Identification of cyclin protein using gradient boost decision tree algorithm

Graphical abstract

Highlights

• •An computational method was developed to identify Cyclin Protein.
• •Several kinds of features were proposed to formulate protein samples.
• •A fusion feature set could produce a high accuracy.

Abstract

Cyclin proteins are capable to regulate the cell cycle by forming a complex with cyclin-dependent kinases to activate cell cycle. Correct recognition of cyclin proteins could provide key clues for studying their functions. However, their sequences share low similarity, which results in poor prediction for sequence similarity-based methods. Thus, it is urgent to construct a machine learning model to identify cyclin proteins. This study aimed to develop a computational model to discriminate cyclin proteins from non-cyclin proteins. In our model, protein sequences were encoded by seven kinds of features that are amino acid composition, composition of k-spaced amino acid pairs, tri peptide composition, pseudo amino acid composition, geary correlation, normalized moreau-broto autocorrelation and composition/transition/distribution. Afterward, these features were optimized by using analysis of variance (ANOVA) and minimum redundancy maximum relevance (mRMR) with incremental feature selection (IFS) technique. A gradient boost decision tree (GBDT) classifier was trained on the optimal features. Five-fold cross-validated results showed that our model would identify cyclins with an accuracy of 93.06% and AUC value of 0.971, which are higher than the two recent studies on the same data.

1. Introduction

Cyclin belongs to a group of proteins which are capable to control the cell cycle by triggering Cdk [1]. Cyclin concentration changes on different levels at several stages of the cell cycle. These changes occurred due to the ubiquitin-mediated cyclin degradation [2]. Cyclin combines with cyclin dependent kinases, like cdk1 proteins and p34, to trigger the cyclin dependent kinase active sites. This cdk1, p34 and cyclin combination forms a MPF (maturation-promoting factor) which activates other proteins [3]. However, phosphorylation is needed for the complete activation of cyclin dependent kinase active sites [3]. Therefore, these phosphorylated proteins are liable for the specific movements during the division of cell cycle e.g., chromatin remodeling and the formation of microtubules [3], [4].

After the Human Genome Project (HGP), biological sequence data has progressively shattered [5]. The traditional investigational techniques have not only low efficient and expensive but also are time consuming. Therefore, it is urgent to identify sequences efficiently in a short period of time. However, existing tools such as FASTA [6] and BLAST [7] only compare the sequence with the known protein databases [8], [9], these tools cannot discriminate whether it is a cyclin or non-cyclin. Now, machine learning classifications are popular in this area [10], [11], [12], [13]. In prior methods, StAR [14] and other classifiers using Pseudo-amino acid composition (PseAAC) could identify cyclins with an accuracy of 83.53%. Sun et al. [15] established a cyclin prediction model based on support vector machine (SVM) which could produce an accuracy of 91.90%. Although both cyclin prediction model can produce good outcomes, there is still room for further improvement by extracting more feature information.

To address the aforementioned issues, an ensemble model was established to predict cyclin in multiple eukaryotic genomes. Fig. 1 shows the workflow of the proposed model. First, seven types of feature descriptors, Amino acid composition [16], Tri-peptide composition [17], Composition of K-spaced amino acid composition [18], [19], Geary autocorrelation [20], Normalized moreau-broto autocorrelation [21], C/T/D [22] and PseAAC [23], [24] were used as features to input into a GBDT classifier [25]. After this, ANOVA [26] and the mRMR [27] with IFS [28] technique was utilized to get optimal feature vectors. The outcomes were evaluated by using five-fold cross validation.

Fig. 1.

The flowchart of the whole study.

2. Materials and methods

A reliable and accurate dataset is necessary to establish a prediction model [29], [30], [31], [32], [33], [34], [35]. Therefore, the dataset was obtained from Mohabatkar et al. [14]. They collected 215 cyclins and 204 non-cyclin proteins to train and test the methods for cyclins prediction. To reduce the overfitting derived from high similarity of sequences, we applied a cluster database at high identity with tolerance 90% [36] and discarded the sequences that exhibited more than 90% sequence identity. As a result, we attained the 167 cyclin and 167 non-cyclin proteins. Then we divided into 70/30 ratio in order to training and testing the model.

