors of TLR4-MyD88 binding as possible anti-inflammatory molecules. Targeting the TLR4-MyD88 binding serves as a plausible starting point for designing therapeutic agents against CS, which is implicated in COVID-19 pathophysiology.