PREDICTION AND CLASSIFICATION OF DNA BINDING PROTEINS INTO FOUR MAJOR CLASSES BASED ON SIMPLE SEQUENCE DERIVED
FEATURES USING ANN
AMIYA KUMAR PATEL*, SEEMA PATELa, PRADEEP KUMAR NAIKb Division of Biotechnology, Majhighariani Institute of Technology and Science (MITS), At- Sriram Vihar, Bhujbala, Po- Kolnara, Rayagada, (Pin – 765017), Orissa, India
aDepartment of Computer Science, Sundargarh Engineering College, At/Po- Kirei, Distt.-Sundargarh, (Pin- 770073), Orissa, India.
bDepartment of Biotechnology and Bioinformatics, Jaypee University of Information Technology, Waknaghat, Distt.-Solan, (Pin- 173 215), Himachal Pradesh, India
The problem of predicting the different classes of DNA binding protein from the protein sequence information is still an open problem in bioinformatics. We implemented a two- layered artificial neural network (ANN) of predicting the DNA binding proteins and their classification into four major classes from their amino-acid sequences. Using 61 sequence derived features we are able to achieve 72.99% correct prediction of proteins into DNA binding/non-DNA binding (in the dataset of 1000 proteins). For the complete set of 61 parameters using 5-fold cross-validated classification, ANN model revealed a superior model (accuracy = 72.99 ± 6.86%, Qpred = 73.952 ± 13.12%, sensitivity = 81.53 ± 6.73%
and specificity = 72.54 ± 6.39%). The classification accuracy for predicted DNA binding protein into four sub-classes was 70.73% (on average) using five fold cross validation, indicating that multi-class ANN classification system (61-11-4) may have certain level of unique prediction capability.
(Received February 14, 2010, accepted March 4, 2010)
Keywords: DNA binding proteins, Classification, Artificial neural network, Sequence derived features
1. Introduction
The prediction of protein structure from amino acid sequence has become the Holy Grail of computational molecular biology. The information necessary for protein folding resides completely within the primary structure; molecular biologists have been fascinated with the possibility of obtaining a complete three-dimensional picture of a protein by simply applying the proper algorithm to a known amino acid sequence [1]. The development of rapid methods of DNA sequencing coupled with the straightforward translation of the genetic code into protein sequences has amplified the urgent need for automated methods of interpreting these one-dimensional, linear sequences in terms of three-dimensional structure and function. Advanced and specialized databases are needed to facilitate the retrieval of relevant information from the deluge of sequence data and to provide insight into the protein structure and function. Further, it is clear that rational classification of proteins encoded in sequenced genomes is critical for making the genome sequences maximally useful for functional and evolutionary studies [2].
The family of DNA binding proteins is one of the most populated and studied amongst the various genomes of bacteria, archea and eukaryotes. Most of these proteins, such as the eukaryotic
*Corresponding author: [email protected]
and prokaryotic transcription factors, contain independently folded units (domains) in order to accomplish their recognition with the contours of DNA. It is now clear that the majority of these DNA-binding scaffolds which are in general relatively small, less than 100 amino acid residues, belong to a large number of structural families with characteristic sequences and three-dimensional designs or conformations [3]. Computational biology applying fast and sensitive algorithms strives to extract the maximum possible information from these sequences by classifying them according to their homologous relationships, predicting their likely biochemical activities and/or cellular functions, three-dimensional structures and evolutionary origin. There have been studies to detect [4, 5], design [6] and predict them using a probabilistic recognition code [7]. There have also been works towards analyzing protein–DNA recognition mechanism [8] and binding site discovery [9].
DNA binding proteins represent a broad category of proteins, known to be highly diverse in sequence and structure. Structurally, they have been divided into 54 protein-structural families [10]. With such a high degree of variance, using conventional annotation methods rooted in database searching for sequence similarity [11], profile or motif similarity [12] and phylogenetic profiles [13] may not lead to reliable annotations. In this context, a DNA binding protein prediction protocol that takes into account the structural information and does not depend on sequential or structural homology to proteins with known functions will be very useful.
