TL;DR
This paper introduces PPARgene, a novel database that consolidates experimentally verified PPAR target genes and employs a machine learning approach to predict new ones. By integrating in silico PPRE analysis with high-throughput gene expression data, they significantly improve prediction accuracy, offering a valuable resource for PPAR research.
What Problem Does This Paper Address?
Peroxisome proliferator-activated receptors (PPARs) are crucial transcription factors involved in a wide array of physiological processes, including lipid metabolism and inflammation. While many studies have identified individual PPAR target genes, there's a need for a centralized, comprehensive resource that aggregates this knowledge. Furthermore, predicting novel PPAR targets computationally has been challenging, with existing methods relying solely on sequence motifs (PPREs) showing limited accuracy. This paper addresses the need for a unified database and a more robust prediction method for PPAR target genes.
Methods
The authors developed a two-pronged approach:
- Manual Curation of Verified Genes: They meticulously reviewed PPAR-related publications from PubMed to identify and collect experimentally verified PPAR target genes. This involved ensuring evidence for both gene regulation (mRNA/protein level) and functional PPREs (reporter assays/DNA-binding) were present.
- Machine Learning for Prediction:
- Feature Engineering: They extracted two key features:
- Conserved PPRE Score (CPS): This involved identifying PPREs (PPAR-responsive elements) in conserved genomic regions between mouse and rat, aiming to improve specificity.
- High Throughput Evidence (HTE): This quantified the experimental evidence from PPAR-related microarray datasets, specifically looking at upregulated genes upon PPAR activation.
- Model Development: A binomial logistic regression classifier was trained using the curated experimentally verified genes as positive examples and randomly sampled genes from the background as negative examples. The model integrated both CPS and HTE features.
- Evaluation: The model's performance was assessed using five-fold cross-validation, with precision, recall, F1 score, and AUC as key metrics.
Datasets
- Experimentally Verified PPAR Target Genes: Manually curated from 9046 PPAR-related publications. This dataset forms the basis of the PPARgene database and was used as positive training data for the prediction model.
- PPAR-Relevant Microarray Data Sets: 22 datasets were acquired from the GEO database, focusing on experiments where PPARs were activated or overexpressed. These were used to derive the High Throughput Evidence (HTE) feature.
- Background Gene Set: All protein-coding genes (excluding experimentally verified targets) were used to generate negative training samples for the prediction model.
Key Insights and Ideas
- The Power of Integration: Combining in silico PPRE analysis with experimental gene expression data (HTE) significantly boosts the accuracy of PPAR target gene prediction compared to using PPREs alone. This highlights the importance of integrating different types of biological evidence.
- Conservation as a Filter: Leveraging evolutionary conservation of PPREs (CPS) improves prediction specificity, reducing false positives that might arise from simple sequence matching.
- A Centralized Resource is Crucial: The PPARgene database provides a much-needed, unified repository of both verified and predicted PPAR targets, streamlining research efforts.
- PPARs' Broad Regulatory Role: The database includes a substantial number of target genes for each PPAR subtype (PPARα, PPARβ/δ, PPARγ), underscoring their widespread influence on cellular processes.
Key Results
- PPARgene Database: Contains 225 experimentally verified PPAR target genes (83 for PPARα, 83 for PPARβ/δ, and 104 for PPARγ), with detailed information on tissue, species, and references.
- Prediction Model Performance: The logistic regression model trained on both conserved PPREs (CPS) and high-throughput evidence (HTE) achieved significantly improved performance compared to models using only PPREs.
- PS (PPRE Score only): Precision 0.57, Recall 0.49, F1 0.52, AUC 0.59
- CPS (Conserved PPRE Score only): Precision 0.61, Recall 0.68, F1 0.64, AUC 0.68
- CPS + HTE (Combined): Precision 0.83, Recall 0.59, F1 0.69, AUC 0.82
- Genome-Wide Prediction: The model predicted 2,683 potential PPAR target genes in the mouse genome, classified into high, median, and low confidence levels.
Limitations and Future Work
- Current Focus: The current release of PPARgene focuses on protein-coding target genes.
- Future Directions:
- Non-coding Targets: The authors plan to extend their prediction methods to include non-coding target genes of PPARs.
- Subtype-Specific Prediction: Future work will involve developing methods to predict target genes for each PPAR subtype individually.
- Database Updates: The database of experimentally supported PPAR target genes will be updated every three months.
Summary updated