Training Recurrent Neural Networks for BrdU Detection with Oxford Nanopore Sequencing: Guidance and Lessons Learned

Area of Science:

• Molecular Biology
• Genomics
• Bioinformatics

Background:

• 5 -bromo-2 -deoxyuridine (BrdU) is a thymidine analog crucial for studying cell proliferation and DNA synthesis.
• Accurate detection of BrdU incorporation at single-nucleotide resolution is essential for understanding DNA replication dynamics.
• Existing deep learning methods for BrdU detection via nanopore sequencing lack accessible tutorials for evolving technologies.

Purpose of the Study:

• To provide an accessible protocol for preparing training data and implementing deep learning models for BrdU detection using nanopore sequencing.
• To train and evaluate a deep learning model for detecting BrdU incorporation in DNA sequences.

Main Methods:

• Utilized publicly available synthetic and real nanopore DNA sequencing datasets with and without BrdU.
• Processed data using open-source and custom software tools.
• Trained bidirectional gated recurrent unit (BiGRU)-based recurrent neural networks (RNNs) on Google Colab using TensorFlow.

Main Results:

• Achieved high specificity (>94%) in BrdU detection using BiGRU-based RNNs.
• Moderate sensitivity was observed due to limitations in available BrdU-positive training data.
• Detailed the end-to-end process of model setup, training, testing, and fine-tuning.

Conclusions:

• The developed protocol, though trained on legacy R9 flow cell data, is adaptable to newer R10 flow cells and other base modification detection.
• This work promotes the wider application of deep learning, especially RNNs, in biological research for analyzing sequential data.
• Facilitates advanced studies on DNA replication and cell proliferation using nanopore sequencing technology.