A hybrid neural network system for prediction and recognition of promoter regions in human genome

A hybrid neural network system for prediction and recognition of promoter regions in human genome first-author †E-mail:chuanboc@163.com; ljrlt@public.wh.hb.cn †E-mail:chuanboc@163.com; ljrlt@public.wh.hb.cn chuanboc@163.com †E-mail:chuanboc@163.com; ljrlt@public.wh.hb.cn †E-mail:chuanboc@163.com; ljrlt@public.wh.hb.cn chuanboc@163.com School of Computer Science & Technology, Huazhong University of Science & Technology, Wuhan 430074, China School of Computer Science & Technology, Huazhong University of Science & Technology, Wuhan 430074, China This paper proposes a high specificity and sensitivity algorithm called PromPredictor for recognizing promoter regions in the human genome. PromPredictor extracts compositional features and CpG islands information from genomic sequence, feeding these features as input for a hybrid neural network system (HNN) and then applies the HNN for prediction. It combines a novel promoter recognition model, coding theory, feature selection and dimensionality reduction with machine learning algorithm. Evaluation on Human chromosome 22 was ~66% in sensitivity and ~48% in specificity. Comparison with two other systems revealed that our method had superior sensitivity and specificity in predicting promoter regions. PromPredictor is written in MATLAB and requires Matlab to run. PromPredictor is freely available at http://www.whtelecom.com/Prompredictor.htm http://www.whtelecom.com/Prompredictor.htm . Hybrid neural network Promoter prediction Compositional features CpG islands INTRODUCTION 1 The publication and preliminary analysis of the human genome sequence (Lander et al., 2001 2001 ; Venter et al., 2001 2001 ) marks a significant milestone in the field of molecular biology. One of the main goals of the Human Genome Project is the characterization, annotation−recognition and categorization of genes from human genome to serve as a periodic table for biomedical research (Lander, 1996 1996 ). In the past few years, many efforts have been devoted to gene annotations. The National Center for Biotechnology Information (NCBI), Ensembl and Golden Path, for instance, provided the initial annotations, but the whole process of annotation is expected to go on for many years, and the current gene annotations only refer to protein-coding regions, relatively few tools have been developed to identify the regulatory regions required for the correct transcriptional activity of the genome. This task is particularly difficult in the case of eukaryotic organisms in which regulatory regions represent a small percentage, overwhelmed by presumably non-functional DNA. So prediction and characterization of regulatory regions is still a challenging problem. Here, we focus on detecting promoters, which are in the class of regulatory regions. A promoter is the region of genomic sequence proximal to the transcription start site (TSS) responsible for the initiation of transcription. In spite of the fact that characterizing regulation of gene expression is an important aspect of understanding gene function, for most human genes, promoters have not been defined or studied. Therefore, reliable recognition and characterization of promoters is a high priority goal in human genome study. Knowledge of promoters will be useful in elucidating regulation and expression mechanisms of genes and may shed light on the function of novel and uncharacterized genes. A well-established measure for promoter prediction accuracy scores a TSS prediction as positive if it is within the range of 200 bp upstream to 100 bp downstream of the true TSS (Fickett and Hatzigeorgiou, 1997 1997 ). Several research groups have developed methods for in silico promoter prediction, including knowledge-based methods, comparative genome analysis as well as methods based on statistical-compositional properties of DNA sequences (Fickett and Hatzigeorgiou, 1997 1997 ; Ohler and Niemann, 2001 2001 ). For most methods, the false-positive rate is roughly estimated