Mechanisms of Action (MoA) Prediction

• Data Description

• Notebook link

Quick View

Goal:

• Predict MoA (only 0 and 1) from data of gene and cell vitality, build multiple binary classifier

Metrics:

• Minimum Log loss

Procedure:

• MoA tag are extremely imbalanced, average 89 positive tags in each column from 21K entries, so a special kfold function has beed used

  from iterstrat.ml_stratifiers import MultilabelStratifiedKFold
  

• Label Smoothing has been conducted to help improve accuracy in this multiple output case A great explanation here

• Use Pipeline function to automatically finished column transform
  • Numerical data
    • Quantile transform
    • PCA
  • Categorical
    • One-hot-labeling
• Build DNN model with 3 NN layer with appropriate regulation.
• Use keras.tuner to do hyper parameters tuning and code is in Notebook

Links

Github