High-Quality Conformer Generation with CONFORGE: Algorithm and Performance Assessment

Thomas Seidel, Christian Permann, Oliver Wieder, Stefan M. Kohlbacher, Thierry LangerJournal of Chemical Information and Modeling 2023 · DOI · OpenReview

TL;DR

Researchers have developed CONFORGE, a new open-source tool for generating accurate 3D molecular conformations. CONFORGE outperforms existing open-source tools and matches or exceeds commercial alternatives in both speed and accuracy, particularly for challenging macrocyclic molecules. This advancement is crucial for drug discovery, enabling more reliable predictions of how molecules will interact with biological targets.

What Problem Does This Paper Address?

Predicting the precise 3D shape (conformation) of a molecule is a fundamental step in computer-aided drug design. This is because a molecule's shape dictates how it binds to its target in the body. However, experimentally determining the "bioactive" conformation is difficult and only possible for a small fraction of molecules. Therefore, computational methods are needed to predict these conformations. The challenge lies in efficiently and accurately exploring the vast number of possible shapes a flexible molecule can adopt and identifying the low-energy, biologically relevant ones. While many tools exist, open-source options have historically lagged behind commercial ones in performance. This paper introduces CONFORGE to bridge this gap, aiming to provide state-of-the-art conformer generation capabilities for free.

Methods

CONFORGE employs a hybrid approach to conformer generation:

  • For small molecules: It uses a knowledge-based systematic approach. This involves leveraging pre-generated libraries of molecular fragments and torsion angles derived from experimental 3D structures. This method is efficient for molecules with rotatable bonds and relatively rigid ring systems.
  • For macrocyclic molecules: It utilizes a purely stochastic approach. This involves distance geometry (DG) to generate initial structures, followed by force-field-driven refinement. This method is better suited for the complex, flexible ring systems found in macrocycles.

CONFORGE can automatically detect macrocyclic structures and select the appropriate sampling method, or users can specify it manually. It also handles multi-component molecules (like salts) by generating conformers for each component separately and then combining them.

Datasets

The study used two primary datasets for evaluation:

  • Platinum Diverse Dataset: This dataset consists of 2859 high-quality protein-bound ligand conformations extracted from the Protein Data Bank (PDB). It's specifically designed for benchmarking conformer generators.
  • Prime Macrocycle Dataset: This dataset comprises 208 macrocyclic structures from the Cambridge Structural Database (CSD), PDB, and the Biologically Interesting Molecule Reference Dictionary (BIRD). It includes diverse and challenging macrocyclic structures.

Key Insights and Ideas

  • Bridging the Open-Source Gap: CONFORGE demonstrates that open-source tools can now compete with, and in some cases surpass, commercial software for conformer generation, a critical step in drug discovery.
  • Hybrid Approach is Key: The success of CONFORGE lies in its ability to intelligently switch between systematic and stochastic sampling methods, tailoring the approach to the molecule's complexity.
  • Macrocycle Performance: The tool shows particular strength in handling macrocyclic molecules, a class of compounds gaining increasing importance in drug development due to their unique properties.
  • Efficiency and Accuracy Synergy: CONFORGE achieves a remarkable balance between computational speed and the accuracy of generated conformers, a difficult trade-off to master.

Key Results

  • Small Molecules: CONFORGE outperformed all other tested open-source conformer generators and performed comparably to or better than commercial generators in terms of both speed and accuracy. For example, it achieved a median accuracy of 0.49 Å (for an ensemble size of 50) on the Platinum Diverse Dataset, which was the best among all tested generators. It also boasted the lowest median processing time per molecule.
  • Macrocycles: CONFORGE achieved the best average accuracy among all benchmarked generators on the Prime Macrocycle Dataset, with a mean accuracy of 1.11 Å. While its processing time was higher for macrocycles, the improved accuracy was deemed to outweigh this.
  • Robustness: CONFORGE was able to process all molecules in the tested datasets without failures, unlike some other generators.

Limitations and Future Work

  • Macrocycle Speed: While CONFORGE excels in macrocycle accuracy, its processing time for these complex molecules is higher than some competitors. Future work aims to optimize this speed.
  • Force Field Flexibility: Currently, CONFORGE primarily uses the MMFF94 force field. Future development plans include supporting other general-purpose force fields, such as the Open Force Field.
  • Targeted Sampling: The authors plan to implement functionality that allows users to restrict conformer sampling to specific parts of a molecule.

Summary updated

Abstract

Click to copy section linkSection link copied!
BibTeX
@article{seidel2023high,
  author = {Thomas Seidel and Christian Permann and Oliver Wieder and Stefan M. Kohlbacher and Thierry Langer},
  title = {{High-Quality Conformer Generation with CONFORGE: Algorithm and Performance Assessment}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2023},
  doi = {10.1021/acs.jcim.3c00563},
}
Figure