at one per kilobase. In another aspect, the ratio of true prediction to false prediction is a small percent, with the exception of one method, PromoterInspestor, which shows predicted accuracy of 43% (Scherf et al., 2000 2000 ). In recent years, many efforts have been devoted to improve promoter predicted accuracy by using CpG islands association (Ioshikhes and Zhang, 2000 2000 ; Davuluri et al., 2001 2001 ; Hannenhalli and Levy, 2001 2001 ; Ponger and Mouchiroud, 2002 2002 ), combination with exon/intron/3′-UTR predictions (Bajic et al., 2002 2002 ; 2003 2003 ) and consensus promoter identification that combines several existing methods (Liu and David, 2002 2002 ). Motivated by these methods, we developed a new hybrid neural network system–the PromPredictor for human genome promoter recognition. It is a combination of a novel promoter recognition model, coding theory, feature selection and dimensionality reduction with machine learning algorithm. The method is based on the statistical concept of pentamer distributions in specific functional regions of DNA and selected the most significant pentamer vocabularies from training sequences by an unsupervised learning technique, in addition to CpG islands features. FEATURE EXTRACTION FROM DNA SEQUENCE 1 From a one-dimensional point of view, a DNA sequence contains characters from the 4-letter nucleic acid alphabet α={A, C, T, G}. An important issue in applying neural networks to promoter classification is how to encode DNA sequences, i.e., how to represent the DNA sequences as the input of the neural networks. In fact, sequences may not be the best representation at all. Good input representations make it easier for the neural networks to recognize underlying regularities. Therefore, good input representations are crucial to the success of neural network learning (Hirsh and Noordewier, 1994 1994 ). Compositional features 2 It is well known that genomes are characterized by species-specific compositional features, and that coding and non-coding DNA are distinguishable in terms of their pentamer and hexamer distributions (Claverie et al., 1990 1990 ). In promoter regions except core promoter elements such as TATA boxes, CAAT boxes and transcription initiation sites (INR), there exists a couple of other individual elements or sequence properties that are associated with promoter sequences. Among these are higher CpG content–CpG islands (Shago and Giguere, 1996 1996 ), secondary structure elements like the HIV-1 TAR regions (Bohjanen et al., 1997 1997 ), cruciform DNA structures (Wang et al., 1998 1998 ), or simple direct repeats (Bell et al., 1997 1997 ). Three-dimensional structures, such as curved DNA (Kim et al., 1995 1995 ), also influence promoter function. Most of these elements can be detected by computer-assisted sequence analysis (Chetouani et al., 1997 1997 ; Schuster et al., 1997 1997 ; Nakaya et al., 1995 1995 ; Nielsen et al., 1995 1995 ), but none of them is really promoter specific and can be found frequently outside of promoters. The secret of promoter function lies in the combination of several promoter elements that must cooperate in transcriptional activation, while none of them can achieve alone. This also summarizes the main problem of promoter recognition. It is necessary to compile several individual weak signals into a composite signal which then indicates a potential promoter. As an attempt, we use pentamer frequency coding method to capture core elements as well as weak signals. The pentamer encoding method extracts various patterns of five consecutive nucleic acids in a DNA sequence and counts the number of occurrences of the extracted pentamer. For instance, given a DNA sequence CGAATCG, the pentamer encoding method gives the following results: 1 for CGAAT (indicating CGAAT occurs once), 1 for GAATC and 1 for AATCG. For each DNA sequence, there are 4 5 =1024 possible pentamers in total. If all the 1024 pentamers were chosen as input features of the neural network, it would require many weight parameters and training data, which makes it