Bayesian detection of periodic mRNA time profiles without use of training examples
© Andersson et al; licensee BioMed Central Ltd. 2006
Received: 28 July 2005
Accepted: 09 February 2006
Published: 09 February 2006
Detection of periodically expressed genes from microarray data without use of known periodic and non-periodic training examples is an important problem, e.g. for identifying genes regulated by the cell-cycle in poorly characterised organisms. Commonly the investigator is only interested in genes expressed at a particular frequency that characterizes the process under study but this frequency is seldom exactly known. Previously proposed detector designs require access to labelled training examples and do not allow systematic incorporation of diffuse prior knowledge available about the period time.
A learning-free Bayesian detector that does not rely on labelled training examples and allows incorporation of prior knowledge about the period time is introduced. It is shown to outperform two recently proposed alternative learning-free detectors on simulated data generated with models that are different from the one used for detector design. Results from applying the detector to mRNA expression time profiles from S. cerevisiae showsthat the genes detected as periodically expressed only contain a small fraction of the cell-cycle genes inferred from mutant phenotype. For example, when the probability of false alarm was equal to 7%, only 12% of the cell-cycle genes were detected. The genes detected as periodically expressed were found to have a statistically significant overrepresentation of known cell-cycle regulated sequence motifs. One known sequence motif and 18 putative motifs, previously not associated with periodic expression, were also over represented.
In comparison with recently proposed alternative learning-free detectors for periodic gene expression, Bayesian inference allows systematic incorporation of diffuse a priori knowledge about, e.g. the period time. This results in relative performance improvements due to increased robustness against errors in the underlying assumptions. Results from applying the detector to mRNA expression time profiles from S. cerevisiae include several new findings that deserve further experimental studies.
Several different algorithms for detection of periodically expressed genes in DNA microarray temporal profiles have been proposed [1–7]. Theoretical and algorithmic foundations for the detection algorithms include for example Fourier analysis [1, 2], spline modelling , single-pulse models , and partial least squares classification . One group of algorithms, including those in [1, 3, 5, 6], use supervised learning methods  that exploit labelled expression profiles of genes known to be periodically expressed in the experiment to find other genes that also are periodic. This supervised learning approach precludes many potential applications where labelled training examples are not available e.g. for poorly characterised organisms.
In the subgroup of recently proposed learning-free algorithms which do not rely on supervised learning, prior knowledge in the form of a known angular frequency ω is presumed. For example, in  the power (amplitude) of frequency ω in the expression profile Fourier spectrum is used in creating a score for detection. However, since the period time usually is not exactly known, a novel method was proposed in  to resolve this problem by employing Fisher's g-test  for detection of periodic temporal profiles. This approach seems attractive but is designed to detect any periodic temporal profile, even if its period differs significantly from the period of the cell-cycle. In other words, it does not use any prior knowledge about the period time.
The algorithm used in the work by Spellman et al. (1998)  partly utilises prior knowledge about the cell-cycle period. This is achieved by averaging the power of Fourier spectrum frequencies across a discrete set of frequencies over a frequency interval that is thought to encompass the true frequency (corresponding to the period time as determined by auxiliary experimental techniques). The detector test statistic is modulated by the Pearson correlations between the present temporal profile and a set of labelled temporal profiles for genes known to be periodically expressed. Thus, the algorithm belongs to the subgroup of supervised methods.
The performance of various algorithms used for detection of cell-cycle genes, i.e. genes regulated by and/or regulating the cell-cycle, were recently compared by de Lichtenberg et al. (2005) . Importantly, one should remember that not all genes, required for the cell-cycle are periodically expressed. One example is cdc28 kinase whose activity is phase specific but regulated by the cyclins on the protein level . Moreover, there are many periodically expressed genes that are not part of the cell-cycle machinery. For example, the gene encoding the cell wall mannoprotein TIR1 (ORF id YER011W) is known to be periodically expressed (it is included in set B1 of de Lichtenberg et al. ). In the Gene Ontology  TIR1 is annotated to the biological process of stress-response, but not to any annotation linked to the cell cycle. With this in mind, inferred periodic expression from temporal patterns of mRNA expression may be used to study to what degree cell-cycle genes are periodically expressed, but it is not expected that a detector for periodic expression will recover all cell-cycle genes.
