EZYDeep: A Deep Learning Tool for Enzyme Function Prediction based on Sequence Information

1. INTRODUCTION

Enzymes are proteins that have a specific biological activity: they catalyze biological reactions. Catalysts are tools that can substantially speed up these processes. At the conclusion of the reaction, they are regenerated and act at low concentrations. The vital functions of organisms are crucially maintained by enzymes. They control numerous other chemical processes that are strongly related to the process of life, including metabolism, nutrition, and energy conversion. Enzymes are utilized in a variety of processes, including industrial synthesis, human health, and environmental remediation [1, 2].

Precision in the selection of enzymes for bioindustrial or biomedical applications is therefore crucial and has become a focus of great medical, environmental, and industrial importance. Most enzymes are proteins, while others are ribonucleic acids. Until now, research on the prediction of classes and subclasses of enzymes has always been focused on the enzyme whose chemical nature is the protein; in other words, when computer models are used to classify enzymes, the method of feature extraction adopted is always protein-specific. To facilitate the in-depth study of enzymes, it has become important to classify and name enzymes internationally. “International Union of Pure and Applied Chemistry” (IUPAC) and “International Union of Biochemistry and Molecular Biology” (IUBMB) have created the “IUPAC-IUBMB Joint Commission on Biochemical Nomenclature” (JCBN) and the “Nomenclature Committee of the International Union of Biochemistry and Molecular Biology (NC-IUBMB)”. It is the NC-IUBMB that maintains the nomenclature and classification of enzymes [3].

The latter created an international commission called the Enzyme Commission, which is tasked with developing a nomenclature for enzymes. The Commission has divided the enzymes into 7 major classes depending on their reactions (see Table 1 to look at the classes) [4]. The EC nomenclature is made up of four parts that identify respectively the main class, subclass, sub-subclass, and substrate class of the enzyme [5]. It is worth mentioning that in recent years, many classifiers capable of classifying enzymes have appeared. Most of the classifiers designed by researchers can classify enzymes at the subclass level [6]. As enzymes were divided into 6 major classes according to the type of reaction catalyzed, a seventh class, the translocase, was added in 2018 [4]. Most prediction methods divide the enzymes into just six classes [7-11]. Functional laboratory identification approaches have been used to discover the function and class of enzymes, although this form of experiment is costly and time-consuming [12, 13]. Consequently, it is appropriate to classify enzymes using bioinformatics technologies and deep learning methods.

With the development of bioinformatics and deep learning [14, 15], researchers have designed many models for enzyme class prediction [16]. In 2003, F. Y. Hunt et al [17] introduced an algorithmic approach to facilitate sequence alignments between a query sequence and a database of sequences. The primary aim of this approach was to identify highly similar proteins that are already annotated, under the assumption that information about the biological function or structure of a protein can be gained through such identification. In 2009, Nasibov et al [18] adopted the K-nearest neighbor (KNN) classification method, which is a machine-learning classification algorithm used for classification and regression tasks. It works by finding the K closest data points in the training set to a given test point, and the class of the test point is determined by a majority vote of the K-nearest neighbors. In 2010, Qiu et al [19] used the support vector machine (SVM), achieving good results. In 2018, Yu Li et al [20] used a combination of Convolutional neural network (CNN) and long short-term memory (LSTM) to build their classification model. In addition, in order to achieve better prediction results, researchers usually combine various feature extraction and classification methods in their prediction process. For example, Shen et al [21] combined functional domain (FunD) and scoring matrix (PsePSSM) to extract features in 2009. Wang et al [22] combined composition, transition, and distribution (CTD) and pseudo amino acid composition (PseAAC) to extract features and ranked sequences with the combination of (RAkEL-RF) and multi-label KNN (MLKNN) methods in 2014. In 2018, Yu Li et al [20] combined one hot encoded protein sequence, position-specific scoring matrix (PSSM), solvent accessibility information, secondary structure information predicted by DeepCNF [23], and functional domains. In 2020, Safyan Aman Memon et al [24] combined the protein sequence, PSSM, disordered regions, protein secondary structure, solvent accessibility, amino acid composition, and functional domains.

Our contribution consists in proposing a new approach in which artificial intelligence is the pillar. It consists in using natural language processing (NLP) [25] with deep learning to determine the function of enzymes based on their primary structure. we chose the protein sequence as the only training data in order to avoid involving erroneous data from a false prediction (for example predicted protein secondary structure...) so that the latter is the only data necessary to predict the function. The organization of the rest of the paper is as follows: The research methodology is described in Section 2. Section 3 presents the results and discussion. The last section concludes this paper and discusses future work.

