Predicting binding poses and affinities for protein - ligand complexes in the 2015 D3R Grand Challenge using a physical model with a statistical parameter estimation

Sergei Grudinin, Maria Kadukova, Andreas Eisenbarth, Simon Marillet, Frédéric CazalsJournal of Computer-Aided Molecular Design 2016 · DOI

TL;DR

This paper presents a novel physical model for predicting protein-ligand binding affinities, combining enthalpic and entropic contributions. While the model showed promise on training data, it struggled with blind test sets, highlighting challenges in generalization and the need for better representation of dynamic properties and broader training data coverage.

What Problem Does This Paper Address?

Predicting how strongly a small molecule (ligand) will bind to a protein is a fundamental challenge in drug discovery. Accurate prediction of both the pose (how the ligand sits in the protein's binding pocket) and the affinity (the strength of binding) is crucial for designing effective drugs. The D3R Grand Challenge provides a benchmark for computational methods to test their ability to predict these properties on unseen data. This paper aims to address the difficulty of accurately predicting binding affinities, particularly for diverse chemical series and protein targets, by developing a physically-motivated model and rigorously testing it on the D3R 2015 challenge datasets.

Methods

The authors developed a binding affinity model based on a physical approximation of the binding free energy ($\Delta G_{bind} = \Delta H - T\Delta S$).

  • Enthalpic Contribution ($\Delta H$): This was modeled as a linear functional of pairwise, distance-dependent interactions between atom types in the protein and ligand. They used a knowledge-based approach, where the parameters (scoring potentials) were learned from experimental data. The model defines 48 atom types based on chemical properties and excludes explicit hydrogen atoms.
  • Entropic Contribution ($-T\Delta S$): This was approximated using a set of geometric descriptors that capture conformational, vibrational, and solvent entropy. These descriptors include measures like Inverse Volume-Weighted Internal Path Length (IVW-IPL), non-interacting surface properties (NIS), atomic solvation energy (ATOM_SOLV), and polar surface area (POLAR_SASA).
  • Parameter Estimation: Both the enthalpic and entropic model parameters were trained using regularized regression (Linear Ridge Regression - LRR, and Kernel Ridge Regression - KRR) on the PDBBind database, a large collection of protein-ligand complexes with experimental binding data.

For pose prediction, they used AutoDock Vina, generating multiple docking solutions and then re-scoring them with their developed Convex-PL potential.

Datasets

  • Training Data: The PDBBind database (refined set, 3,706 complexes) was used to train the binding affinity model. This dataset contains experimentally determined binding affinities (Ka, Ki, IC50) and corresponding protein-ligand crystal structures.
  • Test Data: The 2015 D3R Grand Challenge datasets were used for evaluation:
  • HSP90 dataset: 180 ligands with experimental IC50 values and 8 co-crystal structures.
  • MAP4K dataset: 30 ligands with experimental IC50/Ki values and 18 co-crystal structures.

Key Insights and Ideas

  • Physical Model for Affinity: The core idea is to combine physically motivated enthalpic and entropic terms, rather than relying solely on empirical scoring functions. This aims to provide a more robust and generalizable prediction.
  • Importance of Entropic Descriptors: The study highlights the value of geometric descriptors for capturing entropic contributions, which are often overlooked or simplified in other models.
  • Challenges in Generalization: The significant drop in performance from the training set to the blind test sets (especially for HSP90) underscores the difficulty of building models that generalize well across diverse chemical spaces and protein targets.
  • Impact of Training Data Coverage: The authors suggest that the limited coverage of the test set by the training set in the high-dimensional enthalpic space is a major reason for the poor performance on unseen data.
  • Re-scoring Improves Pose Prediction: Re-scoring AutoDock Vina poses with the Convex-PL potential significantly improved the quality of the best predicted poses, suggesting that while AutoDock Vina's default scoring is simplistic, a more sophisticated approach can refine its initial predictions.
  • Template Selection is Crucial: For pose prediction, the choice of the rigid receptor template and the size of the binding pocket search space critically impacts the accuracy of the predicted poses.

Key Results

  • Pose Prediction:
    • For the HSP90 dataset, their best pose predictions achieved a mean RMSD of 0.65 Å, ranking well.
  • For the MAP4K dataset, pose prediction was more challenging, with a mean RMSD of 3.52 Å, partly due to sampling issues with larger ligands and template selection.
  • Affinity Prediction (Training Set):
    • The model achieved good accuracy on the training set, with Pearson correlation coefficients of 0.87 for the linear model (LRR) and 0.92 for the non-linear model (KRR).
  • Affinity Prediction (Test Sets):
    • Performance on the blind test sets was significantly lower than on the training set. For MAP4K, Spearman correlation coefficients were around 0.3-0.4. For HSP90, they were even lower (around 0.05-0.15).
  • The authors noted that the entropic term contributed more to the linear model's accuracy, while the non-linear model could compensate for its absence.

Limitations and Future Work

  • Overfitting: The large number of parameters in the enthalpic model may lead to overfitting of the training data.
  • Test Set Coverage: The limited representation of test set cases within the training data is a significant limitation for generalization.
  • Dynamic Properties: The model relies on static crystal structures, potentially missing crucial dynamic aspects of protein-ligand interactions.
  • Interpretability of IC50/Ki: The authors speculate that the conversion of IC50/Ki to Ka might introduce inaccuracies, as these measures are not always directly comparable to binding free energy.

Future Work: * Explore regression methods more tolerant to sparse training data. * Investigate ways to better represent conformational entropy and other dynamic properties. * Develop methods that can reliably estimate affinity from static poses and, more generally, from conformational ensembles. * Improve sampling procedures to better explore the energy landscape and conformational space.

Summary updated

BibTeX
@article{grudinin2016predicting,
  author = {Sergei Grudinin and Maria Kadukova and Andreas Eisenbarth and Simon Marillet and Frédéric Cazals},
  title = {{Predicting binding poses and affinities for protein - ligand complexes in the 2015 D3R Grand Challenge using a physical model with a statistical parameter estimation}},
  journal = {Journal of Computer-Aided Molecular Design},
  year = {2016},
  doi = {10.1007/s10822-016-9976-2},
}
Figure