RNA amplification is necessary for profiling gene expression from small tissue samples. Previous studies have shown that the T7 based amplification techniques are reproducible but may distort the ...true abundance of targets. However, the consequences of such distortions on the ability to detect biological variation in expression have not been explored sufficiently to define the true extent of usability and limitations of such amplification techniques.
We show that expression ratios are occasionally distorted by amplification using the Affymetrix small sample protocol version 2 due to a disproportional shift in intensity across biological samples. This occurs when a shift in one sample cannot be reflected in the other sample because the intensity would lie outside the dynamic range of the scanner. Interestingly, such distortions most commonly result in smaller ratios with the consequence of reducing the statistical significance of the ratios. This becomes more critical for less pronounced ratios where the evidence for differential expression is not strong. Indeed, statistical analysis by limma suggests that up to 87% of the genes with the largest and therefore most significant ratios (p < 10e(-20)) in the unamplified group have a p-value below 10e(-20) in the amplified group. On the other hand, only 69% of the more moderate ratios (10e(-20) < p < 10e(-10)) in the unamplified group have a p-value below 10e(-10) in the amplified group. Our analysis also suggests that, overall, limma shows better overlap of genes found to be significant in the amplified and unamplified groups than the Z-scores statistics.
We conclude that microarray analysis of amplified samples performs best at detecting differences in gene expression, when these are large and when limma statistics are used.
Integrative clustering is used to identify groups of samples by jointly analysing multiple datasets describing the same set of biological samples, such as gene expression, copy number, methylation ...etc. Most existing algorithms for integrative clustering assume that there is a shared consistent set of clusters across all datasets, and most of the data samples follow this structure. However in practice, the structure across heterogeneous datasets can be more varied, with clusters being joined in some datasets and separated in others. In this paper, we present a probabilistic clustering method to identify groups across datasets that do not share the same cluster structure. The proposed algorithm, Clusternomics, identifies groups of samples that share their global behaviour across heterogeneous datasets. The algorithm models clusters on the level of individual datasets, while also extracting global structure that arises from the local cluster assignments. Clusters on both the local and the global level are modelled using a hierarchical Dirichlet mixture model to identify structure on both levels. We evaluated the model both on simulated and on real-world datasets. The simulated data exemplifies datasets with varying degrees of common structure. In such a setting Clusternomics outperforms existing algorithms for integrative and consensus clustering. In a real-world application, we used the algorithm for cancer subtyping, identifying subtypes of cancer from heterogeneous datasets. We applied the algorithm to TCGA breast cancer dataset, integrating gene expression, miRNA expression, DNA methylation and proteomics. The algorithm extracted clinically meaningful clusters with significantly different survival probabilities. We also evaluated the algorithm on lung and kidney cancer TCGA datasets with high dimensionality, again showing clinically significant results and scalability of the algorithm.
The copy numbers of genes in cancer samples are often highly disrupted and form a natural amplification/deletion experiment encompassing multiple genes. Matched array comparative genomics and ...transcriptomics datasets from such samples can be used to predict inter-chromosomal gene regulatory relationships. Previously we published the database METAMATCHED, comprising the results from such an analysis of a large number of publically available cancer datasets. Here we investigate genes in the database which are unusual in that their copy number exhibits consistent heterogeneous disruption in a high proportion of the cancer datasets. We assess the potential relevance of these genes to the pathology of the cancer samples, in light of their predicted regulatory relationships and enriched biological pathways. A network-based method was used to identify enriched pathways from the genes' inferred targets. The analysis predicts both known and new regulator-target interactions and pathway memberships. We examine examples in detail, in particular the gene POGZ, which is disrupted in many of the cancer datasets and has an unusually large number of predicted targets, from which the network analysis predicts membership of cancer related pathways. The results suggest close involvement in known cancer pathways of genes exhibiting consistent heterogeneous copy number disruption. Further experimental work would clarify their relevance to tumor biology. The results of the analysis presented in the database METAMATCHED, and included here as an R archive file, constitute a large number of predicted regulatory relationships and pathway memberships which we anticipate will be useful in informing such experiments.
