[Paper Review] Efficient Cysteine Conformer Search with Bayesian Optimization

Lincan Fang, Esko Makkonen+3 more|arXiv (Cornell University)|Jun 26, 2020

Computational Drug Discovery Methods43 references4 citations

TL;DR

This paper presents a Bayesian optimization-based active learning method, BOSS, that efficiently identifies low-energy conformers of cysteine using only 1,000 single-point DFT calculations and ~30 geometry optimizations—less than 10% of the computational cost of existing methods—while accurately predicting conformer energies and structures in agreement with experiment and high-level quantum chemistry benchmarks.

ABSTRACT

Finding low-energy molecular conformers is challenging due to the high dimensionality of the search space and the computational cost of accurate quantum chemical methods for determining conformer structures and energies. Here, we combine active-learning Bayesian optimization (BO) algorithms with quantum chemistry methods to address this challenge. Using cysteine as an example, we show that our procedure is both efficient and accurate. After only one thousand single-point calculations and approximately thirty structure relaxations, which is less than 10% computational cost of the current fastest method, we have found the low-energy conformers in good agreement with experimental measurements and reference calculations.

Motivation & Objective

• Address the challenge of high-dimensional conformational space and high computational cost in molecular conformer searches.
• Overcome limitations of systematic and stochastic methods in sampling efficiency and result consistency.
• Reduce the computational burden of quantum chemistry-based conformer ranking by minimizing expensive DFT calculations.
• Develop a method that identifies all relevant low-energy conformers in a single run, including accurate energy ordering.
• Enable accurate prediction of conformer stability and energy landscapes with minimal data points using active learning.

Proposed method

• Employ Bayesian optimization (BO) as an active learning strategy to sequentially select the most informative conformer configurations for DFT energy evaluation.
• Use a 5-dimensional dihedral angle space (d1–d5) to define the conformational search space of cysteine.
• Leverage a Gaussian process surrogate model to predict the potential energy surface (PES) from sparse DFT single-point energy evaluations.
• Apply DFT geometry optimization only to the most promising conformers identified by the BO algorithm, reducing total optimization steps.
• Use the BOSS tool to extract local minima from the learned PES without additional computational cost.
• Refine selected conformers with DFT, vibrational frequency corrections, and CCSD(T) benchmarking for energy accuracy.

Experimental results

Research questions

• RQ1 Can Bayesian optimization significantly reduce the number of DFT calculations required to identify low-energy conformers of cysteine?
• RQ2 How accurately can the BOSS method reproduce experimental and high-level quantum chemical conformer energies and structures?
• RQ3 To what extent does the method outperform traditional stochastic or systematic conformer search approaches in computational efficiency?
• RQ4 Can the method reliably identify all relevant low-energy conformers in a single run without prior knowledge of the energy landscape?
• RQ5 How do different DFT functionals (PBE vs. PBE0) and basis sets affect the accuracy of the predicted conformer energies and relative ordering?

Key findings

• The method identified all experimentally observed low-energy conformers of cysteine within the first 1,000 DFT single-point calculations and ~30 geometry optimizations.
• The total computational cost was equivalent to only ~60 DFT geometry optimizations—less than 10% of the cost of the fastest existing method (e.g., genetic algorithms requiring 20,000–60,000 evaluations).
• The predicted relative energies of the lowest-lying conformers (e.g., IIb, Ib, Ia) were in excellent agreement with CCSD(T) benchmarks, with deviations of only 0.01–0.02 eV.
• DFT geometry relaxation was found to play a critical role in refining the correct energy ordering of conformers, especially for closely spaced minima.
• PBE0 functional yielded slightly better agreement with high-level benchmarks than PBE, though differences in dispersion corrections (TS vs. MBD) were negligible for this system.
• The method successfully reconstructed a physically meaningful 5D potential energy surface with minimal data, enabling full landscape analysis and future barrier calculations.