Compact and highly active next-generation libraries for CRISPR-mediated gene repression and activation.

Horlbeck MA, Gilbert LA, Villalta JE, Adamson B, Pak RA, Chen Y, Fields AP, Park CY, Corn JE, Kampmann M, Weissman JS. · 2016

Read Compact and highly active next-generation libraries for CRISPR-mediated gene repression and activation. by Horlbeck MA, Gilbert LA, Villalta JE, Adamson B, Pak RA, Chen Y, Fields AP, Park CY, Corn JE, Kampmann M, Weissman JS. (2016) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5094855.txt.

Introduction

Max A. Horlbeck and eleven co-authors built improved CRISPR toolkits in eLife in 2016. Their system uses a disabled Cas9 protein guided by small RNAs to turn genes down (CRISPRi) or up (CRISPRa) without cutting DNA. Earlier work showed that folded DNA packaging can block Cas9, so guide choice matters.

The team wrote a computer method that weighs DNA packaging, position, and sequence to pick strong guide RNAs, then built genome-wide libraries for human and mouse cells with five guides per gene. In a screen for genes human leukemia cells need to live, most guides worked well, over 90 percent of essential genes were found, and false hits stayed low. They also saw none of the stray toxicity reported for cutting-based CRISPR.

Catalog record

CREATORHorlbeck MA, Gilbert LA, Villalta JE, Adamson B, Pak RA, Chen Y, Fields AP, Park CY, Corn JE, Kampmann M, Weissman JS.
YEAR2016
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC5094855