Streptococcus pneumoniae is a human pathogen that is a major cause of infant mortality. Identifying the pneumococcal serotype is an important step in monitoring the impact of vaccines used to protect ...against disease. Genomic microarrays provide an effective method for molecular serotyping. Previously we developed an empirical Bayesian model for the classification of serotypes from a molecular serotyping array. With only few samples available, a model driven approach was the only option. In the meanwhile, several thousand samples have been made available to us, providing an opportunity to investigate serotype classification by machine learning methods, which could complement the Bayesian model.
We compare the performance of the original Bayesian model with two machine learning algorithms: Gradient Boosting Machines and Random Forests. We present our results as an example of a generic strategy whereby a preliminary probabilistic model is complemented or replaced by a machine learning classifier once enough data are available. Despite the availability of thousands of serotyping arrays, a problem encountered when applying machine learning methods is the lack of training data containing mixtures of serotypes; due to the large number of possible combinations. Most of the available training data comprises samples with only a single serotype. To overcome the lack of training data we implemented an iterative analysis, creating artificial training data of serotype mixtures by combining raw data from single serotype arrays.
With the enhanced training set the machine learning algorithms out perform the original Bayesian model. However, for serotypes currently lacking sufficient training data the best performing implementation was a combination of the results of the Bayesian Model and the Gradient Boosting Machine. As well as being an effective method for classifying biological data, machine learning can also be used as an efficient method for revealing subtle biological insights, which we illustrate with an example.
Rapid technological innovation for the generation of single-cell genomics data presents new challenges and opportunities for bioinformatics analysis. One such area lies in the development of new ways ...to train gene regulatory networks. The use of single-cell expression profiling technique allows the profiling of the expression states of hundreds of cells, but these expression states are typically noisier due to the presence of technical artefacts such as drop-outs. While many algorithms exist to infer a gene regulatory network, very few of them are able to harness the extra expression states present in single-cell expression data without getting adversely affected by the substantial technical noise present.
Here we introduce BTR, an algorithm for training asynchronous Boolean models with single-cell expression data using a novel Boolean state space scoring function. BTR is capable of refining existing Boolean models and reconstructing new Boolean models by improving the match between model prediction and expression data. We demonstrate that the Boolean scoring function performed favourably against the BIC scoring function for Bayesian networks. In addition, we show that BTR outperforms many other network inference algorithms in both bulk and single-cell synthetic expression data. Lastly, we introduce two case studies, in which we use BTR to improve published Boolean models in order to generate potentially new biological insights.
BTR provides a novel way to refine or reconstruct Boolean models using single-cell expression data. Boolean model is particularly useful for network reconstruction using single-cell data because it is more robust to the effect of drop-outs. In addition, BTR does not assume any relationship in the expression states among cells, it is useful for reconstructing a gene regulatory network with as few assumptions as possible. Given the simplicity of Boolean models and the rapid adoption of single-cell genomics by biologists, BTR has the potential to make an impact across many fields of biomedical research.
Most existing algorithms for the inference of the structure of gene regulatory networks from gene expression data assume that the activity levels of transcription factors (TFs) are proportional to ...their mRNA levels. This assumption is invalid for most biological systems. However, one might be able to reconstruct unobserved activity profiles of TFs from the expression profiles of target genes. A simple model is a two-layer network with unobserved TF variables in the first layer and observed gene expression variables in the second layer. TFs are connected to regulated genes by weighted edges. The weights, known as factor loadings, indicate the strength and direction of regulation. Of particular interest are methods that produce sparse networks, networks with few edges, since it is known that most genes are regulated by only a small number of TFs, and most TFs regulate only a small number of genes.
In this paper, we explore the performance of five factor analysis algorithms, Bayesian as well as classical, on problems with biological context using both simulated and real data. Factor analysis (FA) models are used in order to describe a larger number of observed variables by a smaller number of unobserved variables, the factors, whereby all correlation between observed variables is explained by common factors. Bayesian FA methods allow one to infer sparse networks by enforcing sparsity through priors. In contrast, in the classical FA, matrix rotation methods are used to enforce sparsity and thus to increase the interpretability of the inferred factor loadings matrix. However, we also show that Bayesian FA models that do not impose sparsity through the priors can still be used for the reconstruction of a gene regulatory network if applied in conjunction with matrix rotation methods. Finally, we show the added advantage of merging the information derived from all algorithms in order to obtain a combined result.