difficult to train the neural network–a phenomenon called “curse of dimensionality”. Different methods have been proposed to solve the problem by careful feature selection and by scaling of the input dimensionality (Chuzhanova et al., 1998 1998 ). What we are proposing here is to select relevant features by employing a distance measure to calculate the relevance of each feature (Dash and Liu, 1997 1997 ). Let X be a feature and x be its value. Let p ( x | Class =1) and p ( x | Class =0) denote the class conditional density functions for feature X , where Class_ 1 represents the positive class and Class_ 0 is the negative class. Let D ( X ) denote the distance function between p ( x | Class =1) and p ( x | Class =0), defined as (Bassat, 1982 1982 ) $$D(X) = \int {\left| {p(x|Class = 1) - p(x|Class = 0)} \right|{\text{d}}x}$$ The distance measure prefers feature X to feature Y if D ( X ) > D ( Y ). Intuitively, this means that it is easier to distinguish between Class_ 1 and Class_ 0 by observing feature X than by observing feature Y . That is, X appears often in Class_ 1 but seldom in Class_ 0 or vice versa. In our work, each feature X is a pentamer. Let c denote the occurrence number of the feature X in a sequence S . Let l denote the total number of pentamers in S and len ( S ) represents the length of S . We have l = len ( S )−4. Define the feature value x for the pentamer X with respect to the sequence S as $$x = \frac{c}{{len(S) - 4}}$$ Since a promoter sequence may be short, random pairings can have a large effect on the result. D ( X ) in Eq.( 1 1 ) can be approximated by the Mahalonobis distance (Solovyev and Makarova, 1993 1993 ) as $$D(X) = \frac{{(m_1 - m_0 )^2 }}{{d_1^2 + d_0^2 }}$$ where m 1 and d 1 ( m 0 and d 0 , respectively) are the mean value and the standard deviation of the feature X in the positive (negative, respectively) training data set. Intuitively, in Eq.( 3 3 ), the larger the numerator is (or the smaller the denominator is), the larger the interclass distance is, and therefore the easier to separate Class_ 1 from Class_ 0 (and vice versa). The mean value m and the standard deviation d of the feature X in a set X of sequences are defined as $$m = \frac{1}{N}\sum\limits_{i = 1}^N {x_i }$$ $$d = \sqrt {\frac{1}{{N - 1}}\sum\limits_{i = 1}^N {(x_i - m)^2 } }$$ where x i is the value of the feature X with respect to sequence S i ∈ X , and N is the total number of sequences in X . Let X 1 , X 2 ,…. X Na , N a < 1024, be the top N a features (pentamers) with the largest D ( X ) values. Intuitively, these N a features occur more frequently in the positive training data set and less frequently in the negative training data set. For each DNA sequence S (whether it is a training or an unlabeled test sequence), we examine the N a features in S , calculate their values as defined in Eq.( 2 2 ), and use the N a feature values as input feature values to the HNN for the sequence S . To compensate for the possible loss of information due to ignoring the other pentamers, a linear correlation coefficient ( LCC ) between the values of the 1024 pentamers with respect to the DNA sequence S and the mean value of the 1024 pentamers in the positive training data set is calculated and used as another input feature value for S . Specifically, the LCC of S is defined as $$LCC(S) = \frac{{1024\sum\limits_{i = 1}^{1024} {x_i \overline {x_i } } - \sum\limits_{i = 1}^{1024} {x_i } \sum\limits_{i = 1}^{1024} {\overline {x_i } } }}{{\sqrt {1024\sum\limits_{i = 1}^{1024} {x_i^2 } - \left( {\sum\limits_i^{1024} {x_i } } \right)^2 } \sqrt {1024\sum\limits_{i = 1}^{1024} {\overline {x_i } ^2 } - \left( {\sum\limits_i^{1024} {\overline {x_i } } } \right)^2 } }}$$ where x ̄ i is the mean value of the i th pentamer, 1 < i < 1024, in the positive training dataset, and x i is the feature value of the i th pentamer with respect to S as defined in Eq.