Thermodynamically Guided CRISPR Spacer Optimization: Strategy-Dependent Off-Target Trade-Offs

Most CRISPR-Cas9 design tools select guide RNAs (gRNAs) by ranking predefined candidates rather than actively optimizing them. Here we instead formulate SpCas9 spacer refinement as a constrained search problem, in which a limited number of nucleotide substitutions may be introduced outside the PAM-proximal seed region. The search is driven by a thermodynamically grounded objective that combines CRISPRspec-based specificity with penalties for suboptimal GC content and long homopolymer runs. Within this single setting we compare five search strategies: three reinforcement-learning (RL) policies, which are Proximal Policy Optimization (PPO), Quantile-Regression Deep Q-Network (QR-DQN), and Rainbow Deep Q-Network (Rainbow DQN), together with a nonlearning score-guided search and a prompt-guided large language model (LLM). All five share the same objective, mutable positions, substitution budget, and evaluation protocol, so that only the search strategy differs. On 100 held-out IGF2R guides from a gene and chromosome excluded from training, every strategy significantly reduced the predicted off-target burden (38.6–78.3%) and improved independent specificity metrics without degrading on-target scores. The strategies nevertheless differed sharply in per-guide safety: the LLM achieved the largest aggregate reduction but introduced new off-target sites into 38% of guides that were initially free of them, whereas PPO, an on-policy RL algorithm, retained a strong gain with the lowest regression rate of 15%. These results show that the thermodynamic objective supplies a robust common optimization signal, while the choice of search strategy governs the trade-off between aggregate improvement and regression of already-favorable guides.