Predicting environmentally responsive transgenerational differential DNA methylated regions (epimutations) in the genome using a hybrid deep-machine learning approach

Pegah Mavaie, Lawrence Holder, Daniel Beck, Michael K. SkinnerBMC Bioinformatics 2021 · DOI

TL;DR

This paper introduces a novel hybrid deep-learning and machine-learning (DL-ML) approach to predict environmentally induced epigenetic changes (epimutations) in DNA methylation regions across the genome. This method outperforms traditional DL and ML approaches, offering a more accurate and interpretable way to identify regions susceptible to transgenerational epigenetic alterations, with potential applications in disease diagnosis and prevention.

What Problem Does This Paper Address?

Epigenetic modifications, such as DNA methylation, play a crucial role in regulating gene expression and can be influenced by environmental factors. These changes can even be transmitted across generations, leading to transgenerational inheritance of traits and diseases. Identifying specific regions in the genome that are susceptible to these environmentally induced DNA methylation changes (Differential DNA Methylated Regions or DMRs) is a significant challenge. Traditional methods for identifying DMRs are often labor-intensive, require extensive domain knowledge for feature selection, and struggle with imbalanced datasets where disease-related events are rare. This paper aims to address these challenges by developing a more efficient and accurate computational method for predicting these environmentally responsive DMRs.

Methods

The core of this research is a hybrid DL-ML approach. It begins with a deep neural network (specifically, a convolutional neural network or CNN) that acts as an automated feature extractor. Instead of relying on human-engineered features, the DL network learns complex, abstract representations directly from the DNA sequence. The initial convolutional layer's output is used to re-express the input data, creating a new set of features. These DL-generated features are then fed into a traditional machine learning classifier, XGBoost, which is known for its effectiveness in handling imbalanced datasets. This hybrid approach leverages the DL network's ability to learn intricate patterns and the XGBoost classifier's efficiency and interpretability. The model is trained and validated using a five-fold cross-validation strategy to ensure robust performance estimates.

Datasets

The study utilizes datasets generated from experiments involving rat genomes. Specifically, these datasets contain information on differential DNA methylated regions (DMRs) identified in the F3 generation of rats that were exposed to one of nine environmental toxicants during the FO generation's fetal gonadal development. The toxicants include atrazine, DDT, glyphosate, vinclozolin, permethrin, DEET, dioxin, jet fuel, methoxychlor, bisphenol A, and phthalates. For training, regions were labeled as DMRs if they met a stringent p-value threshold (< 10⁻⁵) in any of the exposure groups. Non-DMRs were defined as regions containing no CpGs to create a balanced training set. The model was also applied to predict DMRs in the human genome.

Key Insights and Ideas

  • Automated Feature Extraction: The DL component eliminates the need for manual feature engineering, which is a significant bottleneck in traditional bioinformatics. This allows the model to discover novel and potentially more informative features.
  • Hybrid Power: Combining DL for feature extraction with a robust ML classifier like XGBoost offers a synergistic advantage, achieving better performance than either method alone.
  • Handling Imbalanced Data: XGBoost's inherent ability to handle imbalanced datasets is crucial for this problem, as DMRs are rare events in the genome.
  • Interpretability: While DL models can be black boxes, using the DL-extracted features with a tree-based classifier like XGBoost allows for feature importance ranking, providing insights into what drives the predictions.
  • Trade-off Management: The hybrid model demonstrates a successful trade-off between high accuracy and avoiding overly general predictions, identifying specific regions of interest without over-predicting.

Key Results

The hybrid DL-ML model achieved high accuracy in predicting DMRs. For instance, the overall accuracy across all chromosomes was 97.53%, with an F1 score of 0.9502, precision of 0.9556, and recall of 0.9488. When compared to standalone DL models like DanQ, DeepSEA, and DeepCpG, the hybrid model consistently outperformed them in accuracy across individual chromosomes. The model also successfully predicted potential DMRs in the human genome, identifying approximately 2.19 × 10⁶ potential DMR sites. Visualization of the predicted DMRs showed a generally even distribution across the genome, with some regions exhibiting higher DMR density.

Limitations and Future Work

The authors acknowledge that while the model achieves high accuracy, it still predicts a significant fraction of regions as DMRs that were not explicitly identified in the training data. This suggests potential for over-generalization. Future work could involve:

  • Specializing the model: This could include identifying additional negative examples (non-DMRs) for training, such as CpG islands, or exploring ensemble learning approaches.
  • Exposure-specific models: Given that some DMRs are unique to specific exposures, developing models tailored to individual toxicants might yield better results.
  • Incorporating biological features: Adding known biological features that correlate with DMRs (e.g., CpG density) alongside DL-generated features could further enhance performance.
  • Hyper-parameter optimization: Further systematic exploration of DL network hyper-parameters using advanced tuning algorithms is suggested.
  • Expanding datasets: Including more data from rat and human genomes as they become available will improve model robustness.

Summary updated

BibTeX
@article{mavaie2021predicting,
  author = {Pegah Mavaie and Lawrence Holder and Daniel Beck and Michael K. Skinner},
  title = {{Predicting environmentally responsive transgenerational differential DNA methylated regions (epimutations) in the genome using a hybrid deep-machine learning approach}},
  journal = {BMC Bioinformatics},
  year = {2021},
  doi = {10.1186/s12859-021-04491-z},
}
Figure