Clustering with positionspecific constraints on variance: Applying redescending Mestimators to labelfree LCMS data analysis
 Rudolf Frühwirth†^{1},
 D R Mani†^{2}Email author and
 Saumyadipta Pyne^{2, 3}Email author
DOI: 10.1186/1471210512358
© Frühwirth et al; licensee BioMed Central Ltd. 2011
Received: 8 June 2011
Accepted: 31 August 2011
Published: 31 August 2011
Abstract
Background
Clustering is a widely applicable pattern recognition method for discovering groups of similar observations in data. While there are a large variety of clustering algorithms, very few of these can enforce constraints on the variation of attributes for data points included in a given cluster. In particular, a clustering algorithm that can limit variation within a cluster according to that cluster's position (centroid location) can produce effective and optimal results in many important applications ranging from clustering of silicon pixels or calorimeter cells in highenergy physics to labelfree liquid chromatography based mass spectrometry (LCMS) data analysis in proteomics and metabolomics.
Results
We present MEDEA (MEstimator with DEterministic Annealing), an Mestimator based, new unsupervised algorithm that is designed to enforce positionspecific constraints on variance during the clustering process. The utility of MEDEA is demonstrated by applying it to the problem of "peak matching"identifying the common LCMS peaks across multiple samplesin proteomic biomarker discovery. Using reallife datasets, we show that MEDEA not only outperforms current stateoftheart modelbased clustering methods, but also results in an implementation that is significantly more efficient, and hence applicable to much larger LCMS data sets.
Conclusions
MEDEA is an effective and efficient solution to the problem of peak matching in labelfree LCMS data. The program implementing the MEDEA algorithm, including datasets, clustering results, and supplementary information is available from the author website at http://www.hephy.at/user/fru/medea/.
Background
Protein or peptide biomarkers offer great promise in early detection, monitoring and targeted treatment of diseases. Two main strategies have been employed in proteomic biomarker discovery, identitybased and patternbased methods. Identitybased methods use high quality tandem mass spectrometry (LCMS/MS) and identify potential biomarkers among the sequenced peptides [1–3]. While identity makes the task of biomarker validation easier, the approach ignores unidentified peaks in the mass spectra resulting in significant information loss, and has limited throughput due to the need for extensive fractionation. Patternbased, or labelfree approaches [4–6], on the other hand, look for discriminating peak patterns in mass spectra, without regard to their identity. While initial attempts at patternbased biomarker discovery using low quality instrumentation and improper validation were met with criticism [7, 8], the approach nonetheless has merit [9]. Indeed the design and implementation of the PEPPeR platform for proteomic biomarker discovery [10] was an attempt to distill the best of both worlds in a robust, high throughput analytical platform for biomarker discovery. It combined both identity and pattern based approaches to capitalize on the merits of each, while exploiting synergies to minimize the drawbacks, enhancing our ability to successfully find and validate biomarkers.
PEPPeR uses high resolution and high mass accuracy liquid chromatographybased mass spectrometry (LCMS) data from stateoftheart mass spectrometers, and appropriately combines patternbased (unidentified peptide peaks) and identitybased (peptides sequenced via MS/MS, or tandem mass spectrometry) information to generate peptide quantitation for biomarker discovery. From a computational standpoint, the uniqueness of this approach stems from the use of: (i) identified peptides to set automatically calculated matching tolerances for guiding the alignment of unidentified peaks; (ii) matching unidentified peaks across multiple samples (peak matching) using mixture model based clustering. In the present study, we introduce a new algorithm MEDEA (MEstimator with DEterministic Annealing) that can enhance the analytical capacity of the PEPPeR platform. Using two reallife LCMS datasets, and a robust statistical approach, we show how MEDEA can provide a more accurate and efficient solution to the problem of peak matching.
