TL;DR
This study introduces a novel computational pipeline that leverages machine learning and protein interaction data to identify key protein features and biological pathways responsible for compound toxicity. By integrating the CANDO platform with the Tox21 dataset, the researchers not only predict toxicity with high accuracy but also provide mechanistic insights into how compounds exert their toxic effects at a proteomic level.
What Problem Does This Paper Address?
The increasing exposure to chemicals necessitates efficient methods for predicting their toxicity. While machine learning has shown promise in predicting toxicity based on chemical structures, understanding the underlying biological mechanisms, particularly at the proteomic level, has been less explored. Existing approaches often treat toxicity as a black box, making it difficult to pinpoint specific molecular players or pathways involved. This paper addresses this gap by developing a method to identify crucial protein features that drive toxicity and then mapping these features to biological pathways, offering a deeper mechanistic understanding.
Methods
The core of this research is a computational pipeline that combines several key techniques:
- Data Pre-processing: The Tox21 dataset, containing over 10,000 compounds and 12 in vitro toxicity assays, was standardized and cleaned. A significant challenge with this dataset is its high imbalance, with far more inactive (non-toxic) compounds than active (toxic) ones.
- Protein Feature Generation (CANDO Platform): Instead of relying solely on chemical descriptors, the study utilized the Computational Analysis of Novel Drug Opportunities (CANDO) platform. CANDO generates protein-compound interaction signatures by predicting how a small molecule interacts with a vast library of human protein structures. These interaction scores serve as the protein-based features for the machine learning model.
- Data Resampling (SMOTE+ENN): To combat the class imbalance issue in the Tox21 data, the researchers employed a hybrid resampling technique called SMOTE+ENN. This method combines oversampling (SMOTE) to create synthetic minority class samples and undersampling (ENN) to remove noisy majority class samples, thereby balancing the dataset for better model training.
- Machine Learning Model (Random Forest with Feature Selection): A Random Forest (RF) algorithm was chosen for its robustness and ability to perform embedded feature selection. The RF model was trained on the resampled data to predict toxicity. Crucially, the RF's Gini importance measure was used to identify the top 100 most influential protein features (interactions) for each toxicity endpoint.
- Enrichment Analysis: The identified top protein features were then subjected to enrichment analysis using the Reactome pathway database. This step maps the selected proteins to biological pathways, revealing which pathways are significantly overrepresented and likely implicated in the observed toxicity.
Datasets
The primary dataset used in this study is the Tox21 dataset. This dataset comprises:
- Over 10,000 compounds.
- 12 in vitro toxicity assays, categorized into Nuclear Receptor (NR) signaling pathways (7 assays) and Stress Response (SR) pathways (5 assays).
- The dataset is characterized by significant class imbalance, with a much higher proportion of inactive (non-toxic) compounds compared to active (toxic) ones for most assays.
The study also utilized the UniProt database for human protein sequences and the Protein Data Bank (PDB) for protein structures, and Reactome for pathway annotations.
Key Insights and Ideas
- Proteomic Features for Toxicity Prediction: This work pioneers the use of protein-compound interaction signatures as features for predicting toxicity, moving beyond traditional chemical descriptors. This offers a more biologically relevant perspective.
- Mechanistic Understanding through Pathway Enrichment: By linking top protein features to biological pathways, the study provides a higher-level mechanistic explanation for toxicity, rather than just a prediction. This is crucial for understanding how a compound causes harm.
- Network Effects in Toxicity: The analysis suggests that toxicity is often not due to a single protein-target interaction but rather a complex interplay within interconnected biological networks. The overlap of selected proteins across multiple pathways highlights this complexity.
- Validation of AhR Pathway: The case study on the Aryl Hydrocarbon Receptor (AhR) demonstrates the pipeline's ability to identify pathways (e.g., DNA damage, cell cycle regulation, immune response) and proteins (e.g., BRCA1, TRAF6) that are well-supported by existing literature, lending strong credibility to the approach.
Key Results
- High Predictive Performance: The model achieved high AUCROC values for several toxicity endpoints, with the SR-MMP (mitochondrial membrane potential) assay reaching 0.92 and NR-AhR (aryl hydrocarbon receptor) reaching 0.90.
- Identification of Key Proteins and Pathways: The pipeline successfully identified specific proteins and enriched pathways associated with toxicity for the 12 Tox21 assays.
- Case Study Validation: The detailed analysis of the NR-AhR pathway provided strong evidence for the pipeline's ability to uncover biologically relevant mechanisms, with identified proteins and pathways aligning with known literature on AhR toxicity, DNA damage, and inflammation.
- Comparable Performance to Existing Models: The study's model performance was comparable to top-performing models in the Tox21 Data Challenge, while offering the added benefit of interpretability through feature extraction.
Limitations and Future Work
- Literature Validation is Ongoing: While the AhR case study was well-validated, further experimental studies are warranted to confirm the identified pathways and proteins for all twelve toxicity endpoints.
- Computational Demands: Generating protein-compound interaction signatures for a large library of proteins can be computationally intensive.
- Future Directions: The authors suggest implementing novel machine learning feature selection techniques and expanding the application of this pipeline to other toxicity datasets and drug discovery efforts. They envision this approach aiding in hypothesis generation and the development of novel therapeutics by elucidating toxicity mechanisms.
Summary updated