Rule-Based Knowledge Acquisition Method for Promoter Prediction in Human and Drosophila Species

The rapid and reliable identification of promoter regions is important when the number of genomes to be sequenced is increasing very speedily. Various methods have been developed but few methods investigate the effectiveness of sequence-based features in promoter prediction. This study proposes a knowledge acquisition method (named PromHD) based on if-then rules for promoter prediction in human and Drosophila species. PromHD utilizes an effective feature-mining algorithm and a reference feature set of 167 DNA sequence descriptors (DNASDs), comprising three descriptors of physicochemical properties (absorption maxima, molecular weight, and molar absorption coefficient), 128 top-ranked descriptors of 4-mer motifs, and 36 global sequence descriptors. PromHD identifies two feature subsets with 99 and 74 DNASDs and yields test accuracies of 96.4% and 97.5% in human and Drosophila species, respectively. Based on the 99- and 74-dimensional feature vectors, PromHD generates several if-then rules by using the decision tree mechanism for promoter prediction. The top-ranked informative rules with high certainty grades reveal that the global sequence descriptor, the length of nucleotide A at the first position of the sequence, and two physicochemical properties, absorption maxima and molecular weight, are effective in distinguishing promoters from non-promoters in human and Drosophila species, respectively.


Introduction
Gene expression is often regulated by the transcription rate, which is largely controlled by the binding of RNA polymerase II (Pol II) to the regulatory regions of DNA sequences in eukaryotic cells [1]. The regulatory regions (called promoters) that contain a transcription factor binding site and a TATA box are immediately upstream of transcription start sites at which transcription factors and Pol II are accumulated to initiate the transcription ( Figure 1) [2,3]. Promoters are extremely diverse and difficult to identify experimentally using specific sequence patterns or motifs [3,4]. Therefore, the identification of promoters is very challenging, especially in the sequencing of eukaryotic genomes. Some methods for predicting promoters have been developed, and these methods may be categorized into the following four classes according their types of sequence features (see Table 1).
(2) Signal-Feature Class. Powerful biological signals contain core-promoter elements [11], some short modular transcription factor binding sites [12], and CpG islands [13,14]. The core-promoter elements that play important roles in the assembly of transcriptional  Figure 1: The promoter of a DNA sequence containing a transcription factor binding site and a TATA box is immediately upstream to a transcription start site. machinery contain the TATA box, the exons region [15,16], the intron region (initiator sequences), [15,[17][18][19], downstream promoter elements [20], a TFIIB recognition element [11], motif ten element [21], and CCAAT box [19,22,23]. The TATA box, initiator sequences, and consensus sequences for transcription factor binding sites are often used in various promoter recognition methods [24]. However, these features have been confirmed to exist only in a small proportion of all human promoters [25,26].
(4) Epigenetic-Feature Class. Few promoter prediction methods utilize epigenetic information [22,33]. For example, HMM-SA is a supervised learning method for predicting promoters and enhancers from their unique chromatin modification signatures [33]. Similarly, CoreBoost HM systematically analyzes different chromatin features for promoter prediction [22]. Table 2 lists some representative methods that use the above four types of features in combination with effective classifiers to predict promoters. These classifiers involve Fisher's linear discriminant algorithm [6], the hidden Markov model [22,23], the AdaBoost algorithm [8], decision trees [18], relevance vector machines [34], the expectation maximization algorithm [35,36], artificial neural networks [12,15,17,19,29,37], support vector machines (SVM) [38,39], artificial immune recognition systems [40], and others. Recently, the use of ensemble classifiers has become popular in promoter prediction systems. For example, CoreBoost [23] and CoreBoost HM [22] used ensemble classifiers to improve prediction performance. These methods apply boosting techniques with stumps to extract sequence features, including the core-promoter elements score, transcription factor binding site density, the DNA flexibility of promoter sequences, Markovian log-likelihood ratio scores, k-mer frequencies, and epigenetic features.