The PEPPeR algorithm
A key challenge in the design of PEPPeR is the implementation of peak matching. An LCMS peak is identified by a masstocharge ratio m/z, its LC retention time RT and its charge z[10]. The presence of a specific peptide in a sample analyzed by LCMS will result in a peak at a given (m/z, RT, z) location. The intensity of the peak reflects the peptide abundance in the sample.
Due to the inherent limits of chromatography and mass spectrometry, repeated measurements of the same peptide, or measurements of the same peptide in multiple samples will result in variations in the determined m/z and RT values. The m/z variation is dictated by the mass accuracy of the mass spectrometer. For successful application of patternbased approaches like PEPPeR, high mass accuracy is required in order to distinguish the many peptides that arise in the analysis of realworld samples. Acceptable m/z variation for PEPPeR ranges from a few parts per million (ppm) to a few 10's of ppm [10], and is easily achieved by instruments such as the LTQ OrbiTrap [11]. It is worth noting that m/z variation is a function of the actual m/z valuea ppm precision specification allows for larger variation when the m/z values are higher. Retention time variation, on the other hand, is based on chromatography and the physicochemical properties of peptides. Typical RT variation for a peptide peak ranges from a fraction of a minute for wellbehaved peptides under high performance chromatography, to several minutes under chromatographic runs extending 90120 minutes [10]. Unlike m/z variation, RT variation can be treated as a constant limit for the entire chromatographic run.
Allowable limits of variation (tolerances) for m/z and RT are determined based on MS/MS sequenced peptides with confident identities obtained by database searching. A process called landmark matching [10] is used to propagate identities ("landmarks") across many samples, so that a subset of confidently identified peptides are present across multiple samples. Tolerances are calculated using these common landmark peptides. The upper limit of the range of variation actually observed (over multiple samples) for m/z and RT values (after excluding outliers) is defined as the variation tolerance for m/z and RT, respectively, for all peaks (sequenced or otherwise) during the peak matching process.
The need for varianceconstrained clustering
Application of MCLUST, or indeed of most current approaches, to the problem of proteomic peak matching requires the enforcement of tolerance constraints in order to limit membership to only those elements that have attributes (m/z or RT) whose values fall within the allowable variation limits (set by m/z and RT tolerances, respectively). To achieve this goal, the repeated splitmerge approach used post hoc in PEPPeR can (i) result in suboptimal clusters (see Subsection "Cluster Quality" below for an example); and can (ii) be computationally expensive and timeconsuming when a large number of peaks are being clustered (see Subsection "Implementation" below).
Suboptimal clustering can result in splitting a given peptide across multiple clusters, or conversely, merging distinct peptides into one cluster. This can result in unnecessary false positive or false negative markers, undermining the process of biomarker discovery (see Subsection "Cluster Quality"). Furthermore, splitmerge based tolerance enforcement in PEPPeR has resulted in significant limitations when applying the platform to the analysis of large datasets arising from studies involving significant (several tens to a few hundred) numbers of patients. An efficient algorithm that can automatically enforce tolerance constraints during the clustering processlike the MEDEA algorithm presented hereproduces more optimal clusters and enables unsupervised analysis of much larger datasets.
The MEDEA varianceconstrained clustering algorithm is described in the following section. The remaining sections address the application of MEDEA to LCMS peak matching, comparative analysis with the currently used method, followed by discussion and conclusions.
