Abstract
Designing RNA molecules has garnered recent interest in medicine, syntheticbiology, biotechnology and bioinformatics since many functional RNA moleculeswere shown to be involved in regulatory processes for transcription,epigenetics and translation. Since an RNA's function depends on its structuralproperties, the RNA Design problem is to find an RNA sequence which satisfiesgiven structural constraints. Here, we propose a new algorithm for the RNADesign problem, dubbed LEARNA. LEARNA uses deep reinforcement learning to traina policy network to sequentially design an entire RNA sequence given aspecified target structure. By meta-learning across 65000 different RNA Designtasks for one hour on 20 CPU cores, our extension Meta-LEARNA constructs an RNADesign policy that can be applied out of the box to solve novel RNA Designtasks. Methodologically, for what we believe to be the first time, we jointlyoptimize over a rich space of architectures for the policy network, thehyperparameters of the training procedure and the formulation of the decisionprocess. Comprehensive empirical results on two widely-used RNA Designbenchmarks, as well as a third one that we introduce, show that our approachachieves new state-of-the-art performance on the former while also being ordersof magnitudes faster in reaching the previous state-of-the-art performance. Inan ablation study, we analyze the importance of our method's differentcomponents.