*3.2. Experimental Setup*

Using two benchmark and one own prepared protein datasets, we analyzed and verified the e fficiency of the proposed model. The model was trained on a Titan Intel Core i5-6600 processor with X (Pascal)/PCLe/SSE2 GPU, having 64GB of memory using 16.4 LTS Ubuntu operating system. The proposed deep learning model was executed in version 3.5 of python, version 2.2.4 of Keras, and version 1.12 of TensorFlow backend along with an Adam employed as an optimizer. To find the most favorable selection of the hyperparameter of each model, several experiments were conducted. At last, we selected 50 epochs to train the model with a batch size of 100. The PF2095 and MPD samples were split into training 70% and testing 30%, and due to fewer numbers of protein samples in PF175 we kept 80% data in training, and 20% data is utilized for model evaluation.

#### *3.3. Evaluation Metrics*

In this study, a couple of assessment measures are used for the evaluation of the proposed model.These parameters include accuracy, sensitivity, and specificity. The mathematical formulas are defined in the Equation (15)–(17).

$$\text{Accuracy} = \frac{\text{TP} + \text{TN}}{\text{TP} + \text{TN} + \text{FP} + \text{FN}} \times 100\tag{15}$$

$$\text{Sensitivity} = \frac{\text{TP}}{\text{TP} + \text{FN}} \times 100 \tag{16}$$

$$\text{Specificity} = \frac{\text{TN}}{\text{TN} + \text{FP}} \times 100 \tag{17}$$

Now, let us assume that the mitochondrial protein is positive, and the non-mitochondrial protein is negative. The true positive (TP) is that value in which predictive and actual value is positive and true negative (TN) is the value in which predicted value and actual value is negative. Similarly, the false positive (FP) is the value in which a machine predicted as positive but actually it is a negative value and the false negative (FN) is that value in which machine predicted as negative class but actually it is related to positive class value.

#### *3.4. Ablation Study on PF2095*

In this subsection, we conduct an ablation study after comprehensive experiments to analyze the three models in terms of accuracy, sensitivity, and specificity on the PF2095 dataset, which is a new mitochondria proteins of Plasmodium dataset comprising 890 positive and 1205 negative samples. In this dataset 70% of total samples are set for training, and the remaining 30% are used for model evaluation. First, we perform our experiments on CNN-GRU which achieved 89.7% training accuracy, 88.0% testing, 90.4% sensitivity, and 88.9% specificity. The next model CNN-LSTM showed better performance compared to previous one. It obtained 93.5% training, 91.2% testing accuracy, 90.6% sensitivity, and 91.7% specificity. The proposed model MPPIF-NET used CNN with integration of MBD-LSTM with the same number of parameters. It is experimentally proved that the last hybrid approach shows supremacy of performance which obtained 98.2% training accuracy, 97.6% is testing performance of the model, 98.1% of its sensitivity, and 97.2% specificity. The detailed experimental evaluation results are depicted in Table 3 and confusion metrics of the MPPIF-NET are shown in Figure 4.


**Table 3.** Training and testing performance of the MPPIF-NET on different models and datasets.

**Figure 4.** Confusion metrics of the proposed method over the PF2095 dataset.

#### *3.5. Experimental Evaluation on PF175*

We used irregular data in our experiments along with a hold-out technique which is the simplest kind of cross validation. The data was divided into training and testing. We trained our proposed model on 80% of the data and the remaining 20% of data were used for evaluation purposes. During experiments we updated different parameters to achieve good performance. After numerous experiments we set these parameters and their value; for example, maximum length of the protein is 1280 which depends upon the dataset, maximum features = 26, embedding size = 8, number of filters in convolutional are 32, pooling length = 2, batch size = 100, dropout = 0.2, and number of epochs is 50. We also checked different numbers of epochs and finally realized that the trained model fits the protein sequences well and predicts accurately on epochs 50.

Our model achieved better performance in terms of 100% training accuracy, 100% sensitivity, 96.2% specificity, and testing accuracy of 97.14%, which is higher than other state-of-the-art approaches. The confusion matrics of correctly and incorrectly predicted proteins are shown in Figure 5.

**Figure 5.** Confusion metrics of the proposed method over the PF175 dataset.

#### *3.6. Experimental Evaluation on MPD*

This dataset is also an unbalanced dataset and used the hold-out method during experiments. The data is divided into training and testing, which is 70% and 30%. For this dataset we also set the same parameters, except the maximum length of the protein which is 1402; maximum features = 26, embedding size = 8, number of filters in convolutional are 32, pooling length = 2, batch size = 100, dropout = 0.2 and the number of epochs is 50. We have done a lot of experiments with different setup parameters, but finally on epochs 50 we achieved better performance. The confusion matrics of correctly and incorrectly predicted proteins are shown in Figure 6.

**Figure 6.** Confusion metrics of the proposed method over the mitochondria protein dataset (MPD) dataset.

Our model go<sup>t</sup> 99.7% training accuracy, 99.3% sensitivity, 100% specificity, and a testing accuracy of 99.5%, which shows that the proposed model is superior in contrast to state-of-the-art techniques.

