*) a ₁, false negatives; a ₂; false positives

A simple recognition method is based on the representation of transcription factor binding sequences as a set of site realizations R={R₀, R₁, ..., R_k-1}. In other words, the set of realizations in the form of oligonucleotides t bases long coded in the IUPAC-IUB 15 single-letter based codes is created by our algorithm for any particular site. This approach avoids the averaging of the nucleotide composition, as occur while describing the functional site by the weight matrix or consensus. The program SiteGroup uses a set U₀={u₁,...,u_m} of experimentally determined binding site sequences extracted from the TFBSC as the initial information to create the set of realizations R. It is determined by two parameters: (1) the length of the oligonucleotide t and (2) the maximally allowable difference (distance) t^(miss) between these oligonucleotides. The main realization is the oligonucleotide R₀ of length t having the highest frequency in the sample U₀. Then, all sequences u_i containing R₀ are removed from the sample U_0, producing the sample U₁ and so on. Analysis of the sample U_r, during which the rth realization R_r is searched for, is performed at the rth step. In this iterative process, all oligonucleotide words of length t are considered. The distance t from the word R₀ is estimated for each word. The word R_r that has the least distance from the R₀ word is selected. If a number of words have the same minimal distance, the word with the maximal frequency in the set U_r is selected. The iteration is stopped when the set U_r either becomes empty or contains only the words that differ from R₀ by the value exceeding t^(miss). Each set of realizations created by the method described may be characterized by parameter f_ij: the proportion of the sequences from the initial set U₀ represented in this set of realizations (covering of the set U₀). The set of realizations providing maximization of the functional is searched for by exhaustion of the pairs (t^(miss),t )=(i,j).

Expert Weights (0-10 are valid; 5 employed automatically)

1. Translation increases with decreasing the Leader length
2. Translation increases with decreasing the G/C ratio
3. Translation increases with increasing the G/C-imbalance
4. Translation increases with decreasing the alt-AUG content
5. Translation increases with decreasing the framed AUG content
6. Translation increases depending on the "-3 position" rule
7. Translation increases with decreasing the AUG inside leader
8. Translation increases with increasing the [C] content
9. Translation increases with increasing the [YM] content
10. Translation increases with increasing the[CnY] content
11. Comparison with the weight matrices for nucl. content in 5’UTRs of high expression mRNAs.
12. Comparison with complex (high to low) weight matrices for nucl. content in 5’UTRs of high and low expression mRNAs

Figure 5. List of 5’UTR mRNA characteristics important for predicting the level of gene expression. Parameters 8-12 were determined for the (-35;-1) 5’UTR fragment.

q_ij=f_ij* [(f_ij-f_ij-1)+(f_ij-f_ij+1)]. (1)

Examples of the first and second type errors in recognition of several transcription factor binding sites are listed in Table 2. For assessing the second type errors, the compilation of eukaryotic non-first exons was used. Small first and second type errors were observed for the SiteScan recognition. It is especially important for analysis of transcription factor binding sites in long genomic sequences.

COMPUTER SYSTEM FOR PREDICTING mRNA TRANSLATION ACTIVITY

This part of the GeneExpress system is designed for prediction of mRNA translation efficiency basing on analysis of structural and contextual features of 5’ untranslated regions (5’UTRs). It has a program (Leader) for mRNA translation rate prediction and the database containing 5’UTR sequences (Leader_Sq) and some information on the effect of several 5’UTR features on mRNA translation efficiency.

Eukaryotic mRNAs differ considerably in their translation efficiency. This has been attributed to different efficiency of translation initiation. The contextual and structural features of 5’UTRs have strong effect on translation initiation. To reveal these characteristics, we have compared the mRNA sequences of several house-keeping gene groups, highly expressed in eukaryotic cells, and some groups of regulatory genes, whose expression is low and under stringent control. The group of highly expressed mRNAs consists of mRNAs of highly abundant proteins such as actins, tubulins, ribosomal proteins, histones, hsp70, etc.

Figure 6. The control results were obtained using a set of independent data. Broken line is the selected threshold to separate low and high expressed mRNA.

Low expression mRNAs include mRNAs of transcription factors, protein kinases, growth factors, protooncogenes, etc. We have found several features that are different for these two groups (Fig. 5): 5’UTR length, nucleotide composition, context of start AUG codon, and presence of AUGs within 5’UTRs (Ischenko I.V. et al., 1996; Kochetov A.V. et al., 1998). These 5’UTR features affect the 40S ribosomal subunit movement along the leader and, therefore, the efficiency of the translation initiation.

The difference in these features was used for discrimination between high and low expressed genes The program calculates the 5’UTR features and evaluates the translation activity of mRNA. We create a simple discrimination function based on Penrose distance. The discrimination between the control samples of the high and low expressed mRNAs of dicot plants showed that 84% of the high and 76% of the low expressed mRNAs were classified correctly (Fig. 6).

