RekerLab / YokedLearning

:woman_teacher: :arrow_right: :woman_student: Active learning teaching other models

Yoked machine learning utilizes a teacher model to guide a student model. We provide an example pipeline to evaluate yoked learning performance on both classical (part 1) and deep (part 2) machine learning models.

Benchmarking Datasets

• Therapeutics Data Commons

  • ADME: Pharmaco-kinetics (from tdc.single_pred import ADME)
    • CYP2C9 Substrate, Carbon-Mangels et al.
      • data = ADME(name = 'CYP2C9_Substrate_CarbonMangels')
    • CYP2D6 Substrate, Carbon-Mangels et al.
      • data = ADME(name = 'CYP2D6_Substrate_CarbonMangels')
    • CYP3A4 Substrate, Carbon-Mangels et al.
      • data = ADME(name = 'CYP3A4_Substrate_CarbonMangels')
    • HIA (Human Intestinal Absorption), Hou et al.
      • data = ADME(name = 'HIA_Hou')
    • Pgp (P-glycoprotein) Inhibition, Broccatelli et al.
      • data = ADME(name = 'Pgp_Broccatelli')
    • Bioavailability, Ma et al.
      • data = ADME(name = 'Bioavailability_Ma')
  • Tox: Toxicity (from tdc.single_pred import Tox)
    • hERG blockers, Wang et al.
      • data = Tox(name = 'hERG')
    • DILI (Drug Induced Liver Injury), Xu et al.
      • data = Tox(name = 'DILI')
    • Skin Reaction, Alves et al.
      • data = Tox(name = 'Skin Reaction')
    • Carcinogens, Lagunin, et al.
      • data = Tox(name = 'Carcinogens_Lagunin')
    • Clintox, Gayvert, et al.
      • data = Tox(name = 'ClinTox')
  • HTS: High-Throughput Screening (from tdc.single_pred import HTS)
    • SARS-CoV-2 3CL Protease, Diamond
      • data = HTS(name = 'SARSCoV2_3CLPro_Diamond')

• MoleculeNet

  • BACE: Quantitative (IC50) and qualitative (binary label) binding results for a set of inhibitors of human β-secretase 1(BACE-1)
  • BBBP: Binary labels of blood-brain barrier penetration(permeability)

Dependencies

• numpy
• scipy
• pandas
• scikit-learn
• DeepChem
• matplotlib
• seaborn
• chemprop v1.5.2
• MolALKit

Files

Part 1: Classic Yoked Learning

• Code and functions to evaluate yoked learning with classical machine learning models (random forest, naive bayes and logistic regression).
  • yoked_machine_learning_pipeline.py contains functions for evaluating yoked learning
  • yoked_learning_main.py contains the main function to run yoked learning
  • example boxplot/lineplot.ipynb contains an example notebook that visualize comparisons between yoked learning, active learning, and passive learning

Part 2: Deep Yoked Learning

• Code and functions to evaluate yoked learning with deep learning models (MLP).
  • Implementation methods include MolALKit, default MLP parameters (ffn_num_layers: 2, ffn_hidden_size: 300, dropout: 0, batch_size:50) or optimized MLP parameters chemprop
  • Single_dataset_comparison.ipynb contains an example notebook that visualizes the output comparisons
• Please refer to MolALKit for details about Deep Yoked Learning implementations
• Example implementation after data split:

molalkit_run --data_public bace --metrics roc-auc --learning_type explorative --model_config_selector RandomForest_RDKitNorm_Config \
    --split_type scaffold_order --split_sizes 0.5 0.5 --evaluate_stride 100 --seed 0 --save_dir bace_rf_yoked_mlp --n_jobs 4 \
    --model_config_evaluators MLP_RDKitNorm_BinaryClassification_Config