建構可解讀模糊規則知識庫來預測與分析DNA結合的蛋白質

行政院國家科學委員會專題研究計畫 成果報告

建構可解讀模糊規則知識庫來預測與分析 DNA 結合的蛋白

質

研究成果報告(精簡版)

計 畫 類 別 ： 個別型 計 畫 編 號 ： NSC 100-2221-E-009-130- 執 行 期 間 ： 100 年 08 月 01 日至 101 年 07 月 31 日 執 行 單 位 ： 國立交通大學生物科技學系（所） 計 畫 主 持 人 ： 黃慧玲 計畫參與人員： 大專生-兼任助理人員：陳力行 大專生-兼任助理人員：黃柏霖 博士班研究生-兼任助理人員：蔡明儒 報 告 附 件 ： 出席國際會議研究心得報告及發表論文 公 開 資 訊 ： 本計畫可公開查詢

中 華 民 國 101 年 10 月 31 日

中 文 摘 要 ： DNA 結合區域/蛋白質在細胞機能中扮演著相當重要的角色， 因其參與許多生物程序，如 DNA 的轉錄調控、修理、複製 等，如果可以對於 DNA 與蛋白質如何作用對於日後研究與應 用會有極大的幫助。因此以生物資訊的方式預測 DNA 結合區 域/蛋白質便成為有趣且重要的議題。目前有許多研究使用不 同的機械學習法作為預測方法，其中以支援向量分類器最多 且效能最好。然而絕大多數以支援向量分類器做為預測 DNA 結合區域/蛋白質的方法，都使用到大量的特徵及數值作為分 類依據，這些方法雖然具有良好的分類效能，但對於所學習 的特徵資料，卻無法為生物學家提供良好的解讀性。因此， 本計畫旨在建立可解讀的模糊邏輯規則，建立一套相關系 統，以增進預測和有效分析 DNA 結合區域分類的物化特性知 識。我們提供一個運用物化特性為特徵的演化式模糊規則分 類器(iFRC)，做為預測 DNA 結合區域/蛋白質的系統。此系統 主要是以智慧型基因演算法(IGA)來完成最佳化的特徵選擇， 利用 IGA 系統化的挑選出最少的特徵、模糊規則數量最為分 類依據，同時最佳化分類模型，以求得最高的辨識率。 本系統所建構的模糊規則用以分類 DNA 結合區域/蛋白質的平 均正確率為 77.46%，測試的正確率達 83.33%。該結果說明本 系統所建構出來的規則可信度極高。進一步分析經由本次研 究結果，我們發現到在物化特性中以帶正電及凡德瓦利的特 徵影響最大，而蛋白質和 DNA 之間的識別在第一步驟中，蛋 白質和 DNA 的電荷之間的互補性，這項結果符合先前研究， 且被認為是重要的。結果也同時顯示，胺基酸結合區的胜 肽，其凡德瓦力值及所帶的正電荷皆高於非結合區。相較於 之前的研究結果分生實驗的結果無法量化及其他生物資訊的 方法無法解讀的缺點，我們提供較為精確的數值可供分子生 物學實驗室作為實驗突變的參考，希望對日後分子生物研究 及應用有所幫助。 中文關鍵詞： 去氧核醣核酸結合區域 特徵選擇 基因演算法 支援向量機 模糊邏輯規則 知識擷取 物化特性 特異位置分數矩陣 蛋白 質功能預測

英 文 摘 要 ： DNA-binding domains/proteins play essential roles in a cell, which are involved in transcription,

replication, packaging, repair and rearrangement. Numerous prediction methods of DNA-binding

domains/proteins were proposed by identifying informative features and designing effective classifiers. These researches reveal that the

DNA-protein binding mechanism is complicated and existing accurate predictors such as support vector machine (SVM) with position specific scoring matrices (PSSMs) are regarded as black-box methods which are not

easily interpretable for biologists. It is desirable to design predictors using interpretable features and classifiers, and the prediction results are

explainable for knowledge acquisition. In this study, we propose an ensemble fuzzy rule base classifier consisting of a set of interpretable fuzzy rule

classifiers (iFRCs) using informative physicochemical properties as features. In designing iFRCs, feature selection, membership function design, and fuzzy rule base generation are all simultaneously optimized using an intelligent genetic algorithm (IGA). IGA maximizes prediction accuracy, minimizes the number of features selected, and minimizes the number of fuzzy rules to generate an accurate and concise fuzzy rule base. Benchmark datasets of DNA-binding domains are used to evaluate the proposed ensemble classifier of 30 iFRCs. Each iFRC has a mean test accuracy of 77.46%, and the ensemble classifier has a test accuracy of 83.33%, where the method of SVM with PSSMs has the accuracy of 82.81%. The physicochemical properties of the first two ranks according to their contribution are positive charge and Van Der Waals volume. Charge complementarity between protein and DNA is thought to be important in the first step of recognition between protein and DNA. The amino acid residues of binding peptides have larger Van Der Waals volumes and positive charges than those of non-binding ones. The proposed knowledge acquisition method by establishing a fuzzy rule-based classifier can also be applicable to predict and analyze other protein functions from sequences.

英文關鍵詞： DNA-binding domains, feature selection, genetic algorithm, support vector machine, fuzzy rules, knowledge acquisition, physicochemical properties, position specific scoring matrix, protein function prediction

FRKAS:

Knowledge Acquisition Using a Fuzzy Rule Base Approach

to Insight of DNA-Binding Domains/Proteins

Hui-Lin Huang1,2, Fang-Lin Chang3, Shinn-Jang Ho4, Li-Sun Shu5, Wen-Lin Huang6, and Shinn-Ying Ho1,2*

1

Department of Biological Science and Technology, National Chiao Tung University, Hsinchu, Taiwan

2

Institute of Bioinformatics and Systems Biology, National Chiao Tung University, Hsinchu, Taiwan

3

Department of Anesthesiology, Tri-Service General Hospital, Taipei, Taiwan 4

Department of Automation Engineering, National Formosa University, Yunlin, Taiwan

5

Department of Information Management, Overseas Chinese University, Taichung, Taiwan

6

Department of Multimedia Entertainment Science, Asia Pacific Institute of Creativity, Miaoli, Taiwan *Corresponding author Email addresses: HLH: [email protected] FLC: [email protected] SJH: [email protected] LSS: [email protected] WLH: [email protected] SYH: [email protected]

Abstract: DNA-binding domains/proteins play essential roles in a cell, which are

involved in transcription, replication, packaging, repair and rearrangement. Numerous prediction methods of DNA-binding domains/proteins were proposed by identifying informative features and designing effective classifiers. These researches reveal that the DNA-protein binding mechanism is complicated and existing accurate predictors such as support vector machine (SVM) with position specific scoring matrices (PSSMs) are regarded as black-box methods which are not easily interpretable for biologists. It is desirable to design predictors using interpretable features and classifiers, and the prediction results are explainable for knowledge acquisition. In this study, we propose an ensemble fuzzy rule base classifier consisting of a set of interpretable fuzzy rule classifiers (iFRCs) using informative physicochemical properties as features. In designing iFRCs, feature selection, membership function design, and fuzzy rule base generation are all simultaneously optimized using an intelligent genetic algorithm (IGA). IGA maximizes prediction accuracy, minimizes the number of features selected, and minimizes the number of fuzzy rules to generate an accurate and concise fuzzy rule base. Benchmark datasets of DNA-binding domains are used to evaluate the proposed ensemble classifier of 30 iFRCs. Each iFRC has a mean test accuracy of 77.46%, and the ensemble classifier has a test accuracy of 83.33%, where the method of SVM with PSSMs has the accuracy of 82.81%. The physicochemical properties of the first two ranks according to their contribution are positive charge and Van Der Waals volume. Charge complementarity between protein and DNA is thought to be important in the first step of recognition between protein and DNA. The amino acid residues of binding peptides have larger Van Der Waals volumes and positive charges than those of non-binding ones. The proposed knowledge acquisition method by establishing a fuzzy rule-based classifier can also be applicable to predict and analyze other protein functions from sequences.

Keywords: DNA-binding domains, feature selection, genetic algorithm, support

vector machine, fuzzy rules, knowledge acquisition, physicochemical properties, position specific scoring matrix, protein function prediction.