Previously, there have been a few bioinformatics methods developed towards automated identification and prediction of DNA binding proteins. The pseudo-amino acid composition is used to identify proteins that bind to RNA, rRNA and DNA [14]. Structural information was integrated with the neural network approach for the prediction of DNA binding proteins [15].
Electrostatic features of proteins were also characterized through an automated approach for DNA binding protein and DNA binding site prediction [16, 17]. Further, the overall charge and electric moment can be used to identify DNA binding proteins [18]. Accuracy rates achieved in these methods varied from 65% to 86% depending on both the features used and the validation method adopted.
Strategically, we have used a neural network, two-layer, fully automated computational method capable of recognizing DNA binding proteins first, and then classifying them into their different classes based on their sequences derived features.
2. Methodology
Data set for prediction of DNA binding/non-DNA binding
A dataset of 500 DNA binding protein sequences were extracted from PDB. A non- redundant treatment was applied to eliminate the sequences which share a high degree of similarity (>90%) with others in order to avoid overtraining. The treatment was carried out using the program BLASTCLUST (http://www.ncbi.nlm.nih.gov/BLAST/), which used the BLAST algorithm to systematically cluster protein sequences on the basis of pair-wise matches. The default values were used for all BLAST parameters: matrix BLOSUM62, gap opening cost of 11, gap extension cost of 1, E-value threshold of 1e-6. These sequences were used as positive examples for prediction as DNA binding proteins. The sequences data on negative examples were obtained from the SWISSPROT database (http://expasy.org/sprot/). DNA binding proteins were removed from the original dataset. A non-redundant treatment was applied (same as for positive datasets) such that no sequence had similarity higher than 25% to any others. Thus, 500 non-DNA binding sequences were optimized as negative examples.
Dataset for classification of DNA binding proteins into four major classes
The above mentioned 500 protein sequences of DNA binding protein were then grouped into four major classes: class 1 (Homeo box domain) consist of 125 sequences, class II (Zinc finger) consist of 125 sequences, class III (Leucine zipper) having 125 sequences and class IV (Helix-Turn-Helix) with 125 sequences. They were used for construction of neural networks
training and validating the model for classification of predicted DNA binding proteins into four classes.
Neural network architecture
The implementation of ANN was realized using the software package SNNS (Version- 4.2) from Stuttgart University [19]. We have used two feed-forward back-propagation neural networks with a single hidden layer. First layer of neural network is used for prediction of DNA binding/non-DNA binding proteins from the protein sequence, whereas, the second layer is used for classifying the predicted DNA binding protein into out of four major classes. The 1st neural network consisting of 61 inputs, 7 hidden nodes and 1 output node. The number of nodes in the hidden layer was varied from 1 to 11 in order to find the optimal network that allows most accurate separation of DNA binding and non-DNA binding proteins in the training sets. The 2nd neural network consisting of 61 inputs, 11 hidden nodes and four output nodes (each node is specified for each class of DNA binding protein) (Figure 1). The number of nodes in the hidden layer was varied from 1 to 15 in order to find the optimal network that allows most accurate classification of DNA binding protein in the training sets. For each sequence in the training and testing sets, we have transformed 61 network input parameters into the normalized values varying from 0 to 1. Similarly, the output parameters from the ANN were in the range of 0 to 1. During the learning phase, a value of 1 was assigned for the DNA binding protein and 0 for non-DNA binding. For configuration of the ANN, 100 independent training runs were performed to evaluate the average predictive power of the network. The corresponding counts of the false/true positive and negative predictions were estimated using 0.1 and 0.9 cut-off values for non-DNA binding and DNA binding proteins respectively. Thus, a protein sequence from the testing set was considered correctly predicted as DNA binding protein by the ANN only when its output value ranged from 0.9 to 1.0. For each non-DNA binding protein of the testing set the correct prediction was assumed if the corresponding ANN output lies between 0 and 0.1. Thus, all network output values ranging from 0.2 to 0.9 have been ultimately considered as incorrect predictions (rather than undetermined or non-defined). If the input protein sequence is predicted as enzyme than it is parsed into the second layer and is classify into its particular class based on the maximum value obtained from the defined out put node for each class. For example to classify the predicted DNA binding into class 1 (Homeo box) the predicted output value is 1, 0, 0, 0 and so on. The input to second filtering network is the same input values used for the first layer.
