Scoring functions, heuristics, and prior data in read alignment

Transcript

1. Ben Langmead
   Assistant Professor, JHU Computer Science
   langmea@cs.jhu.edu, langmead-lab.org, @BenLangmead
   Workshop on the Future of Algorithms in Biology,
   September 29, 2018
   Scoring functions, heuristics, and prior
   data in read alignment
   

   View Slide

2. Read alignment
   "Sequence mapping is the cornerstone of modern genomics"
   Novak AM, Rosen Y, Haussler D, Paten B. Canonical, stable, general mapping
   using context schemes. Bioinformatics. 2015 Nov 15;31(22):3569-76.
   But how do you...
   Predict mapping quality?
   Characterize what the heuristics are missing?
   Select an appropriate scoring scheme?
   See John Kececioglu's talk!
   

   View Slide

3. A C G T op ex
   A 2 -6 -6 -6 -5 -3
   C -6 2 -6 -6 -5 -3
   G -6 -6 2 -6 -5 -3
   T -6 -6 -6 2 -5 -3
   op -5 -5 -5 -5
   ex -3 -3 -3 -3
   A C G T op ex
   A 2 -4 -4 -4 -4 -2
   C -4 2 -4 -4 -4 -2
   G -4 -4 2 -4 -4 -2
   T -4 -4 -4 2 -4 -2
   op -4 -4 -4 -4
   ex -2 -2 -2 -2
   Bowtie 2 minimap 2
   Affine-gap scoring
   A C G T op ex
   A 9 Qu Qu Qu -40 -6
   C Qu 9 Qu Qu -40 -6
   G Qu Qu 9 Qu -40 -6
   T Qu Qu Qu 9 -40 -6
   op -40 -40 -40 -40
   ex -6 -6 -6 -6
   Novoalign Classic: PAM250
   

   View Slide

4. Heuristics
   0
   0 35
   30 35
   30
   0
   0 35
   30 35
   30
   Ref string 1
   Ref string 3
   Ref substring
   Read
   Read substring
   Ref string 1
   Ref string 3
   Ref substring
   ∅
   Read substring
   0
   0 35
   30 35
   30
   0
   0 35
   30 35
   30
   0
   0 35
   30 35
   30
   Read substring
   x 0
   0 35
   30 35
   30
   Basically a good match for affine-gap scoring
   Seed and extend
   

   View Slide

5. Spliced alignment
   Image by Rgocs
   

   View Slide

6. Fair BJ, Pleiss JA. The power of fission: yeast as a tool for understanding
   complex splicing. Curr Genet. 2017 Jun; 63(3):375-380.
   Aspects of splicing not captured by an affine-gap scoring function. (a) nearby motifs,
   notably donors & acceptors. (b) intron length distributions
   

   View Slide

7. Gotoh O. Modeling one thousand intron length distributions with fitild.
   Bioinformatics. 2018 Oct 1;34(19):3258-3264.
   Chicken
   Nematode
   Green alga
   Fungus
   Protist
   Plant
   Tapeworms
   

   View Slide

8. Pass 1: align to genome, make junction calls
   Pass 2: re-align to genome with putative junctions
   Reads:
   Ref:
   Readlets:
   

   View Slide

9. Graph alignment
   Jouni S, Välimäki N, and Mäkinen V. "Indexing graphs for path queries with applications in genome
   research." IEEE/ACM Transactions on Computational Biology and Bioinformatics (TCBB) 11.2 (2014): 375-388.
   Works around the fact that typical scoring
   scheme ignores known ALTs
   REF
   ALT
   ALT
   ALT
   ALT
   

   View Slide

10. Graph alignment
    ALTs to include in graph are selected (carefully)
    offline. Figure: results from FORGe study.
    Jacob Pritt, Nae-Chyun Chen, Ben Langmead
    doi: https://doi.org/10.1101/311720
    

    View Slide

11. Graph alignment
    Phase 1: Select ALTs to
    include and form genome
    representation & index
    containing them
    Phase 2: Align
    with standard
    heuristics and
    scoring function
    FORGe
    Model & score ALTs
    Select top X%
    Build graph
    index
    Align to
    reference with
    ALTs
    VCF FASTA FASTQ
    SAM
    Inputs:
    Outputs: ALT
    subset
    Index w/
    ALTs
    

    View Slide

12. Using prior data
    FORGe
    Model & score ALTs
    Select top X%
    Build graph
    index
    Align to
    reference with
    ALTs
    Junction discovery
    Align readlets
    Call junctions
    Build graph
    index
    Align to
    reference with
    splice junctions
    

    View Slide

13. Questions
    Two variations on read alignment -- spliced and graph
    alignment -- modify the reference to work around the fact
    that typical heuristics and scoring functions aren’t
    appropriate in those settings.
    Is this a paradigm that we can or should improve on?
    What can we do differently?
    Thank you: Jacob Pritt
    Nae-Chyun Chen
    NSF: IIS-1349906
    

    View Slide