AI and the World's Hard Problems · Bioengineering & synthetic biology · arXiv · 2026-10-21
Curriculum-Augmented GFlowNets For mRNA Sequence Generation
De novo design of mRNA sequences encoding target proteins requires optimizing multiple biological objectives over large combinatorial spaces of synonymous codons, a setting where reinforcement learning suffers from mode collapse and GFlowNets face sparse rewards over long horizons. Curriculum-Augmented GFlowNets (CAGFNs) are a multi-objective GFlowNet trained with an adaptive length-based curriculum driven by learning progress, learning from progressively harder tasks with longer target proteins. On CodonDesignEnv, a new codon-level environment that preserves protein identity via dynamic action masking, CAGFNs achieve competitive or stronger Pareto-front coverage and diversity, converge up to 4 times faster than non-curriculum GFlowNets, and maintain 100% sequence uniqueness compared to RL baselines.