STarFish

STarFish predicts protein targets for small molecules to identify protein–small molecule interactions and prioritize targets for experimental validation.

Key Features:

• Input and output: Accepts a chemical structure as input and returns a ranked list of potential protein targets with probability scores.
• Model stacking approach: Uses a stacked ensemble combining k-nearest neighbors, random forest, and multilayer perceptron base models with logistic regression as a meta-classifier.
• Performance metrics (development): Individual model performance assessed by stratified 10-fold cross-validation with AUROC scores of 0.94–0.99 and BEDROC scores of 0.89–0.94.
• Natural product dataset curation: Curated a dataset of 5,589 compound–target pairs from 1,943 unique compounds and 1,023 unique targets by cross-referencing 20 public natural product databases with the ChEMBL bioactivity database.
• Natural product performance and improvement: Initial testing on natural products showed AUROC 0.70–0.85 and BEDROC 0.43–0.59, and the stacking approach improved performance to AUROC 0.94 and BEDROC 0.73.
• Training data (synthetic): Trained on a synthetic dataset of 107,190 compound–target pairs from 88,728 unique compounds and 1,907 unique targets.

Scientific Applications:

• Drug discovery and development: Prioritizes bioactive small molecules, including natural products, for experimental target validation and target deconvolution to guide downstream pharmacological studies.

Methodology:

Training used a stacked ensemble of k-nearest neighbors, random forest, and multilayer perceptron models with logistic regression as the meta-classifier, evaluated by stratified 10-fold cross-validation on synthetic and curated natural product datasets.