MISCORE: a new scoring function for characterizing DNA regulatory motifs in promoter sequences
- Dianhui Wang1Email author and
- Sarwar Tapan1
https://doi.org/10.1186/1752-0509-6-S2-S4
© Wang and Tapan; licensee BioMed Central Ltd. 2012
Published: 12 December 2012
Abstract
Background
Computational approaches for finding DNA regulatory motifs in promoter sequences are useful to biologists in terms of reducing the experimental costs and speeding up the discovery process of de novo binding sites. It is important for rule-based or clustering-based motif searching schemes to effectively and efficiently evaluate the similarity between a k-mer (a k-length subsequence) and a motif model, without assuming the independence of nucleotides in motif models or without employing computationally expensive Markov chain models to estimate the background probabilities of k-mers. Also, it is interesting and beneficial to use a priori knowledge in developing advanced searching tools.
Results
This paper presents a new scoring function, termed as MISCORE, for functional motif characterization and evaluation. Our MISCORE is free from: (i) any assumption on model dependency; and (ii) the use of Markov chain model for background modeling. It integrates the compositional complexity of motif instances into the function. Performance evaluations with comparison to the well-known Maximum a Posteriori (MAP) score and Information Content (IC) have shown that MISCORE has promising capabilities to separate and recognize functional DNA motifs and its instances from non-functional ones.
Conclusions
MISCORE is a fast computational tool for candidate motif characterization, evaluation and selection. It enables to embed priori known motif models for computing motif-to-motif similarity, which is more advantageous than IC and MAP score. In addition to these merits mentioned above, MISCORE can automatically filter out some repetitive k-mers from a motif model due to the introduction of the compositional complexity in the function. Consequently, the merits of our proposed MISCORE in terms of both motif signal modeling power and computational efficiency will make it more applicable in the development of computational motif discovery tools.
Keywords
Background
Gene transcription is controlled by the essential interactions between Transcription Factor Binding Sites (TFBSs, or simply Binding Sites) and Transcription Proteins known as Transcription Factors (TFs) [1]. Understanding these interactions requires a knowledge on all binding sites associated with their TFs and cis-regulatory modules. Hence, discovering unknown motifs (i.e., a collection of binding sites) in co-expressed genes or finding de novo binding sites associated with a known TF is crucial to understand the gene regulatory mechanisms [2–4]. Experimental approaches for finding DNA motifs are laborious and expensive [5, 6]. Additionally, experimental techniques such as ChIP-chip [7], ChIP-seq [8] and micro-array technology are mostly incapable of predicting specific locations of the binding sites.
It was the biological significance of the cost-effective identification of the DNA motifs that computational motif discovery has received considerable attention in the last two decades. In addition to being cost-effective and time-efficient, the nature of computational techniques offers the fastest and usually the easiest means of adopting rapidly emerging new and revised understandings on the biological process to produce more sensible motif discovery results. Despite being enormously attempted, an effective motif discovery performance by the computational approaches still remains challenging [9–11]. This is partly due to the lack of effective characterization on regulatory motifs which helps in distinguishing the functional motifs from the non-functional ones.
Due to the functional significance in gene regulation, motifs are evolutionarily conserved. Hence, motif instances appear to be rather similar to each other despite having variability in their nucleotide compositions [12]. Motif instances are rarely found in the background sequences, which is often termed as the background rareness. Also, functional motifs are often over-represented in the regulatory regions (foreground) compared to the backgrounds [13–16]. Thus, a motif's background-to-foreground appearance ratio should be smaller than the random ones. Over-representation can be similarly interpreted with the rareness characteristic. However, they are typically expressed with different statistical representations. Another useful characteristic of functional motifs is related to the compositional complexity of the nucleotides, which is termed as motif complexity [17].
- 1.
IC evaluates a motif by quantifying the relative entropy of the motif PFM (Positional Frequency Matrix) under assumption of model independence. This assumption on model independence is fundamentally weak as shown in [20–23].
- 2.
MAP, on the other hand, requires a higher order Markov chain model to estimate the background probabilities [24] prior to motif evaluation. Its computational time and cost increases along with the increment of the order of the used Markov chain model. Also, MAP score can not be used to evaluate the similarity between a k-mer and a motif model, which is essential in computational motif discovery exercises.
- 3.
Both IC and MAP score ignore the motif-complexity feature in the evaluation of the candidate motifs. Hence, a complexity score-based filtering [17] has to be used in candidate motif evaluation. The complexity threshold is empirically set in the filtering process that needs human intervention and careful attempts.
- 4.
Computational motif discovery can be guided by some known motif models as useful a priori knowledge (pk). Motif evaluation in terms of ranking then becomes a motif-to-motif similarity task. Unfortunately, IC and MAP score are not able to embed the pk models in scoring.
Motivated by the above issues, this paper introduces a new motif scoring function, termed as MISCORE (mismatch-based matrix similarity scores), to quantify similarity between a k-mer and a motif PFM using a mismatch computation on the nucleotides. By evaluating each instance k-mer (a candidate binding site) of a motif, MISCORE can quantify the likeliness of the candidate motif to be functional by a combined characterization on the model conservation, the background rareness and the compositional complexity. Our proposed MISCORE share the following three remarkable features: (i) computational efficiency due to its simplicity; (ii) free from any assumption on model dependency; and (iii) an embedability of a priori knowledge in motif scoring. An extension of MISCORE, that adopts a biologically inclined pattern localization approach for an improved recognition of the functional motifs, is also reported in this paper.
Experiments on 33 benchmark DNA datasets have been carried out for evaluating the performance of MISCORE with comparison to IC and MAP score. Firstly, we examine how well these metrics can separate the functional motifs from the random ones. Secondly, we are interested in learning how well they can recognize the functional motifs from a set of putative motif models in terms of candidate ranking. Lastly, we evaluate the effectiveness of MISCORE in recognizing the functional motifs using pk models. The experimental results are found promisingly supportive to MISCORE.
Over-representation is a widely recognized numerical feature for characterizing functional motifs [13–15], that typically differs from the statistical quantification of the background rareness property. Due to their common objective of motif characterization, correlating them through a single framework is fundamentally meaningful and it has not been addressed sufficiently in the literature. MISCORE can be utilized as a similarity metric to perform this correlation as detailed in the latter portion of this paper.
Methods
This section describes MISCORE and its localized version in details. For the sake of completeness, some preliminaries are given, including the notations and the k-mer encoding scheme used throughout this paper, followed by a preliminary introduction on the motif complexity score [17], Information Content [18] and the Maximum a Posteriori score [19].
Preliminaries
Model representation
For a given binary encoded set of k-mers, S = {K1, K2, ..., K P }, the motif PFM model M S can be computed by .
Model complexity
where k is the length of k-mers and f(b i , i) is the observed frequency of the base b i at position i in the model M. Here, the complexity score lies in [(1/4) k , 1], where 1 refers to a fully complex motif PFM.
Maximum a posteriori (MAP) score
A higher MAP score indicates a better likeliness of the motif S to be functional.
Information content (IC)
where f(b i , i) is frequency of the base b i at position i in the model M, and p(b i ) is the pre-computed background frequency of the nucleotide base b i . A higher IC score of a candidate motif indicates a better potential of being a functional one.
MISCORE for motif characterization
where f(b i , i) and k(b i , i) are the observed frequencies of base b i at position i in M and K, respectively.
where the complexity is scored according to the distribution of bases (A, C, G, T) in the K. An equal distribution gives the maximum score of 1 and a dominant distribution, i.e., a nucleotide appears at all positions of the K, gives the minimum complexity of 0. In Eq (6), the score range for both d(K, M ref ) and c(K) is 0[1]. The complexity measure given in Eq (6) helps in automatically eliminating the low-complex motifs from the top rank. In this way, an empirical threshold-based filtering [17] for filtering the low-complex candidate motifs can be avoided.
While no pk model associated with the target motifs is available, then we need to employ some searching tools to generate a model that is qualified to be an approximation M of the target motifs. Then, this putative model is essentially derived from the information embedded in the input sequences by the employed search algorithms. For instance, in the clustering type of motif finding algorithms [17, 26, 27], the putative models can be obtained by grouping k-mers based on a similarity metric.