Initial postulates. It is suggested that the site activity F is determined by context-dependent properties of its nucleotide sequence S: statistical, physical, and conformational (M.P. Ponomarenko et al., 1997a, N.A. Kolchanov et al., 1998). These properties are of two types (A.E. Kel et al., 1993): (1) obligatory, which are invariant for all sequences S_n of the site and determine its basal activity; and (2) facultative, which are individual in terms of their “number, size, and location” for each sequence of the site and modulate the site activity with respect to the basal level. Hence, within the framework of the linear-additive approximation, the activity of the site with sequence S may be described by the following equation:

Figure 7. Principal scheme of the ACTIVITY computer system.

; (2)

where F₀(S) is the basal activity level determined by the occurrence of the obligatory properties of this site in the sequence S_n; {X_k}_k=1,K are the facultative properties; and F_k is the contribution of the facultative property X_k to the site activity F. The principal scheme of the ACTIVITY system is shown in Fig. 7.

Figure 8. Description of conformational property “Helical twist angle of B-DNA” in the ACTIVITY system database.

Database on functional site activities compiles the available data on functional sites with the experimentally measured activities: over 240 site samples of different types, such as promoters and binding sites for E. coli regulatory proteins, TATA boxes and binding sites for eukaryotic transcription factors, translation starts, splicing and 3’-processing sites, etc. Site activity characteristics include the association/dissociation rates of DNA-protein complexes, their lifetimes, equilibrium constants, transcription and translation efficiencies, etc.

Database on conformational and physical/chemical DNA properties compiles the information on context-dependent properties which may play a significant role in DNA-protein interactions. The format of this database is illustrated in Fig. 8. The database currently contains over 40 conformational parameters, determined either computationally or by X-ray analysis. Over 10 physical DNA properties -- melting temperature, persistent length, bending-rigidity, entropy, etc. -- are also included. The system for knowledge discovery on site activity contains two blocks. The first is responsible for revealing any site properties significant for predicting the site activity and the second provides generation of C-code programs to predict the activity of a given sequence.

One example of the context characteristics of the sequence S is the positional weighted concentration X_zmw(S) of mono-, di-, tri-, and tetranucleotides (Ponomarenko et al., 1997b):

, (3)

where d _Z is the "1" or "0" indicator function depending on the match or mismatch between the sequence S and the oligonucleotide Z; w(i) is the function of position effect determined according to the rule: “the more important is the position i for the site activity, the higher its assigned weight w(i)”. The activity prediction employs 180 various weight functions w(i). They are stored in the database for contextual characteristics of the ACTIVITY system (Fig. 9). Three examples of weight functions w(i) are shown in Fig. 10. The nucleotide composition of any oligonucleotide is presented in the 15 single-letter based code.

We also use the mean values of site S properties within the region [a;b] as site characteristics:

. (4)

Search for statistically significant characteristics is implemented as analysis of conformational and physical characteristics for all oligonucleotides with lengths from 1 to M, checking each of the 180 position effect functions w(i). The total number of combinations <Z, m, w> is about 10⁷ for m=4. Similarly, for every conformational or physical property q, the analysis of all possible locations of the region [a, b] within the site sequence is performed. X_qab(S_n) is calculated for a fixed combination <q, a, b> for each sequence of the site. The total number of the combinations <q, a, b> is 10⁵.

The significance of a property for site activity prediction is estimated (Ponomarenko et al., 1997b) within the frames of the Utility Theory for Decision Making (P.C. Fishburn et al., 1970). Let’s calculate the fixed property X_zmw(S_n) for each sequence S_n with the known activity F_n. If the resulting pairs {X_zmw(S_n), F_n} meet all the necessary conditions of the linear regression applicability, then the activity F is predictable from an arbitrary sequence S via the feature X_zmw. To test these conditions of linear regression applicability, a simple regression is optimized for the pair {X_zmw(S_n), F_n}:

;(5)

where f₀ and f₁ are the regression optimized coefficients.

To ensure the reliability of the regression between X_zmw(S_n) and F_n, 22 conditions of regression analysis are tested (the presence of linear, sign, and rank correlations between the predicted F_zmw(S_n) and the experimental F_n activities; the equality of X_zmw(S_n) and F_n distributions, etc.).The significance level a _rt at which the rth condition is met is estimated. Then, the partial utility of the feature X_Zmw in predicting the activity F is calculated as follows:

u_rt in the Utility Theory for Decision Making is determined as:

u_rt(X_Zmw,F)=(6)

MI K0000039
CF SEQUENCE-DEPENDENT CONFORMATIONAL FEATURE
CT PROPERTY AVERAGED FOR REGION [A;B]
DP P0000001
PV Twist
AB 10 18
UT 0.234
LC -0.859
FG http://wwwmgs.bionet.nsc.ru/.../USFTWIST.htm
C-CODE
/*USF/DNA-binding increases with Twist decrease*/
double TwistCalc_for_SynthUSFbind (char *s){
double X; char *seq; int i,k, SiteLength=9;
double DinucPar[16]={ 38.90, ........, 34.96 };
seq=&s[0];
if(strlen(seq) < SiteLength+1)return(-1001.);
for (i=0, X=0.;i < SiteLength-1;i++) {
switch (seq[i ]) { case 'A': k= 0; break;
..........................................
switch (seq[i+1]) { case 'A': k+=0; break;
............................................
if (k > 15) return(-1004.); X+=DinucPar[k]; }
return (X/(double)(SiteLength-1));}

Figure 9. Description of a conformational property Helical twist for USF-binding site in the ACTIVITY.