Even though many promoter prediction methods have been developed, the effectiveness of used features in identifying promoters still needs to be explored. However, accurate promoter prediction relies largely on feature extraction and model selection [3]. Currently, Prom-Machine [39] simply selected 128 four-mer motifs and then utilized these motifs in conjunction with SVM to improve prediction sensitivity and specificity toward the DNA sequences of the five following species: plants, Drosophila, human, mouse, and rat. Additionally, one recently published method, FSPP [41], used both filter and wrapper algorithms to select effective feature subsets from 13 kinds of structural features, including DNAbending stiffness, duplex free energy, and duplex disrupt energy to improve further the sensitivity and accuracy of promoter prediction. Our previous method, ProPolyII [42], selects a small number of sequence-based features to improve prediction performance in human species.
These methods motivate this work to focus on feature selection and effectiveness evaluation of the selected features in promoter prediction. This work presents a knowledge acquisition method (named PromHD) based on if-then rules for promoter prediction in human and Drosophila species. The knowledge can be revealed from three aspects: (1) identified informative DNA sequence descriptors (DNASDs), (2) rules of distinguishing promoter from nonpromoter, and (3) further analysis of distinguishable mechanism using DNASDs. PromHD utilizes a reference feature set of 167 DNASDs, comprising three descriptors of physicochemical properties (absorption maxima, molecular weight, and molar absorption coefficient) [43,44] with 128 top-ranked frequency descriptors of 4-mer motifs and 36 global sequence descriptors. To the best of our knowledge, these three descriptors of physicochemical properties are used herein for the first time in identifying promoter DNA sequences and their sequence-based representation differs from the structural profiles of McPromoter [29]. The 128 top-ranked frequency descriptors of 4-mer motifs are extracted from 256 4-mer combinations of nucleotides (4-base-long nucleotide sequences) according to the scores that equal the difference The Scientific World Journal 3 between the occurrence frequencies of the 4-mer motif in the positive and negative datasets [45,46]. PromHD further utilizes an effective feature mining algorithm (called DNASDmining), which is based on an inheritable biobjective genetic algorithm [47,48], to mine informative DNASDs. A total of 1871 human and 1926 Drosophila promoter sequences were downloaded from the Eukaryotic Promoter Database [49], which is a database containing over 4800 promoters from various species. The same numbers of nonpromoters in human and Drosophila species were collected to evaluate the proposed PromHD method. Accordingly, PromHD identifies two subsets of 99 and 74 DNASDs and yields test accuracies of 96.4% and 97.5% in human and Drosophila species, respectively, which are better than those of SVM-4mer (91.0% and 94.6%) and SVM-GSD (93.6% and 89.2%), respectively. Based on each of the 99-and 74dimensional feature vector, PromHD uses the decision tree method C5.0 [50] to generate several if-then rules. The top-ranked rules reveal that the global sequence descriptor, the length of nucleotide A at the first position of the sequence, is efficient in distinguishing human promoters from nonpromoters, consistent with the findings of Wang et al. and Zhao et al. [22,23]. Alternatively, the top-ranked rules in Drosophila species reveal that two physicochemical properties, absorption maxima and molecular weight, are effective in distinguishing promoters from nonpromoters. Further analysis of the two feature subsets shows that 32 features are common including three physicochemical properties, 14 descriptors of 4-mer motifs, and 15 global sequence descriptors. When the three descriptors of physicochemical properties are excluded, PromHD with the remaining 96(= 99 − 3) and 71(= 74 − 3) DNASD features yield test accuracies of 94.4% and 95.5% in human and Drosophila test datasets, respectively. The prediction accuracies fall by 2.0% (= 96.4% − 94.4% and = 97.5% − 95.5%), reconfirming the three physicochemical properties are obviously effective in distinguishing promoters from nonpromoters in human and Drosophila species. The promoter prediction system by using the PromHD method has been implemented at http://iclab.life.nctu.edu.tw/promhd.

