CRISPR off-target prediction stops reading sequence alone: chromatin accessibility joins at nucleotide level, cross-cell AUC-PR reaches 0.587

Off-target prediction no longer takes only the paired 23-nucleotide sequence of guide RNA and candidate site; chromatin accessibility enters the genomic candidate representation as a per-position signal. Most earlier models modeled only the sequence pair of sgRNA and the 23-nt candidate region, without explicitly modeling the broader genomic context in which the site sits. Trained on T-cell data and evaluated on three independent test sets, K562-I, K562-II and WTC-11, the AlphaGenome-predicted ATAC-seq version reached AUC-PR of 0.587, 0.616 and 0.545, against 0.537, 0.546 and 0.470 for the sequence-only version and 0.564, 0.563 and 0.466 for the best competing models.

The chromatin signal is not appended at the output layer; it interacts with the sequence representation position by position inside the model. Late merging, which concatenates epigenetic summary statistics onto the final MLP layer of existing predictors, produced no statistically significant improvement under Bonferroni-adjusted paired Wilcoxon tests. CRISCross encodes the sgRNA and the genomic candidate separately and runs three interaction blocks that alternate independent Transformer encoding with bidirectional cross-attention restricted to the central 23 nucleotides of the candidate, the sgRNA binding site.

Pretraining no longer depends on the target dataset's own sequences. Existing approaches such as CRISPert rely on dataset-specific pretraining, which limits generalization to unseen sgRNAs and cell types. CRISPRAT samples loci directly from the hg38 reference genome and generates artificial sgRNAs by inserting one to six mismatches under a combinatorial distribution, pretraining with masked language modeling plus R-loop free-energy regression; on the unseen K562-I it reached AUC-PR 0.783, against 0.569 for the model pretrained on T-cell data.

The model was trained and evaluated exclusively on SpCas9 in human cells, with tests covering three independent datasets, K562-I, K562-II and WTC-11, spanning two cell types. It is aimed at users who need to prioritize candidate off-target sites in cell types lacking dedicated experimental data, reducing reliance on costly assays such as GUIDE-seq for initial screening. The authors frame cross-cell-type performance as a promising direction rather than a validated substitute for experimental profiling.

Epigenetic features are statistically significant but weak and highly variable, and adding experimentally measured features gave a non-significant gain, while only AlphaGenome-predicted features produced a significant one; the authors therefore do not conclude that measured epigenetic information is generally strongly predictive. After shuffling ATAC tracks, performance still exceeded the sequence-only baseline, which the authors attribute to shuffling resembling a noise-injection technique known to improve model performance. Sites undetected by GUIDE-seq were treated as negatives, though these are putative rather than confirmed negatives, which may affect both training supervision and evaluation, so AUC-PR may underestimate performance relative to the unknown true labels.