<p>Nucleic acid enzymes (NAzymes), including ribozymes, DNAzymes, and XNAzymes, are programmable biocatalysts that catalyze diverse chemical reactions and provide unique platforms for investigating catalysis and engineering molecular function. Their central roles in biology and evolution, together with emerging applications in biotechnology, synthetic biology, and medicine, have created an urgent need for predictive computational design frameworks. Although multiscale computational enzymology integrating ab initio quantum mechanics can, in principle, provide predictive insight into these systems, current approaches remain computationally prohibitive for large, complex NAzymes and high-throughput design workflows. Machine learning potentials have the potential to enhance the accuracy and efficiency of multiscale modeling approaches, but QM/MM contexts in which the MM environment is highly polarizing remain challenging. Nucleic acid enzymes present a particularly demanding case: experimentally relevant predictions require an accuracy that semiempirical QM/MM methods alone cannot reliably deliver in their often large, highly charged, metal-dependent active sites. Building on the range-corrected QM/MM Δ-machine learning potential (ΔMLP) framework, we develop ΔNAzyme, a foundational ΔMLP for nucleic acid enzyme reactions. We generate a training set of QM/MM simulations spanning diverse nucleic acid enzyme classes, with energies and forces computed at both the reference (DFTB3) and target (PBE0/6–31G*) levels of theory, yielding a correction on top of DFTB3 that recovers PBE0 quality results. We train models using both the Deep Potential and MACE architectures. The MACE model exhibits enhanced transferability, which we attribute to the equivariant, many-body message-passing framework. Free energy profiles from case-study reactions validate ΔNAzyme against target-level results and show that the model supports efficient fine-tuning to new nucleic acid enzyme systems with minimal additional data.</p>