Materials and Methods
In this work, a block diagram is used to illustrate the main components of the proposed PromHD method. Figure 2 presents five main components, which are datasets, DNA sequence descriptors, DNASDmining algorithm, estimating appearance-frequency ratios, and the PromHD prediction system.

Physicochemical Properties of Nucleotides.
The physicochemical properties of the DNA structure of eukaryotic genomes are critical to promoter recognition. This study proposes a sequence-based set of three physicochemical properties of nucleotides (http://www.geneinfinity.org) for designing prediction features that are used to distinguish promoters from nonpromoters. The three DNA sequence descriptors, denoted as AM , MW , and MAC , are derived from the three physicochemical properties-absorption maxima, molecular weight, and molar absorption coefficient, respectively-by averaging over a nucleotide sequence [59]. The three descriptors are the attributes of the subvector [ 1 , 2 , 3 ] in the reference set of 167 DNASDs.
The sequence S20 has five As, seven Cs, four Gs, and four Ts. With reference to Table 3, the values of the absorption maxima (determined at pH 7.0) for nucleotides A, C, G, and T are 259, 271, 253, and 267, respectively. Accordingly, the descriptor AM has a value 263. 6

Global Sequence Descriptors.
The global description of promoter/nonpromoter sequences contains four parts, entropy density profile (EDP), composition, transition, and distribution of DNA nucleotides [6]. The EDP model is a global statistical description for a DNA sequence, based on Shannon's artificial linguistic description for a DNA sequence of finite length [60]. Let be the frequencies of occurrence of nucleotides in a promoter/nonpromoter sequence, where is the index that specifies the nucleotides (A, C, G, T). Six EDPs, EH , EQ , EA , EC , EG , and ET , correspond to the six attributes of the subvector [ 4 , . . . , 9 ], defined as follows: where EH is the Shannon entropy and EQ is a statistical quantity.
The composition is used to measure the frequency of occurrence of each kind of letters in the sequences, and thus herein it is the in (1). Additionally, the four frequencies are also called 1-mer motifs of the nucleotides (A, T, C, and G), denoted as C1 (A), C1 (T), C1 (C), and C1 (G), and correspond to the four attributes of [ 10 , . . . , 13 ].
The third part, transition ( , ), characterizes the percent frequency with which is followed by or is followed by . The six transition frequencies, The fourth part of the global description, distribution, measures the chain length within which the first, 25%, 50%, 75%, and 100% of certain type of letters are located, respectively. For example, for the S20 sequence, the first, 25%, 50%, 75%, and 100% of the nucleotide A are located within the second, 4th, 8th, 13th, and 16th nucleotides, respectively.

Frequency Descriptors of 4-Mer
Motifs. The number of 4-mer combinations of nucleotides (4-base-long nucleotide sequences) is 256. Prom-Machine [39] uses top 128 of the 256 4-mer motifs to improve prediction sensitivity and specificity. Our earlier works estimated the scores of amino acids [45] and of GO terms [46] for predicting DNA-binding proteins and nonclassical secretory proteins, respectively. That motives this work, in which a score for each 4-mer motif is calculated and the 128 top-ranked 4-mer motifs based on those scores are identified. The score is the difference between the occurrence frequencies. A detailed description follows. 6 The Scientific World Journal   Step 1. The occurrence frequencies and are those of the th 4-mer motif in all training promoter (positive) and nonpromoter (negative) sequences, respectively, where = 1, 2, . . . , 256. For example, TATA is the 199th 4-mer motif, that is, = 199, and its occurrence frequencies in the positive and negative classes of the HPL dataset are 199 = 84 and 199 = 453, respectively (see Table 4).
Step 2. Calculate the total numbers of occurrences of 256 4mer motifs in the positive and negative classes, Σ and Σ . For example, the total numbers of occurrences of 256 4-mer motifs in the positive and negative classes of the HPL dataset are 29104 and 137017, respectively. Step 4. The score for each 4-mer motif is the absolute value | ⋅ | of the difference between the proportional frequencies of occurrence in the positive class and that in the negative class. For example, the score of TATA is 0.000417 (= |0.00309 − 0.002892|).
Step 5. Normalize scores of all 256 4-mer motif into the range [0, 13000], and represent them as { 1 , 2 , . . . , }. The normalized score is also called the frequency descriptor of the 4-mer. For instance, the score of TATA motif is 111 (see Table 4).
The Scientific World Journal 7 Step 6. All 256 frequency descriptors are ranked in descending order. The top 128 motifs with descriptors are denoted as C4 (⋅) and they correspond to [ 40 , . . . , 167 ]. For example, the descriptor of the well-known TATA box, as shown in Table 4, is C4 (TATA) = 111 and corresponds to 137 for the DPL dataset.