Most of the algorithms tested are successful in reconstructing the connectivity structure as well as the TF profiles. Moreover, we demonstrate that if the underlying network is sparse it is still possible to reconstruct hidden activity profiles of TFs to some degree without prior connectivity information.
Inferring gene regulatory relationships from observational data is challenging. Manipulation and intervention is often required to unravel causal relationships unambiguously. However, gene copy ...number changes, as they frequently occur in cancer cells, might be considered natural manipulation experiments on gene expression. An increasing number of data sets on matched array comparative genomic hybridisation and transcriptomics experiments from a variety of cancer pathologies are becoming publicly available. Here we explore the potential of a meta-analysis of thirty such data sets. The aim of our analysis was to assess the potential of in silico inference of trans-acting gene regulatory relationships from this type of data. We found sufficient correlation signal in the data to infer gene regulatory relationships, with interesting similarities between data sets. A number of genes had highly correlated copy number and expression changes in many of the data sets and we present predicted potential trans-acted regulatory relationships for each of these genes. The study also investigates to what extent heterogeneity between cell types and between pathologies determines the number of statistically significant predictions available from a meta-analysis of experiments.
MEME and many other popular motif finders use the expectation-maximization (EM) algorithm to optimize their parameters. Unfortunately, the running time of EM is linear in the length of the input ...sequences. This can prohibit its application to data sets of the size commonly generated by high-throughput biological techniques. A suffix tree is a data structure that can efficiently index a set of sequences. We describe an algorithm, Suffix Tree EM for Motif Elicitation (STEME), that approximates EM using suffix trees. To the best of our knowledge, this is the first application of suffix trees to EM. We provide an analysis of the expected running time of the algorithm and demonstrate that STEME runs an order of magnitude more quickly than the implementation of EM used by MEME. We give theoretical bounds for the quality of the approximation and show that, in practice, the approximation has a negligible effect on the outcome. We provide an open source implementation of the algorithm that we hope will be used to speed up existing and future motif search algorithms.
Perinatal depression involves interplay between individual chronic and acute disease burdens, biological and psychosocial environmental and behavioural factors. Here we explored the predictive ...potential of specific psycho-socio-demographic characteristics for antenatal and postpartum depression symptoms and contribution to severity scores on the Edinburgh Postnatal Depression Scale (EPDS) screening tool. We determined depression risk trajectories in 480 women that prospectively completed the EPDS during pregnancy (TP1) and postpartum (TP2). Multinomial logistic and penalised linear regression investigated covariates associated with increased antenatal and postpartum EPDS scores contributing to the average or the difference of paired scores across time points. History of anxiety was identified as the strongest contribution to antenatal EPDS scores followed by the social status, whereas a history of depression, postpartum depression (PPD) and family history of PPD exhibited the strongest association with postpartum EPDS. These covariates were the strongest differentiating factors that increased the spread between antenatal and postpartum EPDS scores. Available covariates appeared better suited to predict EPDS scores antenatally than postpartum. As women move from the antenatal to the postpartum period, socio-demographic and lifestyle risk factors appear to play a smaller role in risk, and a personal and family history of depression and PPD become increasingly important.
Motif finding is a difficult problem that has been studied for over 20 years. Some older popular motif finders are not suitable for analysis of the large data sets generated by next-generation ...sequencing. We recently published an efficient approximation (STEME) to the EM algorithm that is at the core of many motif finders such as MEME. This approximation allows the EM algorithm to be applied to large data sets. In this work we describe several efficient extensions to STEME that are based on the MEME algorithm. Together with the original STEME EM approximation, these extensions make STEME a fully-fledged motif finder with similar properties to MEME. We discuss the difficulty of objectively comparing motif finders. We show that STEME performs comparably to existing prominent discriminative motif finders, DREME and Trawler, on 13 sets of transcription factor binding data in mouse ES cells. We demonstrate the ability of STEME to find long degenerate motifs which these discriminative motif finders do not find. As part of our method, we extend an earlier method due to Nagarajan et al. for the efficient calculation of motif E-values. STEME's source code is available under an open source license and STEME is available via a web interface.