Bayesian Multistate Bennett Acceptance Ratio Methods

This paper introduces BayesMBAR, a Bayesian generalization of the multistate Bennett acceptance ratio (MBAR) method that computes free energy posterior distributions to provide more accurate uncertainty estimates and allow for the incorporation of prior knowledge, such as surface smoothness, into free energy calculations.

The Gist

Imagine you are trying to figure out the "cost" (free energy) of different states a molecule can be in, like how much effort it takes to move a protein from one shape to another. In the world of chemistry, scientists use a tool called MBAR (Multistate Bennett Acceptance Ratio) to calculate these costs based on data they collect from computer simulations.

Think of MBAR as a very smart accountant. If you give it a massive pile of receipts (simulation data), it gives you a very accurate total cost. However, if you only give it a few receipts, the accountant might get a bit shaky. It will still give you a number, but it might be wrong about how confident it should be in that number. It might say, "I'm 99% sure," when it's actually only 50% sure, or vice versa.

This paper introduces a new, upgraded accountant called BayesMBAR. Here is how it works, using simple analogies:

1. The "Gut Feeling" vs. The "Hard Data"

The main difference between the old MBAR and the new BayesMBAR is how they handle uncertainty and "gut feelings" (prior knowledge).

• The Old Way (MBAR): Imagine you are guessing the price of a house in a new neighborhood. You only have data on two houses. The old method looks strictly at those two houses and says, "Based on this, the price is $X." It doesn't really know how shaky that guess is if the data is thin.
• The New Way (BayesMBAR): This method is like a seasoned real estate agent. It looks at the two houses (the data), but it also brings in a "prior belief" or a "gut feeling."
  • Scenario A (No Extra Info): If the agent has no extra info, they use a "blank slate" approach. They ignore their gut feeling and just look at the data. In this case, BayesMBAR gives the exact same price as the old MBAR, BUT it is much better at telling you how unsure it is. It's like the agent saying, "The price is $X, and I'm only 60% sure because we don't have enough data," whereas the old method might have said, "I'm 90% sure."
  • Scenario B (With Extra Info): If the agent knows that houses in this neighborhood usually have smooth, predictable price changes (a "smooth free energy surface"), they can use that knowledge. BayesMBAR can say, "Hey, even though we only have two data points, we know prices usually change smoothly. So, let's adjust our guess to fit that smooth curve." This makes the final guess much more accurate when data is scarce.

2. The "Smoothness" Analogy

The paper specifically highlights a feature where you can tell the computer, "Hey, the cost of these states changes smoothly, like a rolling hill, not a jagged mountain."

• Without this: If you have very few data points, the computer might guess a jagged, weird path between them because it's just connecting the dots blindly.
• With this: The computer uses a "smoothness filter." It assumes the path between your data points is a gentle curve. This prevents the computer from making wild, unlikely guesses when it doesn't have enough data to be certain.

3. The "Two Estimates"

When BayesMBAR does its math, it actually gives you two slightly different answers:

1. The "Most Likely" Answer (MAP): This is the single best guess, which matches the old MBAR method exactly.
2. The "Average" Answer (Posterior Mean): This is the average of all possible reasonable guesses.

The paper found that the "Average" answer is often slightly more accurate overall (less error), even though it might be slightly more biased in one direction. It's like averaging out a bunch of guesses to get a more stable result.

4. Why is this better?

The paper tested this on simple math problems (harmonic oscillators) and a real-world chemistry problem (how phenol dissolves in water).

• When data is plentiful: BayesMBAR acts just like the old MBAR. It converges to the same correct answer.
• When data is scarce (the "small sample" problem): This is where BayesMBAR shines.
  • It gives better uncertainty estimates. It doesn't lie to you about how sure it is. It tells you, "I'm not very sure," rather than pretending to be an expert.
  • It gives more accurate answers if you feed it the "smoothness" rule. It uses that rule to fill in the gaps where data is missing.

5. The Cost

The paper admits that BayesMBAR is a bit slower to run than the old MBAR. It has to do more heavy lifting (sampling from a complex distribution) to get that extra accuracy and better uncertainty estimates. However, the author argues that since the most expensive part of these calculations is actually generating the data (running the simulations), the extra time spent analyzing that data is a small price to pay for getting a more reliable result and a better sense of how much you can trust it.

Summary

BayesMBAR is a smarter version of a standard chemistry calculation tool.

• If you have lots of data, it works just like the old tool but tells you more honestly how confident it is.
• If you have very little data, it can use "rules of thumb" (like smoothness) to make better guesses and avoid wild errors.
• It's a tool for when you need to know not just what the answer is, but how much you can trust that answer.

Technical Summary

Technical Summary: Bayesian Multistate Bennett Acceptance Ratio Methods (BayesMBAR)

Problem Statement

Computing free energies of thermodynamic states is a fundamental challenge in computational chemistry and physics, with applications ranging from protein-ligand binding affinities to phase equilibria. The Multistate Bennett Acceptance Ratio (MBAR) method is a standard technique for estimating these free energies from sampled configurations. While MBAR is unbiased and has minimal variance when the number of configurations is large, its performance and uncertainty estimates are less explored in scenarios with small sample sizes. In such data-scarce regimes, the standard asymptotic analysis used by MBAR often yields inaccurate uncertainty estimates (typically overestimating them), and the method lacks a mechanism to incorporate prior physical knowledge (e.g., the smoothness of free energy surfaces) into the estimation process.