2.1. Feature descriptors

Selecting the feature-encodings that are instructive and autonomous is an important step in creating machine learning models [37], [38], [39], [40]. Expressing the protein sequences with a mathematical formulation is key and difficult in functional element identification [41], [42], [43], [44]. Therefore, seven types of feature-encoding approaches were presented to describe the protein sequences.

2.1.1. Amino acid composition descriptor (AAC)

AAC calculates the frequency of single type of amino acids in a protein sequence [16], [45], [46], [47], [48], [49], [50]. The frequencies $f\left(p\right)$ of 20 residues can be calculated as

$$f\left(p\right)=\frac{N\left(p\right)}{N}\mathrm{p}\in \{ACDEFGHIKLMNPQRSTVWY\}$$ (1)

where $N\left(p\right)$ is the number of the p-th residue in a protein sequence with the length of N residues.

2.1.2. Composition of k-spaced amino acid pairs descriptor (CKSAAP)

The encoding technique composition of k-spaced nucleic acid pairs embodies the incidence of nucleotide pairs disconnected by any K nucleotide (K = 0, 1, 2, 3, 4, 5). The CKSAAP [18], [45], [51] is defined as k-spaced residue pairs $Q_{\mathit{xy}}$ which is illustrated as

$Q_{\mathit{xy}}=\frac{N_{\mathit{xy}}}{N-k}(k=0,1,2,3,4,5\mathrm{a}\mathrm{n}\mathrm{d}xy=\mathrm{t}\mathrm{y}\mathrm{p}\mathrm{e}\mathrm{o}\mathrm{f}\mathrm{A}\mathrm{A})$ (2)where $N_{\mathit{xy}}$ is the number of residue pairs and k denotes the number of nucleotides. In this study, k = 3 and the dimension of the composition of k-spaced amino acid pairs feature was 1600.

2.1.3. Pseudo amino acid composition descriptor (PseAAC)

PseAAC describes the occurrence of the amino acid frequency and the correlation of between two residues’ physicochemical properties [23]. It consists of ${\mathit{Ac}}_i$ and ${\mathit{Ac}}_{\partial i}$.

$${\mathit{Ac}}_i=\frac{N_i}{1+\omega \times {\sum }_{i=1}^{20}{\theta }_i}(\therefore {\theta }_i=\frac{{\sum }_{i=1}^{N-d}{{(Q}_i-Q_{i+d})}^2}{N_Q},(\mathrm{i}=1,2,3\cdots ,20))$$ (3)

$${\mathit{Ac}}_{\partial i}=\frac{\omega \times {\theta }_i}{1+\omega \times {\sum }_{i=1}^{20}{\theta }_i},(\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e},\omega =0.05)$$ (4)

where $N_Q$ is the number of properties and $N_i$ is the i-th amino acid occurrence. $Q_i$ is the ith amino acid property value and ${\theta }_i$ is the sequence order factor.

2.1.4. Tri-peptide composition descriptor (TPC)

TPC are three amino acid molecules joined together and reflects hypothetically substantial starting points for the design of small biotic modulators [17]. Tripeptide composition is defined as

$$f_{\mathit{lmn}}=\frac{N_{\mathit{lmn}}}{N-2}(\therefore l,m,n\in (\mathrm{A},\mathrm{B},\mathrm{C}\cdots .,\mathrm{Z}))$$ (5)

where $N_{\mathit{lmn}}$ represents the number of tripeptide amino acid type l, m, n.

2.1.5. Composition\transition\distribution descriptor (C/T/D)

C/T/D defines the global composition of an amino acid sequence and the frequencies of two different adjoining amino acids and the distribution pattern of an amino acid sequence. Sequence scrambling is the first job to compute the composition, transition and distribution [52]. On the basis of their attributes, amino acids are alienated into three classes (class 1, class 2 and class 3) [24], also named reduced or simplified amino acids [53], [54]. Classifications of charge and hydrophobicity are shown in Table 1. C/T/D with composition $C_a$, transition $T_b$ and distribution is defined as $D_{b,z}$.

$$C_a=\frac{N_a}{N}(\therefore a=1,2,3\cdots )$$ (6)

$$T_b=\frac{N_{b,c}+N_{c,b}}{N-1}(\therefore b=1,2,3\cdots ,c\ne b)$$ (7)

$$D_{b,z}=\frac{N_{b,z}}{N}(\therefore b=1,2,3\cdots ,z=1,0.15N\cdots ,N)$$ (8)

where $N_a$ is the class number, $N_{b,c}$ is the adjoining number of class b and c. $N_{b,z}$ is the number of those AA which are in z-th of b-th class.

