Development and benchmarking of improved computational methods for transcript-level expression analysis using RNA-seq data

After sequencing of the human genome was completed, Scientists were surprised to discover that there are far fewer protein-coding genes than was previously predicted. One reason that an organism as complex as human can be built from a relatively small number of genes is that each gene encodes more than one protein. An intermediate molecule, messenger RNA (mRNA), carries the information from the genome in the cell nucleus to ribosomes which create proteins. These mRNA molecules are also known as transcripts and their full complement is termed the transcriptome. Before they mature these transcripts are edited to form the template for different proteins. This editing process is called splicing and different transcripts that result are called splice variants or isoforms. An additional complexity in the transcriptome is due to the fact that each gene has multiple copies (for example 2 in human, 6 in wheat) and these different copies, called alleles, can be expressed differently under different conditions or in different tissues. The transcriptome is a collection of transcripts which includes all the allele-specific gene isoforms that are expressed in the cell along with other non-coding RNA molecules.

Splicing and allele usage are fundamental ways that the function of genes can be modulated in a tissue-specific manner. Therefore developing technologies to accurately measure transcript expression is a necessary step towards understanding and modelling cells and tissues. A recently developed experimental technology called RNA-seq gives unprecedented access to data about the transcriptome. Computational methods are required to interpret these data which are in the form of a list containing millions of short RNA sequence fragments. These fragments are difficult to interpret because, for example, the same fragment could have come from a large number of different gene isoforms. The question is, which one? Computational methods can be used to answer this question and infer the concentration of different gene isoforms in the sample given these data. In this project we will develop a new computational method, implemented in publically available free software, which uses advanced statistical procedures to solve this problem. An important distinguishing feature of the method is the ability to associate inferred concentrations with a degree of uncertainty which captures technical and biological sources of error as well as the inherent difficulty of the problem due to the difficulty of assigning fragments to gene isoforms. We will create benchmark data that allows us to assess the performance or our method and other available published methods, allowing researchers and end-users of different methods to understand their properties. Finally, we will adapt an existing computer program, puma, to work with the processed RNA-seq data in order to identify genes which change between conditions, which have similar expression patterns or which contribute most to the variance in the data.

RNA-seq technology enables the discovery and quantification of multiple transcripts for each gene, including different gene isoforms and different allelic forms. We propose the development of a Bayesian inference approach for inferring the concentration of different transcripts present in a sample by using a probabilistic model of mapped reads. By using a Bayesian inference approach we will capture the level of inherent uncertainty in our estimates of transcript expression levels due to mapping ambiguity, technical noise, read depth limitations and biological noise. We will include the possibility of discovering unannotated isoforms. The use of a read-level probabilistic model will allow us to incorporate information about read density biases and read mapping quality scores. We will apply the model to quantify allele-specific isoform expression which is particularly challenging in complex genomes such as the hexaploid wheat that can express genes from a set of three diploid genomes. We will develop a transcript-level benchmark dataset for method evaluation in which different gene isoforms are spiked in at known concentrations against a natural background. R-code implementing our methods for transcript-level inference and benchmarking will be disseminated through the Bioconductor project. We will extend the existing puma Bioconductor package for noise propagation in microarray analysis so that the methods there can be applied to transcript-level expression data with an associated multivariate uncertainty distribution.

Communication and Engagement: We will publish papers in open access peer-reviewed journals so that the academic community are made aware of developments. Software will be implemented as open source Bioconductor packages. A public benchmark will lead to better practice by allowing a publically available comparison of competing methods on a level playing field. We have close links to TGAC and the other MRC hubs and we will ensure that all of these groups are made aware of the tools developed and their application. The CGR, as a NERC and MRC hub, also works with a large bioinformatics community and will train new users in working with this software.

Collaboration and Co-production: The investigators are also engaged in many other BBSRC projects which can adopt the methodology developed here to add value to those projects. These projects also provide excellent application data for this proposal. Many of these projects involve short read sequencing of economically important species and comparative analysis to model species and we will identify other projects where the software will be deployed and ensure that their feedback is reflected in the development of the software.

Exploitation and Application: As this tool will be deployed primarily for academic research we do not intend to protect its application. It will be made freely available to the user community through a suitable open source license.

Capacity and Involvement: We are involved in supervising BBSRC and MRC funded Ph.D. students who will benefit from this research as they will be directly using the software developed and we regularly employed sixth form students to undertake research activities in the lab. Both SITRAN and the CGR undertake a wide range of outreach activities to industry, the academic community and the general public and actively engage with the media at local, national and international level

Impact Activity Deliverables and Milestones: Computational Biology developments will be presented at international conferences. Four key papers and associated software will be published along with a benchmarking website.

Resource for activity: We request £1,200 per paper for open access journal charges and £5,700 for presentation at five leading conferences over three years.