Binding sites are evolutionarily constrained with limited mutations, hence a K can be a putative motif instance if d(K, M) <d(K, M ref ) holds, which implies a smaller mismatch to the putative model M than the background reference model M ref . Note that the M ref is a PFM that can be constructed by all k-mers from the background sequences. For a large sized background, each column of the M ref approximates the nucleotides background frequency. Thus, the M ref can be conveniently composed of the nucleotides pre-computable background frequency in each column. Large sequence-portions that have a minimal chance of having the true binding sites can be taken as the backgrounds, e.g., random chunks of large genomic portions or a large collection of upstream regions from the relevant species. Note that a smaller r(K, M) score characterizes a higher similarity of that K to M in respect to its dissimilarity to M ref and a better nucleotide complexity in K, which implies a combined characterization on the model conservation, the background rareness and the compositional complexity.
where | * | is the set cardinality and r(*, *) is the MISCORE given in Eq (6). A smaller MMS score indicates a better potential for a candidate motif to be functional.
Remark
Initially, MISCORE was introduced in [28] to quantify a mismatch-based similarity between a K and a model M S , i.e., . A corresponding MMS was defined by , and utilized as a motif scoring function to quantify the conservation property of a motif S. In [29], an improved version of MISCORE, termed as relative-MISCORE, was introduced to characterize a motif's conservation and the rareness properties by introducing a background reference model M ref in the MISCORE computation. Let r(K, M S ) denote a relative similarity between a K and a model M S . Then, it can be computed by r(K, M S ) = d(K, M S )/d(K, M ref ) that results in a relative-MMS: . As a new scoring function, it was employed as a fitness function in our GAPK framework for motif discovery. In this paper, we introduce a compositional complexity term in the relative-MISCORE as shown in Eq (6), which improves our previous work by preventing k-mers with repetitive nucleotides from motif models. This new characterization simultaneously addresses the model conservation, background-rareness and the compositional complexity properties of the regulatory motifs, which makes the present MISCORE functionally advantageous than IC, MAP score and the previous MISCORE versions. It should be pointed out that other forms of characterization on regulatory motifs exist, provided that they can model the motif signals effectively and efficiently.
Conservation and rareness characterization of functional motifs
TF | R(M t ) | Conserved (M c ) models 5000 models | Random (M r ) models 5000 models | ||||
---|---|---|---|---|---|---|---|
E{R(M c )} ± std | p-value | z-score | E{R(M r )} ± std | p-value | z-score | ||
CREB | 0.188 | 0.257 ±0.025 | 0.009 | 02.75 | 0.458 ±0.016 | 0.000 | 16.60 |
SRF | 0.193 | 0.286 ±0.025 | 0.000 | 03.76 | 0.458 ±0.012 | 0.000 | 22.01 |
TBP | 0.134 | 0.243 ±0.027 | 0.000 | 04.04 | 0.493 ±0.008 | 0.000 | 43.79 |
MYOD | 0.104 | 0.195 ±0.036 | 0.004 | 02.54 | 0.467 ±0.016 | 0.000 | 22.22 |
ERE | 0.214 | 0.331 ±0.012 | 0.000 | 10.15 | 0.439 ±0.007 | 0.000 | 31.87 |
E2F | 0.203 | 0.309 ±0.019 | 0.000 | 05.65 | 0.444 ±0.009 | 0.000 | 27.54 |
CRP | 0.307 | 0.380 ±0.006 | 0.000 | 11.48 | 0.422 ±0.005 | 0.000 | 21.45 |
GAL4 | 0.246 | 0.261 ±0.016 | 0.181 | 00.88 | 0.418 ±0.008 | 0.000 | 20.95 |
CREB* | 0.188 | 0.224 ±0.024 | 0.058 | 01.47 | 0.460 ±0.017 | 0.000 | 15.76 |
SRF* | 0.193 | 0.261 ±0.023 | 0.000 | 03.01 | 0.461 ±0.010 | 0.000 | 26.46 |
TBP* | 0.134 | 0.186 ±0.026 | 0.010 | 02.03 | 0.491 ±0.007 | 0.000 | 48.37 |
MYOD* | 0.104 | 0.158 ±0.033 | 0.057 | 01.62 | 0.472 ±0.015 | 0.000 | 24.05 |
Localized-MISCORE
Transcription proteins rarely contact a single nucleotide without interacting with the adjacent bases in the binding process. Hence, the positions with a higher binding energy given by IC (and also a lower binding energy) are usually clustered as local information blocks in the PFM model of functional motifs [30]. Position-specific similarity metrics assign an equal weight to every position in the model and ignore the variability among the local blocks appearing in the motif PFMs. Since, a motif PFM can be regarded as a descriptor of its binding preferences, the underlaying nucleotide blocks are believed to carry useful information that constitutes the overall characterization of the motif. Based on this understanding, we aim to decompose a motif PFM into a set of local blocks and assign a weight to each block according to its potential of being functional.
where β j (K), β j (M S ) and β j (M ref ) are the j th local block in the K, the M S and the background model M ref , respectively. A w-length local block β j (.) can be produced by shifting a small matrix window β[4×w]such that (2 ≤ w <k) in the K, the M S and the M ref so that, k - w + 1 number of blocks can be produced.
where r l (K, M S ) is the localized-MISCORE given by Eq (10).
Note that the localized-MMS aims to improve the discrimination power for weak motifs, while it performs closely to the MMS for the strong motifs.
Results and discussion
In this section, we evaluate the separability and the recognizability performances of MISCORE with comparison to IC and MAP score. The latter portion of the recognizability analysis describes how our MISCORE can perform motif-to-motif similarity computation and incorporate pk models in recognizing functional motifs.
Separability
It is interesting to observe the performance of MISCORE, IC and MAP score in terms of separating functional motifs from the random ones. Hence, a separability performance evaluation on these modeling metrics are conducted, where the separability is considered as a metric to measure the discriminative score-gaps (normalized) between a functional motif model and a large collection of random non-functional ones.
Separability metric
where E{*} represents the mathematical expectation, γ A = [A max - A min ]-1, γ B = [B max - B min ]-1, and is the score-gap produced by metric A for S t and , reads similarly for the metric B. A max (A min ) and B max (B min ) are the metric-specific maximum (minimum), i.e., the best (worst) possible scores, that perform a normalization. Sep(A, B) > 0 score interprets that the metric B outperforms the metric A, and Sep(A, B) < 0 score indicates the opposite case, while Sep(A, B) = 0 score indicates an equal separability performance by the two metrics.
For each dataset, firstly a true motif S t is generated by carefully aligning all known binding sites using CLUSTAL W [31]. Then, N = 5000 random models are generated by collecting random k-mers from the dataset and by carefully avoiding overlap with the true binding sites subject to . The metric bounds, i.e., the best and the worst possible scores, for score normalization is required in Eq (14). The best-possible score (upper bound) of a metric can be obtained by ensuring the maximum quantification of the motif characteristics. To find the upper bound of a metric, we assume that there exist a hypothetical set of k-mers S* that can ensure the best-possible score of a metric. With an assumption of a perfect conservation between the motif instances, i.e., , , where δ(*, *) is a similarity quantification, the upper bound for the metrics can be deduced using their respective equation. However, the lower bound (i.e., the worst-possible score) of the metrics are difficult to be computed since the conservation characteristic of a given motif can not be completely eliminated in any situation. Having no viable solution to compute this, the lower-bound of these metrics are practically approximated by the worst score produced by the metrics over a large collection of random models.