Fig. 1. Configuration of artificial neural network used to develop binary primary sequence descriptor model for DNA binding/non-DNA binding proteins.
Sequence derived parameters calculation
A set of 61 parameters were calculated from the protein sequence alone using PEPSTAT (EMBOSS suite) ftp://emboss.open-bio.org/pub/EMBOSS [20] for all 1000 protein sequences.
The average values of these 61 parameters were independently calculated for DNA binding and non-DNA binding proteins as well as for each class of DNA binding protein and used as input values to the ANN model.
Fivefold cross-validation
A prediction method is often developed by cross-validation or jack-knife method [21].
Because of the size of the dataset, the jack-knife method (individual testing of each enzyme in the data set) was not feasible. So a more limited cross-validation technique has been used, in which the dataset is randomly divided into five subsets, each containing equal number of DNA binding proteins. Each set is a balanced set that consist of 50 percent of DNA binding and 50 percent non- DNA binding proteins. The data set has been divided into training and testing set. The training set consists of five subsets. The network is validated for minimum error on testing set to calculate the performance measure for each fold of validation. This has been done five times to test for each subset. The final prediction results have been averaged over five testing sets.
Table 1. 61 ‘Pepstat(EMBOSS)’ primary sequence descriptors used in the study.
DNA binding Non-DNA
binding DNA binding Non-DNA
binding Sequence derived
parameters
Max Min Max Min
Sequence derived
parameters Max Min Max Min
Molecular Weight 0.207588 0.00182 0.20947 0.00419 N_Mole % 0.7186 0.1200 0.9091 0.2300 Average Residue 0.11811 0.09159 0.1209 0.09186 N_DayhoffStat 0.1671 0.0987 0.2114 0.1078 Isoelectric Point 0.104656 0.0427 0.1288 0.03857 P_Mole % 0.9572 0.3450 3.6556 0.5680 Extinction
Coefficient 0.29032 0.019 0.33257 0.027 P_DayhoffStat 0.1841 0.0089 0.703 0.02908 Extinction
Coefficient (1 mg/ml)
0.275 0.024 0.376 0.036 Q_Mole % 0.585 0.0871 1.5106 0.1098 Improablity /
Proability inclusion bodies
0.928 0.494 0.979 0.41 Q_DayhoffStat 0.15 0.0098 0.3873 0.0129 A_Mole % 0.18828 0.02881 0.21186 0.03 R_Mole % 1.0682 0.0088 2.1256 0.0187 A_DayhoffStat 0.2189 0.0335 0.2464 0.045 R_DayhoffStat 0.218 0.02389 0.434 0.0452 B_Mole % 0.1989 0.0017 0.0902 0.0011 S_Mole % 0.9035 0.1796 2.2034 0.0012 B_DayhoffStat 0.0292 0.001 0.0109 0.0009 S_DayhoffStat 0.1291 0.0257 0.3148 0.0389 C_Mole % 1 0.00659 2.0339 0.0089 T_Mole % 1.0497 0.3091 1.4352 0.1203 C_DayhoffStat 0.3448 0.02154 0.7013 0.0154 T_DayhoffStat 0.1721 0.0507 0.2353 0.0092 D_Mole % 0.8147 0.0154 1.206 0.0015 V_Mole % 0.15 0.04484 0.17647 0.0289 D_DayhoffStat 0.1481 0.0152 0.2193 0.0652 V_DayhoffStat 0.2273 0.0679 0.2674 0.0546 E_Mole % 1.018 0.0147 1.8615 0.0254