INTRODUCTION

DNA-binding domains are functional proteins in a cell, which play a vital role in various essential biological activities, such as DNA transcription, replication, packaging, repair and rearrangement [1]. These transcription factors are mainly DNA-binding proteins (DNA-BPs) coded by 2~3% of the genome in prokaryotes and 6~7% in eukaryotes [2]. DNA-BPs play a pivotal role in variousintra- and extra-cellular activities ranging from DNA replicationsto gene expression control. The researches reveal that the DNA-protein recognition mechanism is complicated and there is no simple rule for this recognition problem [3].

Some researchers have increasingly interests in the prediction and analyse of DNA-BPs [4-6]. Stawiski et al. presented that DNA-binding proteins could be predicted using a neural network trained with features of secondary structures and charged patches [4]. Ahmad and Sarai found that net charge, net dipole moment and quadrupole moment could each distinguish binding and non-binding proteins with known structures well [5]. Kumar et al. proposed a method for predicting DNA-binding proteins using support vector machine (SVM) and position-specific scoring matrices (PSSMs) profiles [6]. The methods [4-6] can fairly analyze and predict DNA-binding proteins, but suffer from obtaining human-interpretable knowledge from sequences.

Leung et al. [7] focus on protein-DNA bindings between transcription factors (TFs) and transcription factor binding sites (TFBSs). A framework to discover associated TF-TFBS binding sequence patterns in the most explicit and interpretable form from TRANSFAC is proposed [7]. Recent mining on exact TF-TFBS-associated sequence patterns (rules) has shown great potentials and achieved very promising results [8]. The approximate rules reveal both the flexible and specific protein-DNA interactions accurately. Huang et al. [9] proposed a systematic approach Auto-IDPCPs to automatically identify a set of physicochemical and biochemical properties in the AAindex database to design SVM-based classifiers for predicting and analyzing DNA-binding domains/proteins from sequences. Auto-IDPCPs identified 23 features of properties from the AAindex database [10] belonging to five clusters such as hydrophobicity, secondary structure, charge, solvent accessibility, polarity, flexibility, normalized Van Der Waals volume, pK (C, N, COOH and pK-a(RCOOH)), etc.

The trend in analyzing DNA-BPs is not only to predict binding proteins well but also to obtain knowledge for biological understanding and finding. It is desirable to design predictors using interpretable features and classifiers, and the prediction results are explainable for knowledge acquisition. Human thinking and reasoning frequently involve fuzzy information originating from inherently inexact human concepts and matching of similar rather than identical experiences. In many applications, rule-based classifiers are created starting from machine learning and fuzzy logic.

In this study, we propose an ensemble fuzzy rule base classifier consisting of a set of interpretable fuzzy rule classifiers (iFRCs) based on the 23 physicochemical properties as features [9]. Because the DNA-BPs have the property of natural clustering, fuzzy classifiers using a scatter partition of feature spaces often have a smaller number of rules than those using grid partitions. In designing iFRCs, feature selection, membership function design, and fuzzy rule base generation are all simultaneously optimized using an intelligent genetic algorithm (IGA) [11]. IGA maximizes prediction accuracy, minimizes the number of features selected, and minimizes the number of fuzzy rules to generate an accurate and concise fuzzy rule

base.

A fuzzy rule-based knowledge acquisition system (FRKAS) using an ensemble fuzzy rule classifier consisting of 30 iFRCs is proposed for prediction and analyse of DNA-BPs. Each iFRCs has two fuzzy rules, one for binding and the other for non-binding prediction. The ensemble classifier using eight physicochemical properties performs well with a test accuracy of 83.33%, compared with an individual SVM with PSSMs (82.81%) [6] and SVM with 22 physicochemical properties (80.73%) [9]. The physicochemical properties of the first two ranks according to their contribution are positive charge and Van Der Waals volume. The amino acid residues of binding peptides have larger Van Der Waals volumes and positive charges than those of non-binding ones.

MATERIALS AND METHODS

The framework of the proposed FRKAS is given in Fig. 1. FRKAS uses an ensemble fuzzy rule classifier consisting of 30 interpretable fuzzy rule classifiers (iFRCs). The design aim of iFRCs is to generate an accurate and concise fuzzy rule base. The following sections present the used datasets, feature sets, the design of iFRCs, the ensemble fuzzy classifier and knowledge acquisition of DNA-binding domains.

Datasets

For comparisons with existing methods [6], [9], the same benchmark dataset DNAset, also called main dataset from Kumar et al. [6] was used to establish an ensemble fuzzy rule classifier. DNAset has 146 non-redundant DNA-binding domains (or protein chains) in which no two domains have the sequence identity of more than 25%. A non-redundant set of 250 non-binding domains was obtained from Stawiski et al. [4]. They used following criteria: 1) no two protein chains have similarity more than 25% and 2) the approximate size and electrostatics are similar to DNA-BPs. An independent data set DNAiset is additionally used, having 92 DNA-binding domains and 100 non-DNA-binding proteins [6].

Feature sets

The method Auto-IDPCPs [9] consists of three tasks: 1) clustering 531 vectors of physicochemical and biochemical properties in the AAindex database into 20 classes using a fuzzy c-means algorithm, 2) utilizing an efficient genetic algorithm based optimization method to select an informative feature set to represent sequences, and 3) analyzing the selected feature vectors to identify the related physicochemical properties which may affect the binding mechanism of DNA-BPs.

Auto-IDPCPs [9] used a systematic approach to automatically identify a set of 23 properties for predicting and analyzing DNA-binding domains/proteins in the dataset DNAset. The 23 properties belonging to five clusters are used to design the proposed ensemble fuzzy rule classifiers, shown in Table 1.

Interpretable fuzzy rule classifiers (iFRCs)

High performance of iFRCs mainly arises from two aspects. One is to simultaneously optimize all parameters in the design of iFRCs where all the elements of the fuzzy classifier design have been transformed into parameters of a large parameter optimization problem. The other is to use an efficient optimization algorithm IGA which is a specific variant of the intelligent evolutionary algorithm

[11]. The intelligent evolutionary algorithm uses a divide-and-conquer strategy to effectively solve large parameter optimization problems. IGA is shown to be effective in the design of accurate classifiers with a concise fuzzy rule base using an evolutionary scatter partition of feature space [12].

Flexible membership functions

The classifier design of iFRCs uses flexible generic parameterized fuzzy regions which can be determined by flexible generic parameterized membership functions (FGPMFs) and a hyperbox-type fuzzy partition of feature space [12]. Each fuzzy region corresponds to a parameterized fuzzy rule. In this study, the value of each physicochemical property is normalized into a real number in the unit interval [0, 1]. An FGPMF with a single fuzzy set is defined as

                      c x b d x c c d x d b x a a b a x d x a x x if 1 if if or if 0 ) (  (1)

where x ∈ [0, 1] and a ≤ b ≤ c ≤ d. The variables a, b, c and d determining the shape of a trapezoidal fuzzy set are the parameters to be optimized. It is well recognized that confining evolutionary searches within feasible regions is often much more reliable than penalty approaches for handling constrained problems [13]. Therefore, five parameters V1, V2,…, V5[0,1] (without constraints among Vi ) instead of a, b, c and d are encoded into a chromosome for facilitating IGA. Let an additional variable L=V1 which determines location of the fuzzy set characterizing the occurrence of training patterns. When Vi are obtained, variables a, b, c, and d can be derived as follows: a=L-(V2+V3), b=L-V3, c=L+V4, and d=L+(V4+V5) where b ≤ L ≤ c. This transformation can always make the derived values of a, b, c and d feasible and reduce interactions among encoded parameters of the IGA’s chromosomes. Some illuminations of FGPMF are shown in Fig. 2 [12].

Fuzzy rule and fuzzy reasoning method

The following fuzzy if–then rule base for n-dimensional classification problems are used in the design of iFRCs:

Rj : If x1 is Aj1 and . . . and xn is Ajn then class CLj with CFj, j = 1, . . . , N. where Rj is a rule label, xi denotes a variable of physicochemical property, Aji is an antecedent fuzzy set, C is a number of classes, CLj∈ {1, . . ., C} denotes a consequent class label, CFj is a certainty grade of this rule in the unit interval [0, 1], and N is a number of initial fuzzy rules in the training phase. In this study, C=2 (two classes for binding and non-binding), n=23 (initial number in the feature set to be selected), and

N=3C (initial number in the rule set to be selected).

