XNA Basecaller

Code for paper: "Direct high-throughput deconvolution of unnatural bases via nanopore sequencing and bootstrapped learning".

In our work we show how to achieve high-throughput sequencing of DNA containing Unnatural Bases (UBs), a.k.a Non-Canonical Bases (NCBs), using Nanopore and de novo basecalling enabled by spliced-based data-augmentation. The code here contains a basecaller architecture modified for learning to also basecall 1 or 2 additional UBs, and includes real-time data-augmentation for generating train data with UBs in all possible sequencing contexts.

More details in the preprint: https://biorxiv.org/cgi/content/short/2024.12.02.625113v1

Setup

Installation

Full installation should take only a few minutes (<5 mins).

Download Minimap or create a symbolic link to it at bin/minimap2:

curl -L https://github.com/lh3/minimap2/releases/download/v2.17/minimap2-2.17_x64-linux.tar.bz2 | tar -jxvf - minimap2-2.17_x64-linux/minimap2
mv -v minimap2-2.17_x64-linux bin

Install python enviroment:

conda env create -f env.yml
conda activate xna_bc

Install UB-Bonito. Recommended to install with xna_bc activated to ensure valid python version (v3.9) is utilized.

cd ub-bonito
python3 -m venv venv3
source venv3/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
# pip install -r requirements-cuda113.txt --extra-index-url https://download.pytorch.org/whl/cu113/
python setup.py develop
deactivate

Download Data

Run ./download_data.sh script from the project root directory to download all the necessary files (Total 11GB). Might take several minutes (>10 mins):

• XNA basecaller baseline model (pre-trained with DNA only)
• Evaluation reads from POC and Complex libraries (fast5 files)
• Pre-processed DNA and XNA train data files (e.g. chunks, references and segmentations)

Usage

In order to run the following scripts, make sure the correct conda enviroment is activated (conda activate xna_bc) and ub-bonito venv3 is deactivated.

HINT: Make sure bash files have execute permission (chmod +x *.sh)

Training

Use train_and_eval.sh to run basecaller training and evaluation. See file for more options available.

Quick runs (around 20 mins or less):

• Fully-synthetic - ~25% UB Acc.:
  • ./train_and_eval.sh POC training/fully_synth-ubs_X-data_0.01-ub_prop_0.10 ub-bonito/bonito/data/dna_r9.4.1/sampled_0.01/ 0.10 -u X -b 98 -e 1 -Z
• Hybrid - ~10% UB Acc.:
  • ./train_and_eval.sh POC training/hybrid-ubs_X-data_0.01-ub_prop_0.10 ub-bonito/bonito/data/dna_r9.4.1/sampled_0.01/ 0.10 -u X -b 98 -e 1
• Spliced - ~15% UB Acc.:
  • ./train_and_eval.sh POC training/spliced-ubs_X-data_0.01-ub_prop_0.10 ub-bonito/bonito/data/dna_r9.4.1/sampled_0.01/ 0.10 -u X -b 98 -e 1 -m per_kmer

Final models (should take a few hours), 70-80% UB Acc.:

for UBS in X Y XY; do
./train_and_eval.sh CPLX training/spliced-ubs_$UBS-data_0.25-ub_prop_0.09-unfr_3 ub-bonito/bonito/data/dna_r9.4.1/sampled_0.25/ 0.09 -u $UBS -W -b 98 -m per_kmer -f -F 3 -E POC;
done

NOTE: Spliced method real-time train data generation requires many workers to avoid being the bottleneck depending on the batch size chosen. Recommended 32 CPUs (argument -w 32) and larger batch size (-b 512) for faster training.

Evaluation

Use eval_model.sh to run ad-hoc basecaller evaluation. See file for more options available.

• ./eval_model.sh POC training/fully_synth-ubs_X-data_0.01-ub_prop_0.10 -b 98 -u X

Additional Tools

Compare basecalling performances src/tools/comp_basecalls_perf.py

• Ex: python src/tools/comp_basecalls_perf.py training/*
• Use -d argument to output detailed performance per template and UB position

Segment train data chunks src/tools/dtw_segmentation.py

• Employs DTW to group signals from the same kmer based on the signal reference model
• Outputs "breakpoints.npy", file required by the train data generator embedded in ub-bonito
• Tool used to estimate breakpoints from pre-processed DNA and XNA train data
• Ex: python src/tools/dtw_segmentation.py ub-bonito/bonito/data/dna_r9.4.1/sampled_0.01/ -p --pool_chunksize 1

Results

Final model performance on POC library:

Method	UB(s)	UB(s) Acc.	DNA Acc.
Spliced	X	77%	91%
Spliced	Y	81%	92%
Spliced	XY	71%	92%

(Potential) Future updates

• UB kmer modeling script and tools
• End-to-end framework/pipeline description and/or script
• More utility tools

Citation

Preprint: https://biorxiv.org/cgi/content/short/2024.12.02.625113v1

@article{perez2024,
    author={Mauricio Perez and Michiko Kimoto and Priscilla Rajakumar and Chayaporn Suphavilai and Rafael Peres da Silva and Hui Pen Tan and Nicholas Ting Xun Ong and Hannah Nicholas and Ichiro Hirao and Chew Wei Leong and Niranjan Nagarajan},
    title={Direct high-throughput deconvolution of unnatural bases via nanopore sequencing and bootstrapped learning},
    journal={bioRxiv}, 
    year={2024},
    doi={10.1101/2024.12.02.625113}
}