A CRISPRi guide-efficacy model nearly fails on a new library; swapping the within-gene percentile label lifts transfer correlation from 0.008 to 0.136
Synopsis
A CRISPRi guide-efficacy model trained in K562 cells barely transferred to the independent Dolcetto library from Sanson et al. 2018 in HT29 and A375 cells, reaching Spearman correlations of only 0.008 and 0.012. Keeping the same 101 features, architecture and hyperparameters but replacing the within-gene percentile label with a continuous cross-study z-score raised transfer to 0.136 in HT29 and 0.158 in A375, and leave-one-chromosome-out performance in K562 from 0.375 to 0.429. The transfer failure therefore traces mainly to label design rather than missing chromatin features.
Interpretation
A CRISPRi guide-efficacy predictor almost stops working when moved to an independently designed genome-wide library, and the main cause is not the feature set. Earlier evaluations either sampled guides from the training distribution or held the same loci fixed across cell lines; transfer to a library targeting largely different genes had not been directly tested. The K562-trained model, applied without retraining to the 100942 guides of the Dolcetto library, reached Spearman 0.008 (95% CI 0.002 to 0.014) in HT29 and 0.012 (0.006 to 0.018) in A375, far below its 0.375 under leave-one-chromosome-out cross-validation in K562.
Replacing the within-gene percentile label with a continuous z-score improves the same model both in the source distribution and on the external library. Within-gene percentile normalization removes signal that ranks efficacy across different target genes, and an independent library consists mostly of genes outside the training set. With the continuous z-score label, leave-one-chromosome-out performance in K562 was 0.429 (0.412 to 0.447), transfer was 0.136 (0.129 to 0.142) in HT29 and 0.158 (0.152 to 0.164) in A375, with a paired permutation test giving p about 0.0001 in both cell lines.
The model feeds 101 features covering guide sequence, chromatin accessibility, histone modifications, 3D genome contacts and local genomic context into a gradient-boosted tree, and SHAP values decompose each feature group's contribution. Sequence features alone gave 0.329 under leave-one-chromosome-out cross-validation in K562; adding chromatin accessibility raised it to 0.360, adding histone marks or 3D contacts changed nothing significant, and adding local genomic context brought it to 0.375. On 19446 training guides, the full 101-feature model's 0.375 (0.360 to 0.389) exceeded Rule Set 3's 0.259 (0.232 to 0.286) on the 5056 Horlbeck guides with canonical NGG PAM contexts, and a distance-to-TSS linear baseline's 0.021 (0.006 to 0.034).
Perspective
The conclusion rests on training in K562 cells, building labels from the Horlbeck and Gasperini screens, and testing externally on the Sanson et al. 2018 Dolcetto library in HT29 and A375 cells; external features still come from K562 chromatin data because ENCODE lacked the needed histone-mark coverage for those two cell lines at the time.
Even with the continuous z-score label, transfer correlation is only about 0.14 to 0.16, still below in-distribution performance, so the label fix narrows but does not close the cross-library, cross-cell-type gap; because the external evaluation mismatches cell type in both label and feature domains, the available data cannot separate their contributions to the residual gap.