TL;DR
This paper presents a novel active learning (AL) workflow that combines molecular dynamics simulations with machine learning to efficiently discover novel inhibitors for the LRRK2 WDR domain, a key target in Parkinson's disease. The approach successfully identified 8 new inhibitors from a large chemical space, demonstrating a significant reduction in computational cost and experimental testing.
What Problem Does This Paper Address?
Parkinson's disease (PD) is a neurodegenerative disorder, and mutations in the Leucine-Rich Repeat Kinase 2 (LRRK2) gene are a major cause of familial PD. While LRRK2 has been a focus for drug discovery, the WDR domain of LRRK2, implicated in disease pathology, remained an underexplored target with no known inhibitors prior to the CACHE Challenge. The challenge was to computationally identify small molecules that bind to this domain. This paper tackles the subsequent step: optimizing initial "hit" compounds into a more potent and diverse series of inhibitors, a crucial but often computationally expensive phase of drug discovery. The core problem is efficiently navigating vast chemical spaces to find molecules with improved binding affinity while minimizing the number of costly simulations and experimental validations.
Methods
The authors developed a sophisticated computational pipeline centered around an Active Learning (AL) guided Relative Binding Free Energy (RBFE) calculation workflow. Here's a breakdown:
- Virtual Screening: They started with the Enamine REAL database, containing approximately 5.5 billion compounds. This was filtered using specific chemical patterns (SMARTS) to identify "closest analogs" (structurally similar to known hits) and "general analogs" (more diverse).
- Docking: Both sets of analogs underwent docking simulations to predict their binding poses within the LRRK2 WDR domain.
- Molecular Dynamics (MD) with Thermodynamic Integration (TI): For promising candidates, they employed MD simulations using the TI framework to calculate RBFE. This is a physics-based method that quantifies the change in binding free energy when a molecule binds to a target.
- Active Learning (AL): This is the core innovation. Instead of running simulations on all candidates, an AL loop was implemented. In each iteration:
- A machine learning (ML) model was trained on the RBFEs computed in previous steps.
- This ML model then predicted the RBFE for a larger set of unlabeled molecules.
- The AL strategy intelligently selected the most promising molecules (those predicted to have the best binding affinity) for the next round of computationally expensive MD TI simulations. This iterative process allows the workflow to learn and focus computational resources on the most promising regions of chemical space.
- RBFE to ABFE Conversion: For ML model training, RBFEs were converted to Absolute Binding Free Energies (ABFEs) to allow for a unified model for both initial hits.
- Perturbation Map: To handle challenging ligand preparations for MD simulations, they introduced a "perturbation map" strategy, using a modified reference ligand to improve simulation convergence and accuracy.
Datasets
- Enamine REAL Database: A large commercial library of approximately 5.5 billion small-molecule compounds was used for virtual screening.
- CACHE Challenge #1 Data: The paper builds upon the initial hit compounds identified in the first phase of the CACHE Challenge, which provided experimentally validated binders.
- LRRK2 WDR Domain Structure: The crystal structure of the LRRK2 WDR domain (PDB ID: 6DLO) was used for docking simulations.
No new benchmark datasets were introduced, but the work leverages the experimental validation from the CACHE Challenge to refine and validate their computational approach.
Key Insights and Ideas
- Active Learning as a Smart Navigator: The AL approach is crucial for efficiently exploring vast chemical spaces. Instead of brute-forcing simulations, it learns from previous results to guide the selection of molecules for further investigation, significantly reducing computational cost.
- Bridging Physics and Machine Learning: The integration of accurate physics-based MD simulations (for RBFE) with ML models (for prediction and selection) creates a powerful synergy. The physics provides the ground truth, while ML accelerates the exploration.
- Handling "Undruggable" Targets: The LRRK2 WDR domain was an underexplored target. This work demonstrates that even for such targets, where initial structural information might be limited, this AL-RBFE approach can be effective.
- Importance of Reference Ligand Choice: The paper highlights how the choice of reference ligand in RBFE calculations can significantly impact simulation accuracy and convergence, introducing a practical solution for challenging cases.
- Experimental Validation is Key: The tight loop between computational prediction and in-house experimental validation (SPR, 19F-NMR, DLS) is critical for confirming the efficacy of the computational workflow and identifying true hits.
Key Results
- 8 Novel Inhibitors Identified: Out of 35 experimentally tested compounds, 8 were confirmed as novel inhibitors of the LRRK2 WDR domain. This represents a hit rate of 23% for the tested compounds.
- Efficient Exploration: The workflow screened approximately 5.5 billion compounds, selected ~25,000 for AL-RBFE calculations, and computed RBFEs for 672 ligands, ultimately leading to the selection of 75 for experimental testing.
- Mean Absolute Error (MAE): The computed binding free energies showed a mean absolute error of 2.69 kcal/mol compared to experimentally measured K<sub>D</sub> values for the hit compounds.
- Improved Affinity: The AL-RBFE workflow identified 102 analogs with predicted improved binding affinity compared to the initial hits.
- Diverse Chemical Series: The identified hits represent a new chemical series, expanding the known chemical space for LRRK2 WDR domain inhibitors.
Limitations and Future Work
- MAE Skew: The authors note that the MAE is skewed by two specific compounds (O2 and O8), suggesting that excluding them would result in a lower MAE (1.66 kcal/mol). This indicates potential areas for further refinement of the prediction model.
- Statistical Significance: While the AL approach showed promise in selecting binders over naive schemes, the authors acknowledge that the small sample sizes involved limit statistically significant conclusions regarding the superiority of AL.
- General Applicability: The authors suggest that this workflow is widely applicable to screening any chemical space for small-molecule analogs with increased affinity, provided the general constraints of RBFE calculations are met. Future work would likely involve applying this methodology to other challenging drug targets.
- Further Optimization: The identified hits, while confirmed binders, have mid-micromolar affinities. Further optimization rounds would be necessary to achieve higher potency, which this workflow is well-suited to facilitate.
Summary updated