RNA design describes the problem of finding an RNA nucleotide sequence, given a representation of the structure.
Since RNA is one of the most important regulators in cells, artificially designed RNAs could for example be used to develop on-demand RNA-based therapeutics.
We develop LEARNA, an automated reinforcement learning (AutoRL) approach for the design of RNA.
Given a target RNA secondary structure, LEARNA generates an RNA sequence that is then evaluated for its folding into the target structure.
Methodolically, LEARNA uses a joint architecture and hyperparameter search (JAHS) approach combined with meta-learning, in order to find the best performing reinforcement learning (RL) system for the given application.
In particular, LEARNA employs an efficient Bayesian Optimization method, BOHB, to jointly optimize a rich configuration space, including elements of recurrent neural networks (RNNs), convolutional neural networks (CNNs), and fully-connected layers, parameters of the MDP like the shape of the reward function and the size of the state space, as well as training hyperparameters like the batch size.
Each configuration defines a specific RL agent and environment used to meta-learn an RNA design policy across thousands of RNA design tasks. The best performing configuration with respect to the validation loss is finally selected for testing.
By sampling RNA sequences from the learned policy for a given RNA structure, LEARNA achieves massive speed-ups and yields better performance compared to other methods in the field.