Week 4
Structure prediction: threading and energy minimization
Below the identity needed for homology modeling, there are two options. Fold recognition asks which known fold a sequence fits best. It works either through sensitive profile-based homology detection or by threading the sequence onto template structures and scoring the fit with a statistical potential. In the midnight zone, where no template is usable, prediction becomes energy minimization over conformational space. That calls for the deterministic and stochastic search methods introduced this week.
Objectives
- Explain how profiles, PSI-BLAST, and hidden Markov models built from multiple alignments detect distant homologs that pairwise alignment misses.
- Formulate threading as aligning a sequence to a template structure and scoring the result with a contact-based energy. Explain why that pairwise, nonlocal energy breaks ordinary dynamic programming.
- Describe the workarounds for nonlocal threading scores (frozen approximation, double dynamic programming, Monte Carlo, branch and bound) and their trade-offs in accuracy and cost.
- Derive a knowledge-based potential of mean force from contact frequencies observed in known structures with the inverse Boltzmann relation, and explain the role of the reference state.
- Explain why prediction without a template is a global minimization over a rugged energy landscape, and why local minimization alone gets trapped in nearby minima.
- Apply dead-end elimination (original and Goldstein criteria) to prune rotamers in side-chain packing, and describe self-consistent mean field as an alternative deterministic method.
- Implement Metropolis Monte Carlo with simulated annealing for a conformational search: propose a move, accept it with probability min(1, exp(−ΔE/kT)), and lower the temperature on a schedule.
- Compare genetic algorithms, biased Monte Carlo, and replica exchange with plain Monte Carlo by how they escape local minima and what each step costs.
- Explain why representative sampling needs a nonredundant structure set, and how rotamer libraries and coarse-grained models shrink the search space.
Reading
| Gu & Bourne 31, 8 | Ch. 31 (Godzik) covers fold recognition from two sides: the biologist's (sequence profiles and distant homology) and the physicist's (threading with potentials of mean force). Focus on the threading approximations and the analysis of where threading errors come from. Ch. 8 (Samish) surveys search and sampling: deterministic methods such as dead-end elimination and mean field, and stochastic ones such as Monte Carlo, simulated annealing, and genetic algorithms. Focus on the Search Methods section and skim the statistics primer. |
Lectures
| 7 | Classical protein structure prediction II transcript | 1:15:04 |
| 8 | Structure prediction in the midnight zone: energy minimization transcript | 1:15:05 |
| 2:30:09 |