Separability results
Description of the used 33 datasets
TF | L seq (bp) | Res | L bs (min, max, round(avg)) | N seq | N bs |
---|---|---|---|---|---|
data group 1 (dg1): 8 real datasets [32] | |||||
CREB | 200 | H | (05, 30, 12) | 17 | 19 |
SRF | 200 | H | (09, 22, 12) | 20 | 35 |
TBP | 200 | H | (05, 24, 07) | 95 | 95 |
MEF2 | 200 | H | (07, 15, 10) | 17 | 17 |
MYOD | 200 | H | (06, 06, 06) | 17 | 21 |
ERE | 200 | M | (13, 13, 13) | 25 | 25 |
E2F | 200 | M | (11, 11, 11) | 25 | 27 |
CRP | 105 | E | (22, 22, 22) | 18 | 24 |
data group 2 (dg2): 20 artificial datasets [10] | |||||
dm01g | 1500 | D | (13, 28, 20) | 04 | 07 |
dm04m | 2000 | D | (10, 26, 15) | 04 | 09 |
hm02r | 1000 | H | (10, 36, 23) | 09 | 11 |
hm03r | 1500 | H | (14, 46, 27) | 10 | 15 |
hm06g | 500 | H | (06, 14, 08) | 09 | 09 |
hm08m | 500 | H | (05, 34, 15) | 15 | 13 |
hm09g | 1500 | H | (07, 26, 16) | 10 | 10 |
hm10m | 500 | H | (07, 09, 08) | 06 | 11 |
hm11g | 1000 | H | (06, 42, 14) | 08 | 19 |
hm16g | 3000 | H | (09, 54, 23) | 07 | 07 |
hm17g | 500 | H | (10, 18, 15) | 11 | 10 |
hm20r | 2000 | H | (06, 71, 17) | 35 | 76 |
hm21g | 1000 | H | (10, 23, 13) | 05 | 07 |
hm24m | 500 | H | (08, 18, 12) | 08 | 08 |
hm26m | 1000 | H | (11, 36, 25) | 09 | 10 |
mus02r | 1000 | M | (10, 33, 19) | 09 | 12 |
mus10g | 1000 | M | (05, 28, 15) | 13 | 15 |
mus11m | 500 | M | (06, 27, 15) | 12 | 15 |
yst08r | 1000 | M | (12, 49, 21) | 11 | 14 |
yst09g | 1000 | Y | (09, 19, 17) | 16 | 13 |
data group 3 (dg3): 5 real datasets [33] | |||||
CREB | 500 | H | (05, 30, 12) | 17 | 19 |
SRF | 500 | H | (09, 22, 12) | 20 | 36 |
TBP | 500 | H | (05, 24, 07) | 95 | 95 |
MEF2 | 500 | H | (07, 15, 10) | 17 | 17 |
MYOD | 500 | H | (06, 06, 06) | 17 | 21 |
Sep(R, R l ) score comparison for different local block length w in R l
Sep(R, R l ) ± E{std} using 5000 random models | ||||
---|---|---|---|---|
TF | w = O (k /3) | w = max{ O (k /3), 3} | w = min{ O (k /2), 5} | w = O (k /2) |
data group 1 (dg1) | ||||
CREB | 0.022 ± 0.047 | 0.022 ± 0.047 | -0.016 ± 0.049 | -0.016 ± 0.049 |
SRF | -0.022 ± 0.034 | -0.022 ± 0.034 | -0.030 ± 0.035 | -0.030 ± 0.035 |
TBP | 0.125 ± 0.020 | 0.128 ± 0.020 | 0.128 ± 0.020 | 0.128 ± 0.020 |
MEF2 | 0.358 ± 0.041 | 0.358 ± 0.041 | 0.367 ± 0.041 | 0.367 ± 0.041 |
MYOD | 0.066 ± 0.037 | -0.089 ± 0.045 | -0.089 ± 0.045 | -0.089 ± 0.045 |
ERE | -0.008 ± 0.028 | -0.008 ± 0.028 | -0.081 ± 0.031 | -0.210 ± 0.038 |
E2F | 0.110 ± 0.027 | 0.110 ± 0.027 | 0.127 ± 0.026 | 0.136 ± 0.026 |
CRP | 0.052 ± 0.028 | 0.052 ± 0.028 | 0.110 ± 0.024 | -0.110 ± 0.039 |
avg | 0.088 ± 0.033 | 0.069 ± 0.034 | 0.065 ± 0.034 | 0.022 ± 0.037 |
data group 2 (dg2) | ||||
dm01g | 0.101 ± 0.035 | 0.101 ± 0.035 | 0.105 ± 0.036 | 0.100 ± 0.037 |
dm04m | 0.053 ± 0.033 | 0.053 ± 0.033 | 0.051 ± 0.035 | 0.051 ± 0.035 |
hm02r | 0.219 ±0.043 | 0.219 ± 0.043 | 0.146 ± 0.050 | 0.146 ± 0.050 |
hm03r | 0.135 ± 0.037 | 0.135 ± 0.037 | 0.146 ± 0.037 | 0.146 ± 0.037 |
hm06g | 0.139 ± 0.051 | 0.062 ± 0.058 | 0.062 ± 0.058 | 0.062 ± 0.058 |
hm08m | 0.084 ± 0.041 | 0.091 ± 0.041 | 0.088 ± 0.042 | 0.088 ± 0.042 |
hm09g | 0.114 ± 0.075 | 0.114 ± 0.075 | 0.141 ± 0.074 | 0.141 ± 0.074 |
hm10m | 0.134 ± 0.038 | 0.134 ± 0.038 | 0.129 ± 0.040 | 0.129 ± 0.040 |
hm11g | 0.168 ± 0.045 | 0.168 ± 0.045 | 0.191 ± 0.044 | 0.191 ± 0.044 |
hm16g | 0.140 ± 0.077 | 0.140 ± 0.077 | 0.007 ± 0.098 | 0.007 ± 0.098 |
hm17g | 0.065 ± 0.045 | 0.065 ± 0.045 | 0.026 ± 0.049 | 0.026 ± 0.049 |
hm20r | 0.322 ± 0.023 | 0.322 ± 0.023 | 0.299 ± 0.024 | 0.299 ± 0.024 |
hm21g | 0.064 ± 0.051 | 0.064 ± 0.051 | 0.060 ± 0.054 | 0.060 ± 0.054 |
hm24m | 0.107 ± 0.042 | 0.107 ± 0.042 | 0.081 ± 0.045 | 0.081 ± 0.045 |
hm26m | 0.265 ± 0.044 | 0.265 ± 0.044 | 0.216 ± 0.049 | 0.216 ± 0.049 |
mus02r | 0.004 ± 0.119 | 0.004 ± 0.119 | -0.273 ± 0.198 | -0.273 ± 0.198 |
mus10g | 0.350 ± 0.056 | 0.354 ± 0.056 | 0.354 ± 0.056 | 0.354 ± 0.056 |
mus11m | 0.340 ± 0.042 | 0.340 ± 0.042 | 0.329 ± 0.043 | 0.329 ± 0.043 |
yst08r | 0.131 ± 0.045 | 0.131 ± 0.045 | 0.118 ± 0.047 | 0.107 ± 0.047 |
yst09g | 0.353 ± 0.056 | 0.353 ± 0.056 | 0.337 ± 0.058 | 0.333 ± 0.059 |
avg | 0.164 ± 0.050 | 0.161 ± 0.050 | 0.131 ± 0.057 | 0.130 ± 0.057 |
data group 3 (dg3) | ||||
CREB | 0.072 ± 0.042 | 0.072 ± 0.042 | 0.049 ± 0.043 | 0.049 ± 0.043 |
SRF | -0.026 ± 0.028 | -0.026 ± 0.028 | -0.032 ± 0.029 | -0.032 ± 0.029 |
TBP | 0.129 ± 0.019 | 0.133 ± 0.019 | 0.133 ± 0.019 | 0.133 ± 0.019 |
MEF2 | 0.372 ± 0.042 | 0.372 ± 0.042 | 0.380 ± 0.042 | 0.380 ± 0.042 |
MYOD | 0.088 ± 0.034 | -0.076 ± 0.042 | -0.076 ± 0.042 | -0.076 ± 0.042 |
avg | 0.127 ± 0.033 | 0.095 ± 0.035 | 0.091 ± 0.035 | 0.091 ± 0.035 |
Result summary: | E{Sep(R, R l )} ± E{std} on each data group | |||