Proposed DNASDmining
Algorithm. An efficient featuremining algorithm, DNASDmining, for identifying a set of informative DNASDs is developed. The DNASDmining algorithm is an expansive version of an inheritable biobjective genetic algorithm, which is based on an intelligent genetic algorithm (called IGA) [47,61], to identify a small number out of = 167 DNASDs. The feature selection is a combinatorial optimization problem Comb( , ) with a huge search space of size Comb( , ) = !/( !( − )!). The IGA, based on an orthogonal experimental design using a divide-and-conquer strategy and systematic reasoning, can efficiently solve the large combinatorial optimization problem to obtain the solution to Comb( , ). The mechanism can efficiently search for the next solution +1 to Comb( , + 1) by inheriting the last solution . DNASDmining obtains all solutions from = start to end one by one using IGA with the inheritable mechanism [47,61].  Figure 2 shows the sequence representation and the IGAchromosome encoding method. Supplementary Table S2 lists the parameter settings of IGA, such as population size pop = 20. In this algorithm DNASDmining, start = 30, end = 100, and max = 60 based on former experience.
Step 1 (initiation). Randomly generate an initial population of pop individuals. All the binary genes in the individual have 1's and − 0's where = start and gen = 0.
Step 2 (evaluation). Evaluate the fitness values fitness( ) of all individuals. The fitness function of this training model is the prediction accuracy of 10-fold cross-validation (see Evaluation Measures) using the SVM classifier with the DNASDs, , and by decoding the IGA-chromosome.
Step 3 (selection). Use the simple ranking selection that replaces the worst ⋅ pop individuals by the best ⋅ pop individuals to form a new population where is the selection probability.
Step 4 (crossover). Select ⋅ pop parents from the mating pool to perform orthogonal array crossover [47,61] on the selected pairs of parents where is the crossover probability.
Step 5 (mutation). Apply the swap mutation operator to the randomly selected ⋅ pop individuals in the new population where is the mutation probability. To prevent a decline in the best fitness value, mutation is not applied to the best individual.
Step 6 (termination test). If = max , then output the best individual as . Otherwise, increase the number gen by one, and go to Step 2.
Step 7 (inheritance). If < end , then randomly change one bit in the binary genes of each individual from 0 to 1; increase the number by one and let gen = 0, and go to Step 2.
Step 8 (decoding chromosome). Let be the most accurate solution with selected DNASDs among all solutions Obtain the informative features and values of the parameters and .
Step 9 (system uncertainty). Perform Steps 1-8 for independent runs to obtain the best solution, , and the associated parameter settings of the SVM classifier (see Section 2.4).

Evaluation Measures.
The independent dataset test, subsampling or -fold (e.g., 5-or 10-fold) cross-validation test, and the jackknife test are often used to examine the accuracy of a statistical prediction method [63]. The jackknife test is deemed to be the least arbitrary method that can always yield a unique result for a given benchmark dataset [64]. The -fold cross-validation test is used to estimate the error that is involved in the predictions and thus it is also used for model selection [65]. In this work, 10-fold cross-validation scheme is used. Additionally, the independent dataset test is also used in this work to avoid overestimating the success rate of the training model.
Overall accuracy (ACC), sensitivity (SN), and specificity (SP) are three quality measures that are widely used to evaluate the performance of promoter prediction methods [66]. This work also utilizes Matthews correlation coefficient (MCC) to measure the overall performance of the prediction models. It takes into account true and false positives and negatives and is generally regarded as a balanced measure which can be used even if the classes are of very different sizes [67]. These measures are defined as below: TP, TN, FP, and FP stand for true positive, true negative, false positive, and false negative, respectively. The MCC returns a value in the range [−1, 1]. A value of 1 indicates a perfect prediction; 0 indicates a random prediction, and −1 indicates an inverse prediction.

