A major challenge to the characterization of intrinsically disordered regions (IDRs), which are widespread in the proteome, but relatively poorly understood, is the identification of molecular features, such as short motifs, amino acid repeats and physicochemical properties that mediate the functions of these regions. Here, we introduce a proteome-scale feature discovery method for IDRs. Our method, which we call "reverse homology", exploits the principle that important functional features are conserved over evolution as a contrastive learning signal for deep learning: given a set of homologous IDRs, the neural network has to correctly choose a randomly held-out homologue from another set of IDRs sampled randomly from the proteome. We pair reverse homology with a simple architecture and interpretation techniques, and show that the network learns conserved features of IDRs that can be interpreted as motifs, repeats, and other features. We also show that our model can be used to produce specific predictions of what residues and regions are most important to the function, providing a computational strategy for designing mutagenesis experiments in uncharacterized IDRs. Our results suggest that feature discovery using neural networks is a promising avenue to gain systematic insight into poorly understood protein sequences.