Methods
A redescending Mestimator with annealing
where r_{ i } = (x_{ i }  μ)/σ and ψ(r) = dρ/dr.
A special class of Mestimators is formed by redescending Mestimators. They are widely used for robust regression and regression clustering, e.g. see [18, 19] and the references therein. According to the definition in [20], the ψfunction of a redescending Mestimators has to disappear outside a certain central interval. In the following, we merely demand that the ψfunction tends to zero for r → ∞. If ψ tends to zero sufficiently fast, observations lying farther away than a certain bound are effectively discarded.
for all c, and the Mestimator degenerates into a leastsquares estimator. If the temperature drops to zero, the weight function converges to a step function, and the Mestimator approaches the skipped mean. For more details about the estimator and its influence function, see [23].
Clustering Algorithm
 1.
Select an unused peak (the seed) as a cluster center, and find all unused peaks of the same charge in a search frame of size ± 3δ_{ i } .
 2.
Set the iteration number to k = 1.
 3.
Set the temperature to T = T_{ k } .
 4.
Compute the weights of all peaks in the frame relative to the current cluster center.
 5.
Compute the new cluster center by the weighted mean of all peaks in the frame and recompute the search frame.
 6.
Set k := k + 1 and go to 3, unless the maximum number of iterations is reached.
 7.
Mark all peaks in the cluster as used and go to 1.
The starting temperature is T_{1} = 8. This temperature is sufficiently high so that the weights are nonnegligible throughout the search frame constructed in Step 1. The cluster center therefore moves toward the centerofmass of the peaks in the search frame. A higher starting temperature would just slow down the annealing without substantially changing the final clustering. In the following steps, the temperature is lowered according to a predefined annealing schedule. We have chosen an approximately exponential schedule (see [21]): T_{2} = 6, T_{3} = 4, T_{4} = 3, T_{5} = 2, T_{6} = 1.5, T_{7} = ⋯ = T_{11} = 1. Several steps at T = 1 are performed to allow convergence of the Mestimator. The final temperature T_{12} = 0.25 is much lower than 1 and results in a sharp cut at the boundaries of the box.
It is possible to use the Mestimator without annealing, i.e., with starting temperature T_{1} = 1. Cohesive clusters that are well separated will still be effectively identified by the algorithm in the absence of annealing. In fact, the application to LCMS peak clustering encounters many such isolated clusters as indicated by the large number of clusters that require only one iteration in Figure 3. But, there are a substantial number of clusters that are harder to identify, and benefit from a larger number of annealing steps. In these cases the clustering with annealing will produce different groupings, as a higher starting temperature initially explores a larger range in the data space and the cluster center is attracted more strongly to the region of highest peak density in the search frame.
Postprocessing
In the low temperature limit the Mestimator is a skipped mean, which means that cluster center is the arithmetic mean of all peaks in the box. Thus it may happen that a cluster with a few outlying points is split into two, although the entire cluster fits into a box of the prescribed size. Another shortcoming of the algorithm described above is due to its sequential nature. As the peaks attached to a cluster are not made available any more to subsequent clusters, there is no globally optimal association of peaks to clusters. In order to compensate for these shortcomings we have designed a postprocessing algorithm that has two stages. In the first stage, in every region of overlapping clusters, peaks are assigned to the closest cluster center in their vicinity. In the second stage, clusters are fused if their union fits into a single box.
Stage 1: Globally optimal assignment
 1.
Set the temperature to T = 1.
 2.
Select an unused cluster i.
 3.
Find the set J of all clusters j with (i, j) ∈ R.
 4.
Find the set K of all peaks in any of the clusters in J.
 5.
