Machine learning for regulatory analysis and transcription factor target prediction in yeast

Landscape of transcriptional deregulation in lung cancer

BACKGROUND

Lung cancer is a very heterogeneous disease that can be pathologically classified into different subtypes including small-cell lung carcinoma (SCLC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC) and large-cell carcinoma (LCC). Although much progress has been made towards the oncogenic mechanism of each subtype, transcriptional circuits mediating the upstream signaling pathways and downstream functional consequences remain to be systematically studied.

RESULTS

Here we trained a one-class support vector machine (OC-SVM) model to establish a general transcription factor (TF) regulatory network containing 325 TFs and 18724 target genes. We then applied this network to lung cancer subtypes and identified those deregulated TFs and downstream targets. We found that the TP63/SOX2/DMRT3 module was specific to LUSC, corresponding to squamous epithelial differentiation and/or survival. Moreover, the LEF1/MSC module was specifically activated in LUAD and likely to confer epithelial-to-mesenchymal transition, known important for cancer malignant progression and metastasis. The proneural factor, ASCL1, was specifically up-regulated in SCLC which is known to have a neuroendocrine phenotype. Also, ID2 was differentially regulated between SCLC and LUSC, with its up-regulation in SCLC linking to energy supply for fast mitosis and its down-regulation in LUSC linking to the attenuation of immune response. We further described the landscape of TF regulation among the three major subtypes of lung cancer, highlighting their functional commonalities and specificities.

CONCLUSIONS

Our approach uncovered the landscape of transcriptional deregulation in lung cancer, and provided a useful resource of TF regulatory network for future studies.

Discriminating between HuR and TTP binding sites using the k-spectrum kernel method

Background

The RNA binding proteins (RBPs) human antigen R (HuR) and Tristetraprolin (TTP) are known to exhibit competitive binding but have opposing effects on the bound messenger RNA (mRNA). How cells discriminate between the two proteins is an interesting problem. Machine learning approaches, such as support vector machines (SVMs), may be useful in the identification of discriminative features. However, this method has yet to be applied to studies of RNA binding protein motifs.

Results

Applying the k-spectrum kernel to a support vector machine (SVM), we first verified the published binding sites of both HuR and TTP. Additional feature engineering highlighted the U-rich binding preference of HuR and AU-rich binding preference for TTP. Domain adaptation along with multi-task learning was used to predict the common binding sites.

Conclusion

The distinction between HuR and TTP binding appears to be subtle content features. HuR prefers strongly U-rich sequences whereas TTP prefers AU-rich as with increasing A content, the sequences are more likely to be bound only by TTP. Our model is consistent with competitive binding of the two proteins, particularly at intermediate AU-balanced sequences. This suggests that fine changes in the A/U balance within a untranslated region (UTR) can alter the binding and subsequent stability of the message. Both feature engineering and domain adaptation emphasized the extent to which these proteins recognize similar general sequence features. This work suggests that the k-spectrum kernel method could be useful when studying RNA binding proteins and domain adaptation techniques such as feature augmentation could be employed particularly when examining RBPs with similar binding preferences.

Differences in local genomic context of bound and unbound motifs

Understanding gene regulation is a major objective in molecular biology research. Frequently, transcription is driven by transcription factors (TFs) that bind to specific DNA sequences. These motifs are usually short and degenerate, rendering the likelihood of multiple copies occurring throughout the genome due to random chance as high. Despite this, TFs only bind to a small subset of sites, thus prompting our investigation into the differences between motifs that are bound by TFs and those that remain unbound. Here we constructed vectors representing various chromatin- and sequence-based features for a published set of bound and unbound motifs representing nine TFs in the budding yeast Saccharomyces cerevisiae. Using a machine learning approach, we identified a set of features that can be used to discriminate between bound and unbound motifs. We also discovered that some TFs bind most or all of their strong motifs in intergenic regions. Our data demonstrate that local sequence context can be strikingly different around motifs that are bound compared to motifs that are unbound. We concluded that there are multiple combinations of genomic features that characterize bound or unbound motifs.

DNA structural properties in the classification of genomic transcription regulation elements

It has been long known that DNA molecules encode information at various levels. The most basic level comprises the base sequence itself and is primarily important for the encoding of proteins and direct base recognition by DNA-binding proteins. A more elusive level consists of the local structural properties of the DNA molecule wherein the DNA sequence only plays an indirect supportive role. These properties are nevertheless an important factor in a large number of biomolecular processes and can be considered as informative signals for the presence of a variety of genomic features. Several recent studies have unequivocally shown the benefit of relying on such DNA properties for modeling and predicting genomic features as diverse as transcription start sites, transcription factor binding sites, or nucleosome occupancy. This review is meant to provide an overview of the key aspects of these DNA conformational and physicochemical properties. To illustrate their potential added value compared to relying solely on the nucleotide sequence in genomics studies, we discuss their application in research on transcription regulation mechanisms as representative cases.

