Gradient-based Optimization for mRNA Sequence Design

Abstract

MotivationOptimization of mRNA sequences presents fundamental challenges to balance multiple physicochemical/biological properties--including accessibility, stability, and translation efficiency--while preserving amino acid sequences. The discrete nature of RNA sequence design hinders direct application of gradient-based methods, despite their potential for leveraging modern deep learning predictors of the properties in biological sequence design. ResultsWe present the Input Data Differentiable Designer (ID3) framework, a unified computational approach for mRNA sequence optimization that enables gradient-based optimization of discrete RNA sequences through innovative mathematical techniques. ID3 framework encompasses 12 constrained variants across four base configurations and three constraint mechanisms: Codon Profile Constraint, Amino Matching Softmax, and Lagrangian multipliers. The ID3 framework treats trained models as fixed differentiable functions while optimizing input data through continuous probability distributions. We also provide convergence analyses from the perspective of trained model input optimization. Availability and implementationhttps://github.com/Li-Hongmin/ID3.git Contactlihongmin@edu.k.u-tokyo.ac.jp, terai@edu.k.u-tokyo.ac.jp, asai@k.u-tokyo.ac.jp