TL;DR
This paper introduces ComBind, a novel method that significantly improves the prediction of how drug-like molecules (ligands) bind to target proteins. By combining traditional physics-based simulations with information from other known binders, ComBind achieves higher accuracy than existing methods, even when structural data for those other binders is unavailable.
What Problem Does This Paper Address?
Designing new drugs relies heavily on understanding how small molecules (ligands) interact with biological targets (proteins). A critical aspect of this is predicting the precise 3D orientation of a ligand when it's bound to its target protein – known as the "binding pose." Current computational methods fall into two main camps:
- Physics-based methods: These use the 3D structure of the target protein and simulate the physical interactions between the protein and ligand. They are powerful but can be computationally expensive and struggle with ligands that are structurally different from those for which experimental structures exist.
- Ligand-based methods: These rely on experimental data (like binding affinity) for a series of similar ligands and use pattern matching to predict properties for new ligands. They don't require a protein structure but are limited to predicting properties similar to those already measured.
The challenge is to bridge these two approaches, leveraging the strengths of both to achieve more accurate and robust predictions, especially for ligands that are structurally diverse or when experimental structural data is scarce.
Methods
ComBind introduces a statistical framework that integrates physics-based and ligand-based information. The core idea is to predict the binding poses of multiple ligands to a target protein simultaneously, rather than one by one.
- Candidate Pose Generation: For each ligand of interest and a set of "helper" ligands (known binders to the same target), standard docking software (like Glide) is used to generate many potential binding poses and assign a score to each.
- ComBind Scoring Function: This is the novel part. It's a statistical potential that considers two factors:
- Per-ligand favorability: How good is each individual pose according to the standard docking score (Edock).
- Likelihood of similarity: How likely is it that a set of poses (one for each ligand) will be correct, based on observed similarities between known correct poses versus randomly chosen poses. This similarity is quantified using metrics like hydrogen bond patterns, salt bridges, hydrophobic contacts, and shared substructure positions.
- Optimization: An algorithm (coordinate descent) is used to find the set of poses that minimizes the ComBind score, effectively finding the most probable collective binding configuration.
Datasets
The study utilizes several datasets:
- Protein-Ligand Complex Structures: A curated set of 385 protein-ligand complex structures from 28 different drug targets was used to learn the similarity statistics (reference and native distributions).
- ChEMBL Database: This public database was used to identify "helper ligands" – other known binders to a target protein – for benchmarking and for illustrative examples.
- Benchmark Set: A dataset of 245 protein-ligand complexes representing major drug target families was used to evaluate ComBind's performance against state-of-the-art docking methods.
Key Insights and Ideas
- The "Medicinal Chemist's Intuition" Quantified: The paper formalizes the observation that different ligands binding to the same target often adopt similar poses due to shared physical interactions. ComBind quantifies this intuition using statistical distributions derived from experimental data.
- Simultaneous Pose Prediction is Key: Instead of optimizing each ligand's pose independently, ComBind optimizes a set of poses together. This allows the method to leverage the collective information from multiple ligands to improve individual predictions.
- Non-structural Data is Highly Valuable: The success of ComBind highlights that readily available "non-structural" data (lists of known binders) can significantly enhance predictions, especially when experimental structures are difficult or impossible to obtain.
- GPCRs are a Tough Nut to Crack (and ComBind Helps): G-protein-coupled receptors (GPCRs) are a major class of drug targets but are notoriously difficult to crystallize. ComBind shows a substantial improvement in pose prediction accuracy for GPCRs, suggesting its utility in this challenging area.
Key Results
- Improved Accuracy: ComBind significantly outperforms state-of-the-art per-ligand docking methods (like Glide) across various drug target families. For GPCRs, it selects a correct binding pose over 60% more frequently, increasing accuracy from 47% to 76%.
- Robustness: ComBind's performance is robust even when helper ligands have noisy affinity data or bind to slightly different pockets. It also performs well even with a single helper ligand.
- Experimental Validation: The paper demonstrates ComBind's predictive power by accurately predicting the binding poses of antipsychotics at the D2 dopamine receptor, which were subsequently validated through mutagenesis experiments.
- Component Importance: Both protein-ligand interaction similarity and substructure similarity contribute substantially to ComBind's accuracy, with interaction similarity being particularly important.
Limitations and Future Work
- Protein Flexibility: The current implementation of ComBind uses rigid protein structures. The authors note that flexible docking methods exist and that ComBind's scoring function might help select correct poses from a larger set generated by flexible docking. This is an area for future research.
- Ligand Affinity Data: While ComBind can be extended to use ligand affinity data, the authors chose not to in this study to maintain broad applicability, as affinity estimates in public databases are often inconsistent. Future work could explore how to effectively incorporate this data.
- Extensibility: The framework is general and could be extended to incorporate other data types, such as multiple experimental structures of the protein with different ligands or effects of protein mutations.
- Broader Applications: The authors suggest that this combined physics-based and ligand-based approach could be applied to predict other ligand properties, such as binding affinities, which are crucial for virtual screening.
Summary updated