MegaProbe Lab Introduction
Samples of DNA/RNA known as sequences can be used to understand the information of nucleotides in biological structures. Small fragments known as reads are produced from DNA by a DNA Sequencer. In the process of sequence assembly these fragments can be joined in order to reconstruct a complete sequence of the organism’s DNA. De-novo, meaning ”from the beginning”, refers to sequence assembly done without a reference genome, and it is used when trying to discover/reconstruct new genome sequences.
A common problem in sequence assembly can occur from errors in the sequencing data used. Reads can contain one or more mismatches from the original genome and could lead to inaccurate sequence assemblies. In de-novo sequence assembly it becomes particularly challenging, due to not having any reference to compare the reads with and verify their integrity.
This project is a collaboration between Jose Agosto Rivera's Lab, Pittsburgh Supercomputing Center, RISE, and MegaProbe.
Red Brick Wall
As sequencing continues to produce more and more data for lower costs, the analysis is falling behind the production of data, and new techniques have to be devised. The following graph illustrates how sequencing has far outstripped computational capacity.
Wetterstrand KA. DNA Sequencing Costs: Data from the NHGRI Genome Sequencing Program (GSP) Available at: http://www.genome.gov/sequencingcostsdata. Accessed 2017-01-17.
De Bruijn Graphs
To represent and assemble next-generation sequencing data, most programs construct k-mer or De Bruijn graphs. Here's an example graph for a simulated set of reads from a 1000 base pair sequence with some sequencing errors (red dots).
Probabilistic structures
As the sequencing data continues to grow, it becomes infeasible to completely store and process the full data set. We are looking at probabilistic data structures to approximate the graphs, and allow some biological questions to be answered.
The figure below represents a portion of a De Bruijn graph for a real data set from Nematostella Embryonic Transcriptome (Starlet sea anemone). https://darchive.mblwhoilibrary.org/handle/1912/5613
Some tools for hashing and probabilistic counting of k-mers have been implemented in the khmer-tools suite.
Crusoe MR, Alameldin HF, Awad S et al. The khmer software package: enabling efficient nucleotide sequence analysis [version 1; referees: 2 approved, 1 approved with reservations]. F1000Research 2015, 4:900 (doi: 10.12688/f1000research.6924.1)
Error-correction
K-mer hashing and counting can be used to efficently and effectively remove sequencing errors from datasets. For example, all errors in the first graph can be removed by selecting all k-mers that appear at least twice in the data.
Shared transcript prediction
The mutual software described in Fu S, Tarone AM, Sze SH. (2015) Heuristic pairwise alignment of de Bruijn graphs to facilitate simultaneous transcript discovery in related organisms from RNA-Seq data. BMC Genomics 16:S5 predicts shared transcripts by searching for sequences derived from two De Bruijn graphs that resemble each other.
In our lab we are working to directly determine shared structure from two graphs, to speed up and improve the results of mutual.
Differential expression
Current techniques for assesment of differential expression assemble transcriptomes from RNAseq data, then count the abundance of each transcript in a set of samples, and infer probabilities of differential expression.
Can we instead infer differential expression directly from k-mer counts?
Downstream analysis
Regulatory network engineering
I have done some work on reverse-engineering gene regualtory networks from microarray data. Ideally, given enough RNAseq data over a time series, we should be able to infer regulatory relationships from the collection of k-mer counts.