Optimizing CRISPR on-target activity to 75% while reducing off-target editing 4-fold

Context

CRISPR specificity—the ratio of on-target to off-target editing—determines the therapeutic safety profile for gene therapy applications. Off-target edits can cause insertional mutagenesis, oncogenic transformation, or other adverse effects. Regulatory agencies require comprehensive characterization of off-target activity, and high off-target rates represent grounds for clinical hold or rejection.

Improving on-target activity while simultaneously reducing off-target editing is particularly challenging because mutations that increase overall catalytic activity often increase editing at both on- and off-target sites. Engineering specificity therefore requires identifying mutations that selectively enhance discrimination between the intended target and sequence-similar off-target loci.

Challenge

A previous optimization campaign utilizing structural and rational design approaches had achieved on-target editing activity of 40% with a worst-site off-target editing activity of 0.4%. While this represented progress, the specificity ratio remained insufficient for therapeutic development. The partner had exhausted conventional protein engineering approaches and turned to Cradle to address the six-property optimization problem (one on-target, five off-target sites).

The challenge was compounded by the requirement to improve both properties simultaneously. Accepting a trade-off—higher on-target with higher off-target, or lower off-target with lower on-target—would not advance the program. Only simultaneous improvement across all six measurements would yield a therapeutically viable candidate.

Approach

Scientists using Cradle performed two rounds of optimization with 96 candidates each. The partner provided all sequence-function data from the prior structural and rational design campaigns, which Cradle consumed as training data for supervised predictors.

The optimization was configured with on-target activity as the primary objective and off-target editing at all five sites as constraints to be minimized. This multi-objective formulation required Cradle to navigate a high-dimensional fitness landscape where improvements along one dimension could not come at the expense of others.

Results

Cradle achieved substantial improvements in both on-target activity and specificity:

The 1.88-fold improvement in on-target activity brought the system from marginal therapeutic utility (40%) into the high-performance regime (75%). Simultaneously, the 4-fold reduction in worst-site off-target editing substantially improved the safety profile.

The specificity ratio—on-target editing divided by worst-site off-target editing—improved 7.5-fold from 100:1 to 750:1. This metric, which captures the therapeutic window between efficacy and safety, demonstrates that Cradle did not simply boost overall catalytic activity (which would improve both on- and off-target) but rather enhanced selectivity.

Examining the off-target editing profile across all five measured sites revealed that Cradle reduced editing at each site, not just the worst case. This consistent improvement across multiple genomic loci suggests that the optimized variants achieved enhanced target discrimination rather than site-specific suppression.

What this means

This result demonstrates Cradle's capability to optimize the most challenging property for therapeutic genome editing: the on-target to off-target ratio. The ability to simultaneously improve activity and specificity—properties that are often antagonistic in practice—validates that machine learning approaches can identify non-obvious mutational combinations that escape the trade-offs encountered via rational design.

The two-round timeline compares favorably to the extended campaigns typically required for CRISPR specificity engineering. Conventional approaches often rely on structural modeling to predict discriminatory mutations or exhaustive mutagenesis of DNA-contacting residues, both of which require many rounds of iterative refinement. Cradle achieved superior results by learning from prior campaign data rather than requiring de novo structural hypotheses.

For the commercial partner, the 75% on-target activity with 0.1% worst-site off-target editing represents a candidate suitable for regulatory submission and clinical development—an outcome that enables program progression from research to therapeutic development.

Methods note

Cradle ingested sequence-function data from the partner's structural and rational design campaigns. Supervised predictors were trained independently for on-target activity and for each of the five off-target sites. Evotuning was performed on the CRISPR system's evolutionary alignment. Editing measurements were performed via the partner's validated cellular assays with sequence-confirmed off-target loci. Generation employed multi-objective optimization with on-target activity as the primary objective and off-target editing as minimization constraints. Full details of the CRISPR system and genomic target sites remain confidential to the commercial partner.