Any detector for periodic transcription transforms the expression time profile into a real number, the test statistic, which is used to make a decision or to rank the genes. The sensitivity and specificity of the detector thus depends on the particular test statistic threshold employed and should therefore be evaluated for all relevant threshold settings in a Receiver Operating Characteristic (ROC) analysis. The ROC is usually displayed in the form of a curve that shows the probability of detection as a function of the probability of false alarm for all possible values of the threshold . Due to the small number of labelled temporal gene expression profiles presently available, it is not possible to perform reliable evaluation of the performance of the detectors on real biological data. For budding yeast S. cerevisiae there are compilations of genes known to be periodically expressed during the cell-cycle  but there are no compilations of genes known to be non-periodic. Thus, using the few genes known to be periodically expressed, a rough estimate of the probability of detection can be obtained for any threshold value but there will be no estimates at all of the probability of false alarm.
In addition to the lack of labelled data for performance validation, another difficulty is that a gene may very well be periodically expressed in one experimental condition but not in another . In conclusion, meaningful evaluations of detectors must be performed on simulated data. Since the exact distribution of the temporal profiles of genes that are periodically expressed in any experiment is not known, it is important to study the robustness of the method when it is applied to time profiles generated from models different from those assumed in the construction of the detector.
In this work we demonstrate how Bayesian inference [13, 14] can use diffuse prior knowledge about the cell-cycle period time to achieve improved detection of periodically expressed genes with a period close to that of the cell-cycle, without help from any labelled training examples provided by a supervisor. In a recently proposed empirical Bayesian approach to detection of periodically expressed genes , Bayesian inference is used to compensate for undesirable phase de-synchronization among the cells in the population during the cell-cycle experiments. However, Bayesian inference is not used in the critical step of designing a likelihood ratio based optimal detector that maximizes the probability of detection P D for a given choice of the probability of false alarm P FA . The new Bayesian detector we introduce here calculates an approximation of the desired likelihood ratio and is based on diffuse but prior knowledge about the cell-cycle period time that has not been employed in earlier approaches.
We compare the performance of the Bayesian detector introduced here to two state-of-the-art algorithms for learning-free detection of periodically expressed genes. The first is a detector based on Fisher's g-test . It is the only detector proposed for this purpose that is learning-free without requiring the period time to be exactly known. The second detector is based on a combination of two separate score values. It requires the period time to be exactly known and has performed well in a recently reported study .
Evaluation on simulated data
Simulated periodic gene expressions were generated by adding random noise drawn from a normal distribution to fixed periodic temporal waveforms (sinus, saw-tooth, square). Simulated non-periodic gene expressions were generated by samples from another normal distribution. We do not believe this choice of waveforms to correspond to naturally occurring time profiles. However, using several different waveforms provides information about the robustness of the detectors against model errors. In addition to the periodic waveforms, we also simulated gene expression profiles that were periodic with a multiplicative amplitude attenuation similar to what has been observed in experimental data [1, 15].
Comparisons of the Bayesian detector to Fisher's g-test [4, 8] and a combination test proposed by de Lichtenberg et al.  were performed. In Fisher's g-test, the sampling distribution of the magnitude I(ω k ) of the strongest Fourier spectrum component ω k in samples from a sequence of independent, identically distributed Gaussian distributions (white noise), is used in a classical hypothesis test. Specifically, the statistic
is calculated for each gene where N is the number of time points and ω k = 2π k/N. The sampling distribution of this statistic for time-series from a white noise Gaussian process has been determined analytically and is used in a hypothesis test.
In the approach by de Lichtenberg et al. , expression profiles are ranked by the combination test statistic
that depends on two p-values obtained via resampling, P regulation and P periodicity . P regulation is intended to reflect the probability that the standard deviation of a particular gene can be obtained by pure chance. P periodicity is intended to reflect the chance of obtaining a random signal with more energy in the user defined frequency component (period time) of interest than the corresponding energy in the particular gene time profile of interest. P total is defined as the product P periodicity ·P regulation .
It has been observed that the time profiles of yeast gene expression in synchronized populations exhibit attenuation [1, 15]. In order to study the effect of this attenuation, we also evaluated the result of including a multiplicative exponential term e-αt to the simulated time profiles from the periodic class. The attenutation coefficient α was chosen such that that the amplitude of expression at 100 min was 70% of that at 0 min. The attenuation coefficient α was selected to reflect the rate of attenuation that has been observed in microarray time profiles of periodically expressed genes in S. cerevisiae [1, 15]. As seen in Figure 1, the Bayesian detector performs well compared to the other two detectors also on these generative models. The performances are of course expected to become worse for all detectors with growing values of α (faster attenuation).