Table 1.

Attribute classification.

C 1	C 2	C 3	Attributes
±tive R, K	Not + tive nor -tive A, N, C, Q, H, I, G, L, F, P, S, M, W, Y, V, T	-tive D, E	Charge
Polar D, E, K, N, Q, R	Not + tive nor -tive A, G, H, P, S, T, Y	Hydrophobic C, F, I M, V, W	Hydrophobicity

2.1.6. Geary descriptor (GD)

Geary descriptor is a kind of correlation descriptor and have a maximum similarity with Moran descriptor [55]. It is well-defined as $Q(r)$:

$$Q\left(r\right)=\frac{N-1}{2\times \left(N-r\right)}\times \frac{{\sum }_{i=1}^{N-r}{\left(P_i-P_{i+r}\right)}^2}{{\sum }_{i=1}^N(P_i{-r)}^2}(\therefore r=1,2,...,20$$ (9)

where $P_i$ is the property value of the i-th amino acids in AA index.

2.1.7. Normalized moreau-broto autocorrelation descriptor (NMBroto)

NMBroto is also a type of autocorrelation [21] and also have a likeness with Moran as shown in below equation.

$$Q(r)=\frac{{\sum }_{i=1}^{N-r}(P_i\times P_{i+r})}{N-r}(\therefore r=1,2,...,20$$ (10)

where $P_i$ is the property value of the i-th amino acids in AA index.

2.2. Feature selection

The noise in feature vector might result in the unsatisfactory performance of a model [56], [57], [58], [59], [60], [61], [62], [63]. Therefore, the selection of features is an obligatory phase to remove the less important features and increase the productivity of a model [37], [64], [65], [66], [67], [68], [69]. Many feature selection and ranking techniques are available, such as ANOVA, F-score [70], mRMR [27], Chi-square [71], LGBM [72], [73]. A high feature dimensions both can create overfitting and information redundancy and produce poor accuracy of the cross -validation prediction. Therefore, ANOVA is good option to tackle these issues because it consumes less time and gave efficient results. The combination of some of the top-executing features does not mean that the top predictive results can be attained. These features are probably to have a high degree of correlation, which leads to additional redundant information in the feature vectors. Therefore, mRMR is a good option to tackle these issues due to less time consuming and efficient results. These techniques are also used in many high dimensional protein features selection. In this study, the ANOVA and mRMR [27] with IFS [56] was applied to obtain the optimal feature subset. The comparison with other state of the art feature selection techniques is given in Fig. 2S in Supplementary file 1.

2.2.1. ANOVA

ANOVA is used for significance test of mean difference between two or more samples. F-value is the ratio of variance between groups and variance within groups [74]. If the F-value will be larger, then the ability of distinguishing positive and negative samples will be better. Therefore, all features can be sorted according to this F-value.

$$Q_m^2(\xi )=\sum _{i=1}^r{l}_i\frac{({\stackrel{-}{x_i}-\stackrel{-}{x})}^2}{{\mathit{df}}_m}$$ (11)

$$Q_n^2(\xi )=\sum _{i=1}^r\sum _{j=1}^{l_i}\frac{(x_{i,j}-{\stackrel{-}{x_i})}^2}{{\mathit{df}}_n}$$ (12)

$$F(\xi )=\frac{Q_m^2(\xi )}{Q_n^2(\xi )}$$ (13)

2.2.2. mRMR with IFS

mRMR is a filter-based selection technique [75] to achieve an optimal model. Compactness functions are described as y and z, and P(y) and P(z) are the two corresponding probabilities. P (y, z) is the possibility of compactness, and the common information between the two functions can be defined as

$$I\left(y;z\right)=\iint P\left(y,z\right)\mathit{log}\frac{P\left(y,z\right)}{P\left(y\right)P\left(z\right)}dydz$$ (14)