dg 1 | 0.088 ±0.033 | 0.069 ± 0.034 | 0.065 ± 0.034 | 0.022 ± 0.037 |
dg 2 | 0.164 ±0.050 | 0.161 ± 0.050 | 0.131 ± 0.057 | 0.130 ± 0.057 |
dg 3 | 0.127 ±0.033 | 0.095 ± 0.035 | 0.091 ± 0.035 | 0.091 ± 0.035 |
avg | 0.126 ±0.039 | 0.108 ± 0.040 | 0.095 ± 0.042 | 0.081 ± 0.043 |
Sep(*, *) score comparison among R, R l , IC and MAP score
Result details: Sep(*, *) ± E{std} on each dataset using 5000 random models | ||||||
---|---|---|---|---|---|---|
dg | TF | Sep (IC, R) | Sep (IC, R l ) | Sep (MAP, R) | Sep (MAP, R l ) | Sep (R, R l ) |
CREB | -0.099 ± 0.051 | -0.080 ± 0.013 | 0.255 ± 0.030 | 0.268 ± 0.014 | 0.022 ± 0.047 | |
SRF | -0.104 ± 0.036 | -0.133 ± 0.008 | 0.313 ± 0.020 | 0.294 ± 0.009 | -0.022 ± 0.034 | |
TBP | -0.088 ± 0.025 | 0.056 ± 0.002 | 0.302 ± 0.014 | 0.395 ± 0.005 | 0.125 ± 0.020 | |
MEF2 | -0.405 ± 0.088 | 0.092 ± 0.020 | 0.144 ± 0.049 | 0.446 ± 0.017 | 0.358 ± 0.041 | |
dg 1 | MYOD | -0.113 ± 0.043 | -0.022 ± 0.010 | 0.299 ± 0.025 | 0.356 ± 0.011 | 0.066 ± 0.037 |
ERE | 0.060 ± 0.027 | 0.057 ± 0.011 | 0.416 ± 0.017 | 0.414 ± 0.012 | -0.008 ± 0.028 | |
E2F | -0.048 ± 0.032 | 0.064 ± 0.012 | 0.350 ± 0.018 | 0.419 ± 0.012 | 0.110 ± 0.027 | |
CRP | 0.013 ± 0.032 | 0.070 ± 0.018 | 0.486 ± 0.018 | 0.516 ± 0.013 | 0.052 ± 0.028 | |
avg | -0.098 ± 0.042 | 0.013 ± 0.012 | 0.321 ± 0.024 | 0.388 ± 0.012 | 0.088 ± 0.033 | |
dm01g | -0.080 ± 0.042 | 0.024 ± 0.027 | 0.294 ± 0.024 | 0.361 ± 0.023 | 0.101 ± 0.035 | |
dm04m | -0.029 ± 0.038 | 0.026 ± 0.025 | 0.350 ± 0.022 | 0.384 ± 0.022 | 0.053 ± 0.033 | |
hm02r | -0.187 ± 0.067 | 0.089 ± 0.029 | 0.320 ± 0.037 | 0.478 ± 0.024 | 0.219 ± 0.043 | |
hm03r | -0.096 ± 0.045 | 0.076 ± 0.017 | 0.276 ± 0.026 | 0.389 ± 0.015 | 0.135 ± 0.037 | |
hm06g | -0.145 ± 0.068 | 0.001 ± 0.031 | 0.227 ± 0.040 | 0.325 ± 0.025 | 0.139 ± 0.051 | |
hm08m | -0.006 ± 0.048 | 0.082 ± 0.024 | 0.277 ± 0.030 | 0.340 ± 0.021 | 0.084 ± 0.041 | |
hm09g | -0.120 ± 0.087 | -0.009 ± 0.041 | 0.211 ± 0.053 | 0.288 ± 0.035 | 0.114 ± 0.075 | |
hm10m | -0.070 ± 0.050 | 0.071 ± 0.027 | 0.290 ± 0.030 | 0.383 ± 0.022 | 0.134 ± 0.038 | |
dg 2 | hm11g | -0.172 ± 0.062 | 0.077 ± 0.016 | 0.224 ± 0.036 | 0.388 ± 0.016 | 0.168 ± 0.045 |
hm16g | -0.218 ± 0.100 | 0.000 ± 0.049 | 0.227 ± 0.056 | 0.364 ± 0.038 | 0.140 ± 0.077 | |
hm17g | -0.076 ± 0.052 | -0.022 ± 0.026 | 0.379 ± 0.029 | 0.409 ± 0.021 | 0.065 ± 0.045 | |
hm20r | -0.344 ± 0.044 | 0.098 ± 0.002 | 0.234 ± 0.022 | 0.486 ± 0.006 | 0.322 ± 0.023 | |
hm21g | -0.183 ± 0.062 | -0.075 ± 0.036 | 0.293 ± 0.035 | 0.357 ± 0.027 | 0.064 ± 0.051 | |
hm24m | -0.082 ± 0.052 | 0.024 ± 0.032 | 0.324 ± 0.031 | 0.390 ± 0.026 | 0.107 ± 0.042 | |
hm26m | -0.114 ± 0.067 | 0.177 ± 0.034 | 0.377 ± 0.039 | 0.540 ± 0.028 | 0.265 ± 0.044 | |
mus02r | -0.034 ± 0.110 | -0.061 ± 0.058 | 0.409 ± 0.062 | 0.393 ± 0.046 | 0.004 ± 0.119 | |
mus10g | -0.630 ± 0.134 | -0.052 ± 0.020 | 0.001 ± 0.076 | 0.355 ± 0.019 | 0.350 ± 0.056 | |
mus11m | -0.623 ± 0.098 | -0.049 ± 0.021 | 0.050 ± 0.054 | 0.386 ± 0.019 | 0.340 ± 0.042 | |
yst08r | -0.019 ± 0.050 | 0.149 ± 0.024 | 0.037 ± 0.040 | 0.196 ± 0.019 | 0.131 ± 0.045 | |
yst09g | -0.253 ± 0.102 | 0.179 ± 0.036 | -0.053 ± 0.073 | 0.310 ± 0.029 | 0.353 ± 0.056 | |
avg | -0.174 ± 0.069 | 0.040 ± 0.029 | 0.237 ± 0.041 | 0.376 ± 0.024 | 0.164 ± 0.050 | |
CREB | -0.102 ± 0.047 | -0.056 ± 0.012 | 0.248 ± 0.028 | 0.280 ± 0.013 | 0.072 ± 0.042 | |
SRF | -0.085 ± 0.029 | -0.131 ± 0.007 | 0.324 ± 0.016 | 0.296 ± 0.008 | -0.026 ± 0.028 | |
dg 3 | TBP | -0.080 ± 0.023 | 0.052 ± 0.002 | 0.307 ± 0.013 | 0.392 ± 0.005 | 0.129 ± 0.019 |
MEF2 | -0.420 ± 0.092 | 0.122 ± 0.020 | 0.132 ± 0.051 | 0.463 ± 0.017 | 0.372 ± 0.042 | |
MYOD | -0.115 ± 0.040 | -0.017 ± 0.009 | 0.297 ± 0.023 | 0.358 ± 0.010 | 0.088 ± 0.034 | |
avg | -0.160 ± 0.046 | -0.006 ± 0.010 | 0.262 ± 0.026 | 0.358 ± 0.011 | 0.127 ± 0.033 | |
Result summary: E{Sep(*, *)} ± E{std} on each data group | ||||||
data group (dg) | Sep(IC, R) | Sep(IC, R l ) | Sep(MAP, R) | Sep(MAP, R l ) | Sep(R, R l ) | |
dg 1 | -0.098 ± 0.042 | 0.013 ± 0.012 | 0.321 ± 0.024 | 0.388 ± 0.012 | 0.088 ± 0.033 | |
dg 2 | -0.174 ± 0.069 | 0.040 ± 0.029 | 0.237 ± 0.041 | 0.376 ± 0.024 | 0.164 ± 0.050 | |
dg 3 | -0.160 ± 0.046 | -0.006 ± 0.010 | 0.262 ± 0.026 | 0.358 ± 0.011 | 0.127 ± 0.033 | |
avg | -0.144 ± 0.052 | 0.016 ± 0.017 | 0.273 ± 0.030 | 0.374 ± 0.015 | 0.126 ± 0.039 |
Recognizability
It is often observed that after evaluating a set of candidate motifs returned by a discovery tool, the top ranked candidates are not necessarily functional. The ineffectiveness of the motif evaluation metric used can be one of the reasons behind this. Therefore, we have conducted a recognizability performance comparison among these metrics.