We also evaluated the performance of the Bayesian detector when the simulated noise was non-gaussian (Laplace and uniform) and obtained in similar results (data not shown). Since the Bayesian detector performance is more robust against the exact waveform than the other detectors, it is expected to work better also on real data.
Periodically expressed genes in S. cerevisiae
Detection of cell-cycle annotated genes
As discussed in the introduction, it is well known that some genes involved in the cell-cycle regulation are periodically expressed as well as some that are known not be periodically expressed. To what extent genes involved in the cell-cycle machinery are periodically expressed remains an open question. We addressed this question by detecting periodically expressed genes using the Bayesian detector and then analyzing how many cell-cycle genes were detected.
The cell-cycle genes were defined as the list of open reading frames annotated to the biological process "cell-cycle" (GO:0007049) or one of its descendants in the Gene Ontology . Out of the 6178 open reading frames (ORFs) on the microarray used, 290 were annotated to "cell-cycle". Selection was stringent in that only ORFs annotated with the evidence code inferred from mutant phenotype (IMP) were used.
Detection of genes with at least one phase-specific sequence motif
Detection of genes containing a specific cell-cycle motif
Discovering novel phase-specific motifs
Since the probability of detecting a gene with at least one known cell-cycle motif is relatively low for a low rate of false alarm (Figure 5) we investigated whether this could be explained by the presence of other motifs, still unknown to have phase-specific expressions. We thus tested for overrepresented motifs in the sets of ORFs detected as periodic in each of the experiments. We found that among the 319 putative motifs, 18 motifs were significantly overrepresented among the genes detected in the α-factor experiment. The putative motif MCM1' was also overrepresented in the cdc28 experiment, see Table 3.
Detector performance and significance for individual motifs Detector performance and probability of overrepresentation of motifs for detection threshold τ = 0.95.
Overrepresented putative motifs in α-factor experiment Only motifs significant at a Bonferroni corrected 0.05 level shown. * = Not significant. Italics indicates motif also significant in cdc28, see the text.
A periodic gene expression during the cell-cycle corresponds to a phase-specific expression that in turn indicates phase-specific regulation of expression. Therefore, detection of periodic expressions in cell-cycle experiments has gained a lot of attention as a computational tool for improved understanding of cell-cycle regulation. Often the investigator knows the cell-cycle period time roughly, and if not, it could be estimated experimentally. However, such estimates will in general be uncertain due to the small size of the sample sets available. In the Bayesian framework presented here, this kind of diffuse prior information about the period time can be taken into account. As shown by means of the simulations, this results in a detector which is more robust than detectors that rely on an almost exact estimate of the cell-cycle period time.
One should note that although the prior on the period time used in the simulations presented in Figure 1 is not very precise it results in superior performance in comparison with the Fisher's g-test which considers all possible periodicities. One explanation for the superior performance of the Bayesian detector on varying wave forms is due to a higher rate of false positive classifications using the g-test, since it is designed to detect all possible single periodicities. When the signals encountered contain multiple peaks in their Fourier spectra, the Bayesian detector still identifies them well, since the dominant term in the Fourier expansion is that of the desired period. However, the Fisher's g-test is designed for detection of one and only one sinusoidal signal with unknown period time. The test statistic used is the ratio of the maximum power in the Fourier spectrum to the sum of powers at all frequencies. Thus, for waveforms with the same fundamental frequency, the g-test statistic will decrease in magnitude if the Fourier spectra contain more peaks.
As expected from theory, the Bayesian detector also outperforms the combination test by de Lichtenberg et al. when the estimate of the period time deviates from the true value. This deterioration of performance might be explained by the decrease in the combination test output caused by the deviation between the estimated and true period time. When the period time estimate is correct, then the Bayesian detector and the combination test have essentially the same performance but having access to the correct period time is not a realistic possibility in most cases.
