NanoCNN – A CNN-Based Tool to Detect RNA m⁶A Methylations Using Oxford Nanopore Sequencing

NanoCNN is an end-to-end deep learning pipeline for detecting RNA modifications (e.g., m⁶A, m⁵C, pseudouridine, inosine) from Oxford Nanopore direct RNA sequencing data. It is designed to operate directly on raw electrical signal data and produce nucleotide-resolution modification predictions.

---

Overview

NanoCNN provides a complete workflow from raw signal processing to model training and evaluation. The pipeline is built to handle large-scale Nanopore datasets efficiently while maintaining flexibility for different modification types and experimental setups.

Key capabilities:

• Extraction of per-event signal features, normalisation parameters, and alignment data from FAST5 files processed with Tombo
• Signal discretisation into fixed-length histogram representations (default: 100 bins)
• Efficient dataset generation using TFRecord format
• Self-supervised pre-training via masked signal prediction
• Supervised fine-tuning using a bidirectional LSTM architecture
• Modification-specific classification at user-defined nucleotide targets (A, C, G, U)
• Generation of per-site predictions and ROC-based performance evaluation

---

Installation

Clone the repository

git clone https://github.com/Madhurananda/RNA_methylation_data_analysis.git
cd RNA_methylation_data_analysis

External tools required

Ensure the following tools are installed and accessible in your environment:

• Tombo – for signal alignment and normalisation
• minimap2 – for generating BAM alignment files

---

Dependencies

Core Python libraries:

• PyTorch – model architecture and training
• TensorFlow – TFRecord handling
• h5py – FAST5 file access
• pysam – BAM file parsing
• NumPy / pandas – data processing
• scikit-learn – evaluation metrics and ROC analysis
• matplotlib – visualization
• tqdm, natsort – utility functions

---

Pipeline Steps

1. FAST5 → TFRecord Generation

Convert FAST5 files and Tombo outputs into TFRecord shards.

Basic usage:

python generate_tfrecord_bp.py /path/to/fast5_folder /output/tfrecords_dir 8 1 /path/to/alignment.bam 200 20 RNA

Parallel processing across directories:

python do_generate_tfrec_BP_FT_MP.py /basecall_folder ALL_IN_DIR /analysis_folder 8 1 200 20 BP

---

2. Pre-training (Self-Supervised)

Train a bidirectional LSTM using masked signal reconstruction:

python TS_preTrainLSTM_bp.py MSE 3 512 0 1500 512

Output:

Bilst_bp.layers3.hs512.lr1500.b512.GPU.ep*.pt

---

3. Fine-tuning (Supervised)

Train the classifier for a specific nucleotide modification:

python TS_finetune_run_MP.py /data/p5/ /ref/IVT_seq.fa model_output 5 3 512 0 F 1500 1024 1 A A1

METHYL_TYPE options: A, C, G, U

---

4. Prediction & Evaluation

Run predictions on new data:

python TS_finetune_pred_run_MP.py /basefolder /pred_output_dir 3 512 0 512 saved_model.pt /ref_genome /test_data_folder A

Generate ROC curves:

python plot_ROC_allTools.py

---

Configuration

Key parameters are defined in TS_global.py:

Variable	Description	Default
dis_size	Number of histogram bins	100
input_len	Sequence window length	31
m_md_norm_limits	Signal normalization range	[-5, 5, 0.1]
tf_record_max_event	Maximum events per TFRecord	500000
g_dropout_rate	Dropout rate	0.2

Adjust these parameters according to your dataset characteristics before running the pipeline.

---

Outputs

The pipeline produces several outputs across different stages:

• TFRecord files – compressed .gz shards containing processed features
• Pre-trained models – .pt checkpoint files
• Fine-tuned models – trained classification models
• Prediction results – .predres files with per-site predictions
• ROC curves – .png performance comparison plots

---

Citation

If you use NanoCNN in your research, please cite:

Pahar, M. and Liu, Q., 2024, November. NanoCNN: a CNN-based tool to detect RNA m⁶A methylations using Oxford Nanopore sequencing. In 2024 4th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME) (pp. 1–6). IEEE.

@inproceedings{pahar2024nanocnn,
  title={Nano{CNN}: a {CNN}-based tool to detect {RNA} m$^{\text{6}}$A methylations using {O}xford {N}anopore sequencing},
  author={Pahar, Madhurananda and Liu, Qian},
  booktitle={2024 4th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME)},
  pages={1--6},
  year={2024},
  organization={IEEE}
}

---

Notes

• This pipeline assumes familiarity with Nanopore direct RNA sequencing workflows.
• Performance depends on data quality, alignment accuracy, and preprocessing consistency.
• Pre-training is optional but strongly recommended for improved downstream accuracy.

---