Recognizability refers to how well a metric can recognize the best candidate motif from a set of putative candidates in terms of ranking, where the best candidate motif is expected to be top ranked. To conduct this evaluation, we need to have a set of putative candidate motifs generated by some motif discovery tools on each dataset. In this study, we employed MEME [34] to generate a set of putative motifs for each dataset. Then, the best candidate motif is identified by the F-measure [35]: F = 2PR/(P + R), where P = TP/(TP + FP), R = TP/(TP + FN), where TP, FP and FN are the number of true positive, false positive and the false negative predictions, respectively. TP refers to the number of the true binding sites overlapped by at least one predicted site. In this study, we considered a true positive count if a true binding site is overlapped by a predicted site with at least 25% of the length of the true site. FP is the number of the predicted sites that do not have more than 25% overlap with any true binding sites; and FN is the number of the true binding sites that are not overlapped by any predicted sites with at least 25% of the length of the binding site.
where Q is the number of the relevant items whose rank orders are to be considered. In our case, only the best motif's rank is considered, hence Q = 1 and Eq (15) becomes μ = 1/rank(best motif).
Recognizability scores for the best candidate motifs
Result details: a 10-run average μ score on each dataset | |||||
---|---|---|---|---|---|
data group (dg) | TF | MAP | IC | R | R l |
CREB | 0.339 | 0.433 | 0.383 | 0.384 | |
SRF | 0.582 | 0.757 | 0.725 | 0.721 | |
TBP | 0.529 | 0.717 | 0.750 | 0.800 | |
MEF2 | 0.362 | 0.763 | 0.742 | 0.757 | |
dg 1 | MYOD | 0.517 | 0.265 | 0.243 | 0.209 |
ERE | 0.512 | 0.750 | 0.875 | 1.000 | |
E2F | 0.383 | 0.800 | 0.800 | 0.700 | |
CRP | 1.000 | 1.000 | 1.000 | 1.000 | |
avg | 0.528 | 0.686 | 0.690 | 0.696 | |
dm01g | 0.107 | 0.195 | 0.151 | 0.127 | |
dm04m | 0.180 | 0.134 | 0.219 | 0.188 | |
hm02r | 0.159 | 0.305 | 0.700 | 0.617 | |
hm03r | 0.257 | 0.179 | 0.225 | 0.255 | |
hm06g | 0.264 | 0.176 | 0.255 | 0.297 | |
hm08m | 0.341 | 0.304 | 0.224 | 0.320 | |
hm09g | 0.156 | 0.299 | 0.304 | 0.307 | |
hm10m | 0.364 | 0.416 | 0.489 | 0.474 | |
dg 2 | hm11g | 0.275 | 0.390 | 0.194 | 0.192 |
hm16g | 0.419 | 0.540 | 0.550 | 0.507 | |
hm17g | 1.000 | 1.000 | 1.000 | 1.000 | |
hm20r | 0.456 | 0.304 | 0.306 | 0.390 | |
hm21g | 0.407 | 0.450 | 0.180 | 0.190 | |
hm24m | 0.198 | 0.172 | 0.263 | 0.266 | |
hm26m | 0.297 | 0.313 | 0.317 | 0.169 | |
mus02r | 0.400 | 0.393 | 0.233 | 0.332 | |
mus10g | 1.000 | 0.867 | 0.900 | 0.800 | |
mus11m | 0.254 | 0.392 | 0.532 | 0.558 | |
yst08r | 0.247 | 0.239 | 0.151 | 0.231 | |
yst09g | 0.389 | 0.460 | 0.344 | 0.314 | |
avg | 0.359 | 0.376 | 0.377 | 0.377 | |
CREB | 0.512 | 0.422 | 0.375 | 0.540 | |
SRF | 0.369 | 0.407 | 0.373 | 0.398 | |
dg 3 | TBP | 0.542 | 0.875 | 0.583 | 0.750 |
MEF2 | 0.533 | 1.000 | 0.467 | 0.433 | |
MYOD | 0.488 | 0.425 | 0.453 | 0.400 | |
avg | 0.489 | 0.626 | 0.450 | 0.504 | |
Result summary: a 10-run average μ score on each data group | |||||
dg 1 | 0.528 | 0.686 | 0.690 | 0.696 | |
dg 2 | 0.358 | 0.376 | 0.377 | 0.377 | |
dg 3 | 0.489 | 0.626 | 0.450 | 0.504 | |
avg{dg1, dg2, dg3} | 0.458 | 0.563 | 0.506 | 0.526 | |
avg{dg1, dg2} | 0.443 | 0.531 | 0.533 | 0.536 |
Recognizability on degenerated motifs
Weak motif characterization and recognition is challenging to all evaluation metrics. Therefore, in order to observe how the considered metrics perform in recognizing degenerated motifs, we first split the 33 datasets into two categories, i.e., strong and weak motif classes, based on the average positional conservation of the motif PFMs, which is defined as , b i ∈ {A, C, G, T}.
Strong/weak motif class-wise average recognizability scores
Strong/weak motif class-wise E{μ}over 10 runs | |||||
---|---|---|---|---|---|
Motif class | apc (S t ) range | MAP | IC | R | R l |
Weak (17/33 datasets) | apc ≤0.75 | 0.373 | 0.412 | 0.409 | 0.436 |
Strong (16/33 datasets) | apc >0.75 | 0.463 | 0.562 | 0.516 | 0.507 |
Motif recognition using priori-known models
If there exists priori known (pk) estimation of the target motif profile during the search in the query sequences, then the motif discovery algorithms can greatly benefit by utilizing such a priori knowledge in finding motifs that have similar characteristics to the pk model. Often a priori estimation of a target motif model can be obtained from the public databases e.g., [37–39], or by collecting a set of binding sites from the sequences that are known to be co-regulated by the target TF [29]. These pk models can only be the estimation of the target motifs in the search, since: (i) the known binding sites in the public databases are usually incomplete, which may cause the pk profiles to have an incomplete representation that may not be able to reliably discriminate a true motif from a false one [40], and (ii) due to the sequence dissimilarity between the query sequences and the sequences that are known to be co-regulated by the target TF.
One plausible use of the pk models is their involvement in the process of motif evaluation, where the putative motifs will be recognized by referring to the pk models. The ranking of the candidate motifs then becomes a motif-to-motif similarity quantification between the putative and the pk models.
MAP score is unable to evaluate the motif-to-motif similarity. IC, on the other hand, is not originally meant for motif-to-motif similarity computation. However, it has been extended as the average log likelihood ratio (ALLR) [41] for this task. Several other metrics can perform motif-to-motif similarity quantification, e.g., Pearson correlation coefficient (PCC) [42], Kullback-Leibler divergence (KLD) [43–45], Euclidean distance (ED) [46] and Sandelin-Wasserman (SW) metric [47]. But, these metrics can only compute a motif-to-motif similarity without considering motif characteristics.
Note that R pk and r pk , characterizing motif signals with assistance of pk models, can be regarded as the supervised counterparts of R and r, respectively. localized-MISOCRE can be expressed to accommodate the pk models in a similar manner. Similarly, MISCORE can be employed to compute the motif-to-motif similarity in order to group similar candidate motifs in the relevant applications.