Compute the weights of all peaks in set K relative to all cluster centers in J.
 6.
Associate each peak to the cluster with the largest weight.
 7.
If the association has changed, recompute all cluster centers and go to 5; if not, mark all clusters in J as used and go to 2.
Stage 2: Cluster fusion
 1.
Select an unused cluster i.
 2.
Find the set J of all clusters that are overlapping with i.
 3.
Compute the weights of all cluster centers in J relative to cluster center i.
 4.
Select the cluster center with the largest weight and call it j.
 5.
If clusters i and j fit into a single box, assign all peaks in cluster j to cluster i and drop cluster j.
 6.
Mark cluster i as used and go to 1.
Results
Dataset generation
In order to validate the effectiveness of MEDEA and to compare it with MCLUST, we have used two LCMS datasets: (i) MitoMix: mitochondrial extracts from mice with a set of proteins spiked in at known levels [10]; and (ii) DarTB: plasma samples from tuberculosis cases and controls collected at Dar es Salaam, Tanzania, as part of the Gates Grand Challenges in Global Health GC13 project on patternbased proteomic characterization of the epidemiology (prevalence and incidence) of diseases of major importance in the developing world.
The MitoMix dataset was generated from mitochondrial extracts from C57BL6/J mice aged either 2 or 6 weeks, as described previously in [24]. The 2week extract was spiked with Variability Mix α, the 6week extract was spiked with Variability Mix β prior to digestion. The variability mixes contain 12 proteins (not found in mouse plasma) spiked in at different levels in the α and β mixes, as set out in [10]. The samples were reduced, alkylated and trypsin digested, followed by desalting and lyophilization. The samples were then reconstituted and analyzed using reverse phase LCMS on a LTQFT (Thermo Scientific) mass spectrometer, with an 85 minute gradient. MS acquisition consisted of a high resolution precursor MS scan (FTMS) followed by three datadependent MS/MS scans (ion trap) [10]. Each of the two samples (2week extract with mix α, and 6 week extract with mix β) was run in six technical replicates to generate a total of 12 raw files that were subsequently analyzed using MCLUST and MEDEA.
The DarTB dataset consists of 20 TB case and 20 healthy control plasma samples collected at Dar es Salaam, Tanzania. The samples were shipped to The Broad Institute where they are run through a sample processing pipeline starting with immunoaffinity depletion of the top 14 abundant human proteins using an Agilent MARS14 depletion column. The depleted plasma is passed through a low molecular weight filter and subjected to reduction, alkylation and trypsin digestion. The digested sample is then fractionated into ten fractions using a basic pH reverse phase column. Fractions 5, 6, and 7 are analyzed via LCMS on a Thermo LTQFT using a 98 min gradient. The resulting 120 raw files are analyzed using MCLUST and MEDEA.
Data processing
Raw files generated for the MitoMix and DarTB samples by LCMS were extracted and interpreted using SpectrumMill (Agilent Technologies, CA) to provide sequence identities for peptides subject to MS/MS. The raw files were also converted to mzXML and processed using msInspect [25] for peak detection, isotope deconvolution and charge state assignment. The peaks identified by msInspect are parameterized by masstocharge ratio, retention time and charge (m/z, RT, z). These peaks are then merged with confident peptide sequence identities extracted by SpectrumMill. The result is a table of peakssome if which are sequence identifiedfor each LCMS run. The landmark matching algorithm [10] is used to propagate confident identities across samples to maximize identified peptides (landmarks) in each sample. These peaks are then subject to m/z and RT correction [10] to minimize runtorun variation. The m/z and RT variation of common landmarks across samples are used to define m/z tolerance and RT tolerance as described in Figure 1. The peak lists from all samples in a dataset are then concatenated to generate the data table that is used for MCLUST and MEDEA analysis.
Cluster quality
The MitoMix data set
MCLUST vs. MEDEA Comparison for MitoMix Data
MCLUST  MEDEA  