Only the properties with U(X_Zmw,F)>0 are selected for the activity prediction and used to choose a limited set of linearly independent properties

Figure 10. Examples of weight functions w(i).

Knowledge base on functional site activity. All selected characteristics are stored in the knowledge base of the ACTIVITY system. The format of the conformational property description is illustrated in Fig. 9. Using these properties, we generate the program to predict the site activity from its nucleotide sequence through optimization of equation (2). C-code programs for such prediction are also stored in the knowledge base (Ponomarenko et al., 1997b).

Analysis of several site samples has demonstrated applicability of this simple approach. For all these samples, the significant features have been identified and the linear-additive approximation for predicting the site activity has been derived. For example, the weighted concentration of the tetranucleotide VUKK downstream of the SV40 pre-mRNA cleavage point was found to be responsible for the 3’-processing efficiency (Fig. 11a). The USF/DNA factor affinity correlates very well with the helical twist of B-DNA (Fig. 11b). The major groove width determines the Cro repressor/DNA affinity (Fig. 11c). The identified characteristics provide a first approximation in predicting the value of the specific activity of functional sites. These properties were used to generate the method for predicting the transcription activity of mouse a A-crystalline gene by analyzing its promoter sequence; the prediction shows a good agreement with the experimental data (Fig. 11d).

a)b)c)d)

Figure 11. (a) Dependence of the mature DNA yield in the 3’-processing of SV40 virus pre-mRNA on the weighted concentration of VUKK tetranucleotide downstream of the mRNA cleavage point; (b) the dependence of the USF/DNA affinity on the helical twist angle of B-DNA; (c) the dependence of the Cro repressor/DNA affinity on the major groove width of B-DNA; (d) correlation of the experimental and calculated transcription activity for TATA/PE1B-containing promoter region of mouse a A-crystalline gene.

Reliability of the functional site activity values predicted by ACTIVITY from their sequences was studied by the authors earlier (Ponomarenko et al., 1997b) as well as ACTIVITY was compared with weight matrices (Stormo et al., 1986) and neural networks (Jonson et al., 1993).

CONCLUSION

The first step in interpreting a human genome sequence involves finding and annotation of the all genes it contains. The second step consists in characterizing the biological function of the individual genes, the way they are controlled, and their possible involvement in human disease. Significant success has been made in predicting and annotating protein coding regions (exons), although the exons account for only a few percents of the genomic sequence. A considerable part of the genome is occupied by regulatory sequences, which specify the tissue, developmental stage, or biochemical context of gene expression. Recognition, interpretation, and annotation of genome regulatory sequences should be one of the major tasks in the future progress of Human Genome Project. We designed the GeneExpress computer system as a first attempt to integrate the variety of information on genomic regulatory sequences and to use this information in developing and running software for their analysis and recognition. GeneExpress integrates TRRD and GeneNet databases and provides references to external databases, such as TRANSFAC, COMPEL, and EMBL using the SRS query system.

It is essential that the GeneExpress can and have to progress and expand continuously to update and integrate new resources for investigating other molecular events of the gene expression, such as splicing, DNA/protein interactions, etc. In the nearest future, a number of new basic modules will be added to the system including programs for recognition of eukaryotic promoters (Solovyev & Salamov, 1997) and composite regulatory elements (Kel et al., 1995). A great number of software and information resources on various aspects of gene expression regulation, developed by the bioinformatics community, are currently existing. However, representation diversity of the data and the results of the data processing hinders the access to these resources. This diversity is and will be the natural trait of bioinformatics, inherent for its development. Hence, the problem is not to develop uniform data formats but to succeed in integration of the already available software and information resources in the formats developed by their authors to make these resources maximally convenient for experimenters. The advent of the SRS (Etzold and Argos, 1993) opens a way to solve this problem. In addition, the users should be provided with the possibility to arrange complex scenarios of step-by-step running programs in the course of data analysis using the integrated WWW resources. This may be realized, for example, by creating a virtual knowledge base, so that the user can accumulate the results of analysis, visualize them, and compare to both one another and the information available in the integrated databases. In the project AUTOGENE (Ptitsyn et al, 1996), the authors have already tried to integrate the analyzing programs into flexible scenarios with input/output transfer of the results using the virtual knowledge base and demonstrated that the approach is promising.

Acknowledgments

This work was supported partially by grants of Russian National Human Genome Project, Russian Ministry of Science and Technical Politics, Siberian Department of Russian Academy of Sciences, and Russian Foundation for Basic Research (97-04-49740-a and 96-04-50006).

References: