Use of artificial intelligence-based algorithms in detecting COVID-19 mutations

Coronavirus disease 2019 (COVID-19) virus is a deadly type of coronavirus that has emerged recently and is highly contagious. The rapid spread of the COVID-19 virus has caused great fear and panic among people. Countries have had to take some measures such as complete closure and curfew to combat the COVID-19 virus. However, despite these measures, the COVID-19 virus continued to spread. Another method to combat the COVID-19 virus is the development of vaccines and drugs. The development of vaccines and drugs is of great importance in combating the COVID-19 virus. The effectiveness of these developed vaccines and drugs has either significantly decreased or disappeared completely as a result of the mutation of the COVID-19 virus. Therefore, it is very important to combat COVID-19 mutations. If future mutations in the structure of the COVID-19 virus can be predicted, vaccines and drugs can be developed more easily. Therefore, infected areas can be quarantined and ultimately the fight against the COVID-19 virus will be easier. Artificial intelligence-based approaches also offer promising results in detecting or predicting the COVID-19 virus. When the literature has been examined, most of the studies on the COVID-19 virus are studies on other aspects of the COVID-19 virus. For this reason, there is a serious gap in the literature in terms of mutation prediction of the COVID-19 virus. In this thesis study, we aim to fill this gap to some extent. In thesis study, three artificial intelligence-based models (TfrAdmCov, StackGridCov and HyperAttCov) have been proposed to predict future mutations in the COVID-19 Spike (S) protein structure. Firstly, the proposed TfrAdmCov model is completely transformer encoder based with Adam optimization algorithm. With the proposed TfrAdmCov model, dependencies between the variables in the input sequence can be easily captured. The proposed TfrAdmCov model can perform parallel calculations simultaneously because it is transformer encoder-based architecture. In addition, in order to increase the performance of the proposed TfrAdmCov model, agglomerative clustering algorithm has been preferred during creation of the training, testing and Kfold datasets. Additionally, the GridSearchCV algorithm has been used to set the best hyperparameter values of machine learning algorithms. The experimental results in detail shows that the proposed TfrAdmCov model achieves better performance than both classical artificial intelligence -based models and several state-of-the-art models. The proposed TfrAdmCov model achieved 99.93% accuracy value, 100.00% precision value, 97.38% recall value, 98.67% F1-score value and 98.65% MCC value on the COVID-19 testing dataset. In the COVID-19 testing dataset, the TfrAdmCov model with the Adam optimization algorithm correctly predicted 335 samples out of 344 samples in the "mutation" class, while it incorrectly predicted only 9 samples out of 344 samples in the "mutation" class. In addition, the proposed TfrAdmCov model with Adam optimization algorithm correctly predicted all samples out of 12386 samples in the "no mutation" class. Similarly, when the average of 10 random experiments have been taken, the proposed TfrAdmCov model achieved better results than other models in terms of accuracy with 99.924%, recall with 97.18%, F1-score with 98.57% and MCC value with 98.54% on the COVID-19 testing dataset. In addition, in order to statistically compare the proposed TfrAdmCov model with the deep learning models, the results obtained have been analyzed by taking the average of 10 random trials with different random seeds. Detailed evaluation has been carried out for each model in terms of accuracy, precision, recall, F1-score and MCC performance measurement metric using statistical measurements such as mean, standard deviation, median, mininum and maximum. The proposed TfrAdmCov model obtained an average of 0.999238, standard deviation of 0.000036, median of 0.999214, minimum of 0.999214 and maximum of 0.999293 among the 10 accuracy values obtained on the COVID-19 testing dataset. We also performed mutation prediction on the influenza A/H3N2 HA dataset to evaluate the performance of the proposed TfrAdmCov model. The proposed TfrAdmCov model achieved better results than other models 96.33% accuracy, 81.55% precision, 52.33% recall, 63.75% F1-score and 63.61% MCC values on the H3N2 HA testing dataset. On the H3N2 HA testing dataset, the proposed TfrAdmCov model correctly predicted 853 samples out of 1630 samples in the "mutation" class, while it incorrectly predicted 777 samples out of 1630 samples in the "mutation" class. In addition, the proposed TfrAdmCov model correctly predicted 24577 out of 24770 samples in the "no mutation" class, while it incorrectly predicted 193 out of 24770 samples in the "no mutation" class. Results on the influenza H3N2 HA testing dataset showed that the proposed TfrAdmCov model is quite robust. Secondly, we propose a robust StackGridCov model for mutation prediction of the COVID-19 virus. The proposed StackGridCov model is based on ensemble learning. The proposed StackGridCov model is a very successful model that maximizes the performance as much as possible by using many machine learning algorithms. The main reason for this can be expressed as the proposed StackGridCov model reduces the possibility of overfitting by combining the strengths