Moving from pymbar version 3

Pymbar v4.0 contains several changes to improve the API longer term. This, however, breaks the API used in 3.x and previous versions.

The main changes include:

  • Making various estimators return dictionaries, not tuples, making it easier to return optional information requested at call time.

  • Standardizing on snake_case for function names.

  • Making the built-in solvers work to have an interface closer to like scipy solvers.

Snake case

Previous version of pymbar had mixed cases in functions. We have standardized on snake case, and tried to make the method names that do similar things more consistent. Specific changes include:

  • getFreeEnergyDifferences is now compute_free_energy_differences

  • computeExpectations is now compute_expectations

  • computeMultipleExpectations is now compute_multiple_expectations

  • computePerturbedFreeEnergies is now compute_perturbed_free_energies

  • computeEntropyAndEnthalpy is now compute_entropy_and_enthalpy

In the submodule timeseries:

  • statisticalInefficiency is now statistical_inefficiency

  • statisticalInefficiencyMltiple is now statistical_inefficiency_multiple

  • integratedAutocorrelationTime is now integrated_autocorrelation_time

  • normalizedFluctuationCorrelationFunction is now normalized_fluctuation_correlation_function

  • normalizedFluctuationCorrelationFunctionMultiple is now normalized_fluctuation_correlation_function_multiple

  • subsampleCorrelatedData is now subsample_correlated_data

  • detectEquilibration is now detect_equilibration

  • statisticalInefficiency_fft is now statistical_inefficiency_fft

  • detectEquilibration_binary_search is now detect_equilibration_binary_search

Additionally, the other estimators such as the Bennett Acceptance Ratio and exponential averaging/Zwanzig equation have different, more consistent, call signatures. All other estimators are now in the other_estimators module.

More consistent return functionality

Previously, different pymbar functions returned different information as tuples. This became problematic when different functions returned different types of information or different numbers of results. We have thus consolidated on an API where all functions return a dictionary.

As an example of both API changes of API, a short bit of code that would load in data and calculate free energies, instead of being

Example of initializing MBAR in 3.0.7
mbar = MBAR(u_kn, N_k)
results, errors = mbar.getFreeEnergyDifferences()

Would now be written as:

Example of initializing MBAR in 4.0
mbar = MBAR(u_kn, N_k)
results = mbar.compute_free_energy_differences()

Other estimators including bar and exp also use a dictionary for return data.

The pymbar.timeseries submodule return patterns have not changed in 4.0, however, and one should refer to the individual function documentations for these return patterns.

results = bar(w_F, w_R)
print(f'Free energy difference is {results['Delta_f']:.3f} +- {results['Delta_f']:.3f} kT')

results = exp(w_F)
print(f"Forward free energy difference is {results['Delta_f']:.3f} +- {results['dDelta_f']:.3f} kT)
results = exp(w_R)
print(f"Reverse free energy difference is {results['Delta_f']:.3f} +- {results['dDelta_f']:.3f} kT)

Simulation output

Previously, pymbar send all messages to standard out when verbose was set to True. pymbar now uses the logging module to output this information. If you wish to set messages, even if the verbose is set to True, you will need to turn on logging for your script by importing the logging module, and adding the lines:

Enabling logging in pybmar
import logging
import sys
logging.basicConfig(stream=sys.stdout, level=logging.INFO)

pymbar generally uses the logging levels info for information that previously was set to standard out. Note that for a given method to produce extensive information, even with logging, the verbose flag still needs to be set to true.

Free energy surfaces

Previously, pymbar had a method PMF that estimated a free energy from a series of umbrella samples using a histogram approach. This was sematically problematin in two ways. First, the term PMF (potential of mean force) is somewhat of an ambiguous term, as the potential of mean force has some dependence on the coordinate system in which the mean force is calculated. Since pymbar does not calculate free energies by integration of mean force, this caused some comfusion. To be clearer, we now have renamed the class FES, for “free energy surface”.

The inclusion of a PMF function also created some confusion where some authors referred to MBAR as a method to calculate a free energy surface. MBAR can only be used to take biased samples an estimate the unbiased weight of each sample. In order to calculate a free energy surface, one must also find a way to take the set of discrete weighted samples and calculate a continous potential of mean force: see Shirts and Ferguson [1] for a further discussion of the separation of these two distinct tasks in the construction of free energy surfaces. The pymbar code more cleanly separates the calculation of biasing weights associated with umbrella samples, and the estimation of the free energy surface.

For more information on the options for computing free energy surfaces with the code, please see: Free energy surfaces with pymbar.


Previous version of pymbar include acceleration using explict C++ inner loops. The C++ interface has become out of date. pymbar optimization routines are now accelerated with jax. This provides approximately a 2x speed up when performed on most CPUs, and additional acceleration when a GPU can be detected (pymbar does not install the appropriate GPU libraries). jax will be installed when pymbar in installed via conda, but pymbar will function with or without jax installed if there are issues with the JAX configuration.

Other changes

Additional changes not affecting the API:
  • Removed legacy support.

  • Moved testing framework to pytest, added significant numbers of tests.

  • Improved code linting using black

  • Bootstrapping for errors in free energies and expectations is now supported; see Strategies for solution for more information.

  • Added a bar_overlap function to find overlap when using just bar

  • Fixed an error in computing expectations of small numbers.

  • Improved automated adaptive choice of samplers; see Strategies for solution for more information.

  • Many instances of code cleanup.

  • Improved docstring documentation.