Number of clusters  23448  20765 
Average cluster size  3.95  4.47 
Average cluster diameter in m/z  6.25E4  8.34E4 
Average cluster diameter in RT  0.0778  0.1060 
Computing time [s]  2342  198 
MCLUST vs. MEDEA Comparison for MitoMix Data
Number of clusters  MCLUST  MEDEA 

1  1788  1955 
2  667  571 
3  83  42 
4  27  13 
5  13  3 
6  6  5 
7  4  0 
8  0  0 
9  1  0 
Mean  1.40  1.28 
The DarTB data set
MCLUST vs. MEDEA Comparison for DarTB Data
MCLUST  MEDEA  

Number of clusters  287838  218098 
Average cluster size  2.27  3.00 
Average cluster diameter in m/z  9.32E4  0.0023 
Average cluster diameter in RT  0.39  0.96 
Computing time [s]  663756  17532 
MCLUST vs. MEDEA Comparison for DarTB Data
Number of clusters  MCLUST  MEDEA 

1  812  1046 
2  489  460 
3  215  127 
4  98  40 
5  46  22 
6  24  7 
7  10  5 
8  7  3 
9  3  4 
10  7  4 
> 10  9  2 
Mean  2.06  1.64 
Discussion
Validation of clustering results
Clearly the MEDEA clusters show significantly less dispersion per peptide than MCLUST as observed by the much higher density of ratios greater than 1 (i.e. for the darker bins in the right half of the histograms) in both datasets.
Implementation
Compared to the MCLUST implementation of peak matching in PEPPeR, MEDEA is significantly more efficient and results in speedups ranging from about 10 for small datasets to about a factor of 40 for large datasets. In order to compare peak matching speed with MCLUST and MEDEA, we time the entire peak matching process which not only includes the actual clustering part, but involves pre and postprocessing steps, many of which require reading and writing large files, in addition to operations on large tables. MCLUSTbased clustering cannot be performed in PEPPeR without the use of some of these processing steps, and in order to perform a fair comparison (from the perspective of efficient PEPPeR implementation), we target the entire peak matching process.
For the MitoMix data MCLUST based peak matching was run on a cluster using 50 processes. The total computing time was calculated as the sum of the time taken by each of the processes, and amounted to 2,343 sec. MEDEA based peak matching took 198 sec on one of the cluster nodes, resulting a speedup of 11.8 (Table 1). For the much larger DarTB data set, MCLUST based peak matching, run on 2,500 processors, had a total computing time of 663,756 sec. MEDEA resulted in a speed up of 37.8 with a single node computing time of 17,532 sec (Table 3). With such speedups on large data sets, MEDEA makes it feasible to process and analyze significantly larger data sets.
For these data sets, the PEPPeR pre and postprocessing adds significant amount of computing time, especially because of large file input/output operations. When MEDEA is run without this overhead, the clustering is even more efficient, with approximate computing times of 45 sec and 1,500 sec on a typical desktop computer for the MitoMix and DarTB datasets, respectively.
Conclusions
Clustering analysis is used for identifying groups of similar points in data in an unsupervised manner. Popular clustering approaches include hierarchical or partitional algorithms such as agglomerative and kmeans clustering. In addition, finite mixture models have been used extensively in biology and medicine [12, 14, 26]. Often semisupervised approaches are used for imposing various types of constraints on clustering [27]. Yet few algorithms can address the challenge of enforcing constraints on the variance of the clusters, especially in an unsupervised manner. The challenge is clearly not addressed with a sliding window approach that cannot identify the group structure inherent in the data. In particular, the problem assumes further importancein various fields ranging from engineering to economics and biotechnologyif the required constraints on the variance of a cluster are dependent on its position. For example, for ChIPSeq experiments, the appearance of tags along the genome could be modeled by a discrete Poisson distribution, thus requiring the variance of a peak to be equal to its mean [28, 29].
