What is the procedure to develop a new force field for molecular simulation?

Question

Suppose, I want to develop a novel force field for proteins.

What is the empirical procedure I should follow to do that?

Answer 1

The most useful answer to "how do I develop a novel force field for proteins" is "don't". (Or collaborate with a group which does.)

Developing a protein force field is exceptionally difficult because protein structure emerges from the interplay of hydrophobic-hydrophilic competition and backbone conformational flexibility. These are collective phenomena which mean that small changes in coefficients result in large changes in observed behaviour, making optimisation difficult.

Furthermore, very few elements are involved, so the protein FF designer must choose different types for (say) carbons in different chemical environments; too many types will lead to overfitting, while too few types will not capture the chemical complexity involved and lead to wrong answers.

The "how" of it is relatively straightforward: you establish references for how proteins should behave, run MD simulations to see how your MD proteins actually behave, and then iteratively improve your parameters until your MD proteins match your reference criteria. These references can be either "top-down" -- that is, based on experimentally-determined protein behaviour, such as NMR (and increasingly cryo-EM) protein structures and folding/unfolding rates -- or "bottom-up" -- from quantum chemistry calculations. But in either case we are talking massive reference datasets (such as a significant fraction of the PDB).

Thus, in practice, fitting a force field successfully requires expertise in: molecular dynamics simulations, the reference domain (so that, for example, dubious protein structures are less heavily weighted), and general large-scale data analytics (for example, it is easier to crystallize ordered proteins than disordered proteins, so most force fields fifteen years or older are biased towards ordered protein structures even for intrinsically-disordered proteins).

But don't take my word for it. This paper describes the parametrisation of the AMBER ff15ipq force field. Pay attention to just how much work it takes -- and the excellent supplementary information is very frank about the process required, including plenty of dead ends and laborious simulations.