Answers are grounded in this paper and cite their sources.
Unsupervised learning is used throughout computational molecular science on unlabeled data such as molecular dynamics trajectories and conformer ensembles. Some questions admit no predefined labels at all: how many conformational states a molecule visits, whether a virtual library covers chemical space, or which structures are unlike anything else in the set. Clustering, dimensionality reduction, and density estimation serve both cases. They are easy to apply, and their output generally appears like a discovery, whether or not it is one. Establishing a new discovery is difficult, because there is no held-out target to check against. This guide is a prescriptive companion to the review literature: it sets out what to do at each stage of the pipeline these analyses share (curate, split, represent, scale, reduce, group, validate), why each step matters, and what goes wrong when it is skipped, with a checklist for each stage. The representation, the distance metric, and the scaling usually decide the outcome, not the chosen algorithm. An apparently structured low-dimensional map is not evidence that structure exists, and null models, stability analysis, and cheap baselines should be used. Statistical practice in molecular machine learning is not widespread, so we give a comparison protocol covering resampling variability, dependence between repeated estimates, multiple comparisons, and post-selection effects. Leakage and splitting are treated in depth, and we provide accompanying notebooks to run each recommendation on molecular data.
No takes yet. Share an insight, caveat, or question.
Barrett et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: