ab initio structure prediction

Question 1

ab initio methods

Answer 1

• template-free
• no template available or can’t be found
• 3 methods:
  
  • all atom molecular dynamics
    
    • simulate structure as it folds
  • fragment approach
  • contact prediction from multiple sequences

Question 2

rosetta

Answer 2

• fragment approach
• match query sequence to small sections of proteins
  
  • 9 residue segments
  • template based search algorithm
• fit sections together into 3D structure
  
  • build up overlapping fragments with predicted structure
  • create trial model
• if structure doesn’t work remodel it

Question 3

rosetta

identification of good trial structures

Answer 3

• initially low resolution energy function
• side chains represented by single centroid pseudoatom
• major contributions:
  
  • hydrophobic burial, beta strand pairing, steric overlap, specific residue interactions
• form a coarse structure
• refine with rotamer based side chain representations

Question 4

rosetta

model choosing

Answer 4

• many possible methods created in ab initio
• can choose most popular:
  
  • calculate distance between each model
  • correct one has largest number of similar structures with smaller distances
  • some aspects usually correct but large regions can be region

Question 5

contact prediction

Answer 5

• create contact map of inter-residue distances
• x or 1 where 2 atoms closer than set cutoff distance
• better if using MSA to find common pattern of complementary changes
  
  • many sequences needed for strong signal
  • aligning thousands of seqeucnes helps remove noise

Question 6

contact prediction

anti-parallel beta strand

Answer 6

• chains connected giving leading diagonal
• if e.g. residue 1 and 12 are close there is an inidcation of 3D space
  
  • produces off diagonal (feature of 3D space)
• use these terms to build up structure

Question 7

contact prediction

limitations

Answer 7

• limited to proteins with large numbers of homologues
  
  • >5000 ideally
  • often means template is available anyway
• can still be useful for solving parts of a structure
• should improve as databases grow
  
  • good for membrane proteins (difficult to crystallise)

Question 8

future of structure prediction

Answer 8

• increasing number of determined structures
  
  • more templates
  • more fragments for ab initio
• increasing number of sequences
  
  • better MSAs and profiles
  • better information capture for template based modelling
  • increasing viability of contact prediction

Question 9

empirical prediction algorithms

Answer 9

• establish training dataset
  
  • e.g. sequences with known structures
• ensure no duplicated due to homology
  
  • non-redundant set, prevents bias
• learn rules and parameters
• evaluate on testing set not used in training
  
  • no homology important

Question 10

jack-knifing

Answer 10

• cross-validation
• split database into training and testing sets
• start with set of non-homologous data
  
  • take out one/several to form testing set
  • learn on rest and evaluate test data
  • repeat with different test data
• get mean and variation of accuracy
  
  • statistical analysis to compare methods (t test)

Question 11

pitfalls of jack-knifing

Answer 11

• multiple methods available
• with proteins, testing and training data often have homology
• more difficult to detect bias as algorithms become more complex
  
  • best to test on new data not available during development