The origin of
free-energy methods

Before any of the free-energy work, there was a prior question: how do you represent water? In the 1970s the answer that looked most promising was what we would now call machine learning — fit a model to a large set of quantum-mechanical calculations and let the fit stand in for the physics.

Enrico Clementi's MCY model, derived from configuration-interaction results for 64 water dimers, was the leading water model from 1976 to 1982. In 1979 Jorgensen reported his own version of the approach: a model fitted to 291 water dimer structures computed at the STO-3G level, in four- and five-site forms, tested by Monte Carlo simulation of the liquid.

It was then given up — for three reasons that have not gone away:

What replaced it was simpler and worked better. TIPS (1981) was the first three-site potential for water and methanol, fitted not to quantum calculations but to experiment. Two years later, once NPT Monte Carlo and fast enough minicomputers allowed the parameters to be optimized iteratively against measured densities and heats of vaporization, TIPS became TIP3P and TIP4P — models still in daily use across biomolecular simulation more than forty years later.

Free-energy perturbation had existed as an equation since Robert Zwanzig wrote it down in 1954. Turning it into a way to compare different molecules in solution was the open problem — and this group solved it.

William Jorgensen and C. Ravimohan computed the difference in hydration free energy between ethane and methanol by exactly this route — one of the earliest alchemical perturbations for a solute change in a condensed phase. It fixed the multi-window FEP protocol still used today. The trick that makes the number useful is a thermodynamic cycle: free energy is a state function, so a hard-to-measure difference can be assembled from transformations a simulation can do.

The method travelled fast, and mostly by hand. Jorgensen sent a preprint to his friend Peter Kollman at UCSF; soon thereafter, free-energy perturbation was running in AMBER on the same single-topology strategy. Kollman's letter of June 27, 1985 — weeks after the paper's submission — reports the code under development. A second letter, from Martin Karplus (May 3, 1985), asks for a preprint. Soon the same ideas entered CHARMM, in the Karplus group joined by Jiali Gao from Jorgensen's lab. Three of the major packages carried the method within a year.

A note on names: Jorgensen drew the approach from Donald McQuarrie's 1976 statistical-mechanics text — chapter 14, “Perturbation Theory of Liquids,” which opens with Zwanzig's 1954 theory. That is why his early papers call it statistical perturbation theory; the name “free-energy perturbation” came later.

The 1985 code began in 1978 as a program for simple liquids; extended with double-wide sampling, it became BOSS — and then MCPRO for biomolecules, with the OPLS force fields supplying the small-molecule parameters early free-energy work lacked. Jorgensen demonstrated FEP's value for lead discovery with MCPRO and all-atom OPLS, and in 2006 MCPRO was licensed by Schrödinger as MCPRO+.

Schrödinger's manual describes MCPRO+ as “a simulation package built around the MCPRO program”, offering non-expert users relative binding free energies “via highly-accurate free energy perturbation (FEP) theory”. The company's FEP+, launched in 2015, is now used across the pharmaceutical industry. Alchemical free-energy calculation is routine in structure-based drug design.