#### *3.7. Comparative Analysis of the MPPIF-NET with Other Models on PF175*

In the post genomic era, functional annotation is one of the major challenges. From the last decade, a vast number of machine learning and bioinformatic techniques have been proposed to predict protein functionality. The statistics of sequences are boosted day by day in the protein databanks. Identification of these biological sequences via laboratory methods was a laborious task. Therefore, we proposed a deep learning model for the accurate prediction of a huge number of proteins. Hence, it is important to evaluate the performance of models in order to compute the realistic performance of the model. For this we compared our proposed model with the state-of-the-art method using the same dataset. In the first attempt Bhasin et al. [35] proposed a model for eukaryotic subcellular localization protein prediction called (Eslpred) using a hybrid approach containing a dipeptide composition and PSI-BLAST. They achieved 69.71% accuracy, 73.33% specificity, and 57.50% sensitivity. Guda et al. [36] developed a new method for genome-scale prediction of the target mitochondria protein based on the composition of the amino acid and the occurrence frequency of each pattern which repeats in sequences. They achieved 80% accuracy, 87.41% specificity, and 55% sensitivity. Bender et al. [6] built a neural network model for the precise prediction of mitochondrial transit peptides which causes malaria. Due to the complex genomic sequence of PF, Chen et al. [11] developed the increment of diversity model in which a reduced amino acid composition was used in order to extract local features from the biological sequence. The prediction performance achieve 100% superior sensitivity rate, 89% specificity, and 92% accuracy as shown in Table (4). Mitochondria are vital organelles of eukaryotic cells which are involved in processing cellular death and human diseases; therefore, Afridi et al. [9] proposed an ensemble model known as Mito-GSAAC in which the main purpose was to examine an e ffective feature extraction approach. They achieved the highest specificity score of 95.56%, 93.21% accuracy, and 87.5% sensitivity. Accurate identification of the mitochondrial protein of Plasmodium falciparum is an essential role in the discovery of anti-malarial drug targets. Ding et al. [10] used a dipeptide composition for protein encoding. They also used the analysis of variance to overcome the issue of overfitting. They attained 97.1% accuracy, 90% sensitivity, and 99.3% specificity. The aforementioned state-of-the-art techniques utilized the machine learning approaches for the protein sequences prediction. We proposed a deep learning strategy for identification of these biological sequences which gave 97.14% superior testing accuracy compared to other discussed methods as shown in Table 4.


**Table 4.** MPPIF-NET comparative analysis with other models on PF175 dataset.

#### *3.8. Comparative Analysis of the MPPIF-NET with Other Models on MPD*

Mitochondria are the center and powerhouse of the eukaryotic cells. Pharmaceutical companies still desire such a system which accurately predicts the mitochondria protein of Plasmodium in order to prepare drugs. Therefore, Tan et al. [34] proposed an algorithm in order to evaluate the pair composition of amino acids. The extracted features are then passed to the support vector machine classifier for prediction of Plasmodium mitochondria proteins. The SVM model was evaluated which achieved 85% accuracy, 89.28% specificity, and 79.16% sensitivity. Jiang et al. [37] developed a new sequence-based

method which is known as the Discrete Wavelet Transform for sequence prediction. They achieved 50.30% sensitivity, 95.74% specificity, and 76.53% accuracy. Afridi et al. [9] used four computational methods such as AAC, DPC, SAAC, and PAAC. Furthermore, they also evaluated the six machine learning algorithms, such as support vector machine, random forest, multilayer perceptron, AdaBoost, and bagging. Finally, on the basis of the ensemble classifier they achieved 92.62% accuracy, 91.52% specificity, and 90.96% sensitivity. Our proposed model performs well compared to the state-of-the-art methods, having 99.5% accuracy, 100% specificity, and 99.33% sensitivity as shown in Table 5.


**Table 5.** MPPIF-NET comparative analysis with other models on the MPD dataset.

#### **4. Conclusions and Future Directions**

For the identification of mitochondria proteins of Plasmodium some biologists are still concentrating on extracting new patterns from biological sequences and are searching for appropriate machine learning algorithms which accurately classify proteins. In this study, we proposed a deep leaning framework MPPFI-Net which is capable of extracting deep features automatically and can discriminate proteins quickly and accurately. We merged the CNN and MBD-LSTM in order to extract the contextual information from amino acids. Later on, we compared MPPFI-Net performance with the state-of-the-art models, and we conclude that the proposed framework speeds up the performance regarding both prediction accuracy and fitting uncharacterized data. In future, we will boost this work by fusing the traditional features and deep features.

**Author Contributions:** Conceptualization, S.U.K. and R.B.; methodology, S.U.K.; writing—original draft preparation, S.U.K.; review and editing, S.U.K. and R.B.; supervision, R.B. All authors have read and agreed to the published version of the manuscript.

**Funding:** The publication fees were supported by Prof. Ran Baik (HONAM University-20190125).

**Acknowledgments:** The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: this study was supported by the research fund from Honam University (2019).

**Conflicts of Interest:** The authors declare no conflict of interest.
