Bayesian optimization for conformer generation

Lucian Chan, Geoffrey Hutchison, Garrett M. Morris

Abstract

Generating low-energy molecular conformers is a key task for many areas of computational chemistry, molecular modeling, and cheminformatics. Most current conformer generation methods primarily focus on generating geometrically diverse conformers rather than finding the most probable or energetically lowest minima. Here, we present a new stochastic search method called the Bayesian optimization algorithm (BOA) for finding the lowest energy conformation of a given molecule. We compare BOA with uniform random search and systematic search as implemented in Confab to determine which method finds the lowest energy. Energetic difference, root-mean-square deviation, and torsion fingerprint deviation are used to quantify the performance of the conformer search algorithms. In general, we find BOA requires far fewer evaluations than systematic or uniform random search to find low-energy minima. For molecules with four or more rotatable bonds, Confab typically evaluates $10^4$ (median) conformers in its search, while BOA only requires $10^2$ energy evaluations to find top candidates. Despite evaluating fewer conformers, 20-40% of the time BOA finds lower-energy conformations than a systematic Confab search for molecules with four or more rotatable bonds.

Introduction

Small molecules are inherently flexible, adopting various 3D shapes known as conformers by rotating around their chemical bonds. Identifying the "global minimum"—the shape with the lowest potential energy—is a fundamental challenge in drug discovery and materials science. A molecule's biological activity and interaction with targets are heavily dependent on its 3D conformation.

However, finding this minimum is a "needle-in-a-haystack" problem. As the number of rotatable bonds increases, the search space grows exponentially—a phenomenon known as the combinatorial explosion. Traditional systematic searches, which sample at fixed intervals, are computationally intensive and often miss the true minimum. Conversely, stochastic methods like random sampling often fail to explore high-dimensional spaces effectively.

Methodology: The Bayesian Optimization Algorithm (BOA)

To address these limitations, the Bayesian Optimization Algorithm (BOA) treats the molecular energy landscape as a "black box" function. Instead of exhaustive enumeration, BOA builds a probabilistic surrogate model—specifically a Gaussian Process (GP)—to approximate the energy landscape.

Technical Components

  • Locally Periodic Kernels: Torsion angles are periodic ($0$ to $2\pi$). BOA employs a specialized kernel combining periodic and squared-exponential functions, ensuring the model recognizes that $0^\circ$ and $360^\circ$ represent the same physical state.
  • Acquisition Functions: Strategies such as Expected Improvement (EI) and Lower Confidence Bound (LCB) guide the search. These functions balance "exploration" of unknown regions with "exploitation" of known low-energy areas.
  • Energy Evaluation: The Merck Molecular Force Field (MMFF94) provides the "ground truth" energy values used to train the surrogate model.
The iterative process of BOA learning the energy landscape of Biphenyl. (a) Biphenyl structure; (b) True energy landscape; (c-f) Successive iterations showing the GP model (blue) and acquisition function (red) converging on the global minimum.

Experimental Setup

The algorithm was evaluated on a benchmark dataset of 576 distinct small molecules from the Astex diverse set. To ensure a fair comparison with systematic search methods, the study focused on molecules with six or fewer rotatable bonds, where systematic enumeration is still feasible.

Key Results and Discussion

BOA demonstrates significant advantages in both efficiency and accuracy:

  • Efficiency: For molecules with four or more rotatable bonds, BOA requires approximately 100 energy evaluations, compared to over 10,000 for systematic searches. This represents a 100x speedup in higher-dimensional spaces.
  • Accuracy: BOA found lower-energy conformations than the systematic tool Confab 20–40% of the time for molecules with 4+ rotatable bonds. This is because BOA samples the torsion space continuously, allowing it to find minima that fall between the fixed grid points of systematic methods.
  • Reliability: BOA consistently outperformed uniform random search, maintaining high performance as molecular complexity (number of rotatable bonds) increased.

Conclusion and Future Directions

Bayesian Optimization provides a powerful framework for efficient conformer generation. While the current implementation uses Gaussian Processes, which have an $O(N^3)$ computational complexity, future work could explore Deep Neural Networks as surrogate models to handle larger molecules and more evaluations.

The true potential of BOA lies in its application to expensive Quantum Mechanical (QM) calculations. By minimizing the number of required evaluations, BOA could make high-accuracy QM-based conformer searches practical for drug discovery pipelines.

Summary updated

Abstract

Generating low-energy molecular conformers is a key task for many areas of computational chemistry, molecular modeling and cheminformatics. Most current conformer generation methods primarily focus on generating geometrically diverse conformers rather than finding the most probable or energetically lowest minima. Here, we present a new stochastic search method called the Bayesian optimization algorithm (BOA) for finding the lowest energy conformation of a given molecule. We compare BOA with uniform random search, and systematic search as implemented in Confab, to determine which method finds the lowest energy. Energetic difference, root-mean-square deviation, and torsion fingerprint deviation are used to quantify the performance of the conformer search algorithms. In general, we find BOA requires far fewer evaluations than systematic or uniform random search to find low-energy minima. For molecules with four or more rotatable bonds, Confab typically evaluates [Formula: see text] (median) conformers in its search, while BOA only requires [Formula: see text] energy evaluations to find top candidates. Despite using evaluating fewer conformers, 20-40% of the time BOA finds lower-energy conformations than a systematic Confab search for molecules with four or more rotatable bonds.
BibTeX
@article{chan2019bayesian,
  author = {Lucian Chan and Geoffrey Hutchison and Garrett M. Morris},
  title = {{Bayesian optimization for conformer generation}},
  journal = {Journal of Cheminformatics},
  year = {2019},
  doi = {10.1186/s13321-019-0354-7},
}
Figure