Estimating the Appearance-Frequency Ratios.
A total of experimental runs are executed to obtain the best solution in this work due to the system uncertainty of the IGA-based feature selection algorithm. That means that solutions are generated and each solution comprises a subset of selected DNASDs for = 1, . . . , . The best solution must have both high prediction accuracy and a high appearance-frequency value. The estimation procedure is further described below.
Step 4. Calculate the appearance-frequency ratio R = / and the mean value R . For instance, Figure 3 displays the mean R = 47.0% for the HPL dataset.
Step 5. Select the candidate solutions from the runs whose appearance-frequency ratios R are larger than the mean value R . For instance, the appearance-frequency ratio R for = 5,9,12,15,16,17,19, 20 exceeds the mean 47.0%, as shown by Figure 3. Thus, these eight solutions are selected as the candidate solutions.
Step 6. The best solution is the candidate solutions with the highest prediction accuracy. For the above example, the 5th candidate solution 5 having the highest accuracy 98.9% is selected as the best solution of the DNASDmining algorithm, where ( , ) = (2 7 , 2 −5 ),

PromHD Prediction
System. The PromHD prediction system is implemented by using a SVM classifier with a subset of m DNASDs, where the parameter settings of SVM and the value of m are determined in the training phase. Figure 2 illustrates the prediction flowchart of PromHD. The input to this prediction system is a query DNA sequence P. The output is the predicted classpromoter or nonpromoter. The prediction procedure is described as follows.
Step 2. The informative DNASDs are selected from P, where = 99 and 74 for human and Drosophila DNA sequences, respectively.
Step 3. The selected features are input to the trained SVM to classify P as a promoter or non-promoter.

Effectiveness of Informative
DNASDs. DNA sequences in this work are represented using 167-dimensional vectors of DNASDs. This work uses an efficient feature selection algorithm not only to select a subset of size from the 167 DNASDs but also to design a SVM-based classifier simultaneously. To determine the candidate solution in the DNASDmining algorithm, the prediction accuracy of 10-CV is used as a fitness function of the IGA. Figure 4 shows the training accuracies of PromHD from = 30 to 100 when applied to the HPL dataset and processed the 5th experimental run. These accuracies exceed those of SVM-RBS using SVM with a number of selected informative DNASDs that are selected by the rank-based selection (RBS) method [68]. The RBS method is described below.

Comparison of Prediction Performance between PromHD
and Other SVM-Based Methods. Two additional SVM-based classifiers, SVM-4mer and SVM-GSD, are applied for comparisons with SVM-RBS. The SVM-4mer and SVM-GSD methods are implemented by using the 128 top-ranked descriptors of 4-mer motifs and 36 global sequence descriptors, respectively, as input features to the SVM classifier without feature selection, respectively. The best values of parameters and that are determined using a stepwise approach are used in the two SVM-based methods, where ∈ {2 −7 , 2 −6 , . . . , 2 8 } and ∈ {2 −7 , 2 −6 , . . . , 2 8 }. Tables 5 and 6 compare the three SVM-based methods in terms of performance when applied to the HP and DP datasets, respectively. SVM-GSD obtains the highest testing accuracy of 93.6% for the human species; SVM-4mer performs the best in Drosophila species. However, these testing accuracies are lower than those of PromHD, 98. and specificity (SP = 0.960 and 0.988) performances when used with HPT and DPT datasets, respectively.