( 2 2 ). CpG islands features 2 In the human genome, many genes were recognized and validated successfully (Lander et al., 2001 2001 ; Venter et al., 2001 2001 ) by using the so-called CpG islands as gene markers. CpG islands are unmethylated segments of DNA longer than 200 bp, with a G+C content of at least 50%, and the number of CpG dinucleotides being at least 60% of what could be expected from the G+C content of the segment (Bird et al., 1987 1987 ; Gardiner and Frommer, 1987 1987 ; Larsen et al., 1992 1992 ; Cross and Bird, 1995 1995 ). CpG islands are found around a gene that starts in approximately half of mammalian promoters (Larsen et al., 1992 1992 ; Cross and Bird, 1995 1995 ) and are estimated to be associated with ~60% of human promoters (Cross et al., 1999 1999 ). For this reason, Pedersen et al.( 1999 1999 ) suggested that CpG islands could represent a good global signal to locate promoters across genomes. At least in mammalian genomes, CpG islands are good indicators of gene presence. In our prediction system, we use two CpG island features−G+C content and ratio of expected to observed CG dinucleotides (i.e. Obs/Exp). Let len represent the length of one segment of a DNA sequence, the G+C content ( GC _ con ) and Obs/Exp ( o / e ) (Gardiner and Frommer, 1987 1987 ) are defined as $$GC\_con = \frac{{{\text{number of C's}} + {\text{number of G's}}}}{{len}}$$ $$o/e = \frac{{{\text{number of CG's}} \times len}}{{{\text{(number of C's}} \times {\text{number of G's) }}}}$$ ARCHITECTURE OF THE PREDICTION SYSTEM 1 The conceptual structure of our system is depicted in Figs. 1 1 and 2 2 . The overall system shown in Fig. 1 1 comprises a collection of four basic classifiers: promoter_classifier, exon_classifier, intron_classifier and 3′-UTR_classifier. Each of the classifiers is a modified BP neural network and has the same structure. The basic classifier of promoter is shown in Fig. 2 2 . Each classifier is trained by different training sets and the parameters for each classifier are optimized independently. An unknown sequence is partitioned into windows 250 bp long, shifted by 1 bp. For each sliding data window, we compute the feature values following procedures in the previous section and these feature values are used as input of the neural network. The prediction system assigns the sequence to the class promoter if three classifiers–exon, intron, 3′-UTR decide that the sequence is not an exon, intron, 3′-UTR respectively, and only promoter_classifier decides that the sequence belongs to this class. Sequence training and test set 2 From the vertebrate section of the Eukaryotic Promoter Database (EPD), V.79.0 (Cavin et al., 1998 1998 ), promoter sequences from 200 bp upstream to 50 bp downstream of the TSS were taken. Vertebrate exon and vertebrate intron sequences of different location, covering a total of 5×10 6 bp in each set were randomly extracted from GenBank. Vertebrate 3′-UTR sequences with a total of 3×10 6 bp were extracted from the UTR database (Pesole et al., 2002 2002 ). Sequences training set for the four basic classifiers (promoter, exon, intron and 3′-UTR) were created by randomly extracting non-overlapping sequences of 250 bp from the four example sets mentioned above. Redundant sequences were deleted by the program CLEANUP (Grillo et al., 1996 1996 ) which resulted in sets consisting of 1837 sequences from promoter regions, 4500 exon sequences, 6500 intron sequences and 5000 3′-UTR sequences. In these four sets, 2/3 of the sequences were used for training, and the rest were used for validation. System training and parameter optimization 2 According to the definitions above, the number of N a , the neuron number of hidden layers and the training algorithm must be determined for each classifier. Furthermore, an optimal assignment threshold must be calculated. The training for each classifier is independent. For example, the positive training set for exon_classifier is exon, and the negative training set includes intron, promoter and 3′-UTR. We tested the exon_classifier’s performance with different parameters and different training algorithms and recorded the optimized parameters based on accuracy and computer time. The training algorithms include Gradient descent