For simplicity, we demonstrate that MISCORE with the use of pk models can help in recognizing putative motifs, and performs favorably against other metrics. To do this, we first generated a pk model for each dataset by extracting the non-redundant known binding sites associated with CREB, E2F, MEF2 and SRF transcription factors from JASPAR [37]; ERE, MYOD and TBP from TRANSFAC (public v7.0) [38]; and CRP from RegulonDB [39] databases. After alignment, the pk models are generated for the datasets in dg1 and dg3 since they share common transcription factors. For the 20 datasets in dg2, we applied a multiple sequence alignment tool GLAM [48] to align the binding sites of each dataset. Then, the longest conserved block from the alignment is extracted to form a pk model for each dataset.
Recognizability scores for the best candidate motifs using pk models
Result details: a 10-run average μ score on each dataset | ||||||||
---|---|---|---|---|---|---|---|---|
data group (dg) | TF | R pk |
| PCC | ALLR | KLD | ED | SW |
CREB | 0.339 | 0.333 | 0.096 | 0.295 | 0.275 | 0.370 | 0.080 | |
SRF | 0.667 | 0.717 | 0.500 | 0.553 | 0.553 | 0.657 | 0.564 | |
TBP | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | |
MEF2 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | |
dg 1 | MYOD | 0.645 | 0.651 | 0.665 | 0.656 | 0.656 | 0.656 | 0.640 |
ERE | 1.000 | 1.000 | 1.000 | 1.000 | 0.917 | 0.875 | 1.000 | |
E2F | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | |
CRP | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.792 | |
avg | 0.831 | 0.837 | 0.783 | 0.813 | 0.800 | 0.820 | 0.760 | |
dm01g | 0.667 | 0.667 | 0.342 | 0.528 | 0.694 | 0.722 | 0.371 | |
dm04m | 0.377 | 0.485 | 0.662 | 0.498 | 0.487 | 0.484 | 0.647 | |
hm02r | 0.800 | 0.700 | 1.000 | 0.547 | 0.447 | 0.447 | 1.000 | |
hm03r | 0.255 | 0.425 | 0.690 | 0.514 | 0.514 | 0.300 | 0.556 | |
hm06g | 0.444 | 0.429 | 0.611 | 0.407 | 0.353 | 0.546 | 0.427 | |
hm08m | 0.861 | 0.861 | 0.852 | 0.854 | 0.771 | 0.857 | 0.857 | |
hm09g | 0.539 | 0.565 | 0.205 | 0.389 | 0.512 | 0.556 | 0.285 | |
hm10m | 0.412 | 0.495 | 0.558 | 0.490 | 0.490 | 0.500 | 0.820 | |
dg 2 | hm11g | 0.302 | 0.329 | 0.829 | 0.335 | 0.285 | 0.333 | 0.829 |
hm16g | 0.690 | 0.767 | 0.105 | 0.617 | 0.767 | 0.900 | 0.100 | |
hm17g | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | |
hm20r | 0.537 | 0.537 | 0.708 | 0.542 | 0.542 | 0.548 | 0.708 | |
hm21g | 0.148 | 0.148 | 0.483 | 0.204 | 0.214 | 0.214 | 0.324 | |
hm24m | 0.573 | 0.650 | 1.000 | 0.592 | 0.592 | 0.725 | 0.867 | |
hm26m | 0.450 | 0.650 | 0.369 | 0.650 | 0.567 | 0.617 | 0.700 | |
mus02r | 0.182 | 0.209 | 0.329 | 0.184 | 0.184 | 0.199 | 0.345 | |
mus10g | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | |
mus11m | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | |
yst08r | 0.567 | 0.633 | 0.524 | 0.567 | 0.583 | 0.580 | 0.767 | |
yst09g | 0.201 | 0.232 | 0.292 | 0.179 | 0.186 | 0.217 | 0.321 | |
avg | 0.550 | 0.589 | 0.628 | 0.555 | 0.559 | 0.587 | 0.646 | |
CREB | 0.642 | 0.642 | 0.556 | 0.657 | 0.657 | 0.667 | 0.476 | |
SRF | 0.667 | 0.667 | 0.523 | 0.707 | 0.650 | 0.667 | 0.822 | |
dg 3 | TBP | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 |
MEF2 | 0.653 | 0.656 | 0.656 | 0.750 | 0.850 | 0.662 | 0.482 | |
MYOD | 0.486 | 0.653 | 0.500 | 0.563 | 0.563 | 0.577 | 0.661 | |
avg | 0.690 | 0.723 | 0.647 | 0.735 | 0.744 | 0.715 | 0.688 | |
Result summary: a 10-run average μ score on each data group | ||||||||
dg 1 | 0.831 | 0.837 | 0.783 | 0.813 | 0.800 | 0.820 | 0.760 | |
dg 2 | 0.550 | 0.589 | 0.628 | 0.555 | 0.559 | 0.587 | 0.646 | |
dg 3 | 0.690 | 0.723 | 0.647 | 0.735 | 0.744 | 0.715 | 0.688 | |
avg | 0.690 | 0.717 | 0.686 | 0.701 | 0.701 | 0.707 | 0.698 |
Background rareness and over-representation
Another key concept in computational motif discovery is over-representation [13–15, 49]. It looks for motifs that have significant occurrences in the query sequences (input promoters) than the background sequences through some statistical quantification [13, 16]. The functionality of this site multiplicity, i.e., 'the shadow appearances of the binding sites', in the regulatory regions could constitute a mechanism for lateral diffusion of the transcription factors along the sequences, and/or the shadow sites might be the fossils from the process of binding site turnover [16, 50]. Even though the biological reasons behind this site multiplicity are yet to be fully understood [16], it is often considered as a useful motif characteristic and well recognized in the working field.
It is interesting to analyze the correlation between a functional motif's background rareness and over-representation, although both can partially characterize the functional motifs. This section tries to make a sensible link between these two key concepts.
Correlation between background rareness and over-representation
where | * | represents the set cardinality, r(K, M S ) is the MISCORE given in Eq (6) and θ is a cut-off threshold that can be defined as θ = R(S) + std(S)λ, where std represents the standard deviation operator, λ is a threshold regulator and R(*) is the MMS given in Eq (9).
where S bg and S fg are the sets of all k-mers produced by window shifting in the background and in the foreground regions, respectively.
The condition ORS r (M S ) < 1 indicates that M S has a higher frequency in the foreground than the background for a given threshold, which implies that there are comparatively less occurrences of that motif in the background (i.e., background rareness) than the foreground. Hence, the background rareness of a motif can be characterized through its over-representation feature, that can be statistically quantified.
Demonstration
ORS r (M t ) scores with several threshold regulators
ORS r (M t ), θ= R(S t ) + std(S t ) λ | |||||
---|---|---|---|---|---|
TF | L fg (bp) | λ = -0.25, | λ = 0.0, | λ = 0.25, | λ = 0.5 |
CREB | 200 | 0.391 | 0.357 | 0.429 | 0.537 |
500 | 0.762 | 0.576 | 0.884 | 0.806 | |
SRF | 200 | 0.040 | 0.048 | 0.055 | 0.059 |
500 | 0.107 | 0.108 | 0.126 | 0.144 | |
TBP | 200 | 0.334 | 0.385 | 0.441 | 0.548 |
500 | 0.671 | 0.778 | 0.793 | 0.803 | |
MEF2 | 200 | 0.041 | 0.050 | 0.065 | 0.100 |
500 | 0.129 | 0.177 | 0.392 | 0.655 | |
MYOD | 200 | 0.292 | 0.289 | 0.289 | 0.289 |
500 | 0.303 | 0.620 | 0.710 | 0.746 |
Correlation between the over-representation and the background rareness. ORS r scores for the functional models M t , the random models , and the conserved models for q = 1, 2, 3, . . ., 1000 are plotted for each dataset with 200bp and 500bp promoters in the left and in the right column, respectively. Threshold θ = R(M t ) + std(M t )λ, λ = 0.0 is used. The figure depicts a rareness interpretable visualization through the statistical over-representation property of the functional motifs by showing that, the ORS r (M r ) scores are found distant from the ORS r (M t ) scores for all cases which implies that the random models have close to zero chance of being over-represented with comparison to the true models. In addition to this, the ORS r (M t ) scores are found to be mostly rare with comparison to the ORS r (M c ) scores, i.e., these non-functional conserved models have a rare chance of having better over-representation scores than the true models, for most of the datasets.
