Abstract
RNA-small molecule binding is a key regulatory mechanism which can stabilize 3D structures and activate molecular functions. The discovery of RNA-targeting compounds is thus a current topic of interest for novel therapies. Our work is a first attempt at bringing the scalability and generalization abilities of machine learning methods to the problem of RNA drug discovery, as well as a step towards understanding the interactions which drive binding specificity. Our tool, RNAmigos, builds and encodes a network representation of RNA structures to predict likely ligands for novel binding sites. We subject ligand predictions to virtual screening and show that we are able to place the true ligand in the 71st-73rd percentile in two decoy libraries, showing a significant improvement over several baselines, and a state of the art method. Furthermore, we observe that augmenting structural networks with non-canonical base pairing data is the only representation able to uncover a significant signal, suggesting that such interactions are a necessary source of binding specificity. We also find that pre-training with an auxiliary graph representation learning task significantly boosts performance of ligand prediction. This finding can serve as a general principle for RNA structure-function prediction when data is scarce. RNAmigos shows that RNA binding data contains structural patterns with potential for drug discovery, and provides methodological insights for possible applications to other structure-function learning tasks. The source code, data and a Web server are freely available at http://rnamigos.cs.mcgill.ca.