27-homology-modeling.md

2.7 Homology Modeling

29.05.2017 | Slides | Wiki

---

1. Ideas and Keywords

Ideas

• Sequence Determination (short: sequencing) of DNA is highly automated today and very cheap

• Computer programs can help identify genes and coding regions

• From coding regions you can infer the protein sequence (1D information)

• The entire sequencing process is cheap and quick, everything after that isn't.

Available types of Data

• sequences (1D information)
• Annotations of already investigated proteins
• (few) protein structures (3D information)
• Goal: clever combination to infer more knowledge about yet unknown protein

Additional Tools / Databases

• UniProtKB, PDB
• Blast, Smith-Waterman
• PSI-Blast, ClustalW/CluastlX, MaxHom, SAM / HMMer, T-Coffee
• HHblits: SSearch, PSI-Search

Homology / Comparative Modeling

• Goal: Direct link from 1D to 3D structure (this would be the ultimate jackpot, but it does not work so far)
• Work around: Borrow structure from already know, sequence-similar proteins
• Tools: Modeller, Swiss-Model
• Modeller
  • uses a set of spatial restraints applied as PDFs (probability density functions)
    • $$C_{\alpha} - C_{\alpha}$$ distances
    • main chain $$N-O$$ distances
    • main-chain and side-chain dihedral angles
  • Which PDFs? Derived from analysis of 17 homologous protein families
  • needs a related template with a known 3D structure
  • Features
    • models non-hydrogen molecules
    • de-novo (?) prediction of loops
    • local installation
  • Typical steps
    • 1) identify templates / fold recognition
    • 2) align
    • 3) model
    • 4) assess
    • 5) refine
• Swiss-Model
  • originally: fully automated, little user interaction
    • 1) selection of templates
    • 2) modeling (copying coordinates)
    • assessment
  • now: more interactive and sophisticated model assessment
  • sever based service
  • convergent evolution with Modeller

Secondary Structure Prediction

• actual secondary structure of amino acid depends on the local sequence (context)
• even identical stretches (up to 5 aa) can occur in different secondary structures
• which structure is preferred depends on the available **hydrogen bond **opportunities
• more hydrogen bonds => more stable
• Chou-Fasman
  • simply look a the frequency an amino acid occurs in each secondary structure
  • search for nucleation regions
    • for helix: 4 out of 6
    • for sheet: 3 out of 5
  • extend until a window of 4 amino acids drops below 1
  • turns also check for Proline and Glycine
  • More info, in case this was not enough to understand: https://en.wikipedia.org/wiki/Chou%E2%80%93Fasman_method
• GOR I
  • 17 amino acid window
  • considers the state of 8 aa neighbors on each side (bayesian)
  • builds on three matrices (17X20) for helix, sheet and coil
  • (the original 'turn-matrix' was removed since it showed too high variability for a window of 17 aa)
  • thresholds:
    • 4 amino acids for helix
    • 2 amino acids for sheet
• GOR III
  • in addition for GOR I it considers all pairs with on the sliding window (= segment)
  • still not good for sheets, since the could be formed by non-local interaction
• PHDxxx
  • usage of local evolutionary information in the form of sequence profiles generated from multiple alignments
  • usage of global features (length, aa composition, ...)
  • the use of redundancy-reduced, balanced data set for training can be useful
• PHD(-acc, -sec, -htm)
  • add a second layer of networks (PHDsec)
    • L1: sequence residue -> secondary structure of that resisude
    • L2: secondary structure state -> secondary structure state for consolidated (smoothened) predictions
  • create a jury between balanced and unbalanced trained networks and different output states

Performance Assessment

• Precision, Recall, Accuracy
• Qx-Measure: For x states, fraction of correct predictions (TrueNegative + TruePositives) of all predictions
• Significance?
  • determine the average Q on you dataset
  • calculate the standard deviation (sigma)
  • calculate the standard error (sigma / sqrt(N) )
  • N is size of test set
• Compare Methods
  • compare always on the same instances
  • test / training split have to be the same for both tools
  • not overlap allowed between test and training set (structures in comparative modeling range violate this)
  • alternative: compare on fresh data published after publication of methods

Membrane Proteins

• hydrophobic stretches, typically 17-21 amino acids long
• Positive Inside Rule (connecting sequences on inside of cell are positively charged) to determine topology
• Signal Peptides (often confused with TMHs)
• in 3D: hydrophobic parts on the outside
• many hydrophobicity indices out there
• optimized scoring matrices
• Nowadays:
  • interrupted TMHs
  • reentrant parts (leave membrane on entry side)
  • coil regions inside membrane

HSSP Curve

Homology-derived Secondary Structure of Proteins

2. Exercises

Questions

Question: How does ClustalW work? How does it differ from BLAST?

Clustal is a series of widely used computer programs used in Bioinformatics for multiple sequence alignment. from Wikipedia All variants of Clustal align sequences by three main steps:

• Do a pairwise alignment
• Create a guide tree (or use a user-defined tree)
• Use the guide tree to carry out a multiple alignment

? BLAST itself is not a multiple sequence alignment tool

Question: What is HHblits?

A multiple sequence allignment tool using HMMs HHblits-Schematic

Question: What is the definition of accuracy? Is it the same as Qx?

?