High-quality annotation of promoter regions for 913 bacterial genomes

MOTIVATION

The number of bacterial genomes being sequenced is increasing very rapidly and hence, it is crucial to have procedures for rapid and reliable annotation of their functional elements such as promoter regions, which control the expression of each gene or each transcription unit of the genome. The present work addresses this requirement and presents a generic method applicable across organisms.

RESULTS

Relative stability of the DNA double helical sequences has been used to discriminate promoter regions from non-promoter regions. Based on the difference in stability between neighboring regions, an algorithm has been implemented to predict promoter regions on a large scale over 913 microbial genome sequences. The average free energy values for the promoter regions as well as their downstream regions are found to differ, depending on their GC content. Threshold values to identify promoter regions have been derived using sequences flanking a subset of translation start sites from all microbial genomes and then used to predict promoters over the complete genome sequences. An average recall value of 72% (which indicates the percentage of protein and RNA coding genes with predicted promoter regions assigned to them) and precision of 56% is achieved over the 913 microbial genome dataset.

AVAILABILITY

The binary executable for 'PromPredict' algorithm (implemented in PERL and supported on Linux and MS Windows) and the predicted promoter data for all 913 microbial genomes are available at http://nucleix.mbu.iisc.ernet.in/prombase/.

A tutorial on support vector machine-based methods for classification problems in chemometrics

This tutorial provides a concise overview of support vector machines and different closely related techniques for pattern classification. The tutorial starts with the formulation of support vector machines for classification. The method of least squares support vector machines is explained. Approaches to retrieve a probabilistic interpretation are covered and it is explained how the binary classification techniques can be extended to multi-class methods. Kernel logistic regression, which is closely related to iteratively weighted least squares support vector machines, is discussed. Different practical aspects of these methods are addressed: the issue of feature selection, parameter tuning, unbalanced data sets, model evaluation and statistical comparison. The different concepts are illustrated on three real-life applications in the field of metabolomics, genetics and proteomics.

Robust gene network analysis reveals alteration of the STAT5a network as a hallmark of prostate cancer

We develop a general method to identify gene networks from pair-wise correlations between genes in a microarray data set and apply it to a public prostate cancer gene expression data from 69 primary prostate tumors. We define the degree of a node as the number of genes significantly associated with the node and identify hub genes as those with the highest degree. The correlation network was pruned using transcription factor binding information in VisANT (http://visant.bu.edu/) as a biological filter. The reliability of hub genes was determined using a strict permutation test. Separate networks for normal prostate samples, and prostate cancer samples from African Americans (AA) and European Americans (EA) were generated and compared. We found that the same hubs control disease progression in AA and EA networks. Combining AA and EA samples, we generated networks for low low (<7) and high (≥7) Gleason grade tumors. A comparison of their major hubs with those of the network for normal samples identified two types of changes associated with disease: (i) Some hub genes increased their degree in the tumor network compared to their degree in the normal network, suggesting that these genes are associated with gain of regulatory control in cancer (e.g. possible turning on of oncogenes). (ii) Some hubs reduced their degree in the tumor network compared to their degree in the normal network, suggesting that these genes are associated with loss of regulatory control in cancer (e.g. possible loss of tumor suppressor genes). A striking result was that for both AA and EA tumor samples, STAT5a, CEBPB and EGR1 are major hubs that gain neighbors compared to the normal prostate network. Conversely, HIF-lα is a major hub that loses connections in the prostate cancer network compared to the normal prostate network. We also find that the degree of these hubs changes progressively from normal to low grade to high grade disease, suggesting that these hubs are master regulators of prostate cancer and marks disease progression. STAT5a was identified as a central hub, with ~120 neighbors in the prostate cancer network and only 81 neighbors in the normal prostate network. Of the 120 neighbors of STAT5a, 57 are known cancer related genes, known to be involved in functional pathways associated with tumorigenesis. Our method is general and can easily be extended to identify and study networks associated with any two phenotypes.

LipocalinPred: a SVM-based method for prediction of lipocalins

BACKGROUND

Functional annotation of rapidly amassing nucleotide and protein sequences presents a challenging task for modern bioinformatics. This is particularly true for protein families sharing extremely low sequence identity, as for lipocalins, a family of proteins with varied functions and great diversity at the sequence level, yet conserved structures.

RESULTS

In the present study we propose a SVM based method for identification of lipocalin protein sequences. The SVM models were trained with the input features generated using amino acid, dipeptide and secondary structure compositions as well as PSSM profiles. The model derived using both PSSM and secondary structure emerged as the best model in the study. Apart from achieving a high prediction accuracy (>90% in leave-one-out), lipocalinpred correctly differentiates closely related fatty acid-binding proteins and triabins as non-lipocalins.

CONCLUSION

The method offers a promising approach as a lipocalin prediction tool, complementing PROSITE, Pfam and homology modelling methods.

In silico regulatory analysis for exploring human disease progression

Background

An important goal in bioinformatics is to unravel the network of transcription factors (TFs) and their targets. This is important in the human genome, where many TFs are involved in disease progression. Here, classification methods are applied to identify new targets for 152 transcriptional regulators using publicly-available targets as training examples. Three types of sequence information are used: composition, conservation, and overrepresentation.