To enhance interpretability of fuzzy rules, linguistic variables in fuzzy rules can be used. Each variable xi has a linguistic set U= {small, medium, large}. Each linguistic value of xi equally represents 1/3 of the domain [0, 1]. An antecedent fuzzy set Aji∈ AU where AU denotes a set of subsets of U. Examples of linguistic antecedent fuzzy sets are shown in Fig. 3. If xi is Aji representing {medium, large}, it means the value of xi (physicochemical property) is belonging to the set of {medium, large}. If xi is Aji representing {small, medium, large}, it means the physicochemical property is

ALL, i.e., don’t care.

In the training phase, all the variables CLj and CFj are treated as parametric genes encoded in a chromosome and their values are obtained using IGA. The following fuzzy reasoning method is adopted to determine the class of an input pattern xp = (xp1, xp2, . . ., xpn) based on voting using multiple fuzzy if–then rules: Step 1: Calculate score SClassv (v = 1, . . . , C) for each class as follows:

SClassv =



  v Class CL FC R j p j j j CF x ) , ( 



  n i pi ji p j x x 1 ), ( ) (   (2)

where FC denotes the fuzzy classifier, and μji(·) represents the membership function of the antecedent fuzzy set Aji.

Step 2: Classify xp as the class with a maximal value of SClassv.

Notably, xp is classified into the binding or non-binding class for one iFRC. The final classification of xp is determined using the proposed ensemble classifier consisting of 30 iFRCs in the study.

Chromosome representation of IGA

A chromosome consists of control genes for selecting useful features (physicochemical properties) and significant fuzzy rules, and parametric genes for encoding the membership functions and fuzzy rules. The control genes comprise two types of parameters. One is parameters for selecting features. The other is parameters for selecting fuzzy rules. The parametric genes determine variables of three types:

t ji

V ∈ [0, 1], t=1, …, 5, for determining the antecedent fuzzy set Aji, CLj for determining the consequent class label of rule Rj, and CFj∈ [0, 1] for determining the certainty grade of rule Rj , where j=1, …, N and i=1, …, n. A rule base with N fuzzy rules is represented as an individual. The detailed explanation of the chromosome representation and implementation can be referred to [12]. The design of an efficient fuzzy classifier is formulated as a large parameter optimization problem. Once the solution of IGA is obtained, an accurate classifier with a concise fuzzy rule base can be obtained.

Fitness function of IGA

We define the fitness function of IGA for designing iFRCs as follows:

max Fit(FC) = ACC − WrNr − WfNf (3) where Wr and Wf are positive weights. In this study, the fitness function is used to optimize the three objectives: 1) to maximize the classification accuracy ACC, 2) to minimize the number Nr of fuzzy rules, and 3) to minimize the number Nf of selected features. For obtaining an easily-interpretable knowledge rule base for each iFRC, the smaller values of Nr and Nf are better. Therefore, we used large values of weights Wr = 0.2 and Wf = 0.1. Since the classification accuracy is not the first priority for iFRC, the ensemble fuzzy rule base classifier (EFRBC) consists of k (e.g., 30) iFRCs is necessary for obtaining high accuracy of predicting DNA-BPs.

Ensemble fuzzy rule base classifier (EFRBC)

There are three opinions for using an ensemble strategy [14]: 1) Statistical: the reason is related to lack of adequate data to properly represent the data distribution; 2)

Computational: the reason is the model selection problem, and 3) Representational: the reason is to address the cases when the chosen model cannot properly represent the sought decision boundary. In this study, the training dataset DNAset is relatively small, compared with the complex of recognition problems in the binding mechanism. For considering interpretability, the same model SVM is used to construct the EFRBC. Since the decision boundary of iFRCs is not complicated, the ensemble approach can advance the prediction accuracy.

The EFRBC is composed of k=30 iFRCs and a voting method.

(1) Classification of iFRCs: The prediction accuracy is highly related to the conciseness of the fuzzy rule base for every iFRC. The optimal design of iFRCs can simultaneously optimize the three objectives using a weighted sum approach: 1) to maximize prediction accuracy, 2) minimize the number of features selected, and 3) to minimize the number of fuzzy rules. However, the trade-off between prediction accuracy and conciseness of the rule base can be determined by tuning the weights Wr and Wf.

(2) Voting method: Different classification results of the query sequences will be obtained from the outputs of the k independent iFRCs, and then these results are integrated using the simple voting method.

VSj =     



 otherwise j Ci k i 0, , 1 , 1   , (4)

where k is the number of iFRC, j=1, 2, …, C is the class label, Ci is the predicted class label by ith iFRC. In this study, for a given query protein with C=2, the final class is determined by argmax {VS1, VS2}.

Four performance measurements were used to evaluate iFRC and EFRBC: sensitivity (SEN), specificity (SPE), accuracy (ACC), and Matthew's correlation coefficient (MCC), defined as follows: SEN = TP/(TP + FN), SPE = TN/(TN + FP), ACC = (TP+TN)/(TP+FP+TN+FN), and MCC = ((TP×TN)-(FN×FP))/SQRT ((TP+FN)(TN+FP)(TP+FP)(TN+FN)), where TP, TN, FP and FN are the numbers of true positive, true negative, false positive and false negative, respectively.

DNA-binding knowledge acquisition

This study proposes a knowledge acquisition approach based on the optimal design of fuzzy rule bases to insight of DNA-binding mechanism. The informative knowledge can be revealed from three aspects: 1) identified informative physicochemical properties, 2) rules of DNA-binding and non-binding mechanism, and 3) further analysis of binding mechanism using physicochemical properties.

Auto-IDPCPs [9] used a systematic approach to automatically identify a set of properties to design accurate SVM-based classifiers for predicting DNA-binding domains/proteins. By analysing the rules of EFRBC, we can further reduce the number of physicochemical properties with great contribution in predicting the binding mechanism. From the appropriate interpretations of fuzzy rules with linguistic variables, it is more understandable for biologists. The rule-based knowledge provides an effective approach to insight of DNA-binding domains.

An illustrating example of iFRC, shown in Fig. 4, and its explanation are given as follows. Fig. 4 shows two rules of one iFRC which uses tow features, H252 (PRAM900101, Hydrophobicity (Prabhakaran, 1990)) and H398 (ZIMJ680101,

Hydrophobicity (Zimmerman et al., 1968)). The rules R1 and R2 with fuzzy sets are binding and non-binding rules, respectively. The descriptions of two rules are given as follows:

R1: if H252 is {medium, large} and H398 is {medium, large}, then binding with CF1=0.196.

R2: if H252 is {small, medium} and H398 is {small, medium}, then non-binding with CF2=0.549.

For example, a query sequence xp has normalized values of H252 and H398, 0.4 and 0.3, respectively. The classification procedure using this iFRC is described as follows:

Step 1: Use equation (1) to calculate the values of membership functions ub() and un() for binding and non-binding, respectively.

The value of ub1(0.4) is 0.537 for H252and the value of ub2(0.3) is 1.0 for H398. The values of un1(0.4) is 0.606 for H252 and the value of un2(0.3) is 1.0 for H398. The binding value of ub(xp) is ub1(0.4) x ub2(0.3)=0.537 and the non-binding value of un(xp)=un1(0.4)* un2(0.3)= 0.606.

Step 2: Use equation (2) to calculate the score for each class.

Because the non-binding score Sclass =un(xp)*CF2 =0.333 is larger than the binding score Sclass =ub(xp)*CF1=0.105, the query sequence by using the single iFRC is classified into the non-binding class.

The fuzzy regions of binding and non-binding are illustrated in Fig. 4. The final classification of the query sequence using the proposed ensemble classifier is determined using the voting result of k=30 iFRCs. Generally, if the query sequence is located near the boundary of some fuzzy regions, the ensemble strategy can improve the prediction accuracy.

Results

The parameter settings of IGA [11] are Npop = 20, Pc = 0.7, Ps =1−Pc, Pm = 0.01 and α = 15. Because the search space of the optimal design of iFRCs is proportional to the number Np of parameters to be optimized, the stopping condition is suggested to use a fixed number 100Np of fitness evaluations.