Although the present work considers detection of periodically expressed genes with a period close to that of the cell-cycle, it is important to note that the suggested Bayesian approach can easily be adjusted for design of other detectors. For example, it would be straightforward to design detectors for genes with periodic expression at a particular phase of the cell-cycle by applying a non-uniform prior on the phase angle. Furthermore, the phenomena of attenuating signals due to de-synchronization of phase in the cell population under study may be incorporated in the model. This would be an interesting direction for future work. However, we noted that the decrease in performance of the detector on sinusoidal waveforms with slow attenuation was small for a low rate of false alarm. Nevertheless we investigated time profiles of genes from set B1 of de Lichtenberg et al.  that were not detected (for high values of the detection threshold) but could not find a predominant signal of periodicity with attenuation for any of those genes. This suggests that those genes would not have been detected even if the Bayesian detector would have been modified to take attenuation into account.
The present work shows that the Bayesian detector detects more genes necessary for regulation of the cell-cycle than a random detector, but that the probability of detection is low for reasonable rates of false alarm. A trivial explanation is that some genes having a mutant cell-cycle phenotype are not periodically expressed as they are needed throughout the entire cell-cycle, or regulated otherwise. Nevertheless, our approach provides quantitative estimates as to what extent genes necessary for cell-cycle regulation are periodically expressed. As an example, for a stringent condition on periodicity, 36 out of 290 cell-cycle genes are detected as periodic in the α-factor experiment. In addition the reasonable but still relatively high level of false alarm rate may have the trivial explanation that many genes not required for the cell-cycle are also periodically expressed.
It is interesting to note the lack of strong coupling between periodic expression and cell-cycle transcription factor motifs in budding yeast. There are many genes having cell-cycle motifs in their promoter regions that are not detected as periodically expressed. This observation can be explained by complexities of gene regulation, e.g. regulation of chromatin accessibility [15, 19]. It could also be expected that a large set of genes are conditionally cell-cycle regulated depending on the environment.
It has been argued that the synchronization methods used induce many side effects not related to the core functionality of the cell-cycle , thus it is to be expected that different genes are periodically expressed in different experiments. Our analysis supports this notion, since we found that there are a number of genes that are clearly periodically expressed in only one experiment. Expression could also be regulated by combinations of transcription factors. One study supporting this idea has shown that genes sharing pairs of motifs show more similar expression patterns than when considering genes sharing single motifs .
It is also interesting to note that many genes are detected as periodically expressed while their promoter regions contain no known cell-cycle regulated elements. Wolfsberg et al. (1999) used a set of genes determined to be periodically expressed from microarray experiments to search for novel cell-cycle sequence motifs in yeast . Consistent with the present study, the presence of most known cell-cycle motifs were not well correlated with periodic expression and a similar pattern can be expected for novel motifs. However, a strong sequence signal for MCB and MCB-like motifs was found . Our results also indicate that the MCB motif is the motif that is most strongly associated with periodic transcription. One interpretation is that genes regulated by the MCB motif have less complex regulation of expression than genes with other motifs. Similar observations were made by Cho et al. who used manual identification of periodically genes in cdc28 synchronized S. cerevisiae and studied the 500 bp upstream region of the periodically expressed genes . They found that the presence of MCB and SCB motifs were associated with periodicity. However, a large number of the periodically expressed genes did not have MCB, SCB or any other of the known phase specific motifs .
We find that several putative sequence motifs  are significantly over-represented in the genes detected as periodically expressed (see Table 3). For instance, genes involved in DNA metabolism, morphogenetic activities and organization of chromosome structure contain some of these motifs. Furthermore, the putative motifs SFF' and MCM1' are associated with the known cell-cycle motifs SFF and MCM1 respectively. Thus, our findings support earlier notions that these putative motifs are important for periodic expression of cell-cycle related genes.
In addition to our simulations and experiments on real data already described, the Bayesian detector has also been applied to the biological test sets described by de Lichtenberg et al. . Since these test sets only contain positive examples, it is important to note that they do not provide any estimate at all of the probability of false alarm which is as important as the probability of detection in a ROC analysis. Therefore the results obtained using these test sets are available only as supplemental information [see Additional file 4, Additional file 5, Additional file 6].
Learning-free detectors that do not rely on a set of training examples are perhaps the only alternative when trying to identify the periodically expressed genes of poorly characterized organisms for which there is also limited knowledge available from closely related organisms. Our simulations indicate that systematic incorporation of a priori knowledge about the cell-cycle period time using Bayesian inference may improve detection performance in comparison with two other recently proposed learning-free detectors.
