14 good questions and answers about "Advanced structure prediction"
Ab initio (knowledge based energies)
both go into fragment based model constructrion
- Homology detection
- sequence search: BLAST
- profile search: PSI-BLAST / HMMer
- profile-profile search: FFAS03/COMPASS/ HHpred
- Fold recognition
- structure profile -sequence search: Fugue
- Threading: THREADER, RAPTOR
- meta-servers (+ insignificant scores)
And charged ones more conserved on outside
- Ab initio prediction
- “Knowledge based”“ab initio”protein structure prediction
- Usage of existing databases of protein structures ( so ab initio not really suitable term)
- use if no suitable template exists
- PDB is used
- a fragment library is constructed
- so small protein fragments, small structure elements
- all possible elements and transitions included
- target sequence is also split up and tried to match with fragment
- can give multiple options, so monte carlo is used
- Fragments are assembled and multiple models are created, minimizing energy
- --> decoys
- from all decoys, you select the one that looks best like protein structure and matches what you know of the sequence the best
- Now we have a many fragments, at each sequence position, with many possibilities to combine them:
- We use Monte Carlo (MC) to search through different combinations.
- Good combinations are those that give a low energy .Each MC run will give you a different model, since it is a stochastic algorithm.
- done by looking at phi/psi angles
- Such models are called decoys.
- If it lowers the energy function
- If random number < Boltzman factor
(if move accepted then the model is accepted).
Now atomistic force field is used and only small changes allowed.
- Protein like features (detailed energy function)
- Clusters of similar structures are more native like
- cf. entropy
So benchmarking: you want to see: lowest rosetta energy = lowest C-alfa RMSD. But it doensn't look like that!
- Higher grades + faster learning
- Never study anything twice
- 100% sure, 100% understanding