CRISPR-Cas9 has moved from laboratory curiosity to clinical reality at a pace that surprised even its developers. In November 2023, the FDA approved Casgevy — a CRISPR-based therapy for sickle cell disease and transfusion-dependent beta-thalassemia — becoming the first approved CRISPR therapy in the United States. The approval was the culmination of years of safety work, much of which centered on the fundamental challenge of CRISPR gene editing: off-target effects, where the guide RNA directs the Cas9 enzyme to cut at unintended locations in the genome. Machine learning has become central to understanding and predicting these off-target events.
Why Off-Target Edits Are a Safety-Critical Problem
The CRISPR-Cas9 system uses a short guide RNA (sgRNA, typically 20 nucleotides) to direct the Cas9 endonuclease to a complementary DNA sequence. The system tolerates mismatches between the guide RNA and the target DNA — a feature that enables some flexibility but also means Cas9 can cut at sites with partial sequence complementarity. An off-target edit in a tumor suppressor gene, a proto-oncogene, or a region critical to normal cell function could theoretically trigger carcinogenesis or disrupt vital cellular processes.
For a clinical therapy, off-target editing must be characterized comprehensively before the therapy can advance to human trials. Traditional experimental methods — GUIDE-seq, DISCOVER-Seq, CIRCLE-seq — can detect off-target sites genome-wide, but they are expensive, time-consuming, and vary in sensitivity. ML models that predict off-target activity from sequence features offer the possibility of prioritizing experimental validation efforts and guiding guide RNA design to minimize off-target risk from the outset.
Machine Learning Approaches and Their Accuracy
Multiple ML architectures have been applied to off-target prediction. A 2022 analysis in Nature Biotechnology by Concordet and Haeussler benchmarked 10 different prediction tools on validated off-target datasets. Key findings from this comparative analysis:
- The best-performing tools (CRISPOR, CFD score, DeepCRISPR) achieved Pearson correlations of 0.6–0.7 between predicted and observed cleavage efficiencies at known off-target sites
- No single tool consistently outperformed others across all guide RNA sequences and genomic contexts
- Prediction accuracy was substantially higher for mismatches at certain positions (particularly positions 1–10 from the PAM-distal end) than for bulges and insertion/deletion mismatches
- Deep learning models outperformed rule-based scoring methods on novel sequence contexts not well-represented in training data
A 2023 paper from the Broad Institute introduced a transformer-based architecture (CHANGEseq-trained) that improved off-target site detection sensitivity, particularly for low-frequency off-target edits that conventional experimental methods might miss.
Integration Into CRISPR Therapy Development
In the development of Casgevy (exagamglogene autotemcel), Vertex Pharmaceuticals and CRISPR Therapeutics used an extensive off-target characterization process that included both experimental genome-wide detection methods and computational prediction. The guide RNA targeting BCL11A (the therapeutic target for reactivating fetal hemoglobin) was selected in part because computational off-target analysis identified it as having a favorable specificity profile compared to alternative guide sequences.
The FDA’s evaluation of Casgevy’s safety data required demonstration that no clinically relevant off-target editing was detected at any of the top computationally predicted sites. This created a regulatory expectation that has since been applied to other CRISPR therapy candidates entering IND-enabling studies.
Ethical Considerations in Germline Applications
The off-target prediction challenge takes on a different dimension when considered in the context of germline editing — modifications to embryos that would be heritable by future generations. The use of ML to predict (and potentially optimize to minimize) off-target edits in germline applications raises questions that extend beyond technical accuracy. A model that predicts low off-target risk does not guarantee zero off-target events, and in germline editing, even rare off-target edits could be transmitted to all subsequent generations. International scientific consensus, including the 2020 report from the International Commission on the Clinical Use of Human Germline Genome Editing, holds that germline editing should not proceed clinically until, among other requirements, “reliable methods of assessing the effects of off-target edits” are available — a standard that current ML prediction tools do not yet meet.
Key Takeaway
ML-based off-target prediction has become a standard tool in somatic CRISPR therapy development, guiding guide RNA selection and prioritizing experimental validation efforts. Prediction accuracy for known off-target site types is meaningful but not perfect, and models perform worse on novel sequence contexts and insertion/deletion mismatches. For germline applications, current prediction tools are insufficient to meet the safety evidence threshold required for clinical use.
Sources
1. Concordet JP, Haeussler M. CRISPOR: intuitive guide selection for CRISPR/Cas9 genome editing experiments and screens. Nucleic Acids Res. 2018;46(W1):W242–W245.
2. Anzalone AV, Koblan LW, Liu DR. Genome editing with CRISPR-Cas nucleases, base editors, transposases and prime editors. Nat Biotechnol. 2020;38(7):824–844.
3. FDA. Casgevy (exagamglogene autotemcel) Approval Letter. December 8, 2023. fda.gov.
4. International Commission on the Clinical Use of Human Germline Genome Editing. Heritable Human Genome Editing. National Academies Press, 2020.
Medical Disclaimer: This article is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional for medical decisions.