Comprehensive profiling of activity and specificity of CRISPR/Cas9 under cellular environment by deep learning

This study aims to predict the activity and specificity of CRISPR/Cas9 by deep learning at genome-scale among different cell lines. Here, we have focused on embracing and modifying a system for evaluating SpCas9 activity of on-target and off-target using >1,000,000 guide RNAs (gRNAs) covering ~20,000 protein-coding genes and ~10,000 non-coding genes in synthetic constructs with a high-throughput manner. With the help of deep learning algorithms in the field of artificial intelligence, three prediction models with the best generalization performance now are constructed: Aidit_Cas9-ON, Aidit_Cas9-OFF, and Aidit_Cas9-DSB. Moreover, through systematically investigating the influence of diverse cellular environment on gRNA activity and specificity, we noticed that distinct features are favored from H1 cell line compared with the other 2 cell lines for on-target activity and the overall distribution of repair outcomes is markedly different across 3 cell lines, especially in Jurkat. Finally, we identify a key effect protein DNTT strongly influences editing outcomes induced by CRISPR/Cas9. We confirm that this study will greatly facilitate CRISPR-based genome editing. All 929,180 gRNAs (“ON-Target” and “OFF-Target” library) were cloned into a lentivirus backbone with 20-nt random nucleotide as barcode sequences (BC). The BC was incorporated to eliminate the confounding issues stemmed from plasmid propagation and PCR amplification. An average of 300x coverage was maintained over the course of library preparation. To establish a uniform system to record the gRNA activities, we expanded cells from mono clones with stable SpCas9 expression to eliminate the heterogeneity (K562-cas9; Jurkat-Cas9; H1-cas9). The gRNA-target paired assessment cassette and BC were packed into lentivirus, transduced into cells, and integrated into the genome of host cells. We applied high MOI transduction to achieve decent cell-to-library coverage with practical amount of cell population, under the rationale that the relative indel frequencies and the indel frequency ranks are preserved under different amount of transduced lentivirus. K562 and Jurkat cells were harvested at day-3.5 post-transduction and H1 cells were collected at day-5. We amplified the gRNA-target cassette together with the BC from both the plasmid libraries and the gDNA collected from the edited cells. Amplicons were undergone high throughput sequencing (HTS) and the editing outcome were determined.