How are different forcefields developed/evaluated?

Question

The Rosetta software can use a variety of different energy functions:

• Ref2015
• GenPot
• Beta-Nov16

How are these energy functions developed and validated? Is it common to do a single protein design with several of these energy functions?

Answer 1

Some force-fields rely solely on theoretical physical principles, while others are tweaked specifically for the goal of accurately depicting certain structures like sheets and helices or movement/diffusion aspects at different theoretical temperatures of a set of well-studied proteins. (Top-Down vs. Bottom-Up)

I haven't done MD myself but I am in an MD group and people seem to normally stick to a single force-field for a given MD project but it happens that they do very short simulation runs with force-fields just as a test. They definitely don't plan on doing the full simulation on different force-fields unless the comparison of those is their main goal.

Answer 2

The Rosetta ref2015 is well described in two papers, which I wish I had read much earlier, see this JChIM paper.

It's a statistical forcefield, where the terms are weighted to best match the empirical total energy in kcal/mol and the paper does a swell job at explaining what they do.

For several Rosetta experiments, you switch between full-atom and centroid (=coarse grain). These have different weights, as you can see from the files in the database folder or via pyrosetta, but majorly they have different "params" (topologies) for the residues as the centroid lacks many atoms as it's a simplification. In some instances, you may switch between dihedral (normal) to cartesian (explained in the FastRelax paper), wherein the former used internal coordinates (e.g. atom A is 1.5 Å away from B and A,B,C is a 107ª angle etc —cf. a params file ICOOR block), while the later is cartesian coordinates ( A is at [x, y, z] —cf. PDB file).

The GenPot uses its own AtomTypes, which means it uses its own params for each residue —like centroid/fa. The topology of a residue (ResidueType) is based on its conformation (ICOOR), bonding and atom definition, which is atom name, gasteiger-marsili charge and AtomType (an amalgamation of element, sp-hybridization, size, H-bond donor/acceptor etc.) These AtomTypes are better for ligands as many groups behave different from peptide atoms —say, a nitrile nitrogen is definitely different from a deprotonated aromatic nitrogen.

Some application, such as hydration, ligand docking etc. have specifically tweaked scorefunctions for use by their movers. FastRelax is a nice example as the cycle weighting changes are spat out by the tracer at vertiginous speed, where the protocol is:

repeat 5
coord_cst_weight 1.0
scale:fa_rep 0.040
repack
scale:fa_rep 0.051
min 0.01
coord_cst_weight 0.5
scale:fa_rep 0.265
repack
scale:fa_rep 0.280
min 0.01
coord_cst_weight 0.0
scale:fa_rep 0.559
repack
scale:fa_rep 0.581
min 0.01
coord_cst_weight 0.0
scale:fa_rep 1
repack
min 0.00001
accept_to_best
endrepeat

As you can see, the repulsion term of the Lenard-Jones is going from 0.04 up. These values were chosen not for thermodynamic reasons, but worked best for the application (energy minimisation of the structure).

However, this switching is all internal. Scores are always reported with the same scorefunction and comparing between different ones is not advised even for the minority of scorefunctions calibrated to be in kcal/mol.