algorithms, Conjugate gradient (CGB) algorithms (Powell, 1977 1977 ), Quasi-Newton algorithm, One Step Sccant (OSS) algorithm (Battiti, 1992 1992 ), Resilient backpropagation (RPROP) algorithm (Riedmiller and Braun, 1993 1993 ) and Levenberg-Marquardt (LM) algorithm (Hagan and Menhaj, 1994 1994 ). Table 1 1 summarizes the default threshold, corresponding training algorithm and the optimized parameters for four basic classifiers. Table 2 2 shows the validation results. From this table, we can see that the combination of four classifiers can improve the sensitivity in predicting promoter regions obviously. Basic classifiers N a Neuron numbers in hidden layer Training algorithm Default threshold Promoter_classifier 900 3 Levenberg-Marquardt algorithm 0.9368 Exon_classifier 800 2 Resilient backpropagation 0.9470 Intron_classifier 800 2 Resilient backpropagation 0.9353 3′-UTR_classifier 800 2 Resilient backpropagation 0.9170 Number of sequences One class of classifier assigns an unknown sequence to this class when the predicted value is above threshold PromPredictor assigns an unknown sequence to the class promoter if three classifiers–exon, intron, 3′-UTR decide that the sequence is not an exon, intron, 3′-UTR respectively, and only promoter_classifier decides that the sequence belongs to this class 55.4%=484/(484+27+85+277) 69.4%=422/(422+11+19+156) 2 Sets 2 N # # 5 Number of predicted sequences above threshold a a Promoter_classifier (threshold=0.8) Exon_classifier (threshold=0.8) Intron_classifier (threshold=0.8) 3′-UTR_classifier (threshold=0.8) PromPredictor b b Promoter 612 484 20 74 81 422 Exon 1500 27 1485 207 214 11 Intron 2166 85 678 2144 912 19 3′-UTR 1666 277 452 523 1489 156 2 Sensitivity: S e (%) 55.4 c c 69.4 d d EXPERIMENTS AND RESULTS 1 In order to compare the performance of our system with two other promoter recognition systems: PromoterInspestor (Scherf et al., 2000 2000 ) and Dragon Promoter Finder (DPF) (Bajic et al., 2002 2002 ; 2003 2003 ), we used the same evaluation set in PromoterInspestor and DPF. The first set (SET 1) consisted of six Genbank genomic sequences with a total length of 1.38 Mb and 35 known TSSs on these sequences. The second set (SET 2) consisted of publicly available sequence for human chromosome 22–the 35 Mb sequence with 377 known genes. The annotation data (Rel. 2.3) for human chromosome 22 were produced by the Chromosome 22 Gene Annotation Group at the Sanger Centre and were obtained from the world wide web ( http://www.sanger.ac.uk/HGP/Chr22/ http://www.sanger.ac.uk/HGP/Chr22/ ). For the former dataset a promoter region was counted as true positive (TP), if a transcription start site (TSS) was located within or up to 200 bp downstream of the predicted promoter region. Or otherwise the promoter region was counted as false positive (FP). For the latter dataset we used the same method as that used by Scherf et al.( 2000 2000 ) with PromoterInspector: all the predictions located in the range −2000~+500 around the 5′ extremity of a known gene were considered as a true positive promoter region (TP). While the FP predictions were considered as those that fall on the annotated part of the human chromosome 22 covered by known genes, but not sufficiently close to the 5′ end of these genes, thus not representing the TP predictions. The main results and comparisons are summarized in Tables 3 3 – 6 6 . Results are given with respect to several criteria related to the maximum allowed distance between the predicted promoter region and the real TSS. In these experiments, PromPredictor, DPF and PromoterInspector were used with their default parameter settings. The coverage describes the percentage of true promoters in a sequence that has been predicted by PromPredictor Sequence accession number Length (bp) Number of TSS TP FP Coverage * * (%) AC002397 227538 17 6 9 35 L44140 219447 11 8 26 73 D87675 301692 1 1 2 100 AF017257 101569 1 1 1 100 AF146793 204625 4 1 4 25 AC002368 324816 1 1 1 100 Sensitivity: S e = TP /( TP + FN ) Specificity: S p = TP /( TP + FP ) Correlation Coefficient ( CC ): $$CC = \frac{{TP \times TN - FP \times FN}}{{\sqrt {(TP + FP) \times (TP + FN) \times (TN + FP) \times (TN + FN)} }}$$ FN : false negative; TN : true negative The number of FP predictions on SET 1 made by PromoterInspector is 55 by our criteria because strand-nonspecific predictions were counted twice (once for each DNA strand) Program TP FP S e (%) a a S p (%) b b CC (%) c c PromoterInspector 15 55 d d 42.8 21.4 30.3 DPF( S e =0.37) 16 79 45.7 16.8 27.7 PromPredictor 18 43 51.4 29.5 38.9 Represents the total number of predictions in the two-strand search and with strand-specific counting of predictions SET 2 performance PromPredictor TP 248 FP 274 Total number of human gene 377 Total number of predictions * * 869 The average length of predicted promoter region 568 Program S e (%) S p (%) PromoterInspector 45 33 DPF 64 33 PromPredictor 66 48 CONCLUSION 1 In this paper we presented a hybrid neural network approach–PromPredictor for predicting promoter regions in large genomic sequences. It is based on a new promoter model with statistical-compositional features and CpG information and integrates multi-classifier via HNN. The prediction accuracy of PromPredictor, achieved on a large and diverse evaluation-set, shows our novel method is promising for modeling biological systems in general, which does not require any specific knowledge about a particular promoter to make a prediction and thus has a big advantage especially when nothing is known about the promoter to be predicted. Future research should include combination of other groups of signals that may characterize some aspects of gene structure, such as translation initiation site, splice sites or polyA site. References VB Bajic , SH Seah , A Chong , G Zhang , JLY Koh , V Brusic . Dragon Promoter Finder: recognition of vertebrate RNA Polymerase II promoters . Bioinformatics , 2002 . 18 198 - 199 . . Dragon Promoter Finder: recognition of vertebrate RNA Polymerase II promoters VB Bajic , SH Seah , A Chong , SPT Krishnan , JLY Koh , V Brusic . Computer model for recognition of functional transcription start sites in RNA polymerase II promoter of vertebrates . Journal of Molecular Graphic and Modeling , 2003 . 21 323 - 332 . . Computer model for recognition of functional transcription start sites in RNA polymerase II promoter of vertebrates MB Bassat . PR Krishnaiah , LN Kanal . Use of Distance Measures, Information Measures and Error Bounds in Feature Evaluation . Classification, Pattern Recognition and Reduction of Dimensionality: Handbook of Statistics , 1982 . Volume 2 : Amsterdam North-Holland Publishing Company , 773 - 791 . . Use of Distance Measures, Information Measures and Error Bounds in Feature Evaluation R Battiti . First and second order methods for learning: Between steepest descent and Newton’s method . Neural Computation , 1992 . 4 ( 2 ): 141 - 166 . . First and second order methods for learning: Between steepest descent and Newton’s method PJL Bell , VJ Higgins , IW Dawes , PH Bissinger . Tandemly repeated 147 bp elements cause structural and functional variation in divergent MAL promoters of Saccharomyces cerevisiae . Yeast , 1997 . 13 1135 - 1144 . . Tandemly repeated 147 bp elements cause structural and functional variation in divergent MAL promoters of Saccharomyces cerevisiae AP Bird , MH Taggart , RD Nicholls , DR Higgs . Non-methylated CpG-rich islands at the human α-globin locus: Implications for evolution of the α-globin pseudogene . EMBO J , 1987 . 6 999 - 1004 . . Non-methylated CpG-rich islands at the human α-globin locus: Implications for evolution of the α-globin pseudogene PR Bohjanen , Y Liu , MA GarciaBlanco . TAR RNA decoys inhibit Tat-activated HIV-1 transcription after preinitiation complex formation . Nucleic Acids Res , 1997 . 25 4481 - 4486 . . TAR RNA decoys inhibit Tat-activated HIV-1 transcription after preinitiation complex formation PR Cavin , T Junier , P Bucher . The Eukaryotic Promoter Database EPD . Nucleic Acids Res , 1998 . 