Fig. 2: Deep learning uncovers similar rules for gRNAs targeting EEJs and exons.
From: Cas13d-mediated isoform-specific RNA knockdown with a unified computational and experimental toolbox
a Schematic of the SEABASS linear mixed model. Input: LFCs for all timepoints and replicates for EEJ gRNAs. Output: slope, where a more negative slope corresponds to a more active gRNA, and standard error. b Summary of TIGER model adaptations. ‘Cell non-specific set’ is defined below the table. MFE, Minimum free energy. c Cross-validated area under the precision-recall curve (AUPRC) values for each model across gene essentiality thresholds. Random gRNA selection baseline shown for comparison. Dotted line marks the 25% essentiality threshold used in (d). d Pearson correlations between predicted and observed gRNA activity at the 25% essentiality cutoff. Observed values are either LFCs or SEABASS slopes. *p < 0.0001 (two-sided). e SHapley Additive exPlanations (SHAP) plots comparing positional nucleotide contributions for TIGER (solid line) and TIGERjunction (dashed line). Colors denote nucleotide identity (e.g. A is blue). SHAP values indicate predicted impact on LFC when that position has the specific nucleotide. f SHAP value profiles comparing TIGERsite (solid) and TIGERjunction (dashed lines) across EEJs. TIGERsite predicts the mean LFC of all eight gRNAs tiled across an EEJ using a 101 bp window (−50/+50). TIGERjunction predicts the LFC of individual gRNAs using only the 23 bp target sequence. SHAP values from TIGERjunction were tiled across the EEJ at the same eight positions used in the screen (tiles; dotted lines), and then positionally averaged (tile mean; dashed lines). This average closely matches the TIGERsite SHAP profile within the tiling window (gray bars). X-axis indicates gRNA target position centered at the splice site. Source data are provided as a Source Data file.
