PorTis

This repository contains the code for the paper "PorTis: A biological phenotype classification dataset for benchmarking nanopore signal-based protein analysis". It contains code for loading, analyzing, and preprocessing the PorTis dataset as well as the tissue classification model training and inference code.

Authors: Cailin Winston, Marc Expòsit, Jeff Nivala

Structure of the repository

• /src contains core functionality for loading/analyzing data and for running benchmarks
  • constants.py: constant definitions to load and save data files
  • kmedoids_mod.py: modified version of scikit-learn-extra's k-medoids function to cluster signals and return centroids
  • knn_classifier.py: functions required to train and test the kNN classifier
  • knn_load_data.py: functions to load and save signals, labels, and the distance matrix for the kNN classifier
  • metrics.py: functions to calculate metrics of the CNN models
  • models.py: model architecture of the CNN classifiers
  • signal_processing.py: functions used to process signals before use in the classifiers
  • utils_data.py: functions used to load data for the CNN classifiers
  • utils_models.py: helper functions to train the CNN classifiers
• /notebooks contains interactive notebooks for the empirical analysis and clustering of the dataset
  • 1.0.DataAnalysis.ipynb: Exploratory data analysis of signal number and length by replicate and tissue type
  • 2.0.DataPrepkNN.ipynb: Balancing data across replicates and tissues to use in kNN classification
  • 2.1.DTWmatrixCalc.ipynb: Calculating the distance matrix between signals using Dynamic Time Warping (DTW)
  • 3.0.ClustAllSignals.ipynb: t-SNE representation of all signals
  • 4.0.kNNinform.ipynb: kNN classification and identification of signals unique to each tissue
  • 5.0.InformativeSignalClustering.ipynb: Analysis and clustering of informative signals
• /scripts contains python scripts to regenerate test indices and to run the CNN classification benchmarks
  • gen_test_indices.py: generate the indices to divide the data between training and testing sets for use in the CNN classifier
  • 1d_cnn.py: training and inference of the 1D CNN classifier
  • 2d_cnn.py: training and inference of the 2D CNN classifier

Installation and setup

First clone the repository:

git clone [email protected]:uwmisl/PorTis.git

Then, setup the Python environment with the packages/dependencies needed to run the code in this repository:

cd PorTis
conda env create -f env.yml

Then, activate the conda environment:

conda activate portis

Ensure that you have Jupyter notebook to run the notebooks in /notebooks.

Usage

To run any of the notebooks, open them in Jupyter notebook and follow the instructions/comments in the cells.

To run the benchmarks for the signal event classification (Task 1) and "real-time" sample classification (Task 2), run either 1d_cnn.py or 2d_cnn.py from the /scripts directory.

Any combination of the flags --to_train, --to_eval, and --to_sim can be specified, as long as at least one is specified. Running with the --to_train flag will train a new model. Running with the --to_eval flag with evaluate the model on the specified test set; either --to_train must also be specified to train a new model to evaluate or a --model_path to an already trained model can be provided. Running with the --to_sim flag will run the "real-time" sample classification simulation.

python 1d_cnn.py [-h] [--base_dir BASE_DIR] [--train_test_split_id TRAIN_TEST_SPLIT_ID]
    [--length_thresh LENGTH_THRESH] [--to_train] [--to_eval] [--to_sim] [--batch_size BATCH_SIZE]
    [--lr LR] [--epochs EPOCHS] [--model_path MODEL_PATH]

python 2d_cnn.py [-h] [--base_dir BASE_DIR] [--train_test_split_id TRAIN_TEST_SPLIT_ID]
    [--length_thresh LENGTH_THRESH] [--to_train] [--to_eval] [--to_sim][--batch_size BATCH_SIZE]
    [--lr LR] [--epochs EPOCHS] [--rescale_dim RESCALE_DIM] [--stack_depth STACK_DEPTH]
    [--model_path MODEL_PATH]