This is the enhanced version of nucleic transformer proposed by He et al. (2021)

You can obtain the datasets from the link below:

1. https://www.kaggle.com/c/stanford-covid-vaccine/data (Original dataset)
2. https://www.kaggle.com/datasets/shujun717/openvaccine-12x-dataset (Augmented dataset)

Afterwards, unzip them all in a single folder "data"

The experiment is conducted under virtual environment created using Anaconda. You can create your own environment and ensure these packages are installed by using pip install <package name> command.

These are the mandatory packages:

torch 1.9.0 or <any version>
tqdm <any version>
ranger-adabelief 0.1.0
adabelief-pytorch 0.2.1
pandas 1.1.2
numpy 1.19.2

Alternatively, You can import my environment directly by using the code below to save your time and avoid any package dependency issue.

conda env create -f environment.yml

then you can activate the environment, but make sure you check the env name by using the code below

conda env list

Here I include the hypeparameters that give the best single model.

1. pretrain with all available sequences: ./pretrain.sh

2. train on targets: ./run.sh

3. to make predictions and generate a csv file for submission: ./predict.sh then you can make a submission at https://www.kaggle.com/c/stanford-covid-vaccine/submissions

Run the script under Anaconda Prompt, with the correct environment.

He, S., Gao, B., Sabnis, R., Sun, Q. (2021). Nucleic Transformer: Deep Learning on Nucleic Acids with Self-Attention and Convolutions. 10.1101/2021.01.28.428629. https://www.biorxiv.org/content/10.1101/2021.01.28.428629v1 Original Code: https://github.com/Shujun-He/Nucleic-Transformer