Week 5

Numerical optimization for structure prediction

Also this week: Protein docking.

This week covers the numerical optimization that energy-based structure prediction relies on. Gradient-based methods (steepest descent, conjugate gradient, Newton and quasi-Newton) refine a structure to a nearby minimum. Gradient-free methods (Monte Carlo, simulated annealing, genetic algorithms) search a rugged landscape for a global one. Rosetta shows how the two combine: fragment assembly in a reduced model, then all-atom refinement and clustering of decoys. Docking applies the same search-and-score approach to binding between proteins and between proteins and ligands, and models flexibility with inverse kinematics, normal modes, and rigidity analysis.

Objectives

  • Run steepest descent with a line search on a smooth energy function, and explain why it zigzags and slows down on ill-conditioned problems.
  • Derive Newton's method from a second-order Taylor expansion, and explain how quasi-Newton methods (BFGS, L-BFGS) and conjugate gradient avoid forming and inverting the full Hessian.
  • Compare gradient-free global methods (Monte Carlo with simulated annealing, genetic algorithms, replica exchange) by how they escape local minima, when they apply to nondifferentiable scores, and what they cost.
  • Explain how lattice and off-lattice reduced models (centroid side chains, discrete torsion states, predicted local structure) shrink the conformational search, and what accuracy they give up.
  • Describe Rosetta de novo prediction: build 3- and 9-residue fragment libraries from the PDB, insert fragments by Monte Carlo under a low-resolution score, refine with all-atom rotamers, and pick the centers of the largest decoy clusters.
  • Break docking into a search over rigid-body pose (grid search, descriptor matching, energy-based) and ligand conformations (systematic, stochastic, incremental), plus a scoring function (force-field, empirical, knowledge-based). Count a predicted pose as correct when its RMSD is under 2 Å.
  • Explain how electrostatics shapes binding affinity and association rates, and what the Poisson–Boltzmann equation models: dielectric boundary, mobile ions, Debye screening, and the linearized form.
  • Predict protein interaction sites and partners from evolutionary signals (conserved and coevolving positions, phylogenetic profiles, gene fusion, gene neighborhood, similar phylogenetic trees) and from structure.
  • State the loop closure problem, explain why a chain with 6 free dihedrals has at most 16 solutions, and describe how Probik and ChainTweak use inverse kinematics to sample backbone motion.
  • Compute normal modes from the Hessian of a harmonic potential and use Gaussian and anisotropic network models to predict residue fluctuations. Find rigid and flexible regions with FIRST's pebble game, and describe how ROCK and FRODA sample the flexible parts for docking.

Reading

Gu & Bourne 32Drew, Chivian, and Bonneau on de novo prediction: reduced-complexity models, low-resolution scoring terms, Rosetta's fragment assembly and refinement, decoy clustering, and CASP. Focus on the Rosetta section and the trade-off between resolution and sampling, and skim the genome-annotation applications.
Donald 20, 21, 22, 23Donald's short lecture-note chapters on protein flexibility: loop closure and inverse kinematics (Probik, ChainTweak), normal mode analysis and rigidity theory (FIRST and the pebble game), sampling flexible regions (ROCK, FRODA), and normal-mode models of binding and receptor flexibility in docking. Focus on the problem statements and algorithm outlines.
Gu & Bourne 24, 25, 26, 27Four chapters on molecular interactions: electrostatics and the Poisson–Boltzmann equation (24), protein–nucleic acid recognition and statistical potentials (25), predicting protein–protein interaction sites and partners from evolutionary information (26), and docking search, scoring, and virtual screening (27). Ch. 27 is the core docking reading; read the others for the energy terms and interface signals that docking scores use.

Lectures

9Gradient-based numerical optimization transcript59:33
10.1Gradient-free numerical optimization transcript43:56
10.2Great optimization ideas for protein structure prediction transcript21:04
2:04:33