RNA Vaccine Degradation Prediction using Deep Neural Networks

13th place solution for OpenVaccine: COVID-19 mRNA Vaccine Degradation Prediction challnge organized by Stanford and Kaggle.

Getting Started

A summarized description of the approach can be found here.

Prerequisites

• Python3
• Arnie
• bpRNA
• Pytorch
• Scikit-learn
• tqdm

Arnie and bpRNA folders should be in the same directory as the source code.

Running

Extra Dataset Generation

python3 generate_extra_dataset.py --data_path ... --package ... --temp ... --n_threads ...

This step generates a number of secondary structures per sequence (default is 5) using a package from Arnie to be used for augmentation. You need to choose a package, a temperature at which bpp matrices are generated and the number of threads to use in parallel.

Single Model Training

python3 main.py --data_path ... --arch ... --package ... --temp ...

This step trains a single model with 5-fold validation and creates a submission file. You need to choose an architecture (--arch) from cnn, cnn_lstm or cnn_lstm_transformer.

Ensemble

To create an ensemble, you need to generate structures from different packages using differnet temperatures, train different architecures with them and finally average the resulted submissions.