In shared information, searching a subset S with m optimum features helps to determine the feature transmission, which majorly depends on the target {yi} class q.

$$\mathit{max}\mathit{d}\left(S,q\right),d=\frac{1}{\left|S\right|}{\sum }_{y_{i\in S}}I\left(y_i,q\right)\left(i=1,2,3\cdots m\right)$$ (15)

Minimum redundancy can be defined as

$$\mathit{min}\mathit{r}\left(S,q\right),r=\frac{1}{{\left|S\right|}^2}{\sum }_{y_{i,}{y}_{j\in S}}I\left(y_i,y_j\right)$$ (16)

Final selection criteria can be articulated as

$$\mathit{max}\varnothing \left(d,r\right),\varnothing =d-r$$ (17)

The principle of the mRMR technique is to use a typical redundancy and relevance to rank features to acquire the best subset. The IFS [28], [76] scheme was applied in the present study to select the best feature. The details about the IFS method can be found in [56].

2.3. Machine learning classifiers

Classification is a type of supervised learning and have an important role in the decision making [77], [78], [79], [80], [81], [82], [83], [84], [85]. In this study, we select GBDT [25] to identifying cyclin and non-cyclin proteins. Another four kinds of machine learning classifiers Naïve Bayes [86], Support Vector Machine [56], [57], [87], and Ada Boost [88] and Random Forest [84] were performed for comparison.

Gradient boost decision tree algorithm is a very important learning algorithm and has been applied by the researchers in many bioinformatics and mathematical and biological applications [89], [90]. It constructs a climbable and authentic model from a non-linear joint of different weak learners. The main idea of the gradient boost decision tree is to establish a base learner which is excellently interrelated with the loss function of negative gradient [25]. Suppose that there are n numbers of samples:

{($x_1,y_1$)…($x_n,y_n$)} ($\therefore x_i∊x\subseteq R_n,and{y}_i∊y\subseteq R$)

$$f_k\left(x\right):=\sum _{k=1}^{k}T\left(x;{\theta }_k\right)$$ (18)

where $T\left(x;{\theta }_k\right)$ is the new decision tree (k = 1,2,3….), and ${\theta }_k$ is the risk minimization parameter of the new decision tree which is shown in below equation.

$$\widehat{{\theta }_k}=argmin\sum _{i=1}^{n}L\left(y_i,f_{k-1}\left(x\right)+T\left(x;{\theta }_k\right)\right)(\therefore Listhelossfunction$$ (19)

Gradient boost decision tree algorithm calculates the final assessment in a forwarding mode.

$f_k\left(x\right)=f_{k-1}\left(x\right)+T\left(x;{\theta }_k\right)$ (20)

Finally, Loss function $f_{k-1}$ of negative gradient is used for residual calculation.

$$R_{\mathit{ki}}=-{\left[\frac{\partial L\left(y_i,f\left(x_i\right)\right)}{\partial f\left(y_i\right)}\right]}_{f\left(x\right)=f_{k-1}\left(x\right)}(\therefore i=1,2,3....n)$$ (21)

At the end, we trained the model by all $R_{\mathit{ki}}$ to calculates the risk minimization parameter${\theta }_k$. This type of decision trees logically models the relations amongst predictor variables. e.g., mapping the parameters input space X in to J split sections $R_1...R_J$, and the output is $Z_J$ for region $R_J$.

$$T(x;\theta )=\sum _{j=1}^J{z}_{j}I(x_j∊R_j)$$ (22)

The pseudo code of gradient boost decision tree is given below in Algo 1.

