key operation for MD Simulation and studying Protein Interactions

  • EMin is finding lowest energy form on a graph, like a regular math problem
  • Central dogma: lowest energy structure is the ‘correct’ one
    • This is where first derivative = 0
      • If 2nd derivative > 0, it’s a minimum
      • if 2nd derivative < 0, it’s a maximum
      • If 2nd derivative = 0, it’s a saddle point
    • So in this problem
      • y is potential energy usually from force field equaitons
      • Xi is coordinate, sometime cartesian but not usually
  • We have to look for global minimum not local minima
    • Newton’s method: most CPU intensive but accurate
      • Approximate y locally using a Taylor series truncated to a parabola
      • Newton-Raphson: generalization of newton to multiple dimensions
    • **Steepest descent: **cheap on CPU works better when far away
      • Draw bunch of lines and use 3 points to make parabolas
      • calculate one with lowest point, go there. Repeat
      • Good at getting in the area quickly on dramatic cliffs but struggles with shallow plains it will wander aimlessly
    • Conjugate gradients: good general more expensive than steepest
      • Use M steps where M is the number of variables
      • Search in one, draw orthogonal vector for each one and proceed
    • Strategy
      • It’s advisable to start with a cheap one like steepest descent then refine search with a more expensive and accurate method