Machine learning-assisted reprogramming of CRISPR/Cas9 guide RNA for multiplex editing and minimized tracrRNA-dependent off-target

While the single-handle fixed guide RNA (gRNA) configuration in CRISPR/Cas9 systems enables straightforward implementation and efficient delivery, it may concurrently hinder flexible adaptability and reprogrammability. Leveraging the sophisticated architecture of bacterial interspaced short palindromic repeats, we develop a separately expressed gRNA (segRNA) system enabling efficient multiplex editing. We further apply machine learning, trained on high-throughput variant screens, to design optimized non-repetitive crRNA-tracrRNA pairs, facilitating functional annotation of lineage-specific core enhancers using one-step generated dual segRNA libraries and simultaneous interference of six Polycomb repressive complex 1/2 (PRC1/2) components. Exploiting segRNA reprogrammability, we redesign sgRNA sequences to minimize sgRNA self-editing during PAM-less SpRY-Cas9-mediated base editing and engineer tracrRNAs complementary to endogenous coding and noncoding RNAs for CRISPRa-based expression detection. Crucially, we reveal that any gRNA possessing tracrRNA can hijack endogenous RNAs to form unintended segRNA complexes, leading to low-frequency but pervasive off-target effects. We also strategically reprogram crRNA-tracrRNA repeats to eliminate tracrRNA-dependent off-targeting. Collectively, we establish a versatile segRNA reprogramming platform enabling multiplex genome editing, functional enhancer screening, RNA-based sensing, and high-fidelity optimization.