Results

Starting with 8817 TF-target interactions we predict an additional 9333 targets for 152 TFs. Randomized classifiers make few predictions (~2/18660) indicating that our predictions for many TFs are significantly enriched for true targets. An enrichment score is calculated and used to filter new predictions.

Two case-studies for the TFs OCT4 and WT1 illustrate the usefulness of our predictions:

• Many predicted OCT4 targets fall into the Wnt-pathway. This is consistent with known biology as OCT4 is developmentally related and Wnt pathway plays a role in early development.

• Beginning with 15 known targets, 354 predictions are made for WT1. WT1 has a role in formation of Wilms' tumor. Chromosomal regions previously implicated in Wilms' tumor by cytological evidence are statistically enriched in predicted WT1 targets. These findings may shed light on Wilms' tumor progression, suggesting that the tumor progresses either by loss of WT1 or by loss of regions harbouring its targets.

• Targets of WT1 are statistically enriched for cancer related functions including metastasis and apoptosis. Among new targets are BAX and PDE4B, which may help mediate the established anti-apoptotic effects of WT1.

• Of the thirteen TFs found which co-regulate genes with WT1 (p ≤ 0.02), 8 have been previously implicated in cancer. The regulatory-network for WT1 targets in genomic regions relevant to Wilms' tumor is provided.

Conclusion

We have assembled a set of features for the targets of human TFs and used them to develop classifiers for the determination of new regulatory targets. Many predicted targets are consistent with the known biology of their regulators, and new targets for the Wilms' tumor regulator, WT1, are proposed. We speculate that Wilms' tumor development is mediated by chromosomal rearrangements in the location of WT1 targets.

Reviewers

This article was reviewed by Trey Ideker, Vladimir A. Kuznetsov(nominated by Frank Eisenhaber), and Tzachi Pilpel.

Classifying transcription factor targets and discovering relevant biological features

Background

An important goal in post-genomic research is discovering the network of interactions between transcription factors (TFs) and the genes they regulate. We have previously reported the development of a supervised-learning approach to TF target identification, and used it to predict targets of 104 transcription factors in yeast. We now include a new sequence conservation measure, expand our predictions to include 59 new TFs, introduce a web-server, and implement an improved ranking method to reveal the biological features contributing to regulation. The classifiers combine 8 genomic datasets covering a broad range of measurements including sequence conservation, sequence overrepresentation, gene expression, and DNA structural properties.

Principal Findings

(1) Application of the method yields an amplification of information about yeast regulators. The ratio of total targets to previously known targets is greater than 2 for 11 TFs, with several having larger gains: Ash1(4), Ino2(2.6), Yaf1(2.4), and Yap6(2.4).

(2) Many predicted targets for TFs match well with the known biology of their regulators. As a case study we discuss the regulator Swi6, presenting evidence that it may be important in the DNA damage response, and that the previously uncharacterized gene YMR279C plays a role in DNA damage response and perhaps in cell-cycle progression.

(3) A procedure based on recursive-feature-elimination is able to uncover from the large initial data sets those features that best distinguish targets for any TF, providing clues relevant to its biology. An analysis of Swi6 suggests a possible role in lipid metabolism, and more specifically in metabolism of ceramide, a bioactive lipid currently being investigated for anti-cancer properties.

(4) An analysis of global network properties highlights the transcriptional network hubs; the factors which control the most genes and the genes which are bound by the largest set of regulators. Cell-cycle and growth related regulators dominate the former; genes involved in carbon metabolism and energy generation dominate the latter.

Conclusion

Postprocessing of regulatory-classifier results can provide high quality predictions, and feature ranking strategies can deliver insight into the regulatory functions of TFs. Predictions are available at an online web-server, including the full transcriptional network, which can be analyzed using VisAnt network analysis suite.

Reviewers

This article was reviewed by Igor Jouline, Todd Mockler(nominated by Valerian Dolja), and Sandor Pongor.

A semi-supervised method for predicting transcription factor-gene interactions in Escherichia coli

While Escherichia coli has one of the most comprehensive datasets of experimentally verified transcriptional regulatory interactions of any organism, it is still far from complete. This presents a problem when trying to combine gene expression and regulatory interactions to model transcriptional regulatory networks. Using the available regulatory interactions to predict new interactions may lead to better coverage and more accurate models. Here, we develop SEREND (SEmi-supervised REgulatory Network Discoverer), a semi-supervised learning method that uses a curated database of verified transcriptional factor-gene interactions, DNA sequence binding motifs, and a compendium of gene expression data in order to make thousands of new predictions about transcription factor-gene interactions, including whether the transcription factor activates or represses the gene. Using genome-wide binding datasets for several transcription factors, we demonstrate that our semi-supervised classification strategy improves the prediction of targets for a given transcription factor. To further demonstrate the utility of our inferred interactions, we generated a new microarray gene expression dataset for the aerobic to anaerobic shift response in E. coli. We used our inferred interactions with the verified interactions to reconstruct a dynamic regulatory network for this response. The network reconstructed when using our inferred interactions was better able to correctly identify known regulators and suggested additional activators and repressors as having important roles during the aerobic-anaerobic shift interface.