E_DayhoffStat 0.1697 0.0215 0.3102 0.0145 W_Mole % 0.4598 0.00245 0.4839 0.0254 F_Mole % 0.9195 0.1277 1.0044 0.0596 W_DayhoffStat 0.3537 0.0021 0.3722 0.0215 F_DayhoffStat 0.2554 0.0355 0.279 0.0101
G_Mole % 0.25 0.00769 0.36923 0.00503 X_Mole % 0.4562 0.025 0.3262 0.0254 G_DayhoffStat 0.2976 0.0092 0.4396 0.006 X_DayhoffStat 0.5263 0.0562 0.3215 0.025 H_Mole % 0.6513 0.00894 1.0271 0.021
H_DayhoffStat 0.3257 0.0456 0.5136 0.0598 Y_Mole % 0.6135 0.0159 2.4615 0.0521 I_Mole % 1 0.2077 1.0377 0.0089 Y_DayhoffStat 0.1804 0.0154 0.724 0.00987 I_DayhoffStat 0.2222 0.0462 0.2306 0.0564
K_Mole % 1.018 0.0591 2.0455 0.00115 Z_Mole % 0.2222 0.0089 0.3262 0.0154 K_DayhoffStat 0.1542 0.00213 0.3099 0.0002 Z_DayhoffStat 0.894 0.1256 0.265 0.03652 L_Mole % 0.19444 0.03139 0.19101 0.0321 Tiny Mole % 0.6 0.15569 0.6389 0.16239 L_DayhoffStat 0.2628 0.0424 0.2581 0.0021 Small Mole % 0.75 0.4012 0.77119 0.32479 M_Mole % 0.5169 0.0456 1.2346 0.0268 Aliphatic Mole
% 0.31481 0.14808 0.32903 0.02542 M_DayhoffStat 0.3041 0.0154 0.7262 0.0158 Aromatic Mole
% 0.24521 0.04918 0.29231 0.08541 Charged Mole % 0.33533 0.05 0.46986 0.01389 Non-polar
Mole % 0.85 0.45521 0.86154 0.31818 Basic Mole % 0.17365 0.05 0.31624 0.00926 Polar Mole % 0.54479 0.15 0.68182 0.13846 Acidic Mole % 0.16168 0.00897 0.25 0.0154
All the results reported for 2nd layer of ANN were obtained by performing a modified five- fold cross-validation procedure [22]. First, a given number of proteins (80) were randomly drawn from the dataset for each of the four families. The sum of these samples constituted the training set (320). All the other proteins were allocated to the evaluation set (180). Then, the neural network was trained and later evaluated using this partition. The accuracy rate on the evaluation set was
computed as the ratio of the number of correctly classified proteins to the total number of proteins, as is standard in the literature. Next, a new sampling was taken from the dataset to form another training set and evaluation set, and the training and evaluation process were repeated. This procedure was repeated five times and the final results were reported as the averaged accuracy rate over these five runs.
Table 2. 61 ‘Pepstat (EMBOSS)’ primary sequence descriptors used in the study.
