Free
energies as a function of a selected set of collective variables
are commonly computed in molecular simulation and of significant value
in understanding and engineering molecular behavior. These free energy
surfaces are most commonly estimated using variants of histogramming
techniques, but such approaches obscure two important facets of these
functions. First, the empirical observations along the collective
variable are defined by an ensemble of discrete observations, and
the coarsening of these observations into a histogram bin incurs unnecessary
loss of information. Second, the free energy surface is itself almost
always a continuous function, and its representation by a histogram
introduces inherent approximations due to the discretization. In this
study, we relate the observed discrete observations from biased simulations
to the inferred underlying continuous probability distribution over
the collective variables and derive histogram-free techniques for
estimating this free energy surface. We reformulate free energy surface
estimation as minimization of a Kullback–Leibler divergence
between a continuous trial function and the discrete empirical distribution
and show that this is equivalent to likelihood maximization of a trial
function given a set of sampled data. We then present a fully Bayesian
treatment of this formalism, which enables the incorporation of powerful
Bayesian tools such as the inclusion of regularizing priors, uncertainty
quantification, and model selection techniques. We demonstrate this
new formalism in the analysis of umbrella sampling simulations for
the χ torsion of a valine side chain in the L99A mutant of T4
lysozyme with benzene bound in the cavity.