26 353 - 357 . . The Eukaryotic Promoter Database EPD F Chetouani , P Monestié , P Thébault , C Gaspin , B Michot . ESSA: an integrated and interactive computer tool for analyzing RNA secondary structure . Nucleic Acids Res , 1997 . 25 3514 - 3522 . . ESSA: an integrated and interactive computer tool for analyzing RNA secondary structure NA Chuzhanova , AJ Jones , S Margetts . Feature selection for genetic sequence classification . Bioinformatics , 1998 . 14 139 - 143 . . Feature selection for genetic sequence classification JM Claverie , I Sauvaget , L Bougueleret . K-tuple frequency analysis from intron/exon discrimination to Tcell epitope mapping . Methods Enzimol , 1990 . 183 237 - 252 . . K-tuple frequency analysis from intron/exon discrimination to Tcell epitope mapping SH Cross , AP Bird . CpG islands and genes . Curr Opin Genet Dev , 1995 . 5 309 - 314 . . CpG islands and genes SH Cross , VH Clark , AP Bird . Isolation of CpG islands from large genomic clones . Nucleic Acids Res , 1999 . 27 2099 - 2107 . . Isolation of CpG islands from large genomic clones M Dash , H Liu . Feature selection for classification . Intelligent Data Analysis , 1997 . 3 1 - 6 . . Feature selection for classification RV Davuluri , I Grosse , MQ Zhang . Computational identification of promoters and first exons in the human genome . Nature Genetics , 2001 . 29 412 - 417 . . Computational identification of promoters and first exons in the human genome JW Fickett , AG Hatzigeorgiou . Eukaryotic promoter recognition . Genome Res , 1997 . 7 861 - 878 . . Eukaryotic promoter recognition GM Gardiner , M Frommer . CpG islands in vertebrate genomes . J Mol Biol , 1987 . 196 261 - 282 . . CpG islands in vertebrate genomes G Grillo , M Attimonelli , S Liuni , G Pesole . CLEANUP: a fast computer program for removing redundancies from nucleotide sequence databases . Comput Applic Biosci , 1996 . 12 1 - 8 . . CLEANUP: a fast computer program for removing redundancies from nucleotide sequence databases MT Hagan , M Menhaj . Training feedforward networks with the Marquardt algorithm . IEEE Transactions on Neural Networks , 1994 . 5 ( 6 ): 989 - 993 . . Training feedforward networks with the Marquardt algorithm S Hannenhalli , S Levy . Promoter prediction in the human genome . Bioinformatics , 2001 . 17 90 - 96 . . Promoter prediction in the human genome H Hirsh , M Noordewier . Using Background Knowledge to Improve Inductive Learning of DNA Sequences . Proceedings of the Tenth Annual Conference on Artificial Intelligence for Applications , 1994 . San Antonio 351 - 357 . . Using Background Knowledge to Improve Inductive Learning of DNA Sequences IP Ioshikhes , MQ Zhang . Large-scale human promoter mapping using CpG islands . Nature Genetics , 2000 . 26 61 - 63 . . Large-scale human promoter mapping using CpG islands J Kim , S Klooster , DJ Shapiro . Intrinsically bent DNA in a eukaryotic transcription factor recognition sequence potentiates transcription activation . J Biol Chem , 1995 . 270 1282 - 1288 . . Intrinsically bent DNA in a eukaryotic transcription factor recognition sequence potentiates transcription activation ES Lander . The new genomics: global views of biology . Science , 1996 . 274 536 - 539 . . The new genomics: global views of biology ES Lander , LM Linton , B Birren , C Nusbaum , MC Zody , J Baldwin , K Devon , K Dewar , M Doyle , W Fitzhugh . Initial sequencing and analysis of the human genome . Nature , 2001 . 409 860 - 921 . . Initial sequencing and analysis of the human genome F Larsen , G Gundersen , R Lopez , H Prydz . CpG islands as gene markers in the human genome . Genomics , 1992 . 13 1095 - 1107 . . CpG islands as gene markers in the human genome RX Liu , J David . Consensus promoter identification in the human genome utilizing expressed gene markers and gene modeling . Genome Res , 2002 . 3 462 - 469 . . Consensus promoter identification in the human genome utilizing expressed gene markers and gene modeling A Nakaya , K Yamamoto , A Yonezawa . RNA secondary structure prediction using highly parallel computers . Comp Appl Biosci , 1995 . 