Reading and writing the code of life with machine intelligence.

CRISPGen: A Deep Learning-Driven Generative Framework for Optimal CRISPR/Cas9 Guide RNA Design

For decades, designing CRISPR guide RNAs has been treated as a filtering problem — enumerate candidates from the genome, then discard the unsafe ones. CRISPGen reframes it as a generative one: by learning the joint distribution of high-efficiency, low-risk sgRNAs from experimental data, it synthesizes novel guides that are intrinsically optimized for both on-target cleavage and off-target safety. The framework unites DNABERT-2 contextual embeddings, a conditional latent-diffusion generator, and a dual-critic reinforcement-learning stage trained on 1.59M experimental off-target events.