TL;DR
This study benchmarks five molecular docking programs for their ability to predict how oxazolidinone antibiotics bind to bacterial ribosomes. While DOCK 6 showed the best performance, the inherent flexibility and poor electron density of ribosomal structures limited accurate predictions, highlighting the need for improved methods and careful validation of virtual screening protocols for RNA targets.
What Problem Does This Paper Address?
Oxazolidinone antibiotics are crucial for combating Gram-positive bacterial infections. They work by binding to the 50S ribosomal subunit, inhibiting protein synthesis. With numerous crystal structures of ribosomes bound to these antibiotics available, structure-based drug design (SBDD) using molecular docking is a promising avenue for discovering new and improved oxazolidinones. However, a significant challenge is understanding how well computational docking methods can accurately predict the binding poses and rank potential drug candidates for RNA targets like ribosomes. This is particularly difficult due to the inherent flexibility and complex nature of RNA, which can lead to poor electron density in experimental structures, making accurate modeling challenging. This paper aims to address this by systematically evaluating and comparing the performance of several popular docking programs for this specific, challenging application.
Methods
The researchers evaluated five molecular docking programs: AutoDock 4 (AD4), AutoDock Vina (Vina), DOCK 6, rDock, and RLDock. They performed two main types of analyses:
- Redocking Validation: They used eleven existing crystal structures of ribosomes bound to oxazolidinone ligands. Each program was used to "redock" the native ligand back into its binding site. The accuracy of the predicted pose was assessed using the Root-Mean-Square Deviation (RMSD) between the predicted and experimental ligand positions. The internal scoring functions of each program were also analyzed.
- Virtual Screening (VS) and Re-scoring: A dataset of 285 oxazolidinone derivatives was screened against the S. aureus ribosome using the best-performing docking program identified in the validation step (DOCK 6). To improve prediction accuracy, they also employed:
- AnnapuRNA: A machine learning-based scoring function specifically designed for RNA-ligand interactions, used to re-score the docked poses.
- Random Forest Classifier: Used to identify potential biases in the DOCK 6 scoring function by classifying derivatives as over- or under-predicted.
- Multiple Linear Regression (Re-scoring Function): Developed a new re-scoring function that combined the absolute DOCK 6 docking score with various molecular descriptors (e.g., molecular weight, topological polar surface area) to better correlate predicted binding affinity with experimental Minimum Inhibitory Concentration (MIC) values.
Datasets
- Ribosomal Crystal Structures: Eleven 50S ribosomal crystal structures containing oxazolidinone ligands were used for redocking validation. These were sourced from various organisms, including D. radiodurans, E. coli, H. marismortui, S. aureus, and MRSA strains.
- Oxazolidinone Derivative Dataset: A dataset of 285 oxazolidinone and linezolid derivatives was compiled from existing literature for virtual screening. This dataset included known MIC values against S. aureus.
Key Insights and Ideas
- RNA Docking is Challenging: The inherent flexibility of the ribosomal RNA binding pocket significantly hinders accurate docking predictions, even with advanced programs. Poor electron density in many experimental structures exacerbates this issue.
- Program Performance Varies Greatly: Not all docking programs are created equal for RNA targets. DOCK 6 emerged as the top performer in terms of accurately reproducing native poses, but even it only succeeded in a minority of cases.
- Scoring Functions Need Improvement for RNA: Standard scoring functions, even those adapted for RNA, struggle to consistently rank oxazolidinone derivatives based on their biological activity. This suggests a need for more specialized scoring mechanisms.
- Molecular Descriptors Enhance Predictions: Incorporating molecular descriptors into a re-scoring function significantly improved the correlation between predicted and experimental activity, offering a more robust approach than docking scores alone.
- Structural Modifications Don't Always Correlate with Activity: Surprisingly, many structural modifications to the oxazolidinone scaffold did not show a clear correlation with improved docking scores or MIC values, suggesting complex interactions at play.
Key Results
- Redocking Performance Ranking (Median RMSD): DOCK 6 > AD4 > Vina > rDock >> RLDOCK.
- DOCK 6 Success Rate: DOCK 6 successfully reproduced native ligand binding poses in only 4 out of 11 crystal structures. Its success rate for top-performing poses under 2.5 Å RMSD was 52%.
- RLDOCK's High Success Rate but Blind Docking Limitation: RLDOCK achieved the highest success rate (85%) for poses under 2.5 Å RMSD, but this was attributed to its "blind docking" approach, which often placed ligands outside the intended binding site, leading to high RMSD values when compared to the native pose.
- Virtual Screening Correlation: The initial DOCK 6 scoring showed a weak correlation (Pearson's r = 0.485) with experimental pMIC values.
- Improved Correlation with Re-scoring: The developed re-scoring function, incorporating molecular descriptors, significantly improved the correlation to Pearson's r = 0.779.
- DOCK 6 Bias: Morgan fingerprint analysis suggested DOCK 6 tended to underpredict molecules with acetamide tail modifications and overpredict those with methylamino groups.
Limitations and Future Work
- Poor Quality of Ribosomal Structures: The authors highlight that the low resolution and poor electron density of many available ribosomal crystal structures are a major limitation, hindering accurate docking.
- RNA Flexibility: The inherent flexibility of the RNA binding pocket makes it difficult for current docking programs to accurately model binding.
- Limited Correlation in VS: No clear trend was observed between structural modifications and activity in the virtual screening of derivatives, suggesting that simple structure-activity relationships are not easily discernible.
- Small Dataset for Re-scoring: The correlation analysis for the re-scoring function was performed on a relatively small test set (30 compounds), and the authors suggest a larger dataset would be beneficial.
- Future Work: The authors recommend:
- Performing VS campaigns with caution.
- Incorporating more comprehensive strategies like molecular dynamics (MD) simulations and relative free energy calculations.
- Further studies on RNA-ligand interactions should consider the effects of water molecules, pH, and metal ions.
- Developing and utilizing more advanced RNA-specific docking and scoring methods.
- Using larger datasets for validation and re-scoring.
Summary updated