When applying our detector to real data from Spellman et al. , we found that genes detected as periodically expressed only contain a small fraction of cell-cycle genes inferred from mutant phenotype. We also found a statistical overrepresentation of cell-cycle regulated sequence motifs among the genes detected as periodically expressed. Moreover, among the genes detected, we found 19 motifs that have not previously been implicated in regulation of periodic expression in budding yeast. Their roles can now be tested by direct experiments and improve our understanding of cis-acting mechanisms underlying periodic transcription.
The detection problem
The Bayesian detector derived here should, after collection of a temporal profile consisting of samples y(t) from a time interval [a, b], determine whether the temporal profile contains a periodic signal with a period close to the cell-cycle time, or not. We express this as the binary hypothesis test
H1 : y t = A0 + A1 cos(ω t + θ) + u(t)
H0 : y t = B + v(t) (1)
where both u(t) and v(t) are white Gaussian stochastic processes (i.e. no time dependencies) with unknown standard deviations σu and σv, respectively. Model H 1 is regarded as a truncated Fourier series of the actual periodic signal. Adding higher order Fourier series terms as in  would be possible but it is expected that for the signals of interest to the investigator, the term containing the fundamental frequency which corresponds to the cell-cycle period time will be much larger than the rest. Support for this assumption comes from the reported success of the detectors used in [1, 2] and from the results reported here. Furthermore, in typical microarray assays of reported cell-cycle experiments, a conventional principal component analysis (PCA) of the expression data covariance matrix shows that the variance in the data can be ascribed to sinusoidal periodic expression with a period time close to that of the cell-cycle [21, 22].
One should note that the terms u and v in Eq. (1) may contain both measurement noise and model errors. We would also like to stress that by employing a Gaussian distribution for u and v, the detector becomes sensitive only to the first and second order moments of the observed errors ε 1 (t)=y t -A 0 -A 1 cos(ω o t+θ) and ε0(t) = y t - B respectively. This partly explains the robustness of our approach against different distributions of measurement noise and model errors. As long as the first and second order statistics are equal for different problems, the Bayesian test statistic will stay unchanged.
In the Bayesian framework the probabilities for model H1 and H0 respectively are obtained as
where i = 0 or i = 1. P(D|H i ) is the likelihood and P(H i ) the prior probability of model H i . The likelihood is obtained by marginalization of the unknown parameters. If model H i is parametrized by ψ i then
where P(ψ i |H i ) reflects the prior information about the parameters. Here we assume that the investigator has no prior information about the parameters except for the period time T = 2π /ω. Furthermore, we assume the parameters are independent. Hence the prior distribution factors with respect to the parameters; for model H0 we have p(B, σ v ) = p(B)p(σ v ), and for H1 we get p(A0, A1, ω, θ, σ u ) = p(A0)p(A1)p(ω)p(θ)p(σ u ). As described before, the prior p(ω) is chosen to reflect the knowledge the investigator has about the cell-cycle period time T. If it is known for certain that T is in the interval [T1, T2], a uniform prior on this interval is suitable. In the present work we focus on the situation when the investigator knows the expected mean and variance of the prior distribution of ω and we therefore use a Gaussian prior distribution for ω (the maximum entropy distribution). The hyperparameters of p(ω), the mean μω and standard deviation σω must thus be chosen to reflect the prior knowledge of the investigator. As mentioned earlier, the prior for the cell-cycle period time is chosen so that the reported interval of corresponding frequencies encompassed 70% of the probability mass in a Gaussian distribution, with its mean value equal to the average of the reported frequency interval
The priors for the remaining parameters are chosen to reflect the lack of information, thus (improper) uniform priors are used for B and A i . The prior for θ is uniform on [0, π]. For the standard deviations σ i we use Jeffrey's prior which is suitable due to its invariance to scaling of the variables involved . Thus the resulting integrals (see Eq. 3) to be computed become
for model H1 and
for model H0. The last double integral is analytically tractable but the former integral over the five parameters is not . To avoid time consuming numerical integration of that integral, we are using the standard Bayesian Information Criterion (BIC) approximation  which is defined as:
where superscript MAP indicates that the log likelihood should be evaluated at the maximum a posteriori (MAP) parameter setting. d is the number of parameters in model H i and N the number of samples available. Now the log odds of model H 1 over H0,log(P(H1|D)/P(H 0 |D)), may be approximated as BIC(H1)- BIC(H0) - τ where τ= P(H i )/P(H 0 ). Since the detector assigns a gene as periodic when this approximation is above some predefined threshold τ ', the detection rule is to assign a gene as periodic if BIC(H1)- BIC(H0) - τ > τ'. Equivalently, assign a gene as periodic if
g(D) = BIC(H1)- BIC(H0) > τ (7)
where g(D) is the detector test statistic and τ now is a redefined arbitrary detector threshold. Also, note that the test statistic g(D) in (7) is an approximation of the log-likelihood ratio which should be used to maximize P D for a given P FA (the Neyman-Pearson criterion) . The output of the detector is a function of the time (t) and gene expression (y) measurements. Finally, in order to obtain a test statistic that is bounded to the interval [0, 1], the value of g(D) was converted into a probability by the monotonous transformation
Overrepresented known motifs Only motifs significant at a Bonferroni corrected 0.05 level shown. * = Not significant. Note that LYS14 is not a known cell-cycle motif.
