What is the difference between supervised and unsupervised learning in a biology project?

Why

The scikit-learn user guide is organized around this split. Supervised methods include classifiers and regressors trained on pairs of inputs and known outputs. A model trained on gene expression from tumors with known treatment response can predict response for a new tumor, and its accuracy can be measured against held-out labels.

Unsupervised methods include clustering, mixture models and dimensionality reduction such as PCA. Clustering single-cell profiles groups cells by similarity, but nothing in the algorithm says the clusters are cell types. They might reflect cell cycle, stress from dissociation or a batch effect, so the result needs marker genes and experiments to interpret.

That difference drives how each is judged. Supervised models have a clear target and an error rate. Unsupervised results have no ground truth, so stability across parameters and subsamples and agreement with known biology matter more.

The categories blur in practice. Semi-supervised learning uses a few labels with many unlabeled samples, and self-supervised pretraining on unlabeled sequences or images underlies many modern models. A common pipeline uses both: unsupervised PCA to explore and spot batch effects, then a supervised model for the actual prediction.