Class 1 Homeo box
Class 2 Zinc finger
Class 3 Leucine zipper
Class 4 Helix-turn-helix Parameters
Max Min Max Min Max Min Max Min Mol. Weight 0.208 0.002 0.177 0.015 0.067 0.004 0.045 0.038 Average Residue 0.118 0.091 0.120 0.101 0.117 0.104 0.116 0.104 Isoelectric Point 0.105 0.043 0.110 0.046 0.086 0.045 0.101 0.045 Extinction Coefficient 0.290 0.016 0.181 0.001 0.085 0.004 0.067 0.020 Extinction Coefficient
(1 mg/ml) 0.275 0.017 0.225 0.006 0.172 0.031 0.172 0.048 Improablity/Proability
inclusion bodies 0.928 0.494 0.881 0.497 0.871 0.497 0.848 0.503 A_Mole % 0.188 0.029 0.241 0.027 0.161 0.028 0.160 0.022 A_DayhoffStat 0.219 0.034 0.280 0.032 0.187 0.032 0.186 0.025 B_Mole % 0.313 0.027 0.164 0.034 0.133 0.098 0.920 0.013 B_DayhoffStat 0.165 0.010 1.132 0.264 0.455 0.027 0.255 0.001 C_Mole % 1.000 0.092 0.345 0.013 0.400 0.024 0.299 0.049 C_DayhoffStat 0.345 0.022 0.119 0.001 0.138 0.028 0.103 0.017 D_Mole % 0.815 0.062 0.735 0.197 0.903 0.175 0.818 0.279 D_DayhoffStat 0.148 0.025 0.134 0.036 0.164 0.032 0.149 0.051 E_Mole % 1.018 0.013 1.310 0.317 1.132 0.416 0.938 0.299 E_DayhoffStat 0.170 0.001 0.218 0.053 0.189 0.069 0.156 0.050 F_Mole % 0.920 0.128 0.725 0.098 0.556 0.013 0.591 0.166 F_DayhoffStat 0.255 0.036 0.201 0.027 0.154 0.001 0.164 0.046 G_Mole % 0.250 0.008 0.112 0.024 0.100 0.053 0.118 0.050 G_DayhoffStat 0.298 0.009 0.133 0.028 0.119 0.063 0.141 0.059 H_Mole % 0.651 0.012 0.455 0.049 0.635 0.160 0.609 0.134 H_DayhoffStat 0.326 0.065 0.227 0.017 0.318 0.080 0.305 0.067 I_Mole % 1.000 0.208 1.215 0.117 1.148 0.172 1.317 0.307 I_DayhoffStat 0.222 0.046 0.270 0.026 0.255 0.038 0.293 0.068 K_Mole % 1.018 0.032 1.089 0.110 0.833 0.080 1.532 0.136 K_DayhoffStat 0.154 0.065 0.165 0.017 0.126 0.012 0.232 0.021 L_Mole % 0.194 0.031 0.167 0.036 0.139 0.034 0.140 0.060 L_DayhoffStat 0.263 0.042 0.226 0.049 0.188 0.047 0.189 0.081 M_Mole % 0.517 0.015 0.448 0.046 0.556 0.248 0.365 0.103 M_DayhoffStat 0.304 0.081 0.264 0.027 0.327 0.069 0.215 0.061 N_Mole % 0.719 0.103 0.611 0.045 0.862 0.258 0.887 0.140 N_DayhoffStat 0.167 0.061 0.142 0.495 0.201 0.060 0.206 0.033 P_Mole % 0.957 0.140 0.840 0.164 0.862 0.160 0.679 0.166 P_DayhoffStat 0.184 0.033 0.162 0.032 0.166 0.031 0.131 0.032 Q_Mole % 0.585 0.166 0.817 0.068 0.874 0.013 0.855 0.059 Q_DayhoffStat 0.150 0.002 0.210 0.136 0.224 0.001 0.219 0.015 R_Mole % 1.068 0.024 1.525 0.193 0.774 0.214 0.868 0.134 R_DayhoffStat 0.218 0.048 0.311 0.040 0.158 0.044 0.177 0.027 S_Mole % 0.904 0.180 1.250 0.280 0.935 0.248 0.893 0.357 S_DayhoffStat 0.129 0.026 0.179 0.040 0.134 0.069 0.128 0.051
Class 1 Homeo box
Class 2 Zinc finger
Class 3 Leucine zipper
Class 4 Helix-turn-helix Parameters