Prediction performance evaluation

The training samples with 23 properties in the dataset DNAset are represented as 23-dimensional feature vectors. This set of 23 physicochemical properties is identified by Auto-IDPCPs [9]. The dataset DNAiset was used for evaluating test performance of iFRCs and the ensemble classifier EFRBC. Due to the non-deterministic characteristic of genetic algorithms, the best iFRC with high training accuracy is selected for testing DNAiset from 30 runs. The average performance of 30 independent iFRCs in EFRBC is given in Table 2.

The SVM-based classifier with PSSMs has the training and test accuracies of 86.62% and 82.81%, respectively. The SVM-based classifier with 22 physicochemical properties identified by Auto-IDPCPs has high training accuracy 87.12% and a relatively small accuracy of 80.73%. The average performance of iFRCs has the training and test accuracies of 74.32% and 77.46%, respectively, without significant over-training problems. The average number of features is Nf

=1.34 and average number of rules is Nr =2.0. It reveals that the selected features are very effective and the rule bases are very concise. The test performance of EFRBC has a high test accuracy of 83.33%, sensitivity SEN=82.0%, specificity SPE=84.8%, and MCC=0.67, shown in Table 3. The ensemble strategy is effective for accurate prediction with an improvement of 5.87%.

The occurrence number of features and their descriptions in the 30 iFRCs are given in Table 4. There are eight features used in EFRBC. The 531 properties in the AAindex database were classified into six groups [9], [10]: 1) Alpha and turn propensities (A), 2) Beta propensity (B), 3) Composition (C), 4) Hydrophobicity (H), 5) Physicochemical properties (P), and 6) Other properties (O). Table 4 reveals that the two top-rank properties are the positive charge (H88, FAUJ880111) and normalized Van Der Waals Volume (P80, FAUJ880103).

Rule-based knowledge

Figure 5 shows seven iFRCs, a selected subset of 30 iFRCs, containing all the eight features in Table 4. We selected two iFRCs, iFRC1 and iFRC2, to illustrate the rules for DNA-binding mechanism. The iFRC1 and iFRC2 have training accuracies of 74.24% and 74.49%, the test accuracies of 72.40% and 82.81%, the feature numbers

Nf of 2 and 1, and the rule numbers Nr of 2 and 2, respectively. The selected physicochemical properties are P80 (normalized Van Der Waals Volume), H88 (positive charge) and H355 (hydrophobicity). The fuzzy rules are linguistically interpretable as follows:

Fuzzy Classifier iFRC1:

R1: if P80 is ALL and H355 is ALL, then binding with CF=0.161. R2: if P80 is {small, medium} and H355 is {small, medium}, then non-binding with CF=0.576.

Fuzzy Classifier iFRC2:

R1: if H88 is {medium, large}, then binding with CF=0.416. R2: if H88 is {small}, then non-binding with CF=0.165. Analysis of binding mechanism

The two top-rank features are positive charge (H88) and normalized Van Der Waals Volume (P80), shown in Table 4. A typical DNA-binding domain sequence in the training dataset DNAset, shown in Fig. 6(a), is used to have an insight into the binding mechanism. The sequence is the chain of the protein with PDBID 1WVL, a multimeric DNA-binding protein using Sac7d and GCN4 as templates, whose FASTA sequence is shown in Fig. 6(b).

We used the tool APBS [15] plugged in VMD 1.9 [16] to get the direct measurement of the charge distribution on 1WVL protein surface. The surface potential on 1WVL at neutral pH was calculated where the negatively and positively charged surfaces are shown in red and blue, respectively, shown as Fig. 7. The DNA-binding pocket, positively charged cavity, is visible in dark blue. Once the domain and DNA are assembled into clusters, hydrophobic molecules are held together by Van Der Waals interactions. Hydrophobic ridge of residues in the minor groove, the side-chain atoms of the hydrophobic ridge residues, is shown as gray spheres. Protein-DNA interaction resulting in the formation of salt bridges between cationic amino acid side chains and the phosphate backbone completely neutralize particular

phosphate anions, eliminating repulsive interactions with fractional negative charges at neighboring phosphates [17].

Discussion

To avoid from overfitting the small-scale datasets in identifying physicochemical properties using an optimization approach, this study proposes a hybrid computational method of combining evidences by considering robust factors from the viewpoints of statistics and biological experiments from literature. It can be expected that the proposed method can effectively discover and rank more informative physicochemical and biochemical properties closely relative to the DNA-binding mechanism if the size of the training dataset is significantly increased. These discovered properties in predicting and analyzing the DNA-binding mechanism can be further investigated by biologists.

Conclusions

This study has proposed a systematic fuzzy rule based knowledge acquisition system (FRKAS) to predict and analyze DNA-binding domains/proteins. The merits of FRKAS can be summarized, described below.

1) The novel interpretable fuzzy rule classifiers (iFRCs) using informative physicochemical properties as features are proposed in this study. The features and classifiers are more helpful for understanding the binding mechanism rather than the SVM with PSSMs.

2) To obtain an accurate and concise fuzzy rule base, an intelligent genetic algorithm is utilized to optimize simultaneously the three objectives: maximizing prediction accuracy, minimizing the number of features selected, and minimizing the number of fuzzy rules. The designers can tune the weights in the weighted sum approach according to the preference to the three objectives.

3) Due to the small size of the training dataset, an ensemble classifier consisting of iFRCs is adopted to compensate the limitation, resulting in high-accuracy and robust performance.

4) The design of FRKAS considers both prediction accuracy and interpretability at the same time. FRKAS can also be applicable to predict and analyze other protein functions from sequences.

Acknowledgements

The authors would like to thank the National Science Council of Taiwan for financially supporting this research under the contract numbers NSC 100-2221-E-009-130- and NSC 100-2221-E-009-143-.

References

[1] Gao, M., Skolnick, J., A threading-based method for the prediction of DNA-binding proteins with application to the human genome. PLoS Comput Biol 2009,

5, e1000567.

[2] Lejeune, D., Delsaux, N., Charloteaux, B., Thomas, A., Brasseur, R., Protein-nucleic acid recognition: statistical analysis of atomic interactions and influence of DNA structure. Proteins 2005, 61, 258-271.

[3] O'Flanagan, R. A., Paillard, G., Lavery, R., Sengupta, A. M., Non-additivity in protein-DNA binding. Bioinformatics 2005, 21, 2254-2263.

[4] Stawiski, E. W., Gregoret, L. M., Mandel-Gutfreund, Y., Annotating nucleic acid-binding function based on protein structure. J Mol Biol 2003, 326, 1065-1079. [5] Ahmad, S., Sarai, A., Moment-based prediction of DNA-binding proteins. J Mol

Biol 2004, 341, 65-71.

[6] Kumar, M., Gromiha, M. M., Raghava, G. P., Identification of DNA-binding proteins using support vector machines and evolutionary profiles. BMC

Bioinformatics 2007, 8, 463.

[7] Leung, K.-S., Wong, K.-C., Chan, T.-M., Wong, M.-H., et al., Discovering protein–DNA binding sequence patterns using association rule mining. Nucleic

Acids Research 2010, 38 (19), 6324–6337.

[8] Chan, T. M., Wong, K. C., Lee, K. H., Wong, M. H., et al., Discovering approximate-associated sequence patterns for protein-DNA interactions.

Bioinformatics 2011, 27, 471-478.

[9] Huang, H.-L., Lin, I.-C., Liou, Y.-F., Tsai, C.-T., et al., Predicting and analyzing DNA-binding domains using a systematic approach to identifying a set of informative physicochemical and biochemical properties. BMC Bioinformatics 2011, 12 Suppl 1.

[10] Kawashima, S., Pokarowski, P., Pokarowska, M., Kolinski, A., et al., AAindex: amino acid index database, progress report 2008. Nucleic Acids Res 2008, 36, D202-205.

[11] Ho, S. Y., Shu, L. S., Chen, J. H., Intelligent evolutionary algorithms for large parameter optimization problems. IEEE Transactions on Evolutionary

Computation 2004, 8, 522-541.

[12] Ho, S.-Y., Chen, H.-M., Ho, S.-J., Chen, T.-K., Design of accurate classifiers with a concise fuzzy rule base using an evolutionary scatter partition of feature space. IEEE Trans. Systems, Man, and Cybernetics─Part B 2004, 34, 14.

[13] Michalewicz, Z., Dasgupta, D., Leriche, R. G., Schoenauer, M., Evolutionary algorithms for constrained engineering problems. Comput Ind Eng 1996, 30, 851-870.

