Fragment HMM: Protein structure prediction

CREATED: 200803240805 Speaker: Li Shuai Cheng, University of Waterloo

** Existing work

• Wet lab methods: X-ray, NMR ** 150k per structure, 1/2 year, require computational analysis
• Computational methods ** homology modeling (Psi-BLAST) ** threading (RAPTOR) ** fragment assembly (ROSETTA), most successful currently ** consensus

** Background

• 40,000 proteins in PDB, humans have 100,000 proteins
• secondary structure: $\alpha$-helix, $\beta$-sheet, loop
• tertiary structure: 3D structure of one domain
• quaternary structure: few domains docked together
• amino acids are similar, except for side chain
• amino acits lose water when forming peptide bond
• bond is planar, rigid with known bond distances and angles
• each bond has two angles, $\phi$ and $\psi$

** Fragment-HMM

• computational methods: a heuristic science
• come up with a simple theory for protein structure prediction
• position specific: the ith amino acid has a set of nodes $H_i$
• each node has two emission probability distribution, secondary structure $S$ and $(\phi, \psi)$
• Build the HMM

generate contiguous subsequence of length $L=9$ and find $n=200$ structural fragments

for each amino acid there are $nL=1800$ fragments covering it, use a mixture of cosine functions to cover the dihedral angles

for each model in the mixture, generate a hidden node

estimate the transition probability

• Sample the HMM to generate structures and compute its energy, accept a new subsequence if the energy is lower
• Feedback the decoys as the new position-specific fragment library to retrain the HMM