TL;DR
Researchers have developed PbsNRs, a user-friendly web server that uses proteochemometric modeling to predict potential drug binders and molecular scaffolds for nuclear receptors (NRs). This tool outperforms existing methods and has been experimentally validated, offering a promising avenue for accelerating the discovery of new NR-targeted therapies for various diseases.
What Problem Does This Paper Address?
Nuclear receptors (NRs) are crucial protein targets for drug development, involved in a wide range of physiological processes and implicated in numerous diseases, including cancer, diabetes, and liver disorders. While a significant percentage of approved drugs target NRs, the process of identifying effective NR binders and novel molecular scaffolds is challenging and time-consuming. Existing computational methods for predicting drug-target interactions often focus on individual protein-ligand pairs or have limited scope, failing to adequately address the specific needs of NR drug discovery. There's a clear need for a more comprehensive and accurate in silico tool that can efficiently prescreen potential binders and identify promising molecular scaffolds for the diverse family of NRs.
Methods
The core of PbsNRs is a proteochemometric modeling (PCM) approach. Unlike traditional quantitative structure-activity relationship (QSAR) models that focus solely on chemical compound properties, PCM integrates information from both the ligand (chemical compound) and the target protein (NR).
Here's a breakdown of their methodology:
- Descriptor Generation:
- Ligands: Chemical and physical properties of small molecules are described using RDKit, generating 187-dimensional descriptors.
- Proteins (NRs): For each NR target, a 30-dimensional descriptor is created by calculating its similarity to 30 "background" NRs (those with available crystal structures in the PDB) using the Smith-Waterman alignment algorithm. This allows the model to leverage information from NRs with known structures and bioactivity data to predict for NRs that lack such data.
- Machine Learning Models:
- Random Forest (RF): This ensemble learning method was chosen for its robustness and good performance on tabular data.
- Deep Learning (DL): A backpropagation neural network was also explored for comparison.
- Model Training and Optimization:
- The model was trained using a dataset of 6554 bioactive compounds with known EC50 values against five NR targets.
- To address the imbalance between positive and negative samples in the training data, oversampling techniques were employed.
- Gaussian noise was added to enhance model robustness.
- Extensive parameter optimization was performed, including varying training set percentages and oversampling numbers.
- Web Server and Python Package: A user-friendly web server (http://pbsnrs.badd-cao.net) and a downloadable Python package were developed to make PbsNRs accessible to the research community.
Datasets
- Training and Internal Validation: The Orphan Nuclear Receptor Ligand Binding Database (ONRLDB) [51] was used, containing 6554 bioactive compounds with corresponding EC50 values against five NR targets.
- External Validation: 713 independent bioactive compounds for six other NRs not included in the initial modeling dataset were used to evaluate the model's generalization ability.
- Screening Databases:
- ChEMBL [45] (5000 randomly selected chemical structures)
- DrugBank [46] (FDA-approved and experimental drugs)
- TCM@Taiwan [52] (natural products)
- Herbal Ingredients' Targets Database (HIT 2.0) [53] and Natural Product Activity and Species Source (NPASS) [54] were used for screening natural products for NR2B1 ligands.
Key Insights and Ideas
- Proteochemometric Modeling for NRs is Powerful: Integrating both ligand and protein features (PCM) significantly improves prediction accuracy for NR-ligand interactions compared to methods that only consider one aspect. This is a key differentiator for PbsNRs.
- Natural Products are Promising NR Ligand Sources: The study found that natural products from the TCM@Taiwan database showed a significantly higher proportion of potential NR binders compared to FDA-approved drugs and experimental compounds. This highlights natural products as a valuable resource for NR drug discovery.
- Scaffold Identification is Possible: PbsNRs not only predicts binders but also helps identify recurring molecular scaffolds within bioactive compound clusters, which can guide the design of new drug candidates.
- RF Outperforms Deep Learning for this Task: For this specific problem with tabular data (molecular fingerprints), the Random Forest classifier demonstrated better robustness and generalization ability than deep learning models, especially on external validation sets.
- Transfer Learning is Effective: By using similarity to known NR structures, PbsNRs can extend its predictive capabilities to NRs for which experimental bioactivity data is scarce.
Key Results
- High Prediction Accuracy:
- Internal Validation: PbsNRs achieved an ROC-AUC of 0.906 using the RF classifier.
- External Validation: PbsNRs achieved an ROC-AUC of 0.783 using the RF classifier, outperforming other state-of-the-art tools.
- Superior Performance: PbsNRs significantly outperformed existing compound-protein interaction (CPI) prediction tools (Tsubaki's model, DeepCPI) and NR-specific predictors (NR-ToxPred, NR-Profiler) on external validation datasets.
- Experimental Validation: Seven predicted NR2B1 ligands from natural product databases were experimentally tested using biosensors. The results showed 100% agreement between PbsNRs' in silico predictions and experimental outcomes.
- Speed Advantage: PbsNRs was at least 400-1100 times faster than popular docking approaches (AutoDock, CDDOCKER, LibDOCK) for virtual screening tasks.
- Broad Application: PbsNRs successfully screened large databases of chemical compounds and natural products, identifying potential binders and scaffolds.
Limitations and Future Work
- Scaffold Similarity Limitations: The authors acknowledge that PbsNRs might provide similar predictions for compounds with very similar scaffolds, which could sometimes lead to opposite bioactivities.
- Future Directions:
- Incorporate 3D Descriptors: Future work could involve adding 3D descriptors to capture more precise molecular information, especially as more NR structures become available.
- Describe Functional Groups: Beyond scaffolds, incorporating descriptions of specific functional groups could enhance screening accuracy.
- Expand NR Coverage: Continuously updating the NR library and incorporating more bioactivity data will further improve the model's performance and scope.
Summary updated