[14] Dietterich, T. G., The Handbook of Brain Theory and Neural Networks, The MIT Press, Cambridge 2002.

[15] Baker, N. A., Sept, D., Joseph, S., Holst, M. J., McCammon, J. A., Electrostatics of nanosystems: application to microtubules and the ribosome. Proc Natl Acad

Sci U S A 2001, 98, 10037-10041.

[16] Humphrey, W., Dalke, A., Schulten, K., VMD: visual molecular dynamics. J Mol

Graph 1996, 14, 33-38, 27-38.

[17] Strauss, J. K., Maher, L. J., 3rd, DNA bending by asymmetric phosphate neutralization. Science 1994, 266, 1829-1834.

Figures

Figure 1. The framework of the proposed fuzzy rule-based knowledge

acquisition system (FRKAS) based on physicochemical properties (PCPs).

Figure 2. Illuminations of FGPMF: (a) a>0 and d< 1; (b) a<0<b, (c) b≦0; (d)

b≦0 and c≧1.

Figure 3. Examples of an antecedent fuzzy set Aji with linguistic values (small,

medium, large): (a) Aji represents {medium}; (b) Aji represents {medium, large}; (c) Aji represents {small, medium} ; (d) Aji represents {small, medium, large} or ALL.

Training data DNAset

Fuzzy Classifier Design

1) Flexible membership functions

2) Fuzzy rule and fuzzy reasoning method 3) Chromosome representation of IGA 4) Fitness function of IGA

iFRC1 iFRC2 ● ● ● iFRCk Test dataset DNAiset Ensemble classifier Auto-IDPCPs [9] PCPs Analysis of binding mechanism Rule-based knowledge of DNA-binding Mining informative PCPs

iFRC Binding Rule CF1 Non-Binding Rule CF2 Class distribution H252 (a,b,c,d) (0.000,0.745,0.792,1.000) 0.196 (0.000,0.000,0.118,0.580) 0.549 H398 (a,b,c,d) (0.000,0.255,1.000,1.000) (0.000,0.000,0.376,0.718)

Figure 4. An illustrating example of iFRC uses tow features, H252 and H398.

Classifier PCP Binding Rule CF1 Non-Binding Rule CF2

IFRC1 P80 _R1 0.161 R2 0.576 H355 IFRC2 H88 R1 0.416 R2 0.165 IFRC3 P80 R1 0.235 R2 0.969 IFRC4 P80 R1 0.949 R2 0.718 A97 IFRC5 H88 R1 0.910 R2 0.286 A237 IFRC6 H252 R1 0.196 R2 0.549 H398 IFRC7 A237 R1 0.867 R2 0.153 H482 H398 H252 Blue: Binding Red: non-Binding

Figure 5. Some typical iFRCs covering the eight physicochemical properties

(PCPs) in Table 4.

(a)

(b)

Figure 6. An illustrating example of PDBID 1WVL. (a) One domain sequence

randomly selected from the training dataset. (b) The FASTA sequence of 1WVL.

Figure 7. The molecular surface accessible for an electron donor protein. The

negatively and positively charged surfaces are shown in red and blue, respectively. The figure was created by using VMD 1.9 [16]. Surface electrostatic potential was calculated by using the APBS tool [15]. The DNA-binding pocket on 1WVL is visible as dark blue, and hydrophobic ridge of residues in the minor groove is visible as gray spheres. >1WVL:A|PDBID|CHAIN|SEQUENCE MVKVKFKYKGEEKEVDTSKIKKVWRVGKMVSFTYDDNGKTGRGAVSEKDAPKELLDMLARAEREKK GVLKKLRAVENELH >1WVL:B|PDBID|CHAIN|SEQUENCE MVKVKFKYKGEEKEVDTSKIKKVWRVGKMVSFTYDDNGKTGRGAVSEKDAPKELLDMLARAEREKK GVLKKLRAVENELH >1WVL:C|PDBID|CHAIN|SEQUENCE CCTATATAGG >1WVL:D|PDBID|CHAIN|SEQUENCE CCTATATAGG” MVKVKFKYKGEEKEVDTSKIKKVWRVGKMVSFTYDDNGKTGRGAVSEKDAPKELLDMLARAEREKK

Tables

Table 1. Physicochemical properties (PCPs) in the five identified

clusters Cid for analyzing DNA-binding domains, obtained from [9]

Cid AAindex ID PCP Cid AAindex ID PCP 7 BHAR880101 Flexibility 10 FASG760105 pK-C

7 BURA740101 Secondary structure 10 JOND750102 pk- (-COOH) 7 CHOC760103 Solvent accessibility 10 RADA880108 Polarity

7 HOPT810101 Hydrophobicity 16 PRAM900101 Hydrophobicity 7 FAUJ880111 Charge 16 FUKS010104 Solvent accessibility 9 KARP850101 Flexibility 16 KUMS000103 Secondary structure 9 PALJ810115 Secondary structure 18 PONP800107 Solvent accessibility 9 ROSM880101 Hydrophobicity 18 GRAR740102 Polarity

9 KUHL950101 Solvent accessibility 18 FASG760104 pK-N

10 ZIMJ680101 Hydrophobicity 18 FAUJ880113 pK-a(RCOOH) 10 EISD860101 Solvent accessibility 18 FAUJ880103 Normalized van der

10 GEIM800101 Secondary structure Waals volume

Table 2. The performance comparisons between the SVM and fuzzy rule based classifiers. The training dataset and test dataset are DNAset and DNAiset, respectively.

DNAset DNAiset

Accuracy (%) Feature no. Rule no. Accuracy (%)

SVM + PSSMs [6] 86.62 400 NA 82.81

SVM + PCPs [9] 87.12 22 NA 80.73

iFRCs 74.32 1.34 2.0 77.46

EFRBC NA* 8 60 83.33

* The ensemble classifier EFRBC consisting of 30 iFRCs has no training accuracy

Table 3. Performances of the proposed FRKAS on DNAiset

Accuracy (%) SEN (%) SPE (%) MCC

83.33 82.0 84.8 0.67

Table 4. The eight features used in the 30 iFRCs

No. Feature ID AAindex No. Property

20 H88 FAUJ880111 Positive charge

12 P80 FAUJ880103 normalized Van Der Waals Volume

3 A237 PALJ810115 Secondary structure

2 A97 GEIM800101 Secondary structure

1 H252 PRAM900101 Hydrophobicity

1 H355 ROSM880101 Side chain hydropathy

1 H398 ZIMJ680101 Hydrophobicity

1 H482 KUHL950101 Solvent accessibility

A: Alpha and turn propensities. B: Beta propensity. C: Composition. H: Hydrophobicity. P: Physicochemical properties. O: Other properties [9].

Table 5. The rule-based knowledge of DNA-binding mechanism corresponding to the fuzzy classifiers, shown in Fig. 5

Fuzzy classifier Rule-based knowledge

iFRC1

R1-b: If P80 is ALL and H355 is ALL then binding (CF= 0.161) R1-n: If P80 is {small, medium} and

H355 is {small, medium}

then non-binding (CF= 0.576)

iFRC2

R2-b: If H88 is {medium, large} then binding (CF= 0.416) R2-n: If H88 is {small} then non-binding (CF= 0.165) iFRC3

R3-b: If P80 is {medium, large} then binding (CF= 0.235) R3-n: If P80 is {small, medium} then non-binding (CF= 0.969)

iFRC4

R4-b: If P80 is {medium, large} and A97 is ALL

then binding (CF= 0.949) R4-n: If P80 is ALL and A97 is

{small, medium}

then non-binding (CF= 0.718)

iFRC5

R5-b: If H88 is {medium, large} and A237 is {small, medium}

then binding (CF=0.910) R5-n: If H88 is {small} and A237 is

ALL

then non-binding (CF= 0.286)

iFRC6

R6-b: If H252 is {medium, large} and H398 is {medium, large}

then binding (CF= 0.196) R6-n: If H252 is {small, medium} and

H398 is {small, medium}

then non-binding (CF= 0.549)

iFRC7

R7-b: If A237 is {small, medium} and H482 is {medium, large}

then binding (CF= 0.867) R7-n: If A237 is ALL and H482 is

{small, medium}

國科會補助教師出席國際會議結案心得報告

報告人姓名 黃慧玲 所屬學校 學系(所) 交通大學 生物科技學系 會議期間 及地點 2011/12/04 至 2011/12/07 南韓 補助項目 及金額 ■ 機票費 ■ 註冊費 □ 生活費 會議名稱 （中文）2011 第二十二屆基因體資訊國際會議 （英文）2011 The 22nd International Conference on

