X-CRISP

A Cas9 double-strand break is not repaired at random. When the cell falls back on mutagenic end-joining rather than a donor template, the resulting spectrum of insertions and deletions is strikingly reproducible for a given target site, which makes it predictable — and makes therapeutic editing strategies that deliberately exploit end-joining, such as programmed frameshifts or exon skipping, feasible to design in silico. X-CRISP is a neural network for predicting those repair outcome frequencies, developed by Colm Seale and Joana Gonçalves at Delft University of Technology and published in Bioinformatics Advances in 2025.

Existing repair-outcome predictors — FORECasT, Lindel, inDelphi and CROTON — are accurate on the cell line they were trained on but transfer poorly elsewhere, and their feature representations are large and opaque enough that it is hard to read biology out of them. X-CRISP attacks both problems at once: it predicts deletion frequencies from just five interpretable features describing the resected ends and the microhomology bridging them, and it is explicitly built to be re-pointed at a new cellular context through transfer learning rather than retrained from scratch.

The task is distinct from CRISPR off-target scoring, addressed in this catalog by models such as crisprSFM. Those models ask whether a guide will cut a given site; X-CRISP takes the cut as given and asks what the repaired sequence will look like.

#Key Features

• Minimal, interpretable deletion features: Deletion frequencies are predicted from five quantities — the left and right resection edges relative to the cut, the gap between them, and the length and GC fraction of the microhomology — rather than thousands of one-hot sequence indicators.
• Microhomology location beats composition: Feature attribution showed the model prioritises where a microhomology sits relative to the cut site over sequence properties such as GC content, a mechanistic reading that the denser prior models could not support.
• Domain adaptation from tiny target datasets: Six transfer strategies were compared; the most permissive (pretrained initialisation with no frozen layers) beat direct training on target data from as few as 50 samples, and from five samples in the TREX2 setting.
• Released checkpoints, not just training code: The repository ships the mESC-pretrained deletion checkpoints and the per-cell-line adapted variants for HAP1, U2OS, TREX2-expressing K562 and NHEJ-deficient mESC at sample budgets from 2 to 500, loadable directly by predict.py.

#Technical Details

X-CRISP is a composite of three submodels. The deletion model is a fully connected network with two hidden layers of 16 nodes and a single output node, sigmoid-activated at every layer, trained in a Kullback-Leibler divergence variant and a mean-squared-error variant. Insertion frequencies come from a softmax regression over one-hot encoded single nucleotides and dinucleotides in the six bases upstream of the PAM, and the deletion-to-insertion ratio from a softmax regression over the 20 bp target sequence. The base models were trained on 5,900 wild-type mESC target sequences from the FORECasT dataset and evaluated on 3,954 held-out FORECasT targets plus 1,961 inDelphi mESC targets, where X-CRISP improved on the prior models across detailed profile similarity, frameshift prediction and precision-at-top-X classification.

Domain adaptation was evaluated against four target datasets — U2OS (962 test sequences), HAP1 (3,950), TREX2-expressing K562 (3,355) and Prkdc/Lig4 double-knockout mESC (985) — each with 500 target training sequences subsampled to smaller budgets. Adapted models improved over models trained directly on the target data, with the size of the advantage largest in the low-sample regime. The code is GPL-3.0 licensed; the preprocessed repair-outcome counts and train/test target splits are deposited on Figshare.

#Applications

The immediate use is guide selection: given a set of candidate protospacers, X-CRISP ranks them by the repair profile they are expected to produce, which matters when the therapeutic goal is a specific frameshift, a clean exon skip, or a high fraction of a single dominant allele. The transfer pathway is the practical contribution for editing groups working outside the handful of cell lines with large published repair screens — a few dozen measured target sites in the cell type of interest are enough to adapt the model, instead of the thousands needed to train one.

#Impact

X-CRISP is a small, peer-reviewed model that argues a useful point against the field's drift toward larger sequence encoders: a five-feature network built on the right mechanistic representation both outperforms heavier predictors and yields an interpretable claim about microhomology-mediated end joining. Its limits are equally clear. It models mutagenic end-joining after Cas9 cutting only, so homology-directed repair, base editing and prime editing are out of scope; nucleotide-level effects outside the modelled window are not captured; and adapting to a new domain still requires measured repair outcomes from that domain, however few.