Abstract
The structure of RNA is deeply related to its function, and information about RNA substructure energy parameters is useful for predicting its structure from its sequence. RNA in cells is often modified, and these various types of modifications affect its structure and function. In recent years, the use of pseudouridine modifications in RNA vaccines has increased the importance of predicting structures that include modified bases. However, energy parameters of substructures involving modified bases have not yet been sufficiently determined. Therefore, in this paper, we propose a method for inversely calculating energy parameters from base-pairing probabilities. This method optimizes energy parameters using the same mechanism as gradient descent in deep learning. We also propose efficient computational approaches, including the calculation of the derivative of the partition function using a dynamic programming method following computations with the McCaskill algorithm. Because base-pairing probabilities can be obtained by adjusting them through chemical probing methods, it is expected that parameter estimation can be performed without relying on labor-intensive experiments or molecular dynamics simulations.