Genomic Informatics (GIW 2011)

發表海報 論文題目

1. 應用於肌肉肝糖分解之新陳代謝路徑建模動態參數的 最佳化方法

Optimization approach to estimation of kinetic parameters for modelling metabolic pathways of muscle glycogenolysis 2. 應用於選擇複雜資訊標記SNP的智慧型三目標基因演 算法

Intelligent triple-objective genetic algorithm for selecting informative Tag SNPs

報告內容：(1、參加會議經過；2、與會心得3、建議4、攜回資料)

吾人發表的海報論文是在12/05~12/06二日下午的Hall B會議廳

1. 參加會議經過

此次星期日由桃園機場搭釜山航空直飛釜山金海國際機場。我由釜山金海國 際機場搭飯店公車抵達飯店，釜山溫度大約 5。 C~-10。C，看著優美的景象，令 疲勞的身體舒解些許。 今年GIW 2011與BIOINFO 2011二會議共同在海雲台大飯店在韓國釜山舉行 這次的年度會議。本次會議有GIW與BIOINFO兩個議程可以選擇，而plenary talks 是合併在一起舉行的。會議第一天開場是由MIT的David Bartel演講MicroRNAs and Their Regulatory Targets；第二場是由韓國生資中心(KOBIC)的Sanghyuk Lee 演講有關Bioinformatics Research and Resources at KOBIC。第二天是西班牙國家 癌症研究中心的Alfonso Valencia演講A Bioinformatics perspective of Cancer Personalize Genome Data；另一場是由東京大學Kiyoshi Asai演講Algorithms for RNA sequence analysis。最後一天是由首爾大學的Jeong-Sun Seo演講Genome-wide map of common and rare variants in Asian population using massively parallel DNA and RNA sequencing – Preliminary results from 1000 Asian Genome Project。

我此次發表兩篇poster，分別在會議第一天下午四點與會議第二天晚上六點進 行。此次所發表的海報論文名稱為Optimization approach to estimation of kinetic parameters for modelling metabolic pathways of muscle glycogenolysis與Intelligent triple-objective genetic algorithm for selecting informative Tag SNPs。海報論文報告

內容豐富，各國學者彼此交換研究心得，獲得參加學者對吾人研究之正面印象， 也讓其他學者多瞭解我們研究的方法與方向。 下圖一為本人至會議報到櫃臺所拍攝的照片；第二張圖片為本人在兩篇所發表之 海報前的留影。

2、與會心得

感謝研發處與國科會補助參加國際會議之出國補助，使本人得以出席跨領域生 物資訊國際會議，開拓眼界及促進國際觀。每次參加國計會議除了努力讓世界知道 臺灣人在研究方面非常認真與相當有能力為心則。本次會議GIW2011 為第二十二

屆 基 因 體 資 訊 國 際 會 議 此 次 主 辦 單 位 是 KSBSB (Korean Society for Bioinformatics and Systems Biology)。第一次的 GIW 會議成立于1990 年在東京 和大多是在日本舉行了後續會議。2006 年以來，GIW 會面，成為真正的國際約 亞洲-太平洋地區的國家每年舉行。現在它是亞洲協會生物資訊學 (AASBi) 的正 式會議。今 年，韓 國 社會生物資 訊學和 系 統生物學 (KSBSB) 將 GIW 與

BIOINFO 二會議同時在海雲台大飯店在韓國釜山共同舉行這次的年度會議。今 年大會共有五個全體會議講座、 5 小型座談會、 32 接受的論文和超過 120 海 報。大會這次邀請五個世界著名的優秀演講者，其中每個人都是在自己的領域的 領導者。因為他們會給一系列刺激全體會議講座，讓我們受益良多。透過這次會 議將科學交流，和促進了解基因體資訊進展的良好機會。 個人覺得生物資訊這領域，由此次舉辦國韓國，這國家對生物資訊投入組織 相當龐大，也可見他們對這領域的企圖心與團結。反觀台灣生物資訊投入與組織 結構發展還需更努力。

3、建議

近年來國科會、教育部和學校積極鼓勵年輕研究人員，除鼓勵教師參與會議外， 特別是博士班學生，參與大型國際會議，及早進入研究領域的核心，吸取國際研究 經驗，以提高國人的研究水準。參加生物資訊國際會議對老師及學生是非常重要的， 會議中不但可以得到相關研究的最新發展資訊，認識結交許多相關領域的學者，彼 此交換研究心得，更可找到跨領域的學者國際合作，在跨領域的生物資訊研究更是 重要。目前研究生已有多管道獲(部份)補助出席國際會議，建議繼續擴大進行。而 國際化的學術交流是往後的趨勢，也能有所激勵國人學界能力與國際觀。

4、攜回資料

1. 期刊（電子檔安裝於贈送的隨身碟中） 2. 隨身袋一只。

行政院國家科學委員會補助國內專家學者出席國際學術會議報告

報告人姓名 黃慧玲 所屬學校 學系(所) 交通大學 生物科技學系 會議期間 及地點 2012/5/17 至 2012/5/20 中國上海 補助項目 及金額 ■ 機票費 □ 註冊費 ■ 生活費 會議名稱 （中文）2012 年第六屆生物資訊與生醫工程國際會議 （英文）The 2012 4th International Conference on

Bioinformatics and Biomedical Engineering (iCBBE 2012)

發表論 文題目

（中文）1.使用物化特性方法設計螢光蛋白質的預測器 2.使用計分卡設計醣結合蛋白質的預測器

（英文）1.Designing predictors of bioluminescence proteins using an efficient physicochemical property mining method

2.Prediction of Carbohydrate-Binding Proteins Using a Scoring Card Method

報告內容：(1、參加會議經過；2、與會心得；3、建議；4、攜回資料)

