We developed an analytical framework for unsupervised learning with coarse-grained sequence features, rather than linear sequence, to predict activities of intrinsically disordered regions (IDRs), which presents a major advance in how we correlate sequence features to function.