Course project for 11-785: Deep Learning Applications in COVID-19.
Creating a sequence model using recurrent neural networks to identify genomic divergence, mutations, and clade in COVID-19 spike protein sequences.
Team members: Abbey Pei, Amelia Kuang, Jeff Tan, Jenny Ding, Sylvia Zhang
TA mentor: Joseph Konan
SARS-CoV-2 is an RNA virus that has caused over 3 million deaths and infected over 150 million worldwide. With thousands of strains sequenced to date, SARS-CoV-2 poses significant challenges to scientists on keeping pace with lab diagnostic design, clinical treatment options, and vaccine development. Therefore, an efficient method of identifying mutations, divergence, and clade of lab samples from patients would have a significant impact on the documentation of SARS-CoV-2 genomics. These developments help researchers observe trends in the pandemic over time and respond accordingly. This study tackles directly the concerns of SARS-CoV-2, using recurrent neural networks to classify protein sequences and predict their mutations.
We use the following Python packages in our project:
- biopython: For data processing and parsing FASTA protein sequence files
- csv: For data processing and parsing metadata files
- imbalanced-learn: For data downsampling using RandomUnderSampler
- matplotlib: For plotting accuracy, loss, and other model metrics
- numpy: For data processing and analysis
- pandas: For data analysis
- scikit-learn: For embedding visualization using T-SNE and PCA
- torch: For deep learning model development
This repository contains the following files:
- data
- amino_list.txt: Mapping from single-letter amino acid codes to integers
- clade_mapping.npy: Mapping from clade labels to integers
- metadata.zip: Metadata file from GISAID, containing data labels for each sample in the dataset
- pango_mapping.npy: Mapping from pango labels to integers
- spikeprot0309.zip: Spike protein sequence file from GISAID, containing spike protein sequences for each sample in the dataset
- data_old: This directory contains our original and much smaller dataset, with metadata sourced from NextStrain instead of GISAID.
- nextstrain_original: Original NextStrain dataset, obtained by mapping between protein sequence, metadata, and timetree files
- nextstrain_updated: Updated NextStrain dataset, obtained by mapping between just protein sequence and metadata files
- data_preprocess_old.ipynb: Old data processing notebook, used to produce nextstrain_updated dataset
- nextstrain_ncov_global_metadata.tsv: Metadata from NextStrain dataset
- nextstrain_ncov_global_timetree.nexus: Timetree from NextStrain dataset
- figures: This directory contains selected figures from our final report
- models
- baseline_model.ipynb: Implementation of our baseline model with CNN, LSTM, and fully connected layers
- model_2cnn_3fc_balancedData.ipynb: Modification of our baseline model with additional CNN and fully connected layers
- model_2cnn_3fc_bert.ipynb: Implementation of our Bert Embedding with Hydrophilicity Encoding model
- model_pyramidal_lstm.ipynb: Implementation of our Pyramidal LSTM model
- data_preprocess.ipynb: Produces raw training / validation data from GISAID protein sequence and metadata files
- embedding_visualization.ipynb: Uses T-SNE and PCA from scikit-learn to visualize the embeddings and input data of our model
- sequence_balance.ipynb: Takes raw training / validation data produced by data_preprocess.ipynb, and downsamples the dataset to produce balanced data with an equal number of items per label