The detector was implemented in the R language statistical processing environment (http://www.r-project.org) and is available upon request from the authors. The algorithm simply involves locating the MAP estimates for the models (the integrands of Eq. 4 and 5) and the evaluation of Eq 8. For H 0 , an analytical solution is used. For H 1 a numerical approximation was found by optimisation using the Nelder-Mead simplex method . The initial conditions were chosen such that the standard deviation σ u and A 0 were equal to the sample standard deviation and mean of the expression measurement. The mean of the prior distribution of ω was used as its initial value. The initial values of A 1 and θ were determined from the data by an analytical least squares fit of the model A 1 cos(ω o t+θ) to the time series where ω o is the initial value of ω.
Receiver operator characteristics
The Receiver Operator Characteristic (ROC) has been developed by the signal detection community to study how the probability of detection (P D ) varies with the probability of false alarm (P FA )  and is also commonly used in medical decision making . For each selected value of the decision threshold τ∈ [0, 1], we obtain different pairs of values(P FA , P D ). A ROC graph displays a curve indicating these pairs as τ is changed continuously. Note that if the detector is inferring signal (hypothesis H 1 ) at random with probability P D , the false alarms will occur at a rate P D resulting in P FA = P D . Thus, a random detector will yield a ROC curve consisting of a diagonal line.
This work was supported by the Swedish Research Council, Wallenberg Consortium North, Carl Tryggers stiftelse (Stockholm), the Göran Gustafsson foundation (Stockholm), and the faculty of science and technology (Uppsala University).
We are grateful to Prof J Komorowski at the Linnaeus Center for Bioinformatics, Uppsala University for providing a stimulating research environment and valuable comments on the manuscript as well as to Dr T.R. Hvidsten for providing us with the compilation of sequence motifs. Finally we would like to thank the anonymous reviewers for their valuable comments.
- Spellman PT, Sherlock G, Zhang MQ, Iyer VR, Anders K, Eisen MB, Brown PO, Botstein D, Futcher B: Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Mol Biol Cell 1998, 9: 3273–3297.PubMed CentralView ArticlePubMedGoogle Scholar
- de Lichtenberg U, Jensen LJ, Fausboll A, Jensen TS, Bork P, Brunak S: Comparison of computational methods for the identification of cell cycle-regulated genes. Bioinformatics 2005, 21: 1164–1171. 10.1093/bioinformatics/bti093View ArticlePubMedGoogle Scholar
- Lu X, Zhang W, Qin ZS, Kwast KE, Liu JS: Statistical resynchronization and Bayesian detection of periodically expressed genes. Nucleic Acids Res 2004, 32: 447–455. 10.1093/nar/gkh205PubMed CentralView ArticlePubMedGoogle Scholar
- Wichert S, Fokianos K, Strimmer K: Identifying periodically expressed transcripts in microarray time series data. Bioinformatics 2004, 20: 5–20. 10.1093/bioinformatics/btg364View ArticlePubMedGoogle Scholar
- Zhao LP, Prentice R, Breeden L: Statistical modeling of large microarray data sets to identify stimulus-response profiles. Proc Natl Acad Sci U S A 2001, 98: 5631–5636. 10.1073/pnas.101013198PubMed CentralView ArticlePubMedGoogle Scholar
- Luan Y, Li H: Model-based methods for identifying periodically expressed genes based on time course microarray gene expression data. Bioinformatics 2004, 20: 332–339. 10.1093/bioinformatics/btg413View ArticlePubMedGoogle Scholar