Conclusions
This paper contributes a mismatch-based fast computational tool for modeling DNA regulatory motifs. It is free from any assumption on the model dependency, and it escapes from the use of background modeling using Markov chain models. Simultaneously, it embeds the compositional complexity in modeling the motif signals. Our proposed MISCORE can be used as a metric to measure the similarity between k-mers and a motif model, also it can be employed to compute the motif-to-motif similarity.
The experimental results on 33 datasets indicate that MISCORE performs favorably with comparison to the well-known IC and MAP score in terms of the separability and the recognizability. These results also show that MISOCRE is functionally effective in recognizing degenerated motifs, and it can embed the pk models to perform candidate motif ranking.
MISCORE has good potential to be employed as a similarity metric in rule-based or clustering-based motif discovery algorithms, it can also be employed as a numerical feature in machine learning approaches for finding motifs. Furthermore, MISCORE-based Motif Score (MMS) can be employed as a fitness function in evolutionary computation approaches for motif discovery, and for candidate motif ranking in computational motif discovery tools.
Declarations
Acknowledgements
The authors are grateful to Dr Nung Kion Lee (UNIMAS, Malaysia) and Dr Xi Li (CSIRO, Australia) for their contributions to the development of MISCORE during their PhD studies at La Trobe University.
This article has been published as part of BMC Systems Biology Volume 6 Supplement 2, 2012: Proceedings of the 23rd International Conference on Genome Informatics (GIW 2012). The full contents of the supplement are available online at http://www.biomedcentral.com/bmcsystbiol/supplements/6/S2.
Authors’ Affiliations
References
- Yeung K, Medvedovic M, Bumgarner R: From co-expression to co-regulation: how many microarray experiments do we need?. Genome Biology. 2004, 5 (7): R48-10.1186/gb-2004-5-7-r48.PubMed CentralView ArticlePubMedGoogle Scholar
- Chin F, Leung HC: DNA motif representation with nucleotide dependency. IEEE/ACM Transactions on Computational Biology and Bioinformatics. 2008, 5: 110-119.View ArticlePubMedGoogle Scholar
- Pavesi G, Mauri G, Pesole G: In silico representation and discovery of transcription factor binding sites. Briefings in Bioinformatics. 2004, 5 (3): 217-236. 10.1093/bib/5.3.217.View ArticlePubMedGoogle Scholar
- Das M, Dai HK: A survey of DNA motif finding algorithms. BMC Bioinformatics. 2007, 8 (Suppl 7): S21-10.1186/1471-2105-8-S7-S21.PubMed CentralView ArticlePubMedGoogle Scholar
- Elnitski L, Jin VX, Farnham PJ, Jones SJM: Locating mammalian transcription factor binding sites: A survey of computational and experimental techniques. Genome Research. 2006, 16 (12): 1455-1464. 10.1101/gr.4140006.View ArticlePubMedGoogle Scholar
- Jones SJM: Prediction of genomic functional elements. Annual Review of Genomics and Human Genetics. 2006, 7: 315-338. 10.1146/annurev.genom.7.080505.115745.View ArticlePubMedGoogle Scholar
- Lee TI, Jenner RG, Boyer LA, Guenther MG, Levine SS, Kumar RM, Chevalier B, Johnstone SE, Cole MF, ichi Isono K, Koseki H, Fuchikami T, Abe K, Murray HL, Zucker JP, Yuan B, Bell GW, Herbolsheimer E, Hannett NM, Sun K, Odom DT, Otte AP, Volkert TL, Bartel DP, Melton DA, Gifford DK, Jaenisch R, Young RA: Control of developmental regulators by Polycomb in human embryonic stem cells. Cell. 2006, 125 (2): 301-313. 10.1016/j.cell.2006.02.043.PubMed CentralView ArticlePubMedGoogle Scholar
- Ren B, Robert F, Wyrick JJ, Aparicio O, Jennings EG, Simon I, Zeitlinger J, Schreiber J, Hannett N, Kanin E, Volkert TL, Wilson CJ, Bell SP, Young RA: Genome-wide location and function of DNA binding proteins. Science. 2000, 290 (5500): 2306-2309. 10.1126/science.290.5500.2306.View ArticlePubMedGoogle Scholar
- Hu J, Li B, Kihara D: Limitations and potentials of current motif discovery algorithms. Nucleic Acids Research. 2005, 33 (15): 4899-4913. 10.1093/nar/gki791.PubMed CentralView ArticlePubMedGoogle Scholar
- Tompa M, Li N, Bailey T, Church G, De Moor B, Eskin E, Favorov A, Frith M, Fu Y, Kent J, Makeev V, Mironov A, Noble W, Pavesi G, Pesole G, R M, Simonis N, Sinha S, Thijs G, van Helden J, Vandenbogaert M, Weng Z, Workman C, Ye C, Zhu Z: Assessing computational tools for the discovery of transcription factor binding sites. Nat Biotech. 2005, 23: 137-144. 10.1038/nbt1053.View ArticleGoogle Scholar
- Marschall T, Rahmann S: Efficient exact motif discovery. Bioinformatics. 2009, 25 (12): i356-364. 10.1093/bioinformatics/btp188.PubMed CentralView ArticlePubMedGoogle Scholar
- Moses A, Chiang D, Kellis M, Lander E, Eisen M: Position specific variation in the rate of evolution in transcription factor binding sites. BMC Evolutionary Biology. 2003, 3: 19-10.1186/1471-2148-3-19.PubMed CentralView ArticlePubMedGoogle Scholar
- Defrance M, Touzet H: Predicting transcription factor binding sites using local over-representation and comparative genomics. BMC Bioinformatics. 2006, 7: 396+-10.1186/1471-2105-7-396.PubMed CentralView ArticlePubMedGoogle Scholar
- Sinha S, Tompa M: Discovery of novel transcription factor binding sites by statistical over-representation. Nucleic Acids Research. 2002, 30 (24): 5549-5560. 10.1093/nar/gkf669.PubMed CentralView ArticlePubMedGoogle Scholar
- Mariño Ramírez L, Spouge JL, Kanga GC, Landsman D: Statistical analysis of over-represented words in human promoter sequences. Nucleic Acids Research. 2004, 32 (3): 949-958. 10.1093/nar/gkh246.PubMed CentralView ArticlePubMedGoogle Scholar
- Frith MC, Fu Y, Yu L, Chen JF, Hansen U, Weng Z: Detection of functional DNA motifs via statistical over-representation. Nucl. Acids Res. 2004, 32 (4): 1372-1381. 10.1093/nar/gkh299.PubMed CentralView ArticlePubMedGoogle Scholar
- Mahony S, Hendrix D, Golden A, Smith TJ, Rokhsar DS: Transcription factor binding site identification using the self-organizing map. Bioinformatics. 2005, 21 (9): 1807-1814. 10.1093/bioinformatics/bti256.View ArticlePubMedGoogle Scholar
- Stormo GD, Fields DS: Specificity, free energy and information content in protein-DNA interactions. Trends in Biochemical Sciences. 1998, 23 (3): 109-113. 10.1016/S0968-0004(98)01187-6.View ArticlePubMedGoogle Scholar
- Liu XS, Brutlag DL, Liu JS: An algorithm for finding protein-DNA binding sites with applications to chromatin-immunoprecipitation microarray experiments. Nature Biotechnology. 2002, 20 (8): 835-839.View ArticlePubMedGoogle Scholar
- Zhou Q, Liu JS: Modeling within-motif dependence for transcription factor binding site predictions. Bioinformatics. 2004, 20 (6): 909-916. 10.1093/bioinformatics/bth006.View ArticlePubMedGoogle Scholar
- Bulyk ML, Johnson PLF, Church GM: Nucleotides of transcription factor binding sites exert inter-dependent effects on the binding a affinities of transcription factors. Nucleic Acids Research. 2002, 30 (5): 1255-1261. 10.1093/nar/30.5.1255.PubMed CentralView ArticlePubMedGoogle Scholar