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Compact and highly active next-generation libraries for CRISPR-mediated gene repression and activation By Max A Horlbeck, Luke A Gilbert, Jacqueline E Villalta, Britt Adamson, Ryan A Pak, Yuwen Chen, Alexander P Fields, Chong Yon Park, Jacob E Corn, Martin Kampmann, Jonathan S Weissman, Karen Adelman Published in: eLife DOI: 10.7554/eLife.19760 ABSTRACT Abstract We recently found that nucleosomes directly block access of CRISPR/Cas9 to DNA (Horlbeck et al., 2016). Here, we build on this observation with a comprehensive algorithm that incorporates chromatin, position, and sequence features to accurately predict highly effective single guide RNAs (sgRNAs) for targeting nuclease-dead Cas9-mediated transcriptional repression (CRISPRi) and activation (CRISPRa). We use this algorithm to design next-generation genome-scale CRISPRi and CRISPRa libraries targeting human and mouse genomes. A CRISPRi screen for essential genes in K562 cells demonstrates that the large majority of sgRNAs are highly active. We also find CRISPRi does not exhibit any detectable non-specific toxicity recently observed with CRISPR nuclease approaches. Precision-recall analysis shows that we detect over 90% of essential genes with minimal false positives using a compact 5 sgRNA/gene library. Our results establish CRISPRi and CRISPRa as premier tools for loss- or gain-of-function studies and provide a general strategy for identifying Cas9 target sites. DOI: http://dx.doi.org/10.7554/eLife.19760.001 Research Organism: Human INTRODUCTION Highly multiplexed pooled genetic screening methodologies have emerged as powerful and broadly accessible tools for systematically profiling gene function at the scale of mammalian genomes (Paddison et al., 2004). Recently, a number of pooled screening platforms have been developed that utilize the bacterial CRISPR (Clustered Regularly Interspaced Palindromic Repeats)-associated nuclease Cas9 paired with libraries of single-guide RNAs (sgRNAs) to disrupt targeted genes (reviewed in Shalem et al., 2015). We and others have developed tools based on nuclease-dead Cas9 (dCas9) (Qi et al., 2013) to programmably interfere with (CRISPRi) or activate (CRISPRa) transcription (Gilbert et al., 2014, 2013; Konermann et al., 2015; Maeder et al., 2013; Perez-Pinera et al., 2013; Tanenbaum et al., 2014), and used these to systematically manipulate gene expression at genome scale (Gilbert et al., 2014; Konermann et al., 2015). Together, these screening platforms represent a powerful toolkit for unbiased forward gain-of-function and loss-of-function genetic screens in mammalian cells. A key step in implementing CRISPR genetic screens is selecting sgRNAs that mediate high Cas9 activity. We and others recently found that nucleosomes provide a direct and profound impediment to Cas9 access to DNA (Hinz et al., 2015; Horlbeck et al., 2016; Isaac et al., 2016), an observation we expected to be particularly important for applications such as CRISPRi and CRISPRa, which require sustained binding of dCas9 to DNA. We found that nucleosome occupancy was predictive of Cas9 activity complementary to and independent of previously described sgRNA sequence features (Chari et al., 2015; Doench et al., 2014; Xu et al., 2015), adding an additional dimension to the set of parameters expected to influence Cas9 activity. These observations, along with the strong nucleosome-dependent phasing observed downstream of the FANTOM consortium-annotated transcription start site (TSS) (FANTOM Consortium and the RIKEN PMI and CLST (DGT) et al., 2014; Horlbeck et al., 2016), suggested that a quantitative model incorporating all of these features could greatly enhance our ability to predict highly active sgRNAs for CRISPRi and CRISPRa. To test this, we developed a comprehensive machine learning pipeline trained on data collected from 30 CRISPRi and 9 CRISPRa screens. We found that the resulting models were highly predictive of sgRNA efficacy and strongly weighted nucleosome positioning and specific sequence features. We used these models to design and generate CRISPRi and CRISPRa version 2 (v2) libraries, targeting human and mouse genomes, which are greatly enriched for sgRNAs with high predicted activity. These libraries include several additional improvements, including the option to screen with either 10 sgRNAs per gene or a compact half-library containing the top 5 predicted sgRNAs for each gene. To benchmark this new algorithm, we validated the human CRISPRi v2 (hCRISPRi-v2) library with a screen designed to identify genes essential for robust cell growth. In this experiment, essential genes represent a large class of expected positive controls. We identified over 2100 essential genes with high statistical confidence, significantly improving upon our CRISPRi v1 library (Gilbert et al., 2014), and precision-recall analysis showed increased discrimination of gold standard essential and non-essential genes with both 10 sgRNA/gene and 5 sgRNA/gene hCRISPRi-v2 libraries (Hart et al., 2014). A large majority of the hCRISPRi-v2 sgRNAs targeting known essential genes produced robust growth phenotypes, a key advance over previous CRISPRi libraries (Evers et al., 2016), and our algorithm can accurately predict sgRNA activity for data from screens performed with an independently designed library and in different cell types. Furthermore, we observed that CRISPRi lacked any detectable non-specific toxicity associated with genomic DNA breaks and repair, enabling sensitive detection of genes with subtle growth phenotypes. We also conducted a screen for genes that modify growth rates upon overexpression with our hCRISPRa-v2 and found this identified 60% more genes, with greater enrichment for previously established classes of hit genes, than our version 1 CRISPRa screen. Our results suggest that the CRISPRi and CRISPRa v2 libraries have numerous favorable properties relative to alternate approaches as a resource for targeted or genome-scale loss-of-function and gain-of-function studies in mammalian cells. RESULTS An integrated machine learning approach predicts highly active sgRNAs for CRISPRi and CRISPRa We sought to improve upon our first generation CRISPRi and CRISPRa libraries by taking a comprehensive approach that incorporated nucleosome positioning, sequence features, refinement of our original sgRNA design rules, and other potentially informative factors. In order to quantitatively model the contribution of these features to CRISPRi activity, we turned to our recently described CRISPRi activity dataset (Horlbeck et al., 2016) in which we integrated data from 30 CRISPRi screens to select 1,539 high-confidence hit genes, and normalized the phenotypes for sgRNAs targeting each gene to the strongest sgRNAs for that gene, resulting in 'activity scores' for 18,380 sgRNAs. We used this set as training data for elastic net linear regression (Figure 1A) (Hui Zou, 2005). As many of the features included in the model were nonlinear with activity, we first adapted each feature set according to its relationship with activity. Categorical and non-linear parameters were binned prior to linear regression. Because the relationship between CRISPRi activity and target site distance from the TSS was highly periodic and asymmetric, as we had recently shown (Horlbeck et al., 2016), we fit sgRNA positioning features using support vector regression (SVR) to predict a continuous function for any target site (Supplementary Figure 1). An important improvement was the use of FANTOM consortium annotations instead of Ensembl/GENCODE to define the TSS (Cunningham et al., 2015; FANTOM Consortium and the RIKEN PMI and CLST (DGT) et al., 2014; Harrow et al., 2012) (Supplementary file 2), a finding also recently reported by Radzisheuskaya and colleagues (Radzisheuskaya et al., 2016). [Figure: A machine learning approach for identifying highly active sgRNAs for CRISPRi.(A) Schematic of machine learning strategy and datasets. 808 features were calculated for each sgRNA, linearized as indicated, and z-standardized. A linear regression model was then generated using these features to fit to the activity scores of the CRISPRi training set (Horlbeck et al., 2016). 20% of the genes in the training set were reserved to test

Read the complete text