Methodology

The authors introduce BayesMBAR, a Bayesian generalization of the MBAR method. The development proceeds through the following steps:

1. Probabilistic Formulation: The authors reformulate MBAR using the reverse logistic regression model. In this framework, free energies ($F$) are treated as parameters within a likelihood function derived from retrospective conditional probabilities of state indices given configurations.
2. Bayesian Generalization: To create BayesMBAR, free energies are treated as random variables rather than fixed parameters. A prior distribution, $p(F; \theta)$, is placed over the free energies. The posterior distribution, $p(F|Y, X)$, is then computed using Bayes' theorem, combining the likelihood from the reverse logistic regression with the chosen prior.
3. Prior Distributions:
   • Uniform Prior: Used when no specific prior knowledge is available. This choice ensures that the Maximum A Posteriori (MAP) estimate of BayesMBAR recovers the standard MBAR estimate exactly.
   • Gaussian Prior: Used when prior knowledge about the system exists, specifically the smoothness of the free energy surface along collective coordinates. The authors employ a Gaussian Process prior, which, when projected onto discrete states, results in a multivariate Gaussian distribution. The covariance function (e.g., squared exponential) encodes the assumption that free energies at nearby collective coordinates are correlated.
4. Inference and Optimization:
   • Point Estimates: The MAP estimate is found by maximizing the posterior density (using L-BFGS-B or Newton's method). The posterior mean is also computed as an alternative point estimate.
   • Uncertainty Quantification: Uncertainty is derived from the posterior covariance matrix. For systems with more than two states, where analytical integration is infeasible, the authors use the No-U-Turn Sampler (NUTS), a variant of Hamiltonian Monte Carlo, to sample from the posterior distribution.
   • Hyperparameter Optimization: Hyperparameters of the prior (e.g., length scales and variance) are automatically optimized by maximizing the Bayesian evidence (marginal likelihood). This is achieved using a variational inference approach with an Evidence Lower Bound (ELBO) and a Gaussian proposal distribution.

Key Contributions

• BayesMBAR Framework: The development of a rigorous Bayesian framework for free energy estimation that generalizes MBAR.
• Improved Uncertainty Estimates: The method provides posterior-based uncertainty estimates that are shown to be more accurate than standard asymptotic analysis, particularly in low-data regimes where asymptotic methods tend to significantly overestimate uncertainty.
• Incorporation of Prior Knowledge: The ability to integrate physical priors, such as the smoothness of free energy surfaces, directly into the estimation procedure. This leads to more accurate free energy estimates when data is limited.
• Dual Estimators: The introduction of both MAP and posterior mean estimators, with the latter offering a trade-off between bias and variance that can result in lower Root Mean Squared Error (RMSE) in certain small-sample scenarios.

Results

The authors validated BayesMBAR using three benchmark systems:

1. Two Harmonic Oscillators:

   • BayesMBAR with a uniform prior recovered the MBAR (BAR) estimate as the MAP.
   • The posterior mean estimate exhibited lower RMSE than the MAP estimate due to a reduction in standard deviation (SD), despite a slight increase in bias.
   • Uncertainty estimates from BayesMBAR were significantly more accurate than those from asymptotic analysis (which overestimated) and the bootstrap method (which underestimated) for small sample sizes ($n < 100$).

2. Three Harmonic Oscillators:

   • Similar trends were observed in this multistate system. The posterior mean estimate showed lower RMSE than the MBAR estimate for small sample sizes.
   • BayesMBAR's uncertainty estimates avoided the underestimation seen in bootstrap methods and the excessive overestimation of asymptotic analysis.

3. Hydration Free Energy of Phenol:

   • Uniform Prior: When using a uniform prior, BayesMBAR matched MBAR performance in terms of RMSE for large datasets but provided superior uncertainty estimates for small datasets ($n=5$).
   • Normal Prior: By incorporating a Gaussian prior encoding the smoothness of the free energy surface along alchemical variables, BayesMBAR achieved significantly lower RMSE than MBAR when the number of configurations was small ($n < 100$). As the sample size increased, the BayesMBAR estimates converged to the MBAR results, demonstrating that the prior acts as a regularizer when data is insufficient but does not bias the result when data is abundant.

Significance and Claims

The paper posits that BayesMBAR is an essential tool for free energy calculations, particularly in scenarios where:

1. Data is scarce: It provides more reliable uncertainty estimates than standard MBAR, preventing premature termination of sampling (due to underestimation) or unnecessary oversampling (due to overestimation).
2. Prior knowledge is available: It offers a systematic way to incorporate physical constraints (like surface smoothness) or results from cheaper calculations (e.g., docking, MM/GBSA) to improve accuracy without sacrificing convergence to the true value as data volume increases.

The authors acknowledge that BayesMBAR is computationally more expensive than MBAR due to the need for sampling from the posterior distribution. However, they argue that this cost is justified given the improved accuracy of both free energy estimates and uncertainty quantification, especially since the majority of computational cost in free energy calculations typically lies in the initial sampling of configurations rather than the post-processing analysis. The authors have released an open-source Python package to facilitate adoption.