2. MATERIALS AND METHODS

2.6. Web Application

The website built is basically one web page where the user can input a protein sequence. A backend function checks if the input represents a real protein sequence by checking many criteria: sequence size, the first amino acid of the sequence, and the amino acids composing the sequence. The inputted sequence is encoded the same way our dataset sequences were encoded. The encoding output is then passed to the models so the prediction can be made. The results are then displayed to the user in a table (Fig. 1).

2.6.1. Model Overview

EZYDeep, the tool presented in this paper, is a deep learning method based on convolutional neural networks. The experiments were performed on 4 levels which are: level 0: Enzyme and non-enzyme, level 1: first digit EC number prediction, level 2: second digit EC number prediction, level 3: third digit level EC number prediction. We performed some tests on level 4 too which is the fourth and last digit of the EC number but as the majority of 4^th-level classes have a number of sequences inferior to 5, the experiment did not yield satisfactory results.

2.6.2. Evaluation Measure

To assess the quality of the predictions, we employed several commonly adopted metrics [27, 28]. These metrics provide a comprehensive evaluation of the performance of our enzyme classification models. The evaluation metrics used are as follows:

1- Accuracy: This metric calculates the ratio of correctly classified instances (True Positives and True Negatives) to the total number of instances evaluated (True Positives, True Negatives, False Positives, and False Negatives). It is defined as:

2- Recall: Also known as the true positive rate, recall measures the proportion of true positive instances correctly identified by the model. It is defined as:

3- Precision: Precision measures the proportion of true positive instances out of all instances predicted as positive. It is defined as:

4- F1-score: The F1-score is the harmonic mean of precision and recall, providing a balanced measure of the model's performance. It is defined as:

Where:

True Positive (TP): means the actual value and predicted value are the same.

True Negative (TN): means the sum of values of all columns and rows except the values of that class that we are calculating the values for.

False Positive (FP): means the sum of values of the corresponding column except for the TP value.

False Negative (FN): means the sum of values of corresponding rows except for the TP value.

In our experimentation, we conducted a two-step approach. First, we performed a simple test by randomly splitting the dataset into 90% for training and 10% for testing. This initial test provided a baseline assessment of the model's performance.

Subsequently, we employed cross-validation with 10 folds on the entire dataset. This process involved dividing the data into ten equal parts, iteratively training the model on nine parts, and testing the remaining part. By repeating this process ten times, we ensured that each part of the data was used for both training and testing. This cross-validation approach enabled us to obtain a more robust evaluation of the model's performance.

By employing this experimental setup, we obtained comprehensive insights into the effectiveness and generalizability of our enzyme classification model.

3. RESULTS AND DISCUSSION

As mentioned before, the models are tested on 10% of the dataset. The test method is based on the confusion matrix.

Table 6 shows the performance of the EZYDeep models on four different levels of classification, with evaluation metrics including accuracy, precision, recall, and F1-score. At Level 0, the model achieved an accuracy of 97.04% with high precision, recall, and F1-score. For Level 1 classification, the model achieved an accuracy of 97.35% with high precision, but lower recall and F1-score compared to Level 0. For Level 2 classification, the model achieved a lower accuracy of 96.41% with decreased precision, recall, and F1-score. At the highest level of classification (Level 3), the model achieved an accuracy of 95.83% with a high precision, but a significant decrease in recall, resulting in a lower F1 score compared to the other levels. These results suggest that the model performs well for broader classifications.

In addition to the test results obtained from the 10% dataset, we further conducted 10-fold cross-validation to assess the performance of our EZYDeep models. The evaluation metrics used were accuracy, precision, recall, and F1-score. Table 7 presents the performance of the models at four different levels of classification.

The performance measures in Table 6 demonstrate a strong correlation between the evaluation metrics. At Level 0 classification, the model achieved an accuracy of 97.01%, with a precision of 97.04%, recall of 97.01%, and an F1-score of 96.90%. Similarly, at Level 1 classification, the model achieved an accuracy of 97.06%, with a precision of 97.08%, recall of 96.06%, and an F1-score of 97.06%. Moving to Level 2 classification, the model achieved an accuracy of 96.13%, with a precision of 96.22%, a recall of 96.13%, and an F1-score of 96.14%. Finally, at the highest level of classification (Level 3), the model achieved an accuracy of 95.46%, with a precision of 95.79%, recall of 95.46%, and an F1-score of 95.53%.

These results suggest a consistent trend, indicating that as the classification becomes more specific (higher levels), the model encounters challenges in accurately identifying instances belonging to more specific enzyme classes. This leads to a decrease in precision, recall, and subsequently affects the F1-score, while the accuracy remains relatively high.

