Conformer Generation for Structure-Based Drug Design: How Many and How Good?

Andrew T. McNutt, Fatimah Bisiriyu, Sophia Song, Ananya Vyas, Geoffrey R. Hutchison, David Ryan Koes

TL;DR

This paper investigates how the number and quality of molecular conformers generated impact structure-based drug discovery tasks like pharmacophore screening and molecular docking. They found that while larger ensembles generally improve results, there are diminishing returns, and filtering for diversity is crucial, often outperforming simply selecting low-energy conformers.

What Problem Does This Paper Address?

Structure-based drug design relies heavily on understanding how molecules interact with biological targets. A critical first step is generating realistic 3D shapes (conformations) of drug candidates. However, it's not always clear how many conformers are needed or what makes a "good" set of conformers for downstream tasks like virtual screening or docking. This paper aims to empirically determine the optimal size and quality of these conformational ensembles to maximize performance in these key drug discovery applications.

Methods

The authors compared two conformer generation approaches:

  • RDKit ETKDGv3: A well-established, geometry-based method that combines distance geometry with knowledge-based torsional preferences.
  • DMCG (Direct Molecular Conformation Generation): A state-of-the-art deep generative model trained on a large dataset of molecular conformations.

They explored the impact of:

  • Ensemble Size: Varying the number of conformers generated.
  • Energy Minimization: Applying the Universal Force Field (UFF) to refine generated conformers.
  • Ensemble Construction: Different strategies for selecting conformers from a larger pool, including:
  • Random sampling
    • Selecting the lowest energy conformers
    • Filtering by root-mean-square deviation (RMSD) to ensure structural diversity.

The performance of these ensembles was evaluated in two primary tasks:

  • Pharmacophore Search: Using the DUDE benchmark to assess how well ensembles enable the identification of active compounds.
  • Molecular Docking: Using the PDBbind and Wierbowski datasets to evaluate the ability to reproduce known binding poses.

Datasets

The study utilized the following datasets:

  • Platinum 2017: A dataset specifically designed for evaluating conformer generation, focusing on the quality of fit to electron density maps.
  • PDBbind 2020 Refined: A curated dataset of high-quality protein-ligand structures with known binding affinities, used for both bioactive conformation retrieval and redocking.
  • DUDE (Directory of Useful Decoys, Enhanced): A benchmark for virtual screening, containing active compounds and decoys for various protein targets.
  • Wierbowski et al. dataset: Used for cross-docking evaluations.

Key Insights and Ideas

  • Diversity Trumps Energy Alone: Simply selecting the lowest energy conformers from a large ensemble is often insufficient for finding bioactive conformations. The energy of an isolated molecule doesn't fully capture interactions with a protein.
  • RMSD Filtering is Key: Imposing an RMSD cutoff when constructing smaller ensembles from larger ones significantly improves performance, especially for tasks like pharmacophore screening. This ensures structural diversity within the ensemble.
  • Diminishing Returns for Ensemble Size: While larger ensembles generally increase the chance of finding a bioactive conformation, the improvement plateaus relatively quickly. For many tasks, ensembles of 25 conformers or fewer are sufficient.
  • Larger Ensembles Can Hurt Screening: In pharmacophore screening, increasing ensemble size without proper filtering can lead to more false positives, ultimately reducing performance.
  • Energy Minimization is Beneficial: Applying energy minimization (e.g., with UFF) generally improves conformer quality, particularly for deep learning models like DMCG, which may generate less physically realistic initial geometries.
  • Classical Methods Still Hold Up: Despite the rise of deep learning, the established RDKit ETKDGv3 method often outperformed the deep generative DMCG model in recovering bioactive conformations and for docking tasks. This suggests that current deep learning models might be trained on metrics that don't perfectly align with practical drug discovery needs.
  • Docking Benefits from Ensembles: Even for docking methods that treat ligands as partially flexible, providing an ensemble of conformers (even just five) can improve performance over a single conformer, highlighting the importance of exploring non-torsional degrees of freedom.

Key Results

  • Bioactive Conformer Retrieval: RDKit generally outperformed DMCG in retrieving experimental crystal structures within a given RMSD threshold. Energy minimization provided a small but consistent benefit.
  • Pharmacophore Screening: For smaller ensembles (<10 conformers), RMSD-filtered ensembles performed best. For larger ensembles, performance varied significantly by target, with many benefiting from 25 or fewer conformers. Filtering by RMSD was crucial for achieving high F1 scores.
  • Molecular Docking: Providing an ensemble of conformers (even just 5) improved docking performance (fraction of systems with <2 Å RMSD pose) compared to a single conformer for both RDKit and DMCG. RDKit consistently showed better performance than DMCG.

Limitations and Future Work

  • Deep Learning Training Metrics: The authors speculate that DMCG's performance gap with RDKit might stem from its training on metrics (like GEOM coverage) that don't perfectly align with the goal of recovering bioactive conformations. Future work could involve adapting deep learning models to be trained on more relevant metrics.
  • Computational Cost: While larger ensembles can improve results, the computational cost of generation and screening increases. The paper suggests that smaller, well-constructed ensembles are often sufficient and more efficient.
  • Generalizability of DMCG: The study focused on one specific deep generative model (DMCG). Further evaluation of other deep learning approaches is warranted.
  • Future Work: The authors suggest that adapting deep learning models to better capture the nuances of protein-ligand interactions and exploring more sophisticated ensemble construction methods could be fruitful avenues for future research. They also provide scripts for reproducing their analysis, encouraging further community investigation.

Summary updated

Abstract

Conformer generation, the assignment of realistic 3D coordinates to a small molecule, is fundamental to structure based drug design. Conformational ensembles are required for rigid-body matching algorithms, such as shape-based or pharmacophore approaches, and even methods that treat the ligand flexibly, such as docking, are dependent on the quality of the provided conformations due to not sampling all degrees of freedom (e.g. only sampling torsions). Our goal here is not to comprehensively evaluate the expansive suite of available conformer generation tools, but to empirically elucidate some general principles about the size, diversity and quality of conformational ensembles needed to get the best performance in common structure-based drug discovery tasks. In many cases our findings may parallel ``common knowledge'' well-known to practitioners of the field. Nonetheless, we feel it is valuable to quantify these conformational effects while reproducing and expanding upon previous studies. Specifically, we investigate the performance of a state-of-the-art generative deep learning approach versus a more classical geometry based approach, the effect of energy minimization as a post-processing step, the effect of ensemble size (maximum number of conformers), and construction (filtering by RMSD for diversity) and how these choices influence the ability to recapitulate bioactive conformations and perform pharmacophore screening and molecular docking.
BibTeX
@article{mcnutt2023conformer,
  author = {Andrew T. McNutt and Fatimah Bisiriyu and Sophia Song and Ananya Vyas and Geoffrey R. Hutchison and David Ryan Koes},
  title = {{Conformer Generation for Structure-Based Drug Design: How Many and How Good?}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2023},
  url = {https://openreview.net/forum?id=8uTMBPxfCw},
}
Figure