11 685 - 692 . . RNA secondary structure prediction using highly parallel computers DA Nielsen , A Novoradovsky , D Goldman . SSCP primer design based on single-strand DNA structure predicted by a DNA folding program . Nucleic Acids Res , 1995 . 23 2287 - 2291 . . SSCP primer design based on single-strand DNA structure predicted by a DNA folding program U Ohler , H Niemann . Identification and analysis of eukaryotic promoters: recent computational approaches . TRENDS Genet , 2001 . 17 56 - 60 . . Identification and analysis of eukaryotic promoters: recent computational approaches AG Pedersen , P Baldi , Y Chauvin , S Brunak . The biology of eukaryotic promoter prediction–A review . Comput Chem , 1999 . 23 191 - 207 . . The biology of eukaryotic promoter prediction–A review G Pesole , S Liuni , G Grillo , F Licciulli , F Mignone , C Gissi , C Saccone . UTRdb and UTRsite: specialized database of sequences and functional elements of 5′ and 3′ untranslated regions of eukaryotic mRNAs. Update 2002 . Nucleic Acids Res , 2002 . 30 335 - 340 . . UTRdb and UTRsite: specialized database of sequences and functional elements of 5′ and 3′ untranslated regions of eukaryotic mRNAs. Update 2002 L Ponger , D Mouchiroud . CpGProD: identifying CpG islands associated with transcription start sites in large genomic mammalian sequences . Bioinformatics , 2002 . 18 631 - 633 . . CpGProD: identifying CpG islands associated with transcription start sites in large genomic mammalian sequences MJD Powell . Restart procedures for the conjugate gradient method . Mathematical Programming , 1977 . 12 241 - 254 . . Restart procedures for the conjugate gradient method M Riedmiller , H Braun . A Direct Adaptive Method for Faster Backpropagation Learning: The RPROP Algorithm . Proceedings of the IEEE International Conference on Neural Networks , 1993 . San Francisco . A Direct Adaptive Method for Faster Backpropagation Learning: The RPROP Algorithm M Scherf , A Klingenhoff , T Werner . Highly specific localization of promoter regions in large genomic sequences by PromoterInspector: a novel context analysis approach . J Mol Biol , 2000 . 297 599 - 606 . . Highly specific localization of promoter regions in large genomic sequences by PromoterInspector: a novel context analysis approach P Schuster , PF Stadler , A Renner . RNA structures and folding: from conventional to new issues in structure predictions . Curr Opin Struct Biol , 1997 . 7 229 - 235 . . RNA structures and folding: from conventional to new issues in structure predictions M Shago , V Giguere . Isolation of a novel retinoic acid-responsive gene by selection of genomic fragments derived from CpG-island enriched DNA . Mol Cell Biol , 1996 . 16 4337 - 4348 . . Isolation of a novel retinoic acid-responsive gene by selection of genomic fragments derived from CpG-island enriched DNA VV Solovyev , KS Makarova . A novel method of protein sequence classification based on oligopeptide frequency analysis and its application to search for functional sites and to domain localization . Computer Applications in the Biosciences , 1993 . 9 ( 1 ): 17 - 24 . . A novel method of protein sequence classification based on oligopeptide frequency analysis and its application to search for functional sites and to domain localization JC Venter , MD Adams , EW Myers , PW Li , RJ Mural , GG Sutton , HO Smith , M Yandell , CA Evans , RA Holt . The sequence of the human genome . Science , 2001 . 291 1304 - 1351 . . The sequence of the human genome WD Wang , TH Chi , YT Xue , S Zhou , A Kuo . Architectural DNA binding by a high-mobility-group/kinesin-like subunit in mammalian SWI/SNF-related complexes . Proc Natl Acad Sci USA , 1998 . 95 492 - 498 . . Architectural DNA binding by a high-mobility-group/kinesin-like subunit in mammalian SWI/SNF-related complexes 10.1631/jzus.2005.B0401 Project (No. 2001AA231071) supported by the Hi-Tech Research and Development Program (863) of China 2004-10-08 2005-03-07 2005-04-10 2022-01-05 Copyright © 2005, Journal of Zhejiang University Science 2005 Journal of Zhejiang University-SCIENCE B (Biomedicine and Biotechnology) 05 6B