inDecay

Cutting a locus with Cas9 is the easy part; which indel the cell leaves behind is the hard one, and it is not a property of the target sequence alone. The break is resolved by the cell's own repair machinery, and the balance between non-homologous end joining, microhomology-mediated end joining and templated insertion shifts with cell state. A mouse zygote and a K562 cell given the same guide RNA return visibly different indel profiles, so a predictor fitted on one immortalized line mis-ranks the outcomes in the other — exactly the setting that matters for making knockout animals or a therapeutic cell product.

inDecay separates two layers that earlier predictors folded together. A ratio module first estimates how reads will divide between deletions and insertions for the cell type in question. An event module then scores each candidate indel from a short feature vector, and a softmax across the candidates turns those scores into a probability over the near-full repair spectrum. Because one scoring function is applied to every event rather than a per-class output head, the model stays tiny — which is what lets it be re-fitted to an unfamiliar cell state from a few dozen measured guides instead of a genome-scale screen.

inDecay was built by Weizhong Zheng, Lu Yu and colleagues in Yuanhua Huang's group at The University of Hong Kong, with the Centre for Translational Stem Cell Biology and animal-science partners in mainland China. It appeared on bioRxiv in November 2024 and in Genome Biology in 2026, and sits beside crispAIPE, for prime-editing efficiency, and crisprSFM, for off-target specificity.

#Key Features

• Event-centered probability: Each candidate indel is scored independently and normalized against its competitors, so the output is a distribution over the repair profile rather than a summary statistic such as frameshift rate.
• Decay features from first principles: Deletion frequency falls off with deletion length and with distance from the cut site, and microhomology pulls it back up. Those three quantities enter through explicit decay and power transforms with tunable exponents, which is why so few learned parameters are needed.
• Diagonal microhomology detection: The target is split at the cut site and the fragments laid out as a pairwise nucleotide identity matrix; convolving a small identity kernel over it turns the search for microhomology tracks into diagonal-line detection.
• Few-shot adaptation: Fine-tuning starts from a released checkpoint and is anchored to it by an L2 penalty on the weight change, so 30 to 50 measured guides move the model to a new cell line without erasing what pretraining learned.

#Technical Details

Sixty-one features describe each event: 14 specific to deletions, 8 to insertions, and 39 shared across a target, including the ratio module's output, GC content, overall microhomology strength, and a one-hot encoding of the 9 bp upstream of the PAM. They feed a perceptron with two hidden layers of 128 and 64 units — roughly 16,000 weights — whose scalar outputs are softmaxed with a temperature term, and Lindel's 557-class indel coding is extended to 912 classes so deletions up to 37 bp are represented. Pretraining uses the FORECasT self-targeting screen, which profiled over 35,000 guide RNAs in K562, mESC, CHO, iPSC and HAP1 cells, with 1,133 highly covered guides held out for testing. On that held-out set, frameshift-ratio R² in mESC reaches 0.859 against 0.804 for FORECasT, 0.77 for inDelphi and 0.622 for Lindel, and top-5 indel-length KL divergence is 10% and 32% below inDelphi and FORECasT respectively.

#Applications

The practical use is guide RNA prioritization when the intended outcome is specific — a frameshift of a particular length, or avoidance of a dominant in-frame deletion. Returning the whole spectrum lets a designer rank candidates on the probability of the outcome they want rather than on aggregate efficiency. The embryo work extends this to animal model generation and livestock editing, where attempts are expensive and few guides can be tested: the authors generated their own mouse zygote profiles by Sanger-sequencing electroporated embryos, fine-tuned with leave-one-guide-out cross-validation, and carried the model to goat, cattle and porcine embryos.

#Impact

inDecay fills a gap self-targeting screens left open: those assays run in immortalized lines and cannot be performed in zygotes. Code is Apache-2.0, with per-cell-type checkpoints for all five pretrained lines, archived weights and data, and a hosted predictor. Its limits follow from its design: coverage out of the box is confined to the five cell types with released checkpoints, and a new cell state wants the fine-tuning step. Features are hand-engineered rather than learned, and extracting them needs the external SelfTarget toolkit, so inference is not self-contained in the Python package. It is a specialist predictor for template-free Cas9 repair, not a general DNA sequence representation, and the large deletions seen in embryos stay outside what it models.