The interplay of DNA repair context with target sequence predictably biasses Cas9-generated mutations.

The genome engineering capability of the CRISPR/Cas system depends on the DNA repair machinery to generate the final outcome. Several genes can have an impact on mutations created, but their exact function and contribution to the result of the repair are not completely characterised. This lack of knowledge has limited the ability to comprehend and regulate the editing outcomes. Here, we measure how the absence of 21 repair genes changes the mutation outcomes of Cas9-generated cuts at 2,812 synthetic target sequences in mouse embryonic stem cells. Absence of key non-homologous end joining genes Lig4, Xrcc4, and Xlf abolished small insertions and deletions, while disabling key microhomology-mediated repair genes Nbn and Polq reduced frequency of longer deletions. Complex alleles of combined insertion and deletions were preferentially generated in the absence of Xrcc6. We further discover finer structure in the outcome frequency changes for single nucleotide insertions and deletions between large microhomologies that are differentially modulated by the knockouts. We use the knowledge of the reproducible variation across repair milieus to build predictive models of Cas9 editing results that outperform the current standards. This work improves our understanding of DNA repair gene function, and provides avenues for more precise modulation of CRISPR/Cas9-generated mutations.

Predicting Mutations Generated by Cas9, Base Editing, and Prime Editing in Mammalian Cells.

The first fruits of the CRISPR-Cas revolution are starting to enter the clinic, with gene editing therapies offering solutions to previously incurable genetic diseases. The success of such applications hinges on control over the mutations that are generated, which are known to vary depending on the targeted locus. In this review, we present the current state of understanding and predicting CRISPR-Cas cutting, base editing, and prime editing outcomes in mammalian cells. We first provide an introduction to the basics of DNA repair and machine learning that the models rely on. We then overview the datasets and methods created for characterizing edits at scale, as well as the insights that have been derived from them. The predictions generated from these models serve as a foundation for designing efficient experiments across the broad contexts where these tools are applied.

Prediction of prime editing insertion efficiencies using sequence features and DNA repair determinants.

Most short sequences can be precisely written into a selected genomic target using prime editing; however, it remains unclear what factors govern insertion. We design a library of 3,604 sequences of various lengths and measure the frequency of their insertion into four genomic sites in three human cell lines, using different prime editor systems in varying DNA repair contexts. We find that length, nucleotide composition and secondary structure of the insertion sequence all affect insertion rates. We also discover that the 3' flap nucleases TREX1 and TREX2 suppress the insertion of longer sequences. Combining the sequence and repair features into a machine learning model, we can predict relative frequency of insertions into a site with R = 0.70. Finally, we demonstrate how our accurate prediction and user-friendly software help choose codon variants of common fusion tags that insert at high efficiency, and provide a catalog of empirically determined insertion rates for over a hundred useful sequences.

Predicting base editing outcomes using position-specific sequence determinants.

CRISPR/Cas base editors promise nucleotide-level control over DNA sequences, but the determinants of their activity remain incompletely understood. We measured base editing frequencies in two human cell lines for two cytosine and two adenine base editors at ∼14 000 target sequences and find that base editing activity is sequence-biased, with largest effects from nucleotides flanking the target base. Whether a base is edited depends strongly on the combination of its position in the target and the preceding base, acting to widen or narrow the effective editing window. The impact of features on editing rate depends on the position, with sequence bias efficacy mainly influencing bases away from the center of the window. We use these observations to train a machine learning model to predict editing activity per position, with accuracy ranging from 0.49 to 0.72 between editors, and with better generalization across datasets than existing tools. We demonstrate the usefulness of our model by predicting the efficacy of disease mutation correcting guides, and find that most of them suffer from more unwanted editing than pure outcomes. This work unravels the position-specificity of base editing biases and allows more efficient planning of editing campaigns in experimental and therapeutic contexts.