Max Min Max Min Max Min Max Min T_Mole % 1.050 0.309 0.851 0.117 0.874 0.175 0.867 0.224 T_DayhoffStat 0.172 0.051 0.140 0.019 0.143 0.029 0.142 0.037 V_Mole % 0.150 0.045 0.129 0.038 0.152 0.041 0.105 0.040 V_DayhoffStat 0.227 0.068 0.196 0.058 0.230 0.062 0.160 0.061 W_Mole % 0.460 0.128 0.364 0.351 0.255 0.070 0.253 0.033 W_DayhoffStat 0.354 0.867 0.280 0.186 0.196 0.218 0.195 0.166 X_Mole % 0.512 0.142 0.957 0.097 0.519 0.051 0.853 0.002 X_DayhoffStat 0.265 0.105 0.184 0.076 0.332 0.148 0.198 0.024 Y_Mole % 0.614 0.160 0.597 0.032 1.035 0.254 0.544 0.207 Y_DayhoffStat 0.180 0.253 0.176 0.010 0.304 0.075 0.160 0.061 Z_Mole % 0.155 0.195 0.545 0.065 0.519 0.306 0.519 0.006 Z_DayhoffStat 1.231 0.022 0.335 0.208 0.332 0.044 0.332 0.497 Tiny Mole % 0.600 0.156 0.408 0.173 0.364 0.139 0.387 0.175 Small Mole % 0.750 0.401 0.519 0.387 0.597 0.389 0.590 0.368 Aliphatic Mole % 0.315 0.148 0.332 0.165 0.306 0.172 0.276 0.207 Aromatic Mole % 0.245 0.049 0.175 0.039 0.167 0.076 0.166 0.071 Non-polar Mole % 0.850 0.455 0.688 0.460 0.649 0.525 0.659 0.512 Polar Mole % 0.545 0.150 0.540 0.312 0.475 0.351 0.488 0.341 Charged Mole % 0.335 0.050 0.344 0.170 0.278 0.186 0.323 0.168 Basic Mole % 0.174 0.050 0.202 0.084 0.143 0.097 0.188 0.090 Acidic Mole % 0.162 0.000 0.173 0.057 0.151 0.076 0.147 0.071
Performance measures
The prediction results of 1st layer of ANN model developed in the study were evaluated using the following statistical measures.
1. Accuracy of the methods: The accuracy of prediction for neural network models were calculated as follows:
T N
QACC = P+ , where T = (P+N+O+U)
Where P and N refer to correctly predicted DNA binding and non-DNA binding proteins, and O and U refer to over and under predictions, respectively.
2. The Matthews correlation coefficient (MCC) is defined as:
( ) ( )
(
P U) (
P O) (
N U) (
N O)
U O N MCC P
+
× +
× +
× +
×
−
= ×
3. Sensitivity (Qsens) and specificity (Qspec) of the prediction methods are defined as:
U P Qsens P
= + O N Qspec N
= + 4. QPred (Probability of correct prediction) is defined as:
×100
= + O P Qpred P
3. Results and discussion
The 1st layer of ANN model develop in this study (61-7-1) is trained with the sequence derived features (61 parameters) calculated using PEPSTAT. The number of nodes in the hidden layer was varied from 1 to 11 in order to find the optimal network that allows most accurate separation of DNA binding/non-DNA binding proteins in the training sets (Table 3). When applying a fivefold cross-validation test using five data sets, we found that the network reached an overall accuracy of 72.99 ± 6.86%. The prediction results are presented in Table 4. The other performance measures were: Qpred = 73.95 ±13.12%, sensitivity = 81.53 ± 6.73% and specificity
= 72.54 ± 6.39%. The value of the learning parameter was set to 0.1.
Table 3. Parameters of specificity, sensitivity, accuracy and positive predictive values for prediction of DNA binding and non-DNA binding from the protein sequence by the 1st layer of artificial neural networks with the varying number of hidden nodes. The cut-off values of 0.1 and 0.9 have been used for negative and positive predictions respectively.
Hidden Nodes Accuracy Specificity Sensitivity Q(Pred)
1 0.5869 0.6523 0.7423 65.23
3 0.6213 0.6452 0.5013 72.13
5 0.5522 0.5864 0.5123 55.23
7 0.6976 0.6878 0.7535 68.32
9 0.6435 0.6020 0.7632 65.18
11 0.6235 0.6425 0.7123 69.25
Table 4. Performance measure of 1st neural network for the prediction of DNA binding/non-DNA binding proteins using five fold cross validation based on sequence derived features.