- Johansson D, Lindgren P, Berglund A: A multivariate approach applied to microarray data for identification of genes with cell cycle-coupled transcription. Bioinformatics 2003, 19: 467–473. 10.1093/bioinformatics/btg017View ArticlePubMedGoogle Scholar
- Fisher RA: Tests of Significance in Harmonic Analysis. Proceedings of the Royal Society 1929, A: 54–59.View ArticleGoogle Scholar
- Arellano M, Moreno S: Regulation of CDK/cyclin complexes during the cell cycle. Int J Biochem Cell Biol 1997, 29: 559–573. 10.1016/S1357-2725(96)00178-1View ArticlePubMedGoogle Scholar
- Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, Harris MA, Hill DP, Issel-Tarver L, Kasarskis A, Lewis S, Matese JC, Richardson JE, Ringwald M, Rubin GM, Sherlock G: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet 2000, 25: 25–29. 10.1038/75556PubMed CentralView ArticlePubMedGoogle Scholar
- Zweig MH, Campbell G: Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clin Chem 1993, 39: 561–577.PubMedGoogle Scholar
- Shedden K, Cooper S: Analysis of cell-cycle gene expression in Saccharomyces cerevisiae using microarrays and multiple synchronization methods. Nucleic Acids Res 2002, 30: 2920–2929. 10.1093/nar/gkf414PubMed CentralView ArticlePubMedGoogle Scholar
- Gelman A: Bayesian data analysis. In Texts in statistical science. 2nd edition. Boca Raton, Fla., Chapman & Hall/CRC; 2004:xxv, 668.Google Scholar
- Jaynes ET, Bretthorst GL: Probability theory : the logic of science. New York, Cambridge University Press; 2003:xxix, 727.View ArticleGoogle Scholar
- Cho RJ, Campbell MJ, Winzeler EA, Steinmetz L, Conway A, Wodicka L, Wolfsberg TG, Gabrielian AE, Landsman D, Lockhart DJ, Davis RW: A genome-wide transcriptional analysis of the mitotic cell cycle. Mol Cell 1998, 2: 65–73. 10.1016/S1097-2765(00)80114-8View ArticlePubMedGoogle Scholar
- Hughes JD, Estep PW, Tavazoie S, Church GM: Computational identification of cis-regulatory elements associated with groups of functionally related genes in Saccharomyces cerevisiae. J Mol Biol 2000, 296: 1205–1214. 10.1006/jmbi.2000.3519View ArticlePubMedGoogle Scholar
- Pilpel Y, Sudarsanam P, Church GM: Identifying regulatory networks by combinatorial analysis of promoter elements. Nat Genet 2001, 29: 153–159. 10.1038/ng724View ArticlePubMedGoogle Scholar
- Breeden LL: Periodic transcription: a cycle within a cycle. Curr Biol 2003, 13: R31–8. 10.1016/S0960-9822(02)01386-6View ArticlePubMedGoogle Scholar
- Lemon B, Tjian R: Orchestrated response: a symphony of transcription factors for gene control. Genes Dev 2000, 14: 2551–2569. 10.1101/gad.831000View ArticlePubMedGoogle Scholar
- Wolfsberg TG, Gabrielian AE, Campbell MJ, Cho RJ, Spouge JL, Landsman D: Candidate regulatory sequence elements for cell cycle-dependent transcription in Saccharomyces cerevisiae. Genome Res 1999, 9: 775–792.PubMed CentralPubMedGoogle Scholar
- Holter NS, Maritan A, Cieplak M, Fedoroff NV, Banavar JR: Dynamic modeling of gene expression data. Proc Natl Acad Sci U S A 2001, 98: 1693–1698. 10.1073/pnas.98.4.1693PubMed CentralView ArticlePubMedGoogle Scholar
- Holter NS, Mitra M, Maritan A, Cieplak M, Banavar JR, Fedoroff NV: Fundamental patterns underlying gene expression profiles: simplicity from complexity. Proc Natl Acad Sci U S A 2000, 97: 8409–8414. 10.1073/pnas.150242097PubMed CentralView ArticlePubMedGoogle Scholar
- Bretthorst GL: Bayesian spectrum analysis and parameter estimation. New York, Springer-Verlag; 1988:xii, 209.View ArticleGoogle Scholar
- Van Trees HL: Detection, estimation, and modulation theory. New York,, Wiley; 1968:3 v..Google Scholar
- Nelder JA, Mead R: A simplex algorithm for function minimization. Computer Journal 1965, 7: 308–313.View ArticleGoogle Scholar
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.