Rule-Based
Knowledge. This work presents a knowledge acquisition method based on if-then rule for insight of promoter prediction mechanism. The knowledge can be revealed from three aspects: (1) identified informative DNASDs, (2) rules of distinguishing promoters from nonpromoters, and (3) further analysis of distinguishable mechanism using DNASDs. This rule-based knowledge acquisition method uses decision tree method C5.0 [50] to develop if-then rules of the 99-and 74-dimensional DNASD feature vectors in human and Drosophila species. Each if-then rule has two types, one for promoter (Ri-p) and the other for nonpromoter (Ri-n) prediction, where is the rule number index. The selected DNASDs are (A, 1st) (the length of nucleotide A at the first position of the sequence), (C, 100%) (the length of nucleotide C at the last position of the sequence), C4 (GCTC) (the frequency descriptor of 4mer GCTC), and MW (the physicochemical property of molecular weight) in human species. Table 7 shows the interpretable rules as follows.
Rules The CF is a certainty grade of this rule in the unit interval [0, 1]. The R1-p rule has a certainty grade of 0.928 to predict 935 (= 50%×1871) human promoters by using the (A, 1st) feature. With the same (A, 1st) feature, the second rule, R2n, with a certainty grade of 0.999 can identify 864(= (96.2% − 50%) × 1871) nonpromoters. When applying these two rules, the rule-based classifier yields a prediction accuracy of 96.2%, reconfirming that the global sequence descriptor, the length of nucleotide A at the first position of the sequence, is an efficient feature in distinguishing human promoters from nonpromoters. When adding the third rule, PromHD further enhances the prediction accuracy up to 99.5%. For example, a query sequence P has normalized values of 0.0179, 0.9218, 0.2499, and 0.2823 for (A, 1st), (C, 100%), C4 (GCTC), and MW , respectively. The classification procedure using the third rule R1-p (0.0179 > 0.0177542) predicts this query sequence to be a promoter.
Alternatively, the selected DNASDs are MW (the physicochemical property of molecular weight) and AM (the physicochemical property of absorption maxima) in Drosophila species. The interpretable rules, as shown in Table 7, are as follows.

10
The Scientific World Journal The rule-based classifier uses the first rule to predict 961 (= 50% × 1922) Drosophila promoters. The first two rules make the rule-based classifier have a prediction accuracy of 84.7% in Drosophila species. For example, a query sequence has normalized values of MW and AM , 0.4 and 0.3, respectively. The classification procedure using the third rule R3-n MW (= 0.4) > 0.280113 predicts this query DNA sequence to be a nonpromoter. Table 4 lists that the C4 (TGAA) and C1 (AAAG) descriptors have the maximum scores when applied to the HPL and DPL datasets, respectively. A comparison between the two sets of the top 20 descriptors of 4-mer motifs reveals two common 4-mer motifs. One is TGAA, which has scores of 1000 and 732; the other is TGAT, at ranks of 2 and 12 when used with HPL and DPL datasets, respectively. The descriptors of the well-known TATA motif are ranked at the 199th and 98th when applied for the HPL and DPL datasets, respectively.

Top 20 Descriptors of 4-Mer Motifs.
The former descriptor 4 (TATA) ranking at the 199th is excluded out of the reference feature set due to the fact that only 128 top-ranked descriptors of 4-mer motifs are allowed to be included. This agrees closely with the findings of Gershenzon and Ioshikhes [11], who found that the TATA motif exists only in a small proportion of all human promoters. Additionally, only nine descriptors are included in each of the two feature subsets of = 99 and 74 DNASD feature, which are marked with "+" Table 4. This main reason is that the DNASDmining feature selection algorithm considers a set of informative DNASDs at once, rather than individual DNASDs.

