Notes 20110214 Base Calling (Phred)

From SnOwy - Ed's Wiki Notebook

Contents 1 Base Calling 2 Early Phred Paper -- Error Estimation, Confident Base Assignment 3 General Notes 4 Support Vector Machine (SVM) on Next Generation Sequencing (NGS) Data 5 Statistical Learner for Illumina 6 Need for new classifier

Base Calling

• automated reading of fluorescent nucleotides
• original software designed is called Phred

Early Phred Paper -- Error Estimation, Confident Base Assignment

doi:10.1101/gr.8.3.186 - 1998

• high speed sequencing limited by data processing
• phred estimates probability of error
• works with phrap (assembly) and consed (finishing program)
• estimate reliability of reads
• position-specific error probabilities
• good assembly software →
  • accuracy
  • completeness
  • sequential repeats
  • consensus sequences
• finishing: missing data? aditional edits?
• previous work exists for base-calling
• phred -- novel properties
  1. deriving confidence without a bunch of assumptions
  2. assignment of confidence given a trace window (not just a single peak)
  3. maximize discrimination only in confident regions of trace (not equally over entire trace)
• log-transformed error probability
• q = -10log₁₀(p) -- p is estimated probability of error
• error probabilities
  • must be derivable without knowing actual sequence
  • must be valid -- based on observed error
• instead of treating each base position with the same probability of error
• use several classes of error probabilities
  • leads to more accurate results (varies with lower quality reads)
• quality value distribution
• more accurate reads are more important →
  • consensus reads
• error probability (at rate) = expected error (at rate) / number of base callsd
• error probability (at rate) -- provides confidence in estimation of error at a rate class
• study assigning r ≤0.01 (low) derived from expert knowledge (imitates human observer)
• encoded this decision making into algorithm with these parameters ...
• continue at Trace Parameters ...
  1. Peak spacing
  2. Uncalled:Called amplitude ratio at one position
  3. Uncalled:Called amplitude ratio at three adjacent positions
  4. Peak resolution -- related to distance between most recent unresolved base and current base undergoing resolution

General Notes

• base calling requires
  1. guessing at a given base
  2. providing an estimation of error
• additional parameters hinted by Dr. Kremer
  • knowing when to provide no answer rather than a wrong answer

Support Vector Machine (SVM) on Next Generation Sequencing (NGS) Data

ISBN = {978-1-4244-8302-0} - 2010

• previous work ...
  • sequence correction for SAGE data
    • SAGE?
    • previous software unable to handle the larger amounts of data
  • genome read correction
    • assume coverage large, thus rare sequences are incorrect
    • not as useful for transcriptome, metagenome sequencing
    • correct sequences often rare
• require sequence correction based on entire ensembles of reads
• previous work ...
  • FreClu -- iteratively checks errors in clusters of reads
  • RECOUNT -- sequencer error modelled as multinomial thinning process
    • Expectation-Maximization (EM) framework
    • infers true counts → maximize likelihood of observations
• this work ...
  • propose method to classify true and false reads
  • binary classification problem
  • function learned -- feature sets mapped to {0, 1}
• input features ...
  1. observed count
  2. estimated true count (← RECOUNT)
  3. log likelihood (+ entropy penalty)
  4. log likelihood ratio
  5. expectation matching score
  6. self-correctness coefficient
• SVM classifier on ⟨ 6 ⟩ yields 96% accuracy.
• continue at II. Data Simulator ...

Statistical Learner for Illumina

• doi:10.1186/gb-2009-10-8-r83 - 2009
• start at Statistical learner for Illumina base calling ...
• base caller design
  • should mitigate cycle-dependent problems
  • could model sequence chemistry
    • T accumulation
    • intensity fading
    • first and last cycle quirks of sequencing
• problems:
  • unlikely to know all the reasons for errors
  • over complication → data overfitting (bad generalization)
• solution:
  • estimate chemistry from data
• this approach does not use intensity correction before learning
  • allows statistical learner to work with raw values (for better or for worse)
  • trade dependence on intensity correcting model for cognitive overhead
• continue at Statistical learner for Illumina base calling ...

Need for new classifier

• need to eliminate the human being at the end
• increase the throughput
• removes subjectivity
• the person at the top looks at the predictions and the graphs (peaks)
• overrides the software in some cases
• we require examples of the human making the overriding
• put the pdf for the file format (ab1 = applied biosystems one)
• ab1 currently gives us ...
  • raw signals
  • the estimation of the sequence
  • probability assigned
  • understand the transformation between the raw and cooked signal