GeneQC Statistical Model General Idea Reads can be mapped to - - PowerPoint PPT Presentation

GeneQC Statistical Model

General Idea

• Reads can be mapped to

multiple gene loci

• Leads to varying degrees
• f mapping uncertainty
• Potentially causes issues

with inferences based on read counts

• Differentially expressed

genes

• Co-expression patterns
• Various network analyses

Options

• Exclude ambiguous reads
• Multiple assignment
• Random assignment
• Probabilistic assignment
• Only considering local

information

Co-expressed Genes

• Co-expressed genes

provided additional level

• f information
• Global data for more solid

statistical evaluation

Goal

• Create statistically sound model for assignment of ambiguous

reads

• Use co-expression of genes
• Develop method that produces p-value or probability score for

each ambiguous read assignment

• Provide a p-value signifying the confidence of each gene’s read

count

Previous Publications

• Faulkner, G.J., et al., A rescue strategy for multimapping short

sequence tags refines surveys of transcriptional activity by CAGE. Genomics, 2008. 91(3): p. 281-288.

• Hashimoto, T

., et al., Probabilistic resolution of multi-mapping reads in massively parallel sequencing data using MuMRescueLite. Bioinformatics, 2009. 25(19): p. 2613-2614.

• Wang, J., Huda, A., Lunyak, V. V., & Jordan, I. K., A Gibbs

sampling strategy applied to the mapping of ambiguous short- sequence tags. Bioinformatics, 2010. 26(20): p.2501-2508

Overall Direction

• Assign all unambiguous reads
• Use co-expression information of unambiguous reads to make first

probabilistic assignment of ambiguous reads

• Based on assignments, recalculate probabilities for ambiguous

reads

• Continue iterative procedure until no/minimal changes occur

Additional parameters

• Similarity between a given read and each potential gene locus
• Differences generally very minute
• Co-expression rate between genes and co-expressed genes

Concerns & Limitations

• Requires accurate co-expression information
• Limited sample size of co-expression information could skew

probability distribution

• Potentially highly computationally intensive
• Local optimization may occur
• Does not currently consider dependence of read assignment

Our Future Plans

• Collect test data to verify increased performance using statistical

model

• Run model with various validated probability assumptions
• Normal, Poisson, etc.
• Develop R package with statistical model implementation