Algo 1: Gradient Boosting Decision Tree Algorithm
Input: Training Data: = ${\left(x_i,y_i\right)}_{i=1}^n$
Where, $x_i$ is a data point and $y_i$ is the label for $x_i$
Loss function: = L ($y_i$, f ($x$))
1.Initialize the model $f\left(x\right)$: = argmin${\sum }_{i=1}^{n}L\left(y_i,z\right)$ 2.for k = 1,2, 3…, K Do 3.for I = 1,2, 3…, n Do 4.By Calculating the Pseudo residual error: $R_{\mathit{ki}}$=$-{\left[\frac{\partial L\left(y_i,f\left(x_i\right)\right)}{\partial f\left(y_i\right)}\right]}_{f\left(x\right)=f_{k-1}\left(x\right)}$ 5.End 6.End 7.By Constructing a new Decision Tree $T_k\left(x;{\theta }_k\right)$, based on $R_{\mathit{ki}},{\theta }_k$= {$R_{\mathit{kj}}j=\left[1,2,3\cdots J\right]$} 8.for j = 1, 2, 3…., J Do 9.$z_{\mathit{kj}}$ = argmin${\sum }_{x_i\in R_{\mathit{kj}}}^{n}L\left(y_i,f_{k-1}\left(x\right)+z\right)$ 10.End 11.Updating the model $f_k\left(x\right)$ = $f_{k-1}\left(x\right)$ + ${\sum }_{j=1}^j{z}_{\mathit{kj}}I\left(x\widehat{I}{R}_{\mathit{kj}}\right)$ 12.f (x) = ${\sum }_{k=1}^K{\sum }_{j=1}^J{z}_{\mathit{kj}}I\left(x\widehat{I}{R}_{\mathit{kj}}\right)$
Output: The decision tree function f (x)

Scikit - learn package (v – 0.22.1) [91] was used to execute the random forest classifiers. Firstly, we used randomized search cross-validation and then grid search cross-validation to tune hyperparameter. The best tuned parameters of the proposed model are given in Table 2.

Table 2.

Best parameters of the proposed model.

Best Parameters
‘Max-depth’	20
‘Max-features’	05
‘Min-samples-leaf’	03
‘Learning-rate’	0.05
‘Min-samples-split’	02
‘N-estimators’	80
‘Mean square error’	0.1287

2.4. Evaluation metrics

Sensitivity (Sn), specificity (Sp), accuracy (Acc), and matthews correlation coefficient (MCC) [92], [93], [94], [95], [96], [97], [98], [99], [100], [101], [102], [103], [104], [105], [106] were used in this study to check the overall efficiency of the model defined as Equation (23).

$$\left\{\begin{array}{c}\mathit{Sn}=\frac{\mathit{TP}}{\mathit{TP}+FN}\hfill \\ \mathit{Sp}=\frac{\mathit{TN}}{\mathit{TN}+FP}\hfill \\ \mathit{Acc}=\frac{\mathit{TP}+TN}{\mathit{TP}+FP+TN+FN}\hfill \\ \mathit{MCC}=\frac{\mathit{TP}\times TN-FP\times FN}{\sqrt{\left(\mathit{TP}+FN\right)\times \left(\mathit{TN}+FN\right)\times \left(\mathit{TP}+FP\right)\times \left(\mathit{TN}+FP\right)}}\hfill \end{array}\right)$$ (23)

where TP represents the overall cyclins sequences in benchmark data and FP signifies the cyclins sequences false-classified as non-cyclins. Likewise, TN represents the overall non-cyclins sequences in the data and FN signifies the non-cyclin sequences, which were false-classified as cyclins. Consequently, the receiver operating characteristic (ROC) curve was used to illustrate the efficiency of the model graphically. The ROC curvature could assess the projecting ability of the proposed model on the whole assortment of resultant values. The area under the curve (AUC) was premeditated to check the efficiency of the model. A good classifier gave AUC = 1, and the arbitrary performance gave AUC = 0.5.

