Core supervised ML methods — bias-variance trade-off, model selection, and hyperparameter optimisation — applied to chemical property prediction.
Along the way: a full worked case study on reaction-yield prediction, a real example of data bias becoming model bias, missing data in a chemistry context, and three baseline models (linear regression, Gaussian process regression, decision trees) worth always trying.
Takeaway 1. Supervised learning is learning from labeled chemistry data — how you evaluate it matters as much as the model itself.
Takeaway 2. Split by scaffold, not randomly — interpolation is easy, extrapolation isn't. Molecules cluster into structural families, and a random split puts near-duplicates of the same scaffold on both sides of the split, making performance look better than it really is.
Takeaway 3. Prediction error decomposes into bias² + variance + irreducible noise — learning curves (training vs. validation error vs. training-set size) tell you which one is hurting your model.
Takeaway 4. Parameters are learned from data; hyperparameters are tuned on validation data, never on the test set. Touch the test set exactly once, at the very end, for an honest final number.
Takeaway 5. Data bias becomes model bias — and it's invisible in a standard train/test split because the bias is present on both sides. In the NiCOlit reaction-yield case study, a trivial baseline that just knows which table a reaction came from performed almost as well as the trained model, because the model was mostly learning the data-collection process, not the chemistry.