- Roulet E, Busso S, Camargo AA, Simpson AJ, Mermod N, Bucher P: High-throughput SELEX SAGE method for quantitative modeling of transcription-factor binding sites. Nature Biotechnology. 2002, 20 (8): 831-835.View ArticlePubMedGoogle Scholar
- Wang C, Xie J, Craig B: Context dependent models for discovery of transcription factor binding sites. Statistical Methodology. 2006, 3: 55-68. 10.1016/j.stamet.2005.09.007.View ArticleGoogle Scholar
- Thijs G, Lescot M, Marchal K, Rombauts S, De Moor B, Rouze P, Moreau Y: A higher-order background model improves the detection of promoter regulatory elements by Gibbs sampling. Bioinformatics. 2001, 17 (12): 1113-1122. 10.1093/bioinformatics/17.12.1113.View ArticlePubMedGoogle Scholar
- Shannon CE: A mathematical theory of communication. Bell Systems Technical Journal. 1948, 27: 379-423. 623-658View ArticleGoogle Scholar
- Wang D, Lee NK: Computational discovery of motifs using hierarchical clustering techniques. In Proceedings of the 8th IEEE Int'l. Conference on Data Mining. 2008, Pisa, Italy, 1073-1078.Google Scholar
- Lee N, Wang D: SOMEA: self-organizing map based extraction algorithm for DNA motif identification with heterogeneous model. BMC Bioinformatics. 2011, 12 (Suppl 1): S16-10.1186/1471-2105-12-S1-S16.PubMed CentralView ArticlePubMedGoogle Scholar
- Wang D, Lee NK: MISCORE: mismatch-based matrix similarity scores for DNA motifs detection. In Proceedings of the 15th Int'l Conference on Neural Information Processing. 2008, Auckland, New Zealand, 478-485.Google Scholar
- Wang D, Li X: iGAPK: improved GAPK algorithm for regulatory DNA motif discovery. In Proceedings of the 17th Int'l Conference on Neural Information Processing. 2010, Sydney, Australia, 217-225.Google Scholar
- Eisen MB: All motifs are not created equal: structural properties of transcription factor - DNA interactions and the inference of sequence specificity. Genome Biology. 2005, 6 (P7): 277-284.Google Scholar
- Thompson JD, Higgins DG, Gibson TJ: CLUSTAL W: improving the sensitivity of progressive multiple sequence alignment through sequence weighting, position-specific gap penalties and weight matrix choice. Nucleic Acids Research. 1994, 22 (22): 4673-4680. 10.1093/nar/22.22.4673.PubMed CentralView ArticlePubMedGoogle Scholar
- Wei Z, Jensen ST: GAME: detecting cis-regulatory elements using a genetic algorithm. Bioinformatics. 2006, 22 (13): 1577-1584. 10.1093/bioinformatics/btl147.View ArticlePubMedGoogle Scholar
- Blanco E, Farre D, Alba MM, Messeguer X, Guigo R: ABS: a database of Annotated regulatory Binding Sites from orthologous promoters. Nucleic Acids Research. 2006, 34 (suppl1): D63-67.PubMed CentralView ArticlePubMedGoogle Scholar
- Bailey TL, Elkan C: Unsupervised learning of multiple motifs in biopolymers using expectation maximization. Machine Learning. 1995, 21: 51-80.Google Scholar
- Fawcett T: An introduction to ROC analysis. Pattern Recognition Letters. 2006, 27 (8): 861-874. 10.1016/j.patrec.2005.10.010.View ArticleGoogle Scholar
- Wang D, Ma X: A hybird image retrieval system with user's relevance feedback using neurocomputing. Informatica. 2005, 29 (3): 271-280.Google Scholar
- Sandelin A, Alkema W, Engström P, Wasserman WW, Lenhard B: JASPAR: an open-access database for eukaryotic transcription factor binding profiles. Nucleic Acids Research. 2004, 32 (Database issue): D91-94.PubMed CentralView ArticlePubMedGoogle Scholar
- Matys V, Kel-Margoulis OV, Fricke E, Liebich I, Land S, Barre-Dirrie A, Reuter I, Chekmenev D, Krull M, Hornischer K, Voss N, Stegmaier P, Lewicki-Potapov B, Saxel H, Kel AE, Wingender E: TRANSFAC® and its module TRANSCompel®: transcriptional gene regulation in eukaryotes. Nucleic Acids Research. 2006, 34 (suppl 1): D108-D110.PubMed CentralView ArticlePubMedGoogle Scholar
- Salgado H, Gama-Castro S, Martinez-Antonio A, Diaz-Peredo E, Sanchez-Solano F, Peralta-Gil M, Garcia-Alonso D, Jimenez-Jacinto V, Santos-Zavaleta A, Bonavides-Martinez C, Collado-Vides J: RegulonDB (version 4.0) transcriptional regulation, operon organization and growth conditions in Escherichia coli K-12. Nucleic Acids Research. 2004, 32 (suppl_1): D303-306.PubMed CentralView ArticlePubMedGoogle Scholar
- Li L, Liang Y, Bass RL: GAPWM: a genetic algorithm method for optimizing a position weight matrix. Bioinformatics. 2007, 23 (10): 1188-1194. 10.1093/bioinformatics/btm080.View ArticlePubMedGoogle Scholar
- Wang T, Stormo GD: Combining phylogenetic data with co-regulated genes to identify regulatory motifs. Bioinformatics. 2003, 19 (18): 2369-2380. 10.1093/bioinformatics/btg329.View ArticlePubMedGoogle Scholar
- Pietrokovski S: Searching databases of conserved sequence regions by aligning protein multiple alignments. Nucleic Acids Research. 1996, 24 (19): 3836-3845. 10.1093/nar/24.19.3836.PubMed CentralView ArticlePubMedGoogle Scholar
- Roepcke S, Grossmann S, Rahmann S, Vingron M: T-Reg Comparator: an analysis tool for the comparison of position weight matrices. Nucleic Acids Research. 2005, 33: W438-441. 10.1093/nar/gki590. Web Server issuePubMed CentralView ArticlePubMedGoogle Scholar
- Thijs G, Marchal K, Lescot M, Rombauts S, De Moor B, Rouzé P, Moreau Y: A Gibbs sampling method to detect overrepresented motifs in the upstream regions of coexpressed genes. Journal of Computational Biology. 2002, 9 (2): 447-464. 10.1089/10665270252935566.View ArticlePubMedGoogle Scholar
- Aerts S, Van Loo P, Thijs G, Moreau Y, De Moor B: Computational detection of cis -regulatory modules. Bioinformatics. 2003, 19 (Suppl 2):Google Scholar
- Choi IG, Kwon J, Kim SH: Local feature frequency profile: a method to measure structural similarity in proteins. Proceedings of the National Academy of Sciences of the United States of America. 2004, 101 (11): 3797-3802. 10.1073/pnas.0308656100.PubMed CentralView ArticlePubMedGoogle Scholar
- Sandelin A, Wasserman WW: Constrained binding site diversity within families of transcription factors enhances pattern discovery bioinformatics. Journal of Molecular Biology. 2004, 338 (2): 207-215. 10.1016/j.jmb.2004.02.048.View ArticlePubMedGoogle Scholar
- Frith MC, Hansen U, Spouge JL, Weng Z: Finding functional sequence elements by multiple local alignment. Nucleic Acids Research. 2004, 32: 189-200. 10.1093/nar/gkh169.PubMed CentralView ArticlePubMedGoogle Scholar
- Mitrophanov AY, Borodovsky M: Statistical significance in biological sequence analysis. Briefings in Bioinformatics. 2006, 7: 2-24. 10.1093/bib/bbk001.View ArticlePubMedGoogle Scholar
- Dermitzakis ET, Clark AG: Evolution of transcription factor binding sites in mammalian gene regulatory regions: conservation and turnover. Molecular Biology and Evolution. 2002, 19 (7): 1114-1121. 10.1093/oxfordjournals.molbev.a004169.View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.