Learning to fold a random RNA chain

Abstract

We implement a neural network associating a two-state unit to each pair of bases in a random RNA sequence. A Watson—Crick base-pair becomes the structural realization of an activated unit. Accordingly, a secondary structure is associated to a neutral pattern. The coupling tensor is constructed taking into account the cooperativity principles of secondary structure formation, including nucleation effects for initiation of intra-chain duplex formation and stacking of base pairs. The dynamics of the exploration of metastable patterns by the network is shown to follow a random energy model. This model has already been supported experimentally and has been established as a paradigm to explain the initial stages of exploration of conformation space by a random RNA molecule.