3. Results and discussion

3.1. Performance evaluation

First, the training data were converted into feature vectors using feature descriptors (amino acid composition, composition of k-spaced amino acid pairs, tri peptide composition, pseudo amino acid composition, geary correlation, normalized moreau-broto autocorrelation and composition/transition/distribution), and the feature vectors of each encoding model were evaluated by gradient boost decision tree algorithm using a five-fold CV test. Firstly, the ANOVA and mRMR with IFS were used to pick the best feature subset for the sake of better prediction accuracy. Fig. 2(A) and (B) shows the incremental feature selection curve of optimal features and comparison of single encodings and fusion on different machine learning classifiers on the basis of AUC. Table 3 shows the efficiency of the optimized single-encoding models and the feature fusion model on different machine learning methods. The performance of single-encoding models and the fusion model on different machine learning classifiers before feature selection is recorded in Table 1S in Supplementary file 1. We also visualized the single-encoding features and fusion features using t-SNE (t- distributed Stochastic Neighbor Embedding) method before and after feature selection. The t-SNE visualization of single-encoding and fusion before feature selection is available in Fig. 1S in Supplementary file 1 and the t-SNE visualization of the optimized single-encodings and the fusion is shown in Fig. 3. The AUCs of single-encoding models are 0.827, 0.526, 0.825, 0.506, 0.896, 0.854, and 0.890, respectively for AAC, CKSAAP, PseAAC, TPC, C/T/D, GD, and NMBroto. The AUC of composition/transition/distribution was around 0.6% – 39% higher as compared with those of the other encodings. On the contrary, the Acc, Sp, Sn, MCC, and AUC of the feature fusion model were 93.06%, 94.00%, 92.00%, 0.862% and 0.971, respectively. The Acc, Sp, Sn, MCC, and AUC on independent data were 89.36%, 90.10%, 89.45% and 0.823%. ROC with the AUC of 0.954 is given in Fig. 3S in Supplementary file 1. In order to check the better performance and reliability of our model, we further randomly extracted 50 non-cyclin sequences from the public databases and checked the performance by running our model. We found quite reasonable results. The Accuracy, specificity, sensitivity and matthews correlation coefficient were 90.09%, 91.11%, 89.45%, and 0.829%.

Fig. 2.

Plot showing the Incremental Feature Selection (IFS) procedure for identifying Cyclins in 5-fold cross-validation. (A) Firstly, 5711 features were selected from a total of 10,200 features by ANOVA. (B) 304 optimal features were further obtained from the 5711 features by using mRMR. The Acc increases from 88.92% to 93.06%. (C) Feature descriptor contribution in GBDT-based fusion model to predict cyclins. (D) Comparison between single-encodings and fusion features on different machine learning classifiers.

Table 3.

Performance of optimized single-encodings and fusion models on different machine learning classifiers.

GBDT SVM NB AB RF
Descriptor	Acc	Sp	Sn	MCC	AUC	Acc	Sp	Sn	MCC	AUC	Acc	Sp	Sn	MCC	AUC	Acc	Sp	Sn	MCC	AUC	Acc	Sp	Sn	MCC	AUC
AAC	76.34	74.70	79.60	0.520	0.827	55.39	55.20	57.50	0.108	0.551	74.55	77.00	70.10	0.493	0.800	76.35	74.70	79.60	0.528	0.811	76.05	74.65	77.68	0.518	0.824
CKSAAP	51.19	52.00	79.60	0.029	0.526	50.29	50.20	60.50	0.006	0.503	52.09	54.00	24.20	0.051	0.523	50.09	50.00	80.20	0.001	0.505	53.22	53.12	79.88	0.032	0.551
C/T/D	82.33	81.40	83.80	0.647	0.896	58.68	55.00	91.00	0.231	0.587	72.45	68.00	83.80	0.461	0.808	75.44	73.50	79.50	0.511	0.820	82.16	81.10	83.24	0.644	0.892
Geary	78.74	77.30	81.40	0.576	0.854	50.39	52.70	47.30	0.048	0.524	74.55	75.90	71.90	0.492	0.816	69.76	70.10	68.90	0.395	0.762	79.71	78.10	81.77	0.582	0.876
NMBroto	81.13	80.60	82.00	0.623	0.890	79.04	80.50	76.60	0.582	0.790	79.94	81.60	77.20	0.600	0.877	68.26	67.00	71.90	0.366	0.765	80.93	79.83	82.00	0.621	0.890
PAAC	76.04	74.90	78.40	0.522	0.825	70.65	67.50	79.60	0.420	0.707	62.87	58.60	88.00	0.298	0.745	69.16	67.00	75.40	0.386	0.754	75.88	74.40	78.32	0.521	0.825
TPC	50.89	80.00	24.00	0.074	0.506	49.40	49.50	59.30	−0.01	0.492	50.59	51.40	21.60	0.015	0.502	49.40	49.30	39.50	−0.01	0.493	50.28	78.90	24.00	0.071	0.498
Fusion All	93.06	94.00	92.00	0.862	0.971	79.34	78.40	80.91	0.574	0.794	87.00	88.90	84.70	0.752	0.936	82.33	76.30	87.50	0.628	0.888	91.08	92.55	91.11	0.837	0.956