The inclusion of cross-validation allows for a more comprehensive evaluation of the model's performance, taking into account variations in the dataset splits and providing a more robust assessment of its generalization capabilities. The strong correlation observed between the evaluation metrics further validates the model's performance across different levels of classification.

The consistent trend observed in the performance measures indicates the model's ability to achieve high accuracy at broader classification levels. However, as the classification becomes more specific, the model encounters challenges in accurately identifying instances belonging to more specific enzyme classes. To gain further insights into the model's limitations, we analyzed the misclassified sequences, which shed light on potential factors contributing to these errors.

When analyzing the misclassified sequences, several factors could have contributed to the incorrect classification. One possible reason is the presence of ambiguous or overlapping features within the dataset. Enzymes with similar characteristics or functional properties may share common sequence patterns, making it challenging to distinguish between them accurately. Additionally, the model's performance may be influenced by the quality and representativeness of the training data. As certain enzyme classes are underrepresented or inadequately represented in the training set, the model may struggle to learn their distinguishing features effectively. Furthermore, the complexity of the classification task and the inherent variability within enzyme families can also pose challenges. Some enzymes may exhibit diverse functional properties within the same class, leading to misclassification due to subtle differences in their sequences.

Addressing these challenges could involve refining the feature representation used by the model, augmenting the training data with more diverse and representative samples, and exploring more advanced machine-learning techniques or ensemble approaches. By investigating and understanding the reasons for misclassifications, we can enhance the accuracy and robustness of the classification model in future iterations.

Overall, the results obtained from the evaluation metrics and the analysis of misclassified sequences provide valuable insights into the performance and limitations of our EZYDeep models. These findings can guide further research efforts toward developing more accurate and reliable enzyme classification models, ultimately contributing to a deeper understanding of enzyme functionality and facilitating various biotechnological applications.

4. COMPARISON WITH PREVIOUS METHODS

In our study, we aimed to develop a tool for predicting non-enzymes and enzyme functions based solely on their sequence information. To validate its performance, we conducted a comparison with two existing tools, HECNet [24] published in 2020, and DEEPre [20,28] published in 2018, which incorporate additional information, such as protein secondary structure, in addition to non-enzymes and enzymes sequences for prediction.

To evaluate the performance of these tools, we used the HECNet July 2019 dataset, consisting of 12,889 enzymes and 30,374 non-enzymes instances. To compare the performance of the tools, we used the accuracy and F1 score as evaluation measures. The results, shown in Table 8, demonstrate that our tool outperformed both HECNet and DEEPre, achieving higher accuracy and F1 score on the July 2019 dataset.

This comparison (Table 8) highlights the strengths of our tool, which is able to effectively classify and predict non-enzymes and enzyme functions based solely on their sequence information. However, it also highlights some of the limitations of our tool in comparison to previous methods that incorporate additional information. In the future, it may be worth exploring the integration of complementary data sources, to further improve the accuracy of the prediction.

Overall, our results demonstrate the potential of using deep learning for non-enzymes and enzyme functions prediction based on sequence information, and suggest promising avenues for future research in this area.

Moreover, our results suggest that deep learning approaches could be used to predict enzyme functions even for sequences that have not yet been experimentally characterized, paving the way for more efficient enzyme annotation and drug discovery.

CONCLUSION

In conclusion, this study presents a deep learning tool for classifying enzymes based solely on sequence information. The tool, designed using a sequential model with convolutional and dense layers in TensorFlow, showed exceptional performance when evaluated against two existing methods, HECNet and DEEPre, on the HECNet July 2019 dataset. The results revealed that our tool outperformed both HECNet and DEEPre in terms of accuracy and F1 score, underscoring the potential of using sequence information alone to accurately classify enzymes. Notably, our tool achieved a very high accuracy rate of over 95%, indicating its effectiveness in predicting enzyme function.

Our comparison with state-of-the-art enzyme function prediction tools proved that our tool performs significantly better than other methods at the four levels of prediction. This indicates that our tool has the potential to become a valuable resource for predicting enzyme functions, potentially speeding up the process of enzyme annotation.

This work demonstrates the capability of deep learning to capture functional information contained within enzyme sequences and highlights the significance of sequence information in predicting enzyme function. While our tool showed promising results, there is still room for improvement, and future studies may consider incorporating additional information to enhance the accuracy of enzyme classification.

In summary, this study has significant implications for the field of enzyme classification and for the advancement of biomedical research. The success of our tool and its availability through a web application highlights the potential of using machine learning techniques to aid in the prediction and understanding of enzyme function and provides a foundation for future research in this area.

REFERENCES