TL;DR
This paper introduces GRADE and X-GRADE, novel protein-ligand interaction fingerprints (IFPs) that quantify nonbonding interactions using floating-point scores, capturing both presence and quality. These new IFPs demonstrate competitive performance in dimensionality reduction, binding affinity prediction, and 3D-QSAR modeling, offering a valuable new tool for computational drug design.
What Problem Does This Paper Address?
Accurately representing the complex nonbonding interactions between proteins and ligands is crucial for computational drug design methods like docking and binding affinity prediction. While existing interaction fingerprints (IFPs) exist, they often rely on simple bit strings or integer counts, failing to fully capture the nuanced "quality" of these interactions (e.g., specific distance and angle requirements). This limitation can hinder the performance of machine learning models that rely on these descriptors. The authors aim to address this by developing a more sophisticated IFP that encodes both the presence and quality of interactions using continuous values.
Methods
The core innovation is the development of GRADE and X-GRADE, two versions of a novel IFP. These descriptors are built upon GRAIL scores, which are derived from a pharmacophoric representation of both the ligand and receptor.
- GRADE: A 35-element descriptor.
- X-GRADE: An extended 177-element descriptor offering finer granularity.
Both GRADE and X-GRADE consist of two parts:
- Pose-Independent Ligand Features: These capture general ligand properties relevant for binding, such as physicochemical properties (TPSA, XlogP), molecular graph properties, and counts of pharmacophoric features (H-bond acceptors/donors, aromatic rings, etc.). X-GRADE provides a more detailed breakdown of HBA/HBD features based on Sybyl atom types.
- Pose-Dependent Features: These quantify observed nonbonding interactions between the ligand and the binding site. This includes HBA/HBD coverage, GRAIL-based feature interaction scores, and physics-based electrostatic and van der Waals contributions. X-GRADE again offers a more detailed representation of HBA/HBD interactions.
The calculation of these features involves:
- Electrostatic Potential and Pairwise Forces: Calculated using partial charges derived from the MMFF94 force field.
- Van der Waals Energies: Modeled using a Morse potential, which is less sensitive to deviations from ideal distances than the Lennard-Jones potential.
- GRAIL Interaction Scores: Quantify nonbonding interactions between complementary pharmacophoric features, considering distance and angle dependencies.
- Binding Site HBA/HBD Atom Coverage: Assesses how ligand atoms block potential H-bonds with the environment.
Datasets
The authors utilized several established datasets for their evaluations:
- PDBbind v.2020: Used for Case Studies 1 (dimensionality reduction) and 2 (binding affinity estimation). The dataset was split into refined, core, and general sets for training, validation, and testing.
- PL-REX dataset: An external, high-quality dataset used for Case Study 2 to further validate binding affinity prediction models.
- Proprietary insecticide dataset (BASF SE): Used for Case Study 3 (3D-QSAR).
- Cathepsin S dataset (Janssen): A publicly available dataset also used for Case Study 3 (3D-QSAR).
Key Insights and Ideas
- Beyond Binary: Quantifying Interaction Quality: The most significant insight is the move from binary (present/absent) or simple count-based IFPs to continuous scores that capture the "quality" of interactions. This richer representation can provide more nuanced information to machine learning models.
- X-GRADE's Enhanced Detail: The extended X-GRADE descriptor, with its finer granularity in pharmacophoric feature representation, generally showed comparable or slightly improved performance, suggesting that more detailed interaction information can be beneficial.
- Competitive Performance Across Applications: GRADE and X-GRADE demonstrated strong performance across diverse tasks, including visualizing chemical space, predicting binding affinities, and building 3D-QSAR models, often matching or exceeding established methods.
- Speed and Interpretability: A key advantage highlighted is the computational efficiency of GRADE/X-GRADE, especially when compared to some other advanced methods. Furthermore, the authors suggest that the descriptors offer better interpretability than some black-box models.
- Generalizability: The performance on external test sets, particularly in 3D-QSAR, suggests that GRADE features can improve model generalizability and potentially mitigate overfitting.
Key Results
- Dimensionality Reduction (Case Study 1): GRADE and X-GRADE showed competitive performance with established descriptors (AvinaXGB) in visualizing protein-ligand complexes in chemical space using UMAP.
- Binding Affinity Estimation (Case Study 2):
- On the PDBbind dataset, GRADE and X-GRADE models achieved reasonable accuracy, with Random Forest models performing best.
- When compared to existing scoring functions, GRADE/X-GRADE models were significantly faster and performed comparably or better than most, with only one competitor (AvinaRF20) showing higher accuracy.
- On the high-quality PL-REX dataset, GRADE and X-GRADE models showed strong performance, with Pearson R values ranging from 0.92 to 0.36.
- 3D-QSAR (Case Study 3):
- On proprietary insecticide data, adding GRADE features to Morgan Fingerprints (MFP) significantly improved predictive power on external test sets, outperforming plain MFPs.
- On the Cathepsin S dataset, GRADE features also boosted predictive power on external test sets, with a GRADE-based RF model achieving a notable R² of 0.431 even without Morgan fingerprints.
- For the small PL-REX CDK2 dataset, a dedicated 3D-QSAR model was not feasible, and a generic scoring function trained on PDBbind outperformed it.
- Computational Speed: GRADE and X-GRADE were found to be comparably fast to other descriptors, and significantly faster than some advanced methods, especially for larger datasets.
Limitations and Future Work
- Limitations:
- The performance of 3D-QSAR models was highly dependent on the size of the training dataset, with small datasets (like PL-REX CDK2) proving insufficient for robust models.
- While GRADE/X-GRADE showed good performance, some models (e.g., AvinaRF20) still outperformed them in accuracy for binding affinity estimation.
- The authors note that for specific use cases like docking pose postscoring, hyperparameter optimization and feature selection might be necessary to maximize performance.
- Future Work:
- The authors plan to utilize precalculated GRAIL maps for more efficient descriptor generation.
- Leveraging GPUs for calculations is suggested to further reduce descriptor computation times.
- The authors believe GRADE/X-GRADE can be a valuable contribution to IFP methods and beneficial for downstream IFP applications.
Summary updated