Fig. 3.

t-SNE visualization of optimized single encoding features and fusion feature. From (A) to (G) showing single encodings and (H) showing fusion of the single encodings. In the figure, 0 in blue color represents non-cyclin and 1 in orange color showing cyclins. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

3.2. Performance evaluation of different ML algorithms

Single-encoding AAC, CKSAAP, PseAAC, TPC, C/T/D, GD, NMBroto and feature fusion models were inputted into different machine learning classifiers such as Ada boost, SVM, and Naive bayes algorithm. Their performances were compared with that of gradient boost decision tree classifier-based models. A five-fold cross-validation test was used to evaluate these model performances. Results were shown in Table 3. We may notice that the accuracies of feature fusion models were always higher than those of single-encoding models, indicating that the multiple information was effective to achieve better results. Fig. 2 (C) showed the feature descriptor contribution in GBDT-based fusion model. The optimized fusion model consists of 304 features of seven descriptors. AAC descriptor contributed 3.28 % in final fusion model because their 10 features were participated in the fusion model. CKSAAP descriptor contributed 16.11 % in final model because their 49 features were participated in the fusion model. CTD descriptor contributed 13.15 % in final model because their 40 features were in the final fusion model. Geary descriptor contributed 32.89 % because their 100 features were participated in the fusion model. NMBroto descriptor contributed 26.31 % in the final optimized model due to their 80 best features. PAAC contributed 4.93 % in the model with their 15 features and TPC contributed 3.28 % in the final optimized model with their best 10 features. Fig. 2 (D) exhibited that the GBDT-based fusion model performed best among all methods. Particularly, the AUC of GBDT classifier was almost 3.5% – 17.7% higher than that of the other models, indicating that the GBDT-based model was the best for cyclin identification.

3.3. Comparison with existing models

In recent studies, Mohabatkar et al., [14] and Sun et al., [15] used the similar dataset for training their models by using jackknife cross-validation. The accuracies of their models were 83.53% and 91.90%, respectively. We also used the same dataset and applied GBDT algorithm. Results on jackknife cross-validation and five-fold cross-validation showed that our model is better than the two existing models. The comparison of two existing models with our model has been shown in Table 4 and Fig. 4.

Table 4.

Comparison between proposed model and existing methods.

Method	CV	Acc (%)	MCC	Sn (%)	Sp (%)	AUC	Reference
Mohabatkar et al., model	Jack-knife	83.53	–	87.44	–	0.894	[14]
Sun et al., model	Jack-knife	91.90	–	91.00	92.80	0.915	[15]
iCyclin	Jack-knife	92.74	0.853	91.60	93.21	0.958	This Work
iCyclin	Five-fold	93.06	0.862	92.00	94.00	0.971	This Work

Fig. 4.

ROC curve of proposed model and the two existing methods on the basis of jackknife and five-fold cross-validation. The AUCs of different models have been showed.

4. Conclusions

Cyclin proteins are capable to regulate the cell cycle and forms a complex with cyclin-dependent kinases. This complex activates cell cycle but the full activation requires phosphorylation. Cyclin protein have low similarity between their sequences. To date, numerous predictors have been established to classify cyclins in diverse species [14], [15], [107]. In this study, an advanced ensemble model was established to identify cyclins. In the proposed model, protein sequences were encoded by using AAC, CKSAAP, PseAAC, TPC, C/T/D, GD, and NMBroto. Then, these encoding-features were optimized by using ANOVA and mRMR with IFS technique. On the basis of top feature subset, the finest sorting model was achieved by the gradient boost decision tree classifier using five-fold CV test. The estimated outcomes on training data showed that the proposed model provided outstanding generalization capability. The data and codes are also available in the Supplementary file 2. Further studies will aim to create a user-friendly web server for the projected model. Also, additional feature selection methods and algorithms will be implemented to further improve the efficiency to classify cyclins [108], [109], [110], [111], [112], [113], [114], [115], [116], [117].

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: