TL;DR
This paper demonstrates that integrating Absolute Binding Free Energy (ABFE) calculations into virtual compound screening significantly improves the identification of active drug candidates compared to traditional docking methods alone. By applying ABFE to high-scoring compounds from docking, the researchers achieved better discrimination between true actives and decoys, offering a more accurate and efficient approach for early-stage drug discovery.
What Problem Does This Paper Address?
Drug discovery projects often begin with virtual compound screening, a computational method to identify potential drug molecules that bind to a specific target protein. Traditional docking methods are fast but have limitations in accuracy, leading to a significant number of false positives (decoys) being selected. Relative Binding Free Energy (RBFE) calculations offer higher accuracy but are typically only suitable for chemically similar compounds, making them less applicable to the diverse libraries screened in early-stage discovery. Absolute Binding Free Energy (ABFE) calculations, while more accurate and applicable to diverse compounds, are computationally expensive and have not been thoroughly evaluated as a refinement step for virtual screening. This paper addresses the challenge of improving the accuracy and efficiency of virtual compound screening by investigating the effectiveness of ABFE calculations as a post-docking refinement strategy for diverse compound libraries.
Methods
The study employed a multi-step computational approach:
- Baseline Docking: Standard docking calculations were performed using the Glide SP method on a large dataset of compounds (actives and decoys) for three protein targets: BACE1, CDK2, and thrombin. Ligand preparation involved generating various protonation states, tautomers, and stereoisomers to account for potential binding variations.
- Compound Selection for ABFE: Based on docking scores, two sets of 30 compounds each (Tier 1: top-scoring; Tier 2: slightly lower-scoring) were selected for each target. This selection aimed to mimic a realistic virtual screening scenario.
- MD Equilibration: For each selected compound, ten distinct docked poses were generated. These poses were then subjected to short Molecular Dynamics (MD) simulations to equilibrate the protein-ligand complex and ensure stable ligand conformations within the binding site. Poses that showed significant deviation from the binding site were discarded.
- Absolute Binding Free Energy (ABFE) Calculation: The equilibrated poses were then used as starting points for ABFE calculations using the BAT.py software. This involved alchemical transformations to compute the free energy of decoupling the ligand from the binding site and recoupling it to bulk solvent. Two independent ABFE runs were performed for each pose to assess reproducibility. The overall binding free energy was obtained by combining the results from the best-scoring poses.
Datasets
The study utilized the DUD-E (Directory of Useful Decoys, Enhanced) database. This dataset contains known active compounds and computationally generated decoys for various protein targets. For this study, three specific targets were used:
- BACE1: 283 actives and 18,100 decoys.
- CDK2: 474 actives and 27,850 decoys.
- Thrombin: 461 actives and 27,004 decoys.
The DUD-E dataset is a standard benchmark for evaluating virtual screening methods. The authors compared their docking results on the full DUD-E datasets with previously reported values.
Key Insights and Ideas
- ABFE as a Powerful Refinement Tool: The most significant insight is that ABFE calculations, when applied to high-scoring compounds from docking, provide a substantial improvement in distinguishing true actives from decoys. This suggests ABFE can effectively filter out false positives that might pass initial docking screens.
- Importance of Ligand Pose Quality: The study strongly emphasizes that the quality of the initial ligand pose is critical for accurate ABFE calculations. Imperfect poses, even after MD equilibration, can lead to significant variations in calculated binding free energies and hinder the correlation with experimental data.
- ABFE's Advantage for Diverse Libraries: Unlike RBFE, ABFE is well-suited for screening diverse compound libraries because it doesn't rely on chemical similarity between compounds. This makes it a valuable tool for early-stage drug discovery where chemical diversity is key.
- Computational Bottleneck: While ABFE offers improved accuracy, it remains computationally intensive, particularly the free energy simulation stage. This highlights the need for continued efforts in optimizing ABFE methods for speed and efficiency.
- Challenges in Absolute Correlation: Despite improved enrichment, the study found a lack of strong correlation between computed ABFEs and experimental binding affinities. This is attributed to the inherent difficulty in obtaining perfect ligand poses in a "blind" virtual screening scenario.
Key Results
- Improved Enrichment: ABFE calculations consistently outperformed docking calculations in enriching active compounds across all three targets and both tiers of compounds. The Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curves increased by an average of 0.16 (from 0.67 to 0.83) when moving from docking to ABFE.
- Baseline Docking Performance: The docking calculations themselves performed well, achieving good enrichment metrics (AUCs ranging from 0.82 to 0.87 and 1% enrichment factors of 26-30) that were comparable to or better than previous studies on the same datasets.
- ABFE Differentiates Better: Histograms of computed ABFE values showed a clearer separation between actives and decoys compared to docking scores, indicating better discriminatory power.
- Reproducibility Issues: Two independent ABFE runs for the same docked pose sometimes yielded significantly different results (up to 9 kcal/mol difference), highlighting the sensitivity to initial conformations and the need for careful pose selection or multiple runs.
Limitations and Future Work
- Ligand Pose Accuracy: The primary limitation acknowledged is the difficulty in obtaining highly accurate ligand poses in a virtual screening context without prior structural information. This directly impacts the correlation between computed and experimental binding free energies.
- Force Field Accuracy: The authors suggest that potential inaccuracies in the force field, particularly in nonbonded interactions, might lead to overestimation of binding affinities for both actives and decoys.
- Protein Conformational Changes: The ABFE calculations might not fully capture the relaxation of the protein-ligand complex to its unbound conformational ensemble after ligand decoupling, potentially leading to overly favorable binding free energies.
- Future Directions:
- Developing faster and more reliable methods for predicting accurate ligand binding poses is crucial.
- Investigating improved force fields or water models could address potential inaccuracies in binding affinity predictions.
- Exploring methods to better estimate the relative stabilities of multiple poses identified during MD equilibration could improve ABFE accuracy.
- Further optimization of ABFE calculation speed and efficiency is needed to make it more broadly applicable to large-scale screening.
Summary updated