報告內容應包括下列各項：

一、

參加經過

近年來生物資訊愈來愈受重視，也越來越多專注在生物資訊的研討

會如雨後春筍般在舉行。而本次參加的國際會議為2012年第六屆生物資訊

與生醫工程國際會議(iCBBE 2012)，由國際電子電機學會IEEE Engineering

in Medicine and Biology Society贊助舉辦。

我所發表之論文題目是「Designing predictors of bioluminescence

proteins using an efficient physicochemical property mining method」

，主要是

使用計算智慧的最佳化技術來從531個物化特性中挑選出一組最佳的小集

合，結合support vector machine分類器，設計出螢光蛋白質的預測器，並

分析這些蛋白質序列的物化特性在螢光發光所扮演的功能與腳色。另外一

篇是「Prediction of Carbohydrate-Binding Proteins Using a Scoring Card

Method」

，主要是使用由智慧型基因演算法（Intelligent Gene Algorithm）

從原始蛋白質序列計算相鄰氨基酸（di-peptide）的權重找出最佳化的計分

卡（scoring card），設計出醣結合蛋白質的預測器，並分析這些蛋白質序

列的物化特性在醣結合所扮演的功能與腳色。觀察發展的趨勢而言，生物

資訊與生醫工程國際會議越來越受重視，而計算智慧是計算生物的重要技

術。

今年大會共有4個Room，16個Section，包含口頭報告區、演講區、

與海報展示。本次有四個重要的演講，包括（一）法國國家科學研究院

Athel教授演講Metabolic modeling: A Necessary Tool for Biotechnology；

（二）美國康乃爾大學Ann教授演講Fenton Oxidation of Contaminants using

Nanomagnetite；

（三）美國哈佛醫學院Chou教授演講有關An NMR view of

membrane transporters: application to mitochondrial carriers；

（四）以色列理

工學院Daniel教授演講A general overview of medical robotics，令參與者可

更深入瞭解此項領域的重大研究發展趨勢，提高了大家對這方面研究的瞭

解。很榮幸我們的論文被接受口頭報告。此科學論壇，為助長生物資訊與

生醫工程很重視分析工具、奈米元件與生醫應用的發展目前已經與眾多學

術和科學界的領導組織共同合作，合作學術單位遍及義大利、美國、臺灣、

中國、以色列、泰國、日本、摩洛哥、沙烏地阿拉伯、巴基斯坦、俄國、

捷克、伊朗、德國、土耳其、馬來西亞、波蘭、愛爾蘭、英國、法國、卡

達、加拿大、印尼、西班牙……等。本次投稿被iCBBE 2012接受之國際

會議論文亦有機會被轉投至生物資訊與生醫工程國際期刊。本次會議參加

的人員來自許多國家，包含大陸、美國、日本、澳洲、印度、新加坡、香

港、韓國、台灣、馬來西亞、泰國……等等，其中中國學者人數明顯增加，

大會安排讓來自各個國家的學者互相交流、聯誼，促進了與會學者日後的

學術交流機會。其間大會於第二天安排午宴讓來自各個國家的學者互相交

流、聯誼，期望能促進與會學者日後學術交流的機會。

我的口頭報告時段是19日早上8:30~12:00，我們這會場的發表主要是

蛋白質和蛋白質體在生物資訊的發展，因此大部分都會使用機器學習法來

進一步分析討論。我的論文報告主題是使用基因演算法的最佳化技術從

531個物化特性挑選出一組最佳解，結合support vector machine分類器，設

計出螢光蛋白質的預測器並進一步分析螢光蛋白質序列的物化特性在螢

光發光所扮演的功能。

第二篇論文由共同作者李華錦博士後研究員上台發表，而我在台下

參與全程。主題是使用計分卡（scoring card）設計醣結合蛋白質的預測器，

介紹此計分卡由智慧型基因演算法算出最佳化的一份計分卡，因此，只要

經過此簡單之計分卡即可以達到以往使用繁複計算方法才能達到的準確

度，大幅降低一般研究生物資訊計算所需的複雜度。

由於生物資訊研究需求整合資訊、數學與生物領域的廣泛知識，可

研究的題材也因此相當廣泛並且要有能整合跨領域知識的創意，所以聆聽

來自眾多不同地區的學者的研究成果，可有效率地吸收新知。

圖一：論文發表中

圖二：論文發表後合影（左起黃文玲教授、徐禮燊教授、李華錦博士後研

究員、何信瑩教授、何信璋教授與我）

2、與會心得

感謝國科會補助參加國際會議之出國補助，使本人得以出席跨領域生

物資訊國際會議，開拓眼界及促進國際觀。每次參加國計會議除了努力讓

世界知道臺灣人在研究方面非常認真與相當有能力為心則。我對於本次會

議重點 – 生物資訊與生醫工程特別感興趣，由其對很多研究主題及發展

方向也關心，希望對提高台灣的學術聲望及研究能量提升有所貢獻。

關於跨領域的交流，由此次國際會議舉辦的大陸上海，他們對跨領域

的企圖心及投入組織相當龐大，也可見他們對這領域的企圖心與團結。台

灣對跨領域的投入還需要更加努力，我們可以從他們的經驗作為參考與借

鏡，來促進國內的跨領域和產學合作的發展。

3、建議

近年來國科會、教育部和學校積極鼓勵年輕研究員，不僅是教師，還

有特別是博士班學生，參與大型國際會議。及早進入研究領域的核心，吸

取國際研究經驗，以提高國人的研究水準。參加國際生物資訊或是生醫工

程相關會議對老師及學生是非常重要的，會議中不但可以得到相關研究的

最新發展資訊，認識結交許多相關領域的學者，彼此做研究與心得交換，

更可找到跨領域的國際合作夥伴，在跨領域的生物資訊研究更為重要。希

望國家與學校單位能多在補助年輕學者出國，並且繼續擴大進行，以提升

國內研究的品質。

4、攜回資料

Proceedings of the 2012 6

th

International Conference on Bioinformatics

and Biomedical Engineering (iCBBE 2012), Shanghai, China, May 17-20,

2012. (含紙本與光碟)

Prediction of Carbohydrate-Binding Proteins

Using a Scoring Card Method

Institute of Bioinformatics and Systems Biology, National Chiao Tung University, Hsinchu, Taiwan*1 Department of Automation Engineering, National Formosa University, Yunlin, Taiwan*2 Department of Information Management, Overseas Chinese University, Taichung , Taiwan*3

Corresponding author: 886-3-571-2121, ext: 56910; e-mail: hlhuang @mail.nctu.edu.tw *4

Abstract—Carbohydrate-binding proteins play a pivotal role in a variety of important biological recognition processes. Compared with most studies of predicting binding sites, very few studies investigate prediction of carbohydrate-binding proteins. This paper proposes a highly interpretable scoring card method for predicting carbohydrate-binding proteins. First, a large-scale data set of carbohydrate-binding proteins (CBPDB) collected from three up-to-date databases, CAZy, CGF and Swiss-Prot is utilized. The data set CBPDB consists of 2380 positive and negative proteins with sequence identity 25% by removing sequence redundancy. Secondly, we adopt a novel scoring card method by way of generating an optimized scoring card of dipeptides to predict the carbohydrate-binding proteins. The prediction performance is promising with an independent test accuracy of 78.67%. The dipeptide score is helpful in discovering motif. The scoring card provides an insight of further analyzing the binding mechanism between carbohydrate and proteins.

Keywords—Binding, carbohydrate, genetic algorithm,

scoring card, protein, prediction.

I. INTRODUCTION

C

arbohydrates play an essential role in a variety of important biological recognition processes like infection, immune response, cell differentiation, and neuronal development. All of these biological phenomena may be regulated by the interaction of these carbohydrates with proteins [1-5]. Carbohydrate-binding proteins are becoming extremely useful in curing various illnesses. Experimental work for identifying carbohydrate-binding proteins is costly and time consuming. Therefore, effective computational methods for predicting carbohydrate-binding proteins are desirable.

It is vitally important to develop an automated and efficient method for timely identification of novel carbohydrate-binding proteins. However, some researches of using empirical rules [6] or machine learning methods [7] mainly focused on prediction and analysis of

carbohydrate-binding sites of proteins that are already known as carbohydrate-binding proteins.

Someya et al. [8] first clarified the definition carbohydrate-binding proteins and then constructed positive and negative datasets. Using both informative features and an appropriate classifier is essential to design an effective method for predicting carbohydrate-binding proteins based on the primary sequence only. They developed a carbohydrate-binding protein prediction system by using support vector machines (SVMs) [9], where the prediction of carbohydrate-binding proteins was formulated as a binary classification problem. Someya et al. [8] trained the SVM with three different encoding methods: a direct encoding method (AA-20), and two grouping methods (Levitt-6 and Someya-7). The SVM-based method with AA-20 performs well with a leave-one-out accuracy of 87% for the sequence with a sequence identity 35%.

Kumar et al. [10] developed SVM modules for distinguishing between cancer and non-cancerlectin proteins by using dipeptide composition, split composition, position specific scoring matrix (PSSM) profiles and PSSM with 14 PROSITE domains as input features.

The merits of this study are to 1) utilize a large-scale data set of carbohydrate-binding proteins collected from three up-to-date databases, CAZy, CGF and Swiss-Prot, 2) propose an easily interpretable method rather than the black-box-like SVM for biologists, and 3) obtain a robust performance with accuracy of 78.67% on carbohydrate-binding proteins with sequence identity 25%.

II. MATERIALS

A. Datasets

Carbohydrate binding proteins are obtained from the Consortium for Functional Glycomics (CFG) database and Carbohydrate Active Enzyme (CAZy) database. All the records from these two databases are served as positive protein samples which can bind carbohydrate. The Gene Ontology (GO) annotation terms about carbohydrate-binding functions are obtained from the GOA database. The GO term, carbohydrate binding function, and its child terms are collected. Finally, the number of GO terms which are defined as carbohydrate binding function is 778.

To obtain the negative dataset consisting of non-carbohydrate binding proteins, the Swiss-Prot database of release 2011_06, is also used. The Swiss-Prot database is divided into positive and negative datasets using GO terms mentioned above. If the GO terms of polypeptides in Swiss-Prot contain any GO terms defined as carbohydrate binding protein, the sequences are classified as positive samples. All sequences obtained from the three databases, CAZy, CGF and Swiss-Prot. The sequences from CAZy and CFG are carbohydrate-binding protein. Sequences from Swiss-Prot consist of carbohydrate- and non- carbohydrate-binding proteins.