Analysis of the Identified DNASDs.
The orthogonal experimental design with orthogonal array and factor analysis is an efficient method for simultaneously examining the individual effect of several factors on the evaluative function [47,48]. In this study, the two levels of a factor represent its inclusion and exclusion of the feature in the feature selection using IGA [47,48]. The factor analysis can quantify the effects of individual factors on the evaluation function, rank the most effective factors, and determine the best level for each factor for optimization of the evaluation function. The most effective factor has the largest main effect difference (MED) amongst the levels of a single factor.  Table S3). Specifically, the (A, 1st) has the highest MED value of 93.1, meaning that the length of nucleotide A at the first position of the sequences can be used to distinguish promoters from nonpromoters in the human species. This result is consistent with the first if-then rule, R1p ( Table 7).
As for the four features with the MED values exceeding 30, they are the descriptors of the physicochemical properties ( AM , MW , and MAC ) and C4 (AAGT), revealing that the three physicochemical properties of absorption maxima, molecular weight, and molar absorption coefficient can be used to distinguish promoters from nonpromoters in the Drosophila specie. The C4 (AAGT), a descriptor of the AAGT motif, has the fourth highest MED value of 31.5; however, it only has a score of 73 and a rank of 41, as shown in Table 4, revealing that DNASDmining can consider the internal correlation within relevant features rather than  individual features using an efficient global optimization [45]. The transition descriptor (C, G) of the occurrence frequency of CG or GC has the fifth highest MED value of 26.6. This analytical result agrees with those obtained in other studies in [4,19,22,23], which have found that GC content is effective in identifying promoter regions.
The one EDP descriptor EG is ranked the 13th and 32th for DPL and HPL, respectively, as shown in Supplementary Tables S3 and S4. The two compositions that are related to nucleotide A and G, denoted as C1 (A) and C1 (G), clearly contribute to promoter prediction, consistent with the findings of Wang et al. and Zhao et al. [22,23]. The one transition descriptor (G, T) that characterizes the frequency of occurrence of two nucleotides GT or TG is ranked the 84th and 23th for human and Drosophila species, respectively. Among the 11 common distribution descriptors, (A, 1st), (T, 1st), and (G, 1st) are ranked the first, second, and third for the human species but the 46th, 20th, and 27th for Drosophila species, respectively. Two of the 14 common 12 The Scientific World Journal The abbreviations P , C1 , C4 , E , , and represent the descriptors of physicochemical property (PCP) and the global sequence descriptors (GSDs) of 1-mer motif, 4-mer motif, EDP, distribution, and transition, respectively. descriptors of 4-mer motifs, C4 (GAGC) and C4 (GAAG), not only have high scores of 592 and 699, respectively, but are also identified by PromHD to be informative DNASDs as can be seen by comparing Tables 4, S3, and S4. Although the ranks of these common 32 DNASDs largely differ between human and Drosophila species, they form the six clusters of P , E , C1 , , , and C4 . The scenarios reveal that the six clusters are all useful for distinguishing promoters from nonpromoters in human and Drosophila species.

Conclusions
Promoter prediction is an important problem in elucidating the regulation of gene expression. Therefore, the development of a well-characterized promoter system is vital for synthetic biology applications. This proposed PromHD method presents a reference feature set of 167 DNASDs, utilizes a feature mining algorithm to select a feature subset of informative DNASDs, and acquires rule-based knowledge based on the selected feature subset. The mining algorithm using an optimization approach to feature selection identifies the most informative and discriminating DNASDs among human and Drosophila species. The top-ranked rules reveal that the global sequence descriptor, the length of nucleotide A at the first position of the sequence, and two physicochemical properties, absorption maxima and molecular weight, are efficient in distinguishing promoters from nonpromoters in human and Drosophila species, respectively. Additionally, this work analyzes the contributions of a feature set of DNA sequence descriptors to the promoter prediction using the MED values. The three physicochemical properties of absorption maxima, molecular weight, and molar absorption coefficient have high MED values, meaning the three properties are clearly useful in distinguishing promoters from nonpromoters in human and Drosophila species. Future work on PromHD will develop a well-characterized promoter system for synthetic biology applications. Moreover, we believe that this proposed method will also be effective in designing prediction methods for other DNA sequence-based applications. The promoter prediction system by using PromHD has been implemented at http://iclab.life.nctu.edu.tw/promhd. All used datasets were given in the website.