Designing messenger RNA (mRNA) sequences that encode a desired protein means choosing from a vast space of synonymous codon sequences, each with different biological properties. This is a hard optimization problem: reinforcement learning methods tend to collapse onto a few solutions, while standard GFlowNets struggle with sparse rewards over long sequences.
The authors propose Curriculum-Augmented GFlowNets (CAGFNs), which train a multi-objective GFlowNet using a curriculum driven by learning progress. The model starts with shorter target proteins and gradually moves to longer ones, making the exploration problem more tractable. They also introduce CodonDesignEnv, a codon-level environment that uses dynamic action masking to preserve the target protein's amino-acid identity.
In experiments, CAGFNs converged up to four times faster than non-curriculum GFlowNets and achieved competitive or stronger Pareto-front coverage and diversity. The generated sequences were 100% unique, in contrast to RL baselines that suffered from mode collapse. The paper is an arXiv preprint, so the results have not yet undergone peer review, but the approach points toward more scalable generative methods for mRNA synthesis.