MCatch - Recognition of Composite Elements

Help page

Introduction

MatrixCatch is a tool designed to search potential composite elements (CEs) in DNA sequences. MatrixCatch uses a library of experimentally identified CEs collected in TRANSCOMPELŽ database as well as mononucleotide position weight matrices (PWM) for single binding sites (BS) for transcription factors (TF) collected in TRANSFACŽ.

As an input MatrixCatch requires:

DNA sequence(s) in FASTA, EMBL or plain sequence formats (see below);
Type of multiple testing correction, maximum P-Value or target frequency of CEs;

Maximum decline of PWM score (W) for individual BS and allowable distance variation (D) between them. These values are relative to those defined in a CE model for each CE. The following should be valid:

W >= (Thr_1,2-S_1,2 )/Thr_1,2and D >= abs(L-L_m)/L_m,

where S_1,2are the actual PWM scores and Lis the actual distance between beginnings of BS. Thr_1,2are PWM scores defined by the CE model.
Composite score (CS). This value reflects compensatory effect of cooperative binding. It can be negative, indicating that one or both binding sites of identified CE are better than it is required by the CE model. Composite score (CS) is calculated by the formula:

CS=(Thr₁-S₁ )/Thr₁ + (Thr₂-S₂ )/Thr₂ + abs(L-L_m)/53.62,

where S_1,2are the actual PWM scores and Lis the actual distance between beginnings of BS. Thr_1,2are PWM scores defined by the CE model.

As an output MatrixCatch gives a list of potential CEs, their positions, scores, p-values and respective links to the specific CEs in the database.

Advantages:

Higher accuracy: MatrixCatch performs much better that CompelPatternSearch (Kel-Margoulis et al., 2002) and similar to the method specifically developed for recognition of NFATp/AP-1 (Kel et al., 1999).
Wider scope: MatrixCatch is able to identify 265 types of CEs, what is similar to CompelPatternSearch and much more than just one NFATp/AP-1.

Brief description of the method

The idea behind MatrixCatch was to complete the knowledge about variation of DNA sequences in composite elements by taking into account sequence variation for respective single binding sites. As a typical example, let us discuss CE Sp1/E2F-1. It has been experimentally identified only in two genomic locations (TRANSCOMPELŽ numbers Ń00132 and Ń00226), providing two DNA stretches for analysis. At the same time, single binding sites for Sp1 and E2F-1 (independently from each other) were detected experimentally in 108 and 39 locations respectively. To take this advantage, we constructed an integral model for every CE, which consists of (Figure 1) :

position weight matrix (PWM) for first binding site, score and orientation;
position weight matrix (PWM) for second binding site, score and orientation;
distance between both sites.

Figure 1. Model of composite element.

To find potential CE of the specific type (for example, Sp1/E2F-1 or NFAT/AP-1) MatrixCatch will apply all existing models for this type of CE (respectively two models for Sp1/E2F-1 and 15 for NFAT/AP-1). DNA sequence is reported as potential CE if: both PWMs can recognize binding sites in correct orientation with a scores higher than the respective scores defined in the model, distance between these two sites lies within length variation defined by the model.
To search for "non perfect" matches, user may low the required PWM scores and increase length variation. In addition, a composite score and P-Value parameters are provided. Composite score reflects compensatory effect of cooperative binding, which allows one of the binding sites to be weaker than defined in the model, provided that another is stronger. P-Value shows statistical significance of the results.

Input parameters

DNA sequence formats: plain, FASTA, EMBL.

Plain sequence format:
(All new lines, blanks and characters other than A,T,G,C will be ignored, characters can be in UPPER or lower case, lines can be of different length.)

TAGTGATCCgcgccacccggatcctttgcaATCTTCGCTTCAACACTTCAACAACCCCC
AACGTGGCTTACTTGTCAGTTTTACACCGTgtCG
ACGC
ATATCcaagatGTCATGCCGATTCGTGTAGAATATTGGTACCTCCCTAGACTA
gccagcaacccaacctagataccatgtcatgtagctccatccatatagtactccttagcataggagaaaagtcctgagcgatcacgaa

FASTA format:

>seq1
TAGTGATCCgcgccacccggatcctttgcaATCTTCGCTTCAACACTTCAACAACCCCC
AACGTGGCTTACTTGTCAGTTTTACACCGTgtCG
ACGC
ATATCcaagatGTCATGCCGATTCGTGTAGAATATTGGTACCTCCCTAGACTA
gccagcaacccaacctagataccatgtcatgtagctccatccatatagtactccttagcataggagaaaagtcctgagcgat
cacgaa

EMBL format:
(Only the essential fields to recognize an entry in EMBL format are shown. More fields may be included.)