Here we have presented a new approach based on robust statistics for identifying clusters in continuous data that respect positionspecific constraints on clustervariance. In this regard, we find the application of Mestimators most suitable. In particular, we developed MEDEA as an effective and fast solution to the problem of peak matching in labelfree LCMS data. By analyzing reallife samples, we have shown that MEDEA is not only significantly more efficient (achieving speedups of up to about 40), but also produces clusters that are more coherent. Data points that are known to arise from the same peptide are more consistently grouped into the same cluster when compared to peak matching based on Gaussian mixture model based clustering using MCLUST, as validated by our comparative analysis. Given its potential usefulness for practical proteomic analysis, the MEDEA algorithm has been integrated into the PEPPeR pipeline.
In the context of applying constrained clustering to peak matching in LCMS data, the choice of constraints is critical. Here, the m/z and RT variation tolerance values provide the constraints that MEDEA enforces. If the constraints are too wide, then two different peptides could end up being clustered into a single group. On the other hand, constraints that are too small can cause a peptide to be split into two different clusters, thereby making the analysis of such data for purposes like biomarker discovery much more complex and unreliable. In the PEPPeR platform, the m/z and RT tolerances are determined based on the observed landmark peptides, and set using the m/z and RT variation ranges for the landmark peptides, after removing outliers. In spite of that, there could be situations where an aberrant peptide elutes over a long period of time, or when multiple isobaric peptides elute within the given RT tolerance window. Under these situations, it would be impossible for any clustering algorithm to correctly group the peptides without knowledge of the actual peptide sequence obtained by tandem MS or other means.
While MEDEA enforces the specified constraints on the variation within a cluster, it does so without assuming an explicit model, Gaussian or otherwise, for the distribution of peaks in the cluster. Clusters members are thus identified solely by the proximity of their peaks, independent of any assumed parametric distribution, as long as they fit into a box of the prescribed size (i.e., satisfy required constraints) around the center of gravity of the cluster. Outliers that respect the constraints are integrated into the cluster with systematic postprocessing. While the issue of robustness for cluster outliers, often due to asymmetric or heavy tailed effects, has recently been addressed with new parametric algorithms (e.g., finite mixtures of multivariate skew t distributions [30–32]), such robustness would tend to includerather than excludecluster outliers in the heavytailed distributions. Furthermore, as in the case of MCLUST, such modelbased clustering methods are not capable of enforcing userspecified constraints on cluster extension. Hence, the constraints would again have to be imposed aposteriori, requiring a computationally expensive splitandmerge algorithm, similar to the one outlined in Figure 1. Therefore we believe that MEDEA, with its unique combination of a robust estimator with automatic constraint enforcement, presents a useful and effective approach that fills an important gap in clustering applications.
Notes
List of abbreviations used
 ChIPSeq:

chromatin immunoprecipitation sequencing
 FTMS:

Fourier transform mass spectrometry
 LC:

liquid chromatography
 LCMS:

liquid chromatography based mass spectrometry
 LCMS/MS:

liquid chromatography based tandem mass spectrometry
 LTQ:

linear trap quadrupole
 MCLUST:

model based clustering
 MEDEA:

Mestimator with deterministic annealing
 m/z:

masstocharge ratio
 PEPPeR:

platform for experimental proteomic pattern recognition
 RT:

retention time
 z :

charge
Declarations
Acknowledgements
This work was supported in part by a grant R01CA126219 to D. R. Mani from the National Cancer Institute as part of the Clinical Proteomic Technologies Initiative. The MitoMix data is publicly available as part of the PEPPeR publication [10]. The DarTB data was generated as part of the Gates Grand Challenges in Global Health (GC13) grant to Steven A. Carr (Broad Institute, Cambridge, MA, USA). The sample collection and protocols were conceived by Michael A. Gillette and Rushdy Ahmad, and the samples were processed with help from Margaret Pyle (all from the Broad Institute, Cambridge, MA, USA). This data will be published as part of a separate manuscript under preparation.
Authors’ Affiliations
References
 Adkins JN, Varnum SM, Auberry KJ, Moore RJ, Angell NH, Smith RD, Springer DL, Pounds JG: Toward a Human Blood Serum Proteome: Analysis By Multidimensional Separation Coupled With Mass Spectrometry. Molecular & Cellular Proteomics 2002, 1(12):947–955. 10.1074/mcp.M200066MCP200View ArticleGoogle Scholar
 Rifai N, Gillette MA, Carr SA: Protein biomarker discovery and validation: the long and uncertain path to clinical utility. Nature Biotechnology 2006, 24(8):971–983. 10.1038/nbt1235View ArticlePubMedGoogle Scholar
 Srinivas PR, Verma M, Zhao Y, Srivastava S: Proteomics for Cancer Biomarker Discovery. Clinical Chemistry 2002, 48(8):1160–1169.PubMedGoogle Scholar
 Adam B, Qu Y, Davis JW, Ward MD, Clements MA, Cazares LH, Semmes OJ, Schellhammer PF, Yasui Y, Feng Z, Wright GL: Serum protein fingerprinting coupled with a patternmatching algorithm distinguishes prostate cancer from benign prostate hyperplasia and healthy men. Cancer Research 2002, 62(13):3609–3614.PubMedGoogle Scholar
 Petricoin EF, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM, Mills GB, Simone C, Fishman DA, Kohn EC, Liotta LA: Use of proteomic patterns in serum to identify ovarian cancer. Lancet 2002, 359(9306):572–577. 10.1016/S01406736(02)077462View ArticlePubMedGoogle Scholar
 Villanueva J, Philip J, Entenberg D, Chaparro CA, Tanwar MK, Holland EC, Tempst P: Serum peptide profiling by magnetic particleassisted, automated sample processing and MALDITOF mass spectrometry. Analytical Chemistry 2004, 76(6):1560–1570. 10.1021/ac0352171View ArticlePubMedGoogle Scholar
 Baggerly KA, Morris JS, Coombes KR: Reproducibility of SELDITOF protein patterns in serum: comparing datasets from different experiments. Bioinformatics 2004, 20(5):777–785. 10.1093/bioinformatics/btg484View ArticlePubMedGoogle Scholar
 Sorace JM, Zhan M: A data review and reassessment of ovarian cancer serum proteomic profiling. BMC Bioinformatics 2003, 4: 24. 10.1186/14712105424PubMed CentralView ArticlePubMedGoogle Scholar
 Gillette MA, Mani DR, Carr SA: Place of pattern in proteomic biomarker discovery. Journal of Proteome Research 2005, 4(4):1143–1154. 10.1021/pr0500962View ArticlePubMedGoogle Scholar
 Jaffe JD, Mani DR, Leptos KC, Church GM, Gillette MA, Carr SA: PEPPeR, a platform for experimental proteomic pattern recognition. Molecular & Cellular Proteomics 2006, 5(10):1927–1941. 10.1074/mcp.M600222MCP200View ArticleGoogle Scholar
 Olsen JV, de Godoy LMF, Li G, Macek B, Mortensen P, Pesch R, Makarov A, Lange O, Horning S, Mann M: Parts per million mass accuracy on an Orbitrap mass spectrometer via lock mass injection into a Ctrap. Molecular & Cellular Proteomics 2005, 4(12):2010–2021. 10.1074/mcp.T500030MCP200View ArticleGoogle Scholar
 Banfield JD, Raftery AE: Modelbased Gaussian and nonGaussian clustering. Biometrics 1993, 49(3):803–821. 10.2307/2532201View ArticleGoogle Scholar
 Ganesalingam S, McLachlan GJ: A case study of two clustering methods based on maximum likelihood. Statistica Neerlandica 1979, 33: 81–90. 10.1111/j.14679574.1979.tb00665.xView ArticleGoogle Scholar