Fivefold cross validation
Accuracy Specificity Sensitivity Q(Pred) Prediction range (DNA binding
Prediction range (Non-DNA
binding) C1 0.8.20 0.8632 0.7271 85.12 0.6726 – 1.00 0.00 – 0.5240 C2 0.7430 0.7791 0.8580 70.61 0.5079 – 1.00 0.00 – 0.5658 C3 0.7002 0.6024 0.8001 71.61 0.4257 – 1.00 0.00 – 0.5386 C4 0.7140 0.6567 0.8901 62.28 0.3592 – 1.00 0.00 – 0.6486 C5 0.6906 0.7259 0.8015 80.14 0.4748 – 1.00 0.00 – 0.5836 Mean 0.7299 ±
0.0686
0.7254 ± 0.0639
0.8153 ± 0.0673
73.952 ± 13.123
By applying a modified fivefold cross-validation test using five data sets, we found that the second layer of network (61-11-4) is a superior model for classification of predicted DNA binding proteins into their suitable classes. The number of nodes in the hidden layer was varied from 1 to 15 in order to find the optimal network that allows most accurate classification system of DNA binding proteins in the training sets (Table 5). Out of 500 DNA binding proteins (125 proteins from each class) in each cross validation set 220 to 343 DNA binding proteins were correctly classified. However, the network was more efficiently classify the proteins belonging to Leucine zipper and Helix-turn-helix in compare to other classes (Table 6). The classification accuracy for DNA binding proteins from 4 families is in the range of 73.34% to 80.06% using five fold cross validation with an overall accuracy of 76.74% using a sequence derived features, indicating that multi-class ANN classification system (61-11-4) may have certain level of unique prediction capability.
Table 5. Result for classification of DNA binding proteins into four major classes using 2nd neural network based on protein sequence derived features with the varying number of hidden nodes.
Number of hidden
nodes
Number of DNA binding protein
taken
Correctly predicted DNA binding protein Total
Homeo
box
Zinc finger Leucine zipper
Helix-turn- helix
1 500 65 74 61 65 265
3 500 56 61 51 52 220
5 500 55 70 61 65 251
7 500 65 80 75 78 298
11 500 86 87 81 89 343
15 500 84 83 77 81 325
Table 6. Predicted result of 2nd layer of neural network for classification of DNA binding proteins into corresponding classes using the fivefold cross validation sets.
Accuracy rate (%) using five fold cross validation Family
1 2 3 4 5 Mean ± sd
Homeo box 66.7 70.8 66.7 79.2 83.3 73.34 ± 7.55 Zinc finger 69.2 73.3 78.3 71.7 82.5 75.0 ± 5.35 Leucine zipper 85.0 83.3 80.8 80.0 71.2 80.06 ± 5.34 Helix-turn-helix 70.8 73.2 80.6 85.8 82.4 78.56 ± 6.33
Average 76.74 ± 6.14
The classes of newly found DNA binding proteins are usually determined either by biochemical analysis of eukaryotic and prokaryotic genomes or by microarray chips. These experimental methods areboth time-consuming and costly. With the explosion of proteinentries in databanks, we are challenged to develop an automatedmethod to quickly and accurately determine the enzymatic attributefor a newly found protein sequence: is it a DNA binding or a non-DNA binding protein? If it is, to which class does it belongs? The answers to these questions are important because they may help deduce the mechanism and specificity of the query protein, providing clues to the relevant biological function. Although it is an extremely complicated problem and might involvethe knowledge of three-dimensional structure as well as manyother physicochemical factors, some quite encouraging resultshave been obtained by a bioinformatical method established on the basis of amino acid composition alone [23]. Since the amino acid composition of a protein does not containany of its sequence-order information, a logical step to furtherimprove the method is to incorporate the sequence-order informationinto the predictor. To realize this, the most straightforward way is to represent the sample of a protein by its entire sequence,the so-called sequential form.
The results demonstrate that the developed ANN-based model for binary prediction of DNA binding/non-DNA binding proteins and classification of predicted DNA binding proteins into four major classes is adequate and can be considered an effective tool for ‘in silico’ screening.