ID Sequence 1, example
SQ
TAGTGATCCgcgccacccggatcctttgcaatCTTCGCTTCAACACTTCAACAACCCCCC
AACGtGGCTTACTTGTCAGTTTTACACCGTGTCGGTAGTATAGAAGGAGCCACAGGGGAC
ACGCATATCCAAGatgtcatgccgattcgtGTAGAATATTGGTACCTCCCTAGTTAGAGA
TAGGATTGCCACGCCACTAAAAgacgcccggTGACGGTTAGCATAATTATCCTGAAACTG
GTTCGATTGGCTGCTGCAGATCTAGACGCCACGCCAGCAACCCAACCTAGATACCATGTC
atgtagctccatccatatagtactccttagcataggagaaaagtcctgagcgatcacgaa
//

ID   Sequence 1, example
SQ   Sequence 360 BP; 63 A; 92 C; 97 G; 108 T; 0 other;
     ctgcagcccc ggtttcgcaa agttaataat tttcagccgc gcacgtggtt ggccaaaccg 60
     caccctcctt cccgtcgttt cccatctctt cctcctttag agctaccact atataaatca 120
     gggctcattt tctcgctcct cacaggctca tctcgctttg gatcgattgg tttcgtaact 180
     ggtgagggac tgagggtctc ggagtggatt gatttgggat tctgttcgaa gatttgcgga 240
     ggggggcaat ggcgaccgcg gggaaggtga tcaagtgcaa aggtccgcct tgtttctcct 300
     ctgtctcttg atctgactaa tcttggttta tgattcgttg agtaattttg gggaaagctt 360
//

P-Value multiple testing correction
P-value is calculated by the formula:

P-value = 1 - (1 - p*q)^abs(L-L_m),

where p, q - frequencies of PWM hits with thresholds S₁and S₂respectively. S_1,2are the actual PWM scores and Lis the actual distance between beginnings of BS. Frequency tables were generated in advance on 100 Mb random sequence with equal nucleotide distribution.

Four types of multiple testing correction can be applied. List of CEs first ordered according to P-value with smallest values on top. #CE - rank of CE in the sorted list. After correction of P-Values, the list is again sorted.

Bonferroni: new P-value = P-value * SequenceLength;
Bonferroni step-down: new P-value = P-value * (SequenceLength - #CE);
Benjamini and Hochberg: new P-value = P-value * (SequenceLength / #CE);
No correction.

Relaxation values:

To be able to search for "non-perfect" matches, some parameters of the original CE model may be relaxed. This can be achieved by decreasing PWM score (can be lowered down by maximum 0.25 (25%) of the original score defined in a model) and increasing length variation (maximum is 5: if original distance were 10bp, than 10ą50bp (i.e. from -40bp to 60bp) will be allowed. Please note, distance may have negative value, which in such cases indicates that the second binding site is now starts before the fist. Here we always regard "first" and "second" BS according to the their order in the CE model.

Composite score (CS) reflects mutual effect of cooperative binding and defined by the formula:

CS=(Thr₁-S₁ )/Thr₁ + (Thr₂-S₂ )/Thr₂ + abs(L-L_m)/53.62,

where S_1,2are the actual PWM scores and Lis the actual distance between beginnings of BS. Thr_1,2are PWM scores defined by the CE model.

The lower CS is the better is the overall similarity. Please note: negative values of CS are relevant and indicate that the matched DNA sequence(s) has higher PWM score(s), than it is required by the model. For practical use, it can be suggested (if no other pre-knowledge exists) to use the same value as for PWM scores, which will allow only one of two BSs to be weaker. Length variation can be set to 0.5

Recognition characteristics
To investigate a recognition characteristics of MatrixCatch we tested it on the most investigated type of composite element NFAT/AP-1. 15 different examples of this CE are known. We also apply previously developed approaches CompelPatternSearch (Kel-Margoulis et al., 2002) and method specifically developed for recognition of NFATp/AP-1 (Kel et al., 1999). Using "Jack Knife" method we sequentially removed CEs one by one and tried to detect the whole set of CEs by relaxing search parameters. False positive rate (over-prediction) was calculated using a set of exon sequences from human genome. Resulted values are shown on Figure 2 and Table 1.

Figure 2. Recognition characteristics of three methods.

Table 1. Comparison of two methods - MatrixCatch and CompelPatternSearch.

Additional information about single binding sites significantly improves the quality of recognition.

False negative, FN (%)	75%	60%	50%	40%	20%	0%
MatrixCatch	0,040883	0,302535	0,310711	0,787719	1,880621	11,5372
CompelPatternSearch	0,065735	0,32046	1,0682	8,036154	19,89318	27,31306
fold difference	1,61	1,06	3,44	10,20	10,58	2,37

References

Kel-Margoulis O.V., Kel A.E., Reuter I., Deineko IV, Wingender E. TRANSCompel: a database on composite regulatory elements in eukaryotic genes. Nucleic Acids Res. 2002, 30, 332-334.
Kel A, Kel-Margoulis O, Babenko V, Wingender E. Recognition of NFATp/AP-1 composite elements within genes induced upon the activation of immune cells. J Mol Biol. 1999, 288, 353-376.