Dr Thomas Bartlett

* @ucl.ac.uk

Research Interests

Stochastic networks and processes; high-dimensional statistics and sparse statistical modelling; computational statistics; models of genomic and epigenomic data.

Awards/Prizes

• 2017 MRC Fellowship (skills development, quantitative). Awarded by U.K. research council, to fund three years’ independent work developing novel statistical methodology for biomedical science applications.
• 2015 EPSRC Doctoral prize. Fellowship awarded by U.K. research council, to fund two years’ independent postdoctoral work following my PhD.
• 2013 Bogue Fellowship awarded by University College London, to fund a 4-month research visit to Columbia University, New York, U.S.A.

List of Publications

• 2017 Single-cell Co-expression Subnetwork Analysis. Bartlett TE, Müller S, Diaz A. Nature Scientific Reports 7(1): 15066
• 2017 Network inference and community detection, based on covariance matrices, correlations and test statistics from arbitrary distributions. Bartlett TE. Communications in Statistics - Theory and Methods 46(18): 9150-9165
• 2017 Parenclitic network analysis of methylation data for cancer identification. Karsakov A, Bartlett TE, Ryblov A, Meyerov I, Ivanchenko M, Zaikin A. PloS one, 12(1): e0169661
• 2016 Detection of Epigenomic Network Community Oncomarkers. Bartlett TE, Zaikin A. Annals of Applied Statistics 10(3): 1373-1396. Preprint available: http://arxiv.org/abs/1506.05244
• 2016 Epigenetic reprogramming of Fallopian tube fimbriae in BRCA mutation carriers defines early ovarian cancer evolution. Bartlett TE, Widschwendter M, et al. Nature Communications 7: 11620
• 2015 Intra-gene DNA Methylation Variability is a Clinically Independent Prognostic Marker in Women's Cancers. Bartlett TE, Widschwendter M, et al. PLoS one, 10(12): e0143178
• 2015 Glioblastoma Stem Cells Respond to Differentiation Cues but Fail to Undergo Commitment and Terminal Cell-Cycle Arrest. Carén H, Stricker SH, Bulstrode H, Gagrica S, Johnstone E, Bartlett TE, Feber A, Wilson G, Teschendorff AE, Bertone P, Beck S, Pollard SM. Stem cell reports, 5(5): 829-842
• 2014 A DNA methylation network interaction measure, and detection of network oncomarkers. Bartlett TE, Olhede SC, Zaikin A. PLoS one, 9(1): e84573
• 2013 Corruption of the intra-gene DNA methylation architecture is a hallmark of cancer. Bartlett TE, Zaikin A, Olhede SC, West J, Teschendorff AE, Widschwendter M. PLoS One, 8(7): e68285
• 2013 A beta-mixture quantile normalization method for correcting probe design bias in Illumina Infinium 450K DNA methylation data. Teschendorff AE, Marabita F, Lechner M, Bartlett TE, Tegner J, Gomez-Cabrero D, Beck S. Bioinformatics, 29(2): 189-196

Invited Talks

• 2018 Two-way sparsity for time-varying networks, with applications in genomics. Joint Statistical Meeting. Vancouver, Canada.
• 2017 A dynamic network model for single-cell genomic data. University College London Department of Statistics seminar series. London, U.K.
• 2016 Stochastic network models for `omics applications. Joint Statistical Meeting. Chicago, U.S.A.
• 2015 A Power Variance Test for Nonstationarity in Complex-Valued Signals. IEEE Conference on Machine Learning Applications. Miami, U.S.A.
• 2015 Uni- and Bi-Partite Stochastic Network Models with Applications to 'Omics Data. UC Berkeley Department of Statistics, Statistics and Genomics seminar series. Berkeley, U.S.A.
• 2015 Network inference and community detection, based on covariance matrices, correlations and test statistics from arbitrary distributions. University College London Department of Statistics, Stochastic Processes Group seminar series. London, U.K.
• 2015 Statistical Modelling of Stochastic Processes in Epigenetics. North West Research Associates. Seattle, U.S.A.
• 2014 Statistical Network Methodology for Biomarker Detection. Joint Statistical Meeting. Boston, U.S.A.

Code

File: SBDN_1.0.tar

Description: Implements the sparse Bayesian dynamic network model of Bartlett, et al. (2018), as an R function which calls C++ code.

Usage:

1. Download SBDN_1.0.tar
2. In the terminal, or at the command line or command prompt, navigate to the directory where you downloaded it
3. Type: R CMD INSTALL SBDN_1.0.tar
4. Start R, and load the package by typing require(SBDN)

• The model is implemented as the function dynNetModelSamp
• Syntax and explanation are given in the help file. To view this, type: ?dynNetModelSamp

File: cor2adj.R

Description: Takes as input a correlation matrix estimated from n samples, and outputs an adjacency matrix inferred according to the methodology presented in Bartlett et al. (2016).

Usage: adjMat <- cor2adj(corMat,n)

File: DCSBMcluster.R

Usage: clusters <- DCSBMcluster(adjMat,k,rowNorm,gamma,nrep)

Description: Divides the nodes of a network (as represented by the input adjacency matrix adjMat) into k clusters, by fitting the degree-corrected stochastic blockmodel by spectral clustering, with regularisation and modularity maximisation. Model fitting can sometimes be improved with row-normalisation of the spectral decomposition by setting rowNorm=1. Model fitting can sometimes also sometimes be improved by first clustering only the most important nodes, by setting gamma=1. The k-means step in the spectral clustering is run nrep times, retaining the clustering with the greatest modularity.

File: DCSBMcoCluster.R

Usage: clusters <- DCSBMcoCluster(adjMat,kR,kC,rowNorm,gamma,nrep), where clusters is a list object, containing vectors clustersR and clustersC for the row and column clusters, respectively.

Description: Divides the nodes of a bipartite network (as represented by the asymmetric input adjacency matrix adjMat) into kR row clusters and kC column clusters, by fitting the degree-corrected stochastic co-blockmodel (Bartlett et al., in preparation) by spectral clustering, with regularisation and modularity maximisation. Model fitting can sometimes be improved with row-normalisation of the spectral decomposition by setting rowNorm=1. Model fitting can sometimes also sometimes be improved by first clustering only the most important nodes, by setting gamma=1. The k-means step in the spectral clustering is run nrep times, retaining the clustering with the greatest co-modularity.