X-CRISP: domain-adaptable and interpretable CRISPR repair outcome prediction

Area of Science:

• Molecular Biology
• Bioinformatics
• Genetics

Background:

• Controlling CRISPR editing outcomes is vital for gene therapy success.
• Donor template-based editing is often inefficient, necessitating alternative strategies like mutagenic end-joining repair.
• Existing machine learning models for predicting end-joining repair outcomes lack generalizability and interpretability.

Purpose of the Study:

• To develop a flexible and interpretable neural network, X-CRISP, for predicting CRISPR editing repair outcome frequencies.
• To improve the generalizability and interpretability of machine learning models in this domain.
• To leverage transfer learning for efficient model adaptation to new cell lines and experimental conditions.

Main Methods:

• Developed X-CRISP, a neural network utilizing minimal outcome and sequence features, including microhomologies (MH).
• Evaluated X-CRISP's performance against prior models on detailed and aggregate outcome predictions.
• Employed transfer learning by pre-training X-CRISP on wild-type mouse embryonic stem cell (mESC) data and adapting it to human cell lines (K562, HAP1, U2OS) and modified mESC lines.

Main Results:

• X-CRISP outperformed existing models in predicting CRISPR repair outcome frequencies.
• The model prioritized microhomology location over sequence properties for deletion outcomes.
• Adapted X-CRISP models showed significant improvement with as few as 50 target data samples, demonstrating the efficacy of transfer learning.

Conclusions:

• X-CRISP offers a more accurate, generalizable, and interpretable approach to predicting CRISPR editing outcomes.
• Transfer learning significantly reduces the data requirements for developing predictive models for new cell types or conditions.
• This work facilitates the advancement of gene therapy by improving the control and predictability of CRISPR-based genome engineering.