Supervised machine learning

We have developed a supervised machine learning tool called Target Mate (TM). Code for this tool can be found in the CC repository: tool/targetmate/.

TM uses CC signatures as features, similar to the use of structural descriptors in structure-activity relationship (SAR) studies. TM was first presented in this article.

Machine-learning problem

• Binary classification
• Single-output regression

Algorithms

• Vanilla algorithms
• Automated
  • Randomized search
  • HyperOpt
  • TPOT
  • AutoSklearn

Confidence

Conformal prediction

Featurizers

• Classical Morgan Fingerprint
• Stacked CC signatures
• Ensemble of CC signatures

Train/test splits

• Random
• Stratified
• Scaffold-based

Model explanation

• Shapley analysis

Negative sampling

• Random
• Diversity oriented and reliable negatives

Parsers for tasks

• MoleculeNet
• ChEMBL