CAMPOLINA

Preprint | Weights

Sara Bakić^1,2, Krešimir Friganović¹, Bryan Hooi¹, Mile Šikić^1,3
1Genome Institute of Singapore (GIS), Agency for Science, Technology and Research (A*STAR), Singapore
²School of Computing (SOC), National University of Singapore, Singapore
³Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia

This is the official implementation of the paper "Campolina: A Deep Neural Framework for Accurate Segmentation of Nanopore Signals".

About

Nanopore sequencing provides real-time, long-read sequencing, enabling the direct analysis of nucleic acids. The real-time methods allow parallel execution of sequencing and analysis by processing raw nanopore signals, the immediate product of nanopore sequencing, at the same rate as they are produced. The core step of real-time frameworks, segmentation of raw nanopore signals into events corresponding to individual nucleotide translocations, is traditionally done with an algorithmic approach, which is insufficiently robust in noisy regions of the signal. We propose Campolina, a first deep-learning framework for accurate segmentation of raw nanopore signals. Campolina is a convolution-based model trained to identify event borders in raw nanopore signals. We extensively compare Campolina with the traditional Scrappie algorithm on R9.4.1 and R10.4.1 datasets. To evaluate the quality of the segmentation, we develop an evaluation pipeline that quantifies how well the obtained events match the ground truth and the corresponding reference. Additionally, we test Campolina as a segmentation strategy in the existing real-time frameworks, Sigmoni and RawHash2. Campolina significantly enhances the quality of segmentation compared with the algorithmic approach, and improves the performance of real-time frameworks, all while maintaining time efficiency, thereby increasing the potential of effective real-time analyses.

Quick Start - Inference

Use following commands for the installation (Prerequisites: Python>=3.8 and CUDA>=11.8):

git clone https://github.com/lbcb-sci/Campolina
cd Campolina
pip install .

Installation

1. Clone the repo.

git clone https://github.com/lbcb-sci/Campolina
cd Campolina

2. Create conda environment. All external dependencies should be contained in environment.yml.

# create conda environment for Campolina
conda env create -f environment.yml

# activate Campolina environment
conda activate campolina

3. Download pre-trained weights.

mkdir weights
cd weights
wget https://zenodo.org/records/15626806/files/R10_model.pth # model trained for R10.4.1 nanopore version
wget https://zenodo.org/records/15626806/files/R10_model.pth # model trained for R9.4.1 nanopore version

Usage

We provide pre-trained Campolina models for R10.4.1 and R9.4.1 nanopore version. Both mdoels follow the same architecture design and can be used for predicting event borders from raw nanopore signals using inference.py script. The details of running the inference can be checked with --help.

Evaluation

Within the scope of this paper, we develop an extensive pipeline for assessing the quality of the obtained segmentation. The assessment of the segmentation quality can be with the following steps:

1. Extract ground-truth.

bash extract_ground_truth.sh

Assess the segmentation quality

To run the full segmentation quality assessment pipeline, the following steps need to be done:

1. Extract additional information for the predicted events with convert_parquet_for_analysis.py. The exact usage can be checked with --help.
2. Align the predicted segmentation to the ground truth segmentation with align_events.py. The exact usage can be checked with --help.
3. Assess the quality of predicted segmentation with assess_segmentation_quality.py. The exact usage can be checked with --help.

Contact

If you have any questions, please feel free to email the authors or open an issue.

Acknowledgment

This research is supported by the Singapore Ministry of Health’s National Medical Research Council under its Open Fund – Individual Research Grants (NMRC/OFIRG/MOH-000649-00).