The results also demonstrated that the sequence derived parameters readily accessible from the protein sequences only, can produce a variety of useful information to be used ‘in silico’; clearly demonstrates an adequacy and good predictive power of the developed ANN model. There is strong evidence, that the introduced sequence features do adequately reflect the structural properties of proteins. The structure of a protein is an important determinant for the detailed molecular function of proteins, and would consequently also be useful for prediction of DNA binding proteins and for their classification. This observation is not surprising considering that the calculated parameters should cover a very broad range of proprieties of bound atoms and
molecules related to their size, polarizability, electronegativity, compactness, mutual inductive and steric influence and distribution of electronic density, etc. As it can be seen that the average value for different classes of DNA binding proteins were clearly separated (Table 1 & 2) and, hence, the selected 61 parameters should allow building an effective ANN model for binary prediction as well as their classification further.
Presumably, accuracy of the approach operating by the sequence derived features can be improved even further by expanding the parameters or by applying more powerful classification techniques such as Support Vector Machines or Bayesian Neural Networks. Use of merely statistical techniques in conjunction with the sequence parameters would also be beneficial, as they will allow interpreting individual parameter contributions into “DNA binding/non-DNA binding- likeness”.
References
[1] C. B. Anfinsen, Science 181, 223 (1973).
[2] C. Wu, M. Berry, S. Shivakumar, J. McLarty, J. Mach. Learn. 21 N(1-2), 177 (1995).
[3] C. Branden, J. Tooze, Introduction to Protein Structure, Garland Publishing Co., New York (1991).
[4] S. F. Altschul, W. Gish, W. Miller, E. W. Myers, D. J. Lipman, J. Mol Biol. 215, 403 (1990).
[5] S. F. Altschul, T. L. Madden, A. A. Schaffer, J. Zhang, Z. Zhang, W. Miller, D. J. Lipman, Nucleic Acids Res. 25, 3389 (1997).
[6] A. Krogh, M. Brown, I. S. Mian, K. Sjolander, D. Haussler, J. Mol. Biol. 235, 1501 (1994).
[7] S. Vinga, J. Almeida, Bioinformatics 19, 513 (2003).
[8] J. K. Vries, R. Munshi, D. Tobi, J. Klein-Seetharaman, P. V. Benos, I. Bahar, Appl.
Bioinformatics 3, 137 (2004).
[9] K. Harpreet, G. P. S. Raghava, Protein Science 12, 923 (2003).
[10] E. C. Webb, Enzyme Nomenclature, Academic Press, San Diego CA, (1992).
[11] A. Narayanan, E. C. Keedwell, B. Olsson, Appl. Bioinformatics 1(4), 191 (2002).
[12] M. Bhasin, G. P. S. Raghava, Nucleic Acids Research 32, W383 (2004a).
[13] M. Bhasin, G. P. S. Raghava, Nucleic Acids Research 32, W414 (2004b).
[14] Y. Cai, S. L. Lin, Biochimica et Biophysica Acta. 1648, 127 (2003).
[15] S. Ahmad, M. M. Gromiha, A. Sarai, Bioinformatics 20, 477 (2004).
[16] E. W. Stawiski, L. M. Gregoret, Y. Mandel-Gutfreund, J. Mol. Biol. 326, 1065 (2003).
[17] S. Jones, P. Van Heyningen, H. M. Berman, J. M. Thornton, J. Mol. Biol. 287, 877 (1999).
[18] S. Ahmad, A. Sarai, J. Mol. Biol. 341, 65 (2004).
[19] A. Zell, G. Mamier, Stuggart Neural Network Simulator (Version- 4.2), University of Stuttgart, Stuttgart, Germany (1997).
[20] P. Rice, I. Longden, A. Bleasby, Trends in Genetics 16 (6), 276 (2000).
[21] K. C. Chou, C. T. Zhang, Crit. Rev. Biochem. Mol. Biol. 30, 275 (1995).
[22] D. J. Hand, ‘Construction and assessment of classification rules’. New York: John Wiley and Sons (1997).
[23] K. C. Chou, D. W. Elrod, J. Proteome Res. 2, 183 (2003).
