Notes 20100928 McConkey BIOL 614

From SnOwy - Ed's Wiki Notebook

Contents 1 Hidden Markov Models (HMMs) and PSI-BLAST 1.1 PSI-BLAST and sequence profiles 1.2 PSSMs 1.2.1 Calculating PSSMs 1.2.2 PSSM Corruption 1.3 HMM 2 Course Project 3 Brendan Suggests

Hidden Markov Models (HMMs) and PSI-BLAST

• each represents sequence profiles

PSI-BLAST and sequence profiles

• more remote homologies
  • proteins, genes that have more distant common origins but may be beyond a single BLAST search
  • recognition of highly diverse gene families
  • all homologues within a given species -- splice isoforms
  • novel genes
• performed in five steps
  1. select query, search against protein database
  2. psi-blast constructs MSA, creates profile (PSSM)
  3. PSSM used as query
  4. psi-blast . . .
• redundant sequences are removed to reduce bias in query sequence

PSSMs

• captures position-specific patterns -- see weblogo -- describes conservation, preference at column
• highly conserved items are functionally related (structurally too?)
• those columns which have fewer tokens have higher bitscores -- we infer confidence with this representation

Calculating PSSMs

• technical issues -- building upon existing score matrices
• PSSM scores based on weighted BLOSUM matrix
• m_u,a = Σ_b(f_u,bs_a,b)
• where u is a given position in the alignment
• possible to use log odds instead -- but we just use frequencies seen in the MSA -- allow background frequency
• m_u,a = 1/λ log(q_u,a / p_a) and q_u,a = αf_u,a + (βp_a) / (α + β)

PSSM Corruption

• occurs when a particular non-homologous protein is grouped into the profile
• the erroneous signal is then propagated down into the next iterations
• sequences that may do this -- hydrophobic regions? low complexity regions?
• prevention
  1. filtering of biased composition regions
  2. change e-values -- reduce from 0.005 (default) to 0.0005 etc.
  3. visually inspect output in each iteration -- remove suspicious hits (in PSI-BLAST -- uncheck the box)

HMM

• the mathematical background is a bit different from PSSMs
• Markov assumption / Markov chain
• deals with gaps better than PSSMs
• requires a MSA as input
• example --

GTWYA
GLWYA
GRWYE
GTWYE
GEWFS

• recall: log odds score => sum of ln(probabilities)
• we also include a background probability to allow for low identity sequences

Course Project

• project proposals due this week
• example projects from two years ago ...
  • effectiveness of alignment algorithms on short highly variable 16S gene sequences
  • sequence, structural analyses of major outer membrane protein OmpF (Pseudomonas putida UW4)
  • in silico analysis of uncharacterized arabidopsis thaliana protein At1g22850
  • searching for glucocorticoid-induced matrix metalloproteinase overexpression in stressed adult zebrafish (microarray data)
• in phylogeny
  • phylogenetic reconstruction of Pontoporidea using Bayesian techniques
  • analysis of conserved domains of CHK1 for use in cloning CHK1 in Oncorhynchus mykiss
  • evolutionary diversification of 5′-Methylthioadenosine nucleosidase (MTN), Methylthioadenosine (MTA) generating enzymes
  • annotation of arabidopsis thaliana loci AT1G05620, AT2G36310
• deliverables
  • seminar
  • textual project report

Brendan Suggests

• position specific correlation given a structure and a PSSM
• promoter sequence *something* with HMM