The positive dataset is composed of 57330 polypeptides while the negative dataset is composed of 405046 polypeptides. USEARCH [11] is used to remove the sequence redundancy of the dataset. The threshold of USEARCH is set to 25%. After treating with USEARCH, the positive dataset contains 2380 polypeptides and the negative dataset contains 49647 polypeptides.

To avoid the unbalanced problem, the previous method [12] is used. The size of the negative dataset randomly chosen is equal to that of the positive dataset. Finally, the completed dataset contains 2380 polypeptides, shown in Table 1. The dataset is equally divided into the training and test datasets. Table 1. The numbers of sequences in the dataset CBPDB

stage positive negative Original 57330 405046 Identity threshold 25% 2380 49647

Chosen negative sequences 2380 2380 Final dataset CBPDB 1190 1190

III. METHODS

A novel scoring card method for predicting carbohydrate-binding proteins is proposed.

A. Construction of a scoring card

Figure 1 shows the data structures and experimental flow chart of the proposed scoring card method. The scoring card in arrow figure stands for the average of the statistic scoring cards.

The CBPDB dataset was equally divided into two parts, one for training and the other for independent test (outer loop in Fig. 1). Furthermore, the training data set is split into ten parts randomly (inner loop in Fig. 1) for ten-fold cross-validation. Therefore, the method can obtain ten validation results and ten statistic scoring cards. The statistic procedure of the scoring card method is described as follows:

1) Separate the data set into two classes of carbohydrate- and non-carbohydrate-binding proteins and calculate 400 dipeptide amounts for each class.

2) Due to the variance of sequence lengths in the two classes, the number of each dipeptide in one certain class is divided by the total number of dipeptides in this class. 3) A dipeptide in the carbohydrate-binding class got +1

score; otherwise, a dipeptide in the non-carbohydrate-binding class got -1 score. So the 400 scores in a scoring card can be derived from summation of all the dipeptide scores in two classes.

4) Normalized the scores as positive numbers into the range from 0 to 1000 in the scoring card.

Source Dataset 1/2 1/2 9/10 1/10 Scoring card Optimized scoring card Threshold decision by validation Independent Test Optimized scoring card IGA Scoring card

Fig. 1 Outline of the scoring card method

B. Threshold value determination

In order to find a best threshold to assort the data in two classes, the validation data in the inner loop were used. The amounts of 400 dipeptides of all samples in the validation data set were counted, and then the amounts were multiplied by the counterparts of dipeptide number in the scoring card. Finally, sum the 400 numbers up and the summation is divided by the sequence length to obtain a score of a sequence. The threshold with the highest accuracy in validation is chosen as the threshold value to classify two classes in the independent test.

In the test, the amounts of 400 dipeptides in a sample were multiplied by the counterparts of dipeptide number in the scoring card. Sum the 400 numbers up and the summation is divided by the sequence length of a protein to obtain a score for one sample. Equation 1 shows the calculation process. In Eq. 1, i and j = A, C, ..., Y, the 20 amino acids, W is the weighting of dipeptides of the test sample, and S is the score in the scoring card, and L is protein length. Fig.1 illustrates the process of test sample multiplied by the counterparts of dipeptide number in the scoring card. Then the query protein can be classified according to the threshold determined in the validation step. 20 , 1 Sample score ij ij i j W S L   



(1)

C. Scoring card

The ten statistics scoring cards in inner loop were averaged to one scoring card (the arrow illustration in Fig.1). Afterward the average scoring card was evaluated by ten validation datasets. Ten best thresholds would be derived from ten validations, and then the average of the ten thresholds was used in the classification of independent test. In the final step, there only had one scoring card, one threshold and one test result.

D. IGA-scoring card

The scoring card is further optimized by an intelligent genetic algorithm (IGA) [13] [14]. IGA utilizes an orthogonal array (OA) [15] that was used in the crossover operation to bring the better children than the traditional crossover method, and it can efficiently obtain a high-quality solution set. Orthogonal array is a fractional factorial array, which assures a balanced comparison of levels of any factor.

Ten validation data sets were used for evaluating the fitness function. The algorithm of generating the IGA-scoring card is described as follows:

Step 1: Initial population

The half of initial population in the IGA-scoring card consists of the ten statistic scoring cards in an inner loop, and the other ten individuals in the initial population were randomly generated from scores 0 to 1000. Therefore, the initial population comprises Npop individuals totally and Npop＝20.

Step 2: Evaluation

The fitness function of every individual was appraised via AUC from TPR and FPR.

TPR = TP/ (TP+FN) (2) FPR = FP/ (FP+TN) (3) Where TP is true positive and FP is false positive. In the ROC curve, X axle is FPR and Y axle is TPR. The validation data were used to find the best threshold according to the highest accuracy and it can get both TPR and FPR values from each threshold. TPR and FPR were used to draw the ROC curve and evaluate the fitness function for every individual.

Step 3: Selection

Use the traditional tournament selection that selects the winner from two randomly selected individuals to form a mating pool.

Step 4: Crossover

Select Pc•Npop parents from the mating pool to perform orthogonal array crossover on the selected pairs of parents where Pc is the crossover probability. Pc＝0.9.

Step 5: Mutation

Apply the real number mutation operator to the randomly mutated the gene from 0 to 1000 if the generated digit < Pm, where Pm is the mutation probability. Pm＝0.01. Step 6: Termination

When the IGA come to 100 generations is the stop condition and output the best individual as IGA-scoring card. Otherwise, go to Step 2.

This flow chart is divided into training and test processes. In the training part of the beginning, it used the training dataset to build a scoring card, then the scoring card is optimized by IGA with the validation dataset to get the best scoring card and threshold value. In the test process, the input sample could obtain a score through the scoring card.

A C D E F G H I K L M N P Q R S T V W Y A 568 924 805 686 449 639 496 818 514 1000 753 955 766 484 596 474 660 668 246 670 C 456 972 769 889 439 425 429 12 364 655 389 797 566 659 926 184 147 633 828 79 D 632 590 590 958 590 523 783 443 654 598 413 496 326 444 896 626 458 424 328 299 E 219 532 988 634 458 588 633 848 129 808 990 572 426 799 727 532 42 549 378 232 F 697 826 453 557 557 323 14 431 387 709 196 180 277 22 837 407 807 785 407 327 G 413 396 106 466 611 769 966 665 833 828 803 88 782 768 587 678 391 598 46 295 H 306 260 837 546 6 20 930 891 486 963 973 500 939 639 745 758 394 236 66 397 I 431 137 369 816 261 494 757 947 509 1000 487 689 225 574 499 477 633 338 222 326 K 560 519 447 546 726 778 693 663 583 530 685 469 126 256 691 527 546 433 20 47 L 265 503 723 855 594 656 293 844 665 467 529 348 346 752 787 463 590 666 27 342 M 980 467 621 996 443 969 436 655 654 987 940 907 505 389 831 987 903 230 443 451 100 N 490 438 358 594 893 0 140 531 330 479 206 367 580 299 413 202 348 256 258 250 200 P 295 64 424 465 88 62 617 466 536 327 502 369 877 365 512 459 442 56 343 615 300 Q 535 641 886 619 431 904 403 689 309 683 365 266 869 846 878 704 174 730 430 395 400 R 368 68 583 626 563 386 523 268 857 432 547 405 134 858 875 940 335 570 167 261 500 S 824 835 568 558 303 434 469 304 840 698 470 399 677 976 996 488 200 588 298 352 600 T 508 240 379 632 586 765 924 570 586 646 691 553 613 960 609 450 378 568 102 516 700 V 665 712 125 480 600 625 461 743 791 723 669 149 100 576 411 404 730 485 706 528 800 W 5 0 277 285 658 13 126 525 9 832 97 6 374 497 197 320 20 30 68 600 900 Y 593 278 203 84 24 13 60 137 207 664 342 23 244 352 312 607 126 203 819 52 1000

Fig. 2 The heat map of the IGA-scoring card

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1-Specificity S e n s it iv it y

Train ROC, AUC=0.872 Test ROC, AUC=0.857

Fig. 3 The training and test ROC curves

IV. RESULTS

A. Scoring card results

Figure 2 shows the best IGA-scoring card (with the highest training accuracy) of 150 generation optimizations from 25 independent runs. Figure 3 depicts the training and test ROC curves of the number-24 scoring card method. From the curves, the AUC of training and test are 0.872 and 0.857 and the accuracies of the training and independent test are 80.67% and 78.67%, respectively.