| tags |
|---|
ggg, ggg2023, ggg201b |
Working individually or in groups, please reconstruct the original text for the handouts I’m giving you. You may use the computer as you wish, but I suggest you avoid any attempts at programming...
Specific questions:
-
what techniques or tactics are you using?
-
suppose you had to do this with lots more data; how would you do it?
-
suppose you were given an unlimited supply of undergraduates; what would you tell them to do and how would you split the work up?
what advantages do you have in this exercise over DNA?
what are the pluses and minuses of the different datatypes?
Does the order or organization of reads matter?
if you get to use a reference, how do you know it’s the right reference?
- (what strategies might you use to validate your assembly?)
- (talk about standard computational approach)
what strategy would you use if I told you you could have as many undergrads as you wanted to do this?
3 strategies typically observed:
- Alignment of common words/letters/phrases
- seeded overlap/layout/consensus
- attempt to find sections that align and extend outwards
- OLC assembly
- complicated by severity of errors
- Extension from common words/letters/phrases
- start with one word, expand outwards from that word
- De Bruijn graph assembly
- K-mers
- the worst with severe errors; extension will be halted by areas with low reads/high error
- this approach will store more words, so it will need more memory with more data (?)
- Greedy
- greedy assembly
The best strategy for teams: have every member of a group come up with their own conclusion, then compare and correct This prevents people from building on each other's mistakes
This exercise is easier than real genomic analysis: the language is known, the word boundaries are maintained, the error rate is low, there are no reverse complement reads
Multiple flavors of scale challenges
- size of genome
- bacterial vs. human, for example
- depth of sequencing
- number of reads to align
- throwing away reads that have already been resolved accelerates the process by reducing the overall load
Long vs. short reads
- Long reads critical to overcoming repetition
- Short reads have better fidelity
- Most approaches combine long and short reads
- Currently, only way to rely on long reads to generate a high-quality genome is PacBio
- Nanopore can't
- And PacBio is v expensive
- what strategies?
- greedy
- overlap-layout-consensus
- word-based "pivot"
- average coverage vs actual coverage
- it helps that we know the language - imagine doing this in a language you didn't know!
- more data is not necessarily helpful without a good strategy... welcome to bioinformatics!
- how do you deal with data with different errors, biases, etc?
- what strategies do you have for validation?
- words? (k-mers)
- sentences? (mapping)
- "does it make sense" -> is that a good strategy?
- (what kinds of errors might you expect?)
- errors vs variants
- how do you deal with repeats?
- do you use all the data?
- internal validation vs external validation
- same samples / technical
- different samples / biological
More fundamental q: is there sufficient information to reconstruct the true text?
- Types of samples and assumptions
- genomics: ~single genome
- rnaseq: ~single genome, multiple splice variants
- metagenomics: ~multiple genomes, multiple strain variants
- assumptions of clonality
- Types of data / information content
- short reads
- long reads
- 10x
- long range scaffolding
- hi-c / contacts
- Errors and types of errors in data and effects
- Types of errors in assemble
this is shotgun sequencing: assumption of even sampling of nucleotide content.
fundamentally, three strategies for making use of shotgun sequencing:
- assembly
- construct new seqeunce!
- assumption of ~little strain variation
- mapping
- reference dependent / biased
- quantification
- assumptions about coverage/sampling rate
A final thought about assembling genomes - what do you do next?