A hyperparameter is a configuration value chosen outside model fitting. It controls learning, preprocessing, featurization, regularization, or model complexity. Tuning searches for settings that avoid both underfitting and overfitting while preserving performance on new molecules. section 3.4.3 section 4.4.3
Common molecular machine-learning hyperparameters include:
- Morgan fingerprint radius and bit count, such as radius 2 or 4 and 1,024 or 2,048 bits. section 3.4.3
- Regularization choices and strengths, including L1, L2, elastic net, and the alpha parameter. section 3.4.3 section 5.3.1
- Model-complexity controls, including polynomial degree, SVR epsilon-tube width, and decision-tree depth or leaf limits. section 4.4.3 section 5.3.1
Scikit-Learn pipelines combine transformers for standardization, molecule processing, and feature generation with a final estimator. GridSearchCV or RandomizedSearchCV evaluates candidate combinations through cross-validation, reducing repetitive code and helping prevent preprocessing leakage. section 3.4.3
Searching fingerprint and SGD settings
A pipeline can search fingerprint radius, fingerprint size, penalty, and alpha while scoring five-fold cross-validation with F1. One grid search selected 2,048 bits, radius 2, L2 penalty, and alpha 0.1, achieving a cross-validated score of 0.8809.

