 McLachlan GJ, Basford KE: Mixture Models: Inference and Application to Clustering. Marcel Dekker, New York; 1988.Google Scholar
 Fraley C, Raftery AE: MCLUST: Software for ModelBased Clustering, Density Estimation and Discriminant Analysis. 2003.Google Scholar
 R Development Core Team: R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna, Austria. 2009. ISBN 3–900051–07–0 ISBN 3900051070Google Scholar
 Huber PJ: Robust Statistics: Theory and Methods. John Wiley & Sons, New York; 1981.View ArticleGoogle Scholar
 Garlipp T, Müller Ch: Regression clustering with redescending Mestimators. In Innovations in Classification, Data Science, and Information Systems. Edited by: Baier D, Wernecke KD. Springer, Berlin, Heidelberg, New York; 2005.Google Scholar
 Müller Ch: Redescending Mestimators in regression analysis, cluster analysis and image analysis. Discussiones Mathematicae Probability and Statistics 2004, 24(1):59–75.Google Scholar
 Hampel FR, Ronchetti EM, Rousseeuw PJ, Stahel WA: Robust Statistics: The Approach Based on Influence Functions. John Wiley & Sons, New York; 1986.Google Scholar
 Rose K: Deterministic annealing for clustering, compression, classification, regression, and related optimization problems. Proceedings of the IEEE 1998, 86(11):2210–2239. 10.1109/5.726788View ArticleGoogle Scholar
 Li SZ: Robustizing robust Mestimation using deterministic annealing. Pattern Recognition 1996, 29(1):159–166. 10.1016/00313203(95)000712View ArticleGoogle Scholar
 Frühwirth R, Waltenberger W: Redescending Mestimators and Deterministic Annealing, with Applications to Robust Regression and Tail Index Estimation. Austrian Journal of Statistics 2008, 37(3&4):301–317. [http://www.stat.tugraz.at/AJS/ausg083+4/08306Fruehwirth.pdf]Google Scholar
 Mootha VK, Bunkenborg J, Olsen JV, Hjerrild M, Wisniewski JR, Stahl E, Bolouri MS, Ray HN, Sihag S, Kamal M: Integrated Analysis of Protein Composition, Tissue Diversity, and Gene Regulation in Mouse Mitochondria. Cell 2003, 115(5):629–640. 10.1016/S00928674(03)009267View ArticlePubMedGoogle Scholar
 Bellew M, Coram M, Fitzgibbon M, Igra M, Randolph T, Wang P, May D, Eng J, Fang R, Lin C, Chen J, Goodlett D, Whiteaker J, Paulovich A, McIntosh M: A suite of algorithms for the comprehensive analysis of complex protein mixtures using highresolution LCMS. Bioinformatics 2006, 22(15):1902–1909. 10.1093/bioinformatics/btl276View ArticlePubMedGoogle Scholar
 McLachlan GJ, Peel KE: Finite Mixture Models. Wiley, New York; 2000.View ArticleGoogle Scholar
 Basu S, Davidson I, Wagstaff K: Constrained Clustering Advances in Algorithms, Theory, and Applications. Chapman and Hall/CRC, Boca Raton; 2008.Google Scholar
 Mikkelsen TS, Ku M, Jaffe DB, Issac B, Lieberman E, Giannoukos G, Alvarez P, Brockman W, Kim TK, Koche RP, Lee W, Mendenhall E, O'Donovan A, Presser A, Russ C, Xie X, Meissner A, Wernig M, Jaenisch R, Nussbaum C, Lander ES, Bernstein BE: Genomewide maps of chromatin state in pluripotent and lineagecommitted cells. Nature 2007, 448:, 553–560.PubMed CentralView ArticlePubMedGoogle Scholar
 Zhang Y, Liu T, Meyer CA, Eeckhoute J, Johnson DS, Bernstein BE, Nussbaum C, Myers RM, Brown M, Li W, Liu XS: Modelbased Analysis of ChIPSeq (MACS). Genome Biology 2008 2008, 9: R137.Google Scholar
 Pyne S, Hu X, Wang K, Rossin E, Lin TI, Maier L, BaecherAllan C, McLachlan GJ, Tamayo P, Hafler DA, De Jager PL, Mesirov JP: Automated highdimensional flow cytometric data analysis. Proceedings of the National Academy of Sciences of the United States of America 2009, 106(21):8519–8524. 10.1073/pnas.0903028106PubMed CentralView ArticlePubMedGoogle Scholar
 FrühwirthSchnatter S, Pyne S: Bayesian inference for finite mixtures of univariate and multivariate skew normal and Skew t Distributions. Biostatistics 2010, 11: 317–336. 10.1093/biostatistics/kxp062View ArticlePubMedGoogle Scholar
 Ho H, Pyne S, Lin T: Maximum likelihood inference for mixtures of skew Student t normal distributions through practical EMtype algorithms. Statistics and Computing 2011, in press.Google 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.