Kaggle-OpenVaccine

Kaggle competition "OpenVaccine: COVID-19 mRNA Vaccine Degradation Prediction"

This is my solution (top 19%) to the Kaggle competition "OpenVaccine: COVID-19 mRNA Vaccine Degradation Prediction" https://www.kaggle.com/c/stanford-covid-vaccine

Data

• Attributes: RNA sequence and structure at each position
  • RNA sequence
  • RNA structure
  • RNA predicted loop type
  • RNA pairing probability matrix (BPPS)
• Target: Degradabilities under five conditions at each position

Metrics

• Mean-columnwise RMSE (MCRMSE)

Models

• RNN-based sequence model
• Multi-layer bi-directional GRU/LSTM layers
• Embedding layers to translate RNA sequence and structure (14 tokens) into dense features
• Additional BPPS and RNA-based features
• Model blending from different cross validation training sets

Training

• Weighted training using the signal-to-noise ratio as sample weight
• 50-100 epochs
• 4-fold cross validation

Results

• Public leaderboard: 0.2393 (ranked 312/1636, top 19%)
• Private leaderboard: 0.3556 (ranked 300/1636, top 19%)

Software package

• PyTorch on Google Colab

Ideas tried but not worked

• Add Conv1d layer with batch/layer normalization
• Use k-mer to encode RNA