of several base models. These base models may make errors in different parts of the input sequences. By combining the outputs of these base classifiers, the meta-classifier can compensate for these errors and ultimately make a more accurate prediction. The proposed StackGridCov model is flexible as different machine learning algorithms can be used in both the level-0 layer and the level-1 layer. The proposed StackGridCov model is more robust than other ensemble learning and other artificial intelligence techniques as it is less affected by overfitting. This is because the base learners are trained on the same training dataset and the meta learner is trained on the new large dataset by combining the predictions of these base classifiers on the training dataset, ultimately reducing the possibility of overfitting. In this thesis study, while the base learners at level-0 have been selected as SVM, RF, XGBoost, ANN, DT, GB, ET, AdaBoost learner has been chosen as the meta classifier at level-1. This selection of both base classifiers and meta classifier significantly improved the performance of the proposed StackGridCov model. In addition, we use the GridSearchCV hyperparameter tuning algorithm to improve the performance of the proposed StackGridCov model and other models. To evaluate the performance of the proposed StackGridCov model and other models, the stratified 10-fold cross-validation technique as well as the holdout technique has been used. Additionally, to evaluate the performance of the proposed StackGridCov model, mutation prediction has been performed on the previously emerging influenza A/H1N1 HA virus dataset. The proposed StackGridCov model with GridSearchCV method outperformed other algorithms in terms of accuracy value of 0.6623, F1-score value of 0.6723, MCC value of 0.3273 and AUC value of 0.7018 on the COVID-19 testing dataset. Moreover, the proposed StackGridCov algorithm with GridSearchCV technique outperformed the StackGridCov model without GridSearchCV technique on the COVID-19 testing dataset. The proposed StackGridCov model with GridSearchCV method increased the accuracy value (from 0.6016 to 0.6623), precision value (from 0.5833 to 0.6415), recall value (from 0.6566 to 0.7062), F1-score value (from 0.6178 to 0.6723). ), the MCC value (from 0.2063 to 0.3273) and the AUC value (from 0.6133 to 0.7018). The proposed StackGridCov model with the GridSearchCV method correctly predicted 399 samples out of 565 samples in the "mutation" class on the COVID-19 testing dataset, while it incorrectly predicted only 166 samples out of 565 samples in the "mutation" class. In addition, the proposed StackGridCov model with the GridSearchCV method correctly predicted 223 samples out of 587 samples in the "no mutation" class on the COVID-19 testing dataset, while it incorrectly predicted 364 samples out of 587 samples in the "no mutation" class. Similarly, the proposed StackGridCov outperformed other models in terms of accuracy value of 0.6610, a precision value of 0.6614, an F1-score value of 0.6607 and an MCC value of 0.3226 on the KFold dataset. Moreover, the proposed StackGridCov model outperformed other models in terms of accuracy value of 0.9460, recall value of 0.7969, F1-score value of 0.8093 and MCC value of 0.7780 on the Influenza A/H1N1 HA testing dataset. As a result, it has been observed that using the GridSearchCV hyperparameter technique has been generally increased the performance of the proposed StackGridCov model and other models. Thirdly, the HyperAttCov model, which is based on LSTM, HyperMixer and attention mechanisms, is proposed for COVID-19 virus mutation prediction. Attention mechanisms have been used to maximize the performance of the proposed HyperAttCov model. The proposed HyperAttCov model is able to capture the most relevant input features and long-term temporal dependencies in the input sequence. Additionally, in this thesis study, attention mechanisms (input attention mechanism and temporal attention mechanism) have been used to improve the performance of the proposed HyperAttCov model by focusing on important parts of the COVID-19 dataset. While the input attention mechanism is applied to the entire input dataset, the temporal attention mechanism is applied to the data obtained from the HyperMixer architecture. The proposed HyperAttCov model achieved better performance than many deep learning-based and machine learning models. When the experimental results have been examined in detail, it has been observed that the proposed HyperAttCov model reached 70.0% accuracy value, 92.0% precision value and 46.5% MCC value in the COVID-19 testing dataset. Similarly, the proposed HyperAttCov model achieved 70.2% accuracy value, 90.4% precision value and 46.2% MCC value on the COVID-19 testing dataset when averaged over 10 random trials. In addition, the proposed HyperAttCov model achieved very successful results on the COVID-19 testing dataset compared to the study in the literature. As a result, the proposed TfrAdmCov, StackGridCov and HyperAttCov models can successfully predict mutations that will occur on both the COVID-19 S protein and the influenza datasets. In addition, in this thesis study, it has been observed that the use of agglomerative clustering algorithm and GridSearchCV hyperparameter technique played an effective role in mutation prediction of the COVID-19 virus. The results obtained this thesis study are promising for vaccines and drugs development.