Optimized parameters for Cas9 CRISPR interference library design

This study leverages large-scale CRISPRi datasets to develop an updated scoring scheme and characterize off-target effects, culminating in the design and validation of the optimized, compact Cas9 CRISPRi library named Katsano.

The Gist

The Big Picture: Turning Down the Volume, Not Breaking the Radio

Imagine your cells are a giant, complex radio station. Every gene is a different song playing on a different channel. Sometimes, to understand what a song does, you need to turn the volume down or mute it completely.

In the past, scientists used a "sledgehammer" approach (CRISPR Knockout) to study genes. They would smash the radio speaker (cut the DNA) to stop the music. This works, but it's messy, permanent, and can damage the radio in ways unrelated to the song itself.

CRISPRi (CRISPR Interference) is a better tool. Instead of smashing the speaker, it's like putting a heavy, silent blanket over it. The song is still there, but it's muffled. It's reversible, precise, and doesn't break the radio.

However, there's a problem: The "Blanket" isn't always perfect. Sometimes it doesn't cover the speaker well enough (the gene is still too loud), and sometimes it accidentally covers the next speaker over (off-target effects).

This paper is about designing the perfect blanket and the best instructions for where to place it. The researchers built a new, super-optimized library of instructions called Katsano.

---

The Journey to a Better Blanket

1. Testing the Blanket Materials (The KRAB Domains)

The researchers needed to know which material makes the best "muffling blanket." They tested different versions of a protein called KRAB (which acts as the heavy blanket).

• The Experiment: They tried attaching the blanket to the front (N-terminus) or the back (C-terminus) of the delivery truck (dCas9). They also tested different types of blankets (Kox1 vs. Zim3).
• The Result: They found that putting the Zim3 blanket on the front of the truck worked the best. It muffled the genes most effectively without getting in the way.

2. Learning Where to Place the Blanket (On-Target Activity)

You can't just throw the blanket anywhere; it has to cover the right spot. The "sweet spot" is right at the start of the gene (the Transcription Start Site).

• The Old Way: Previous libraries used simple rules, like "put the guide 50 steps from the start."
• The New Way (Rule Set 3i): The researchers gathered massive amounts of data to build a smart GPS. This GPS considers:
  • Sequence: Is the DNA sequence easy to grab?
  • Location: Is it exactly where the gene starts?
  • Accessibility: Is the DNA "open" and easy to reach, or is it tightly packed like a rolled-up carpet? (They used a new map called ATAC-seq to see this).
• The Outcome: They created a scoring system (RS3i) that tells you exactly which guide will work best, like a weather forecast predicting exactly where the sun will shine.

3. Avoiding the "Wrong House" (Off-Target Effects)

Sometimes, a guide might look at a gene and say, "Hey, that looks like my target!" and accidentally mute the wrong song. This is dangerous.

• The Discovery: They found a specific pattern in the "address" of the guide (the seed sequence). If the address contains too many "GG" pairs (like a PAM motif), the guide gets confused and starts muting random genes nearby.
• The Fix: They added a rule: "If the address has too many 'GG's, throw it out." This simple rule drastically reduced mistakes.

4. The Problem of Multiple Addresses (Alternative TSS)

Some genes are like houses with two front doors (Alternative Transcription Start Sites). If you only know about the main door (MANE Select), you might miss the gene if it's using the back door in a specific cell type.

• The Solution: The new library, Katsano, doesn't just look for the main door. It checks for the most common "back doors" used across different cell types, ensuring no gene is left unmuted.

---

The Result: Introducing "Katsano"

The researchers combined all these lessons into a new library called Katsano. Think of Katsano as a Master Key Ring for the human genome.

• Smarter: It uses the new GPS (RS3i) to pick the absolute best keys.
• Safer: It filters out keys that might open the wrong doors (low "Seed Scores").
• Smaller but Stronger: Because the keys are so good, you don't need as many of them. Katsano is smaller than previous libraries but works better.
• The Proof: When they tested Katsano on cancer cells, it found the "essential" genes (the ones the cancer needs to survive) much faster and more accurately than the old libraries. It missed fewer targets and made fewer mistakes.

Summary Analogy

• Old Libraries: Like hiring a team of 10 people to try and cover a speaker with a blanket, hoping one of them gets it right. Some people use the wrong blanket, some stand in the wrong spot, and some accidentally cover the neighbor's speaker.
• Katsano (New Library): Like hiring one expert who has a high-tech map, the perfect blanket material, and a rulebook that says "never stand near the neighbor's door." This expert covers the speaker perfectly, every single time.

This paper gives scientists a much sharper, more precise tool to study how genes work, which helps in understanding diseases and developing new treatments.

Technical Summary

1. Problem Statement

CRISPR interference (CRISPRi) is a vital tool for studying loss-of-function phenotypes without inducing double-strand breaks (DSBs), offering reversible and titratable gene suppression. However, large-scale CRISPRi screens are costly, necessitating compact libraries with high efficiency. Existing genome-wide CRISPRi libraries (e.g., hCRISPRiv1, hCRISPRiv2, Dolcetto) face several limitations:

• Outdated Annotations: They rely on older transcript annotations, missing updated Transcription Start Sites (TSSs) and alternative promoters.
• Suboptimal Prediction Models: Existing on-target scoring models were often trained on CRISPR knockout (CRISPRko) data, which may not accurately reflect CRISPRi-specific efficacy determinants.
• Off-Target Effects: The specific sequence features driving off-target repression in CRISPRi (where dCas9 binding alone causes silencing) differ from CRISPRko, yet libraries often use similar off-target filters.
• System Architecture: There was a lack of large-scale comparative data on optimal repressor domains (KRAB variants) and their fusion positions (N- vs. C-terminus) for dCas9.

2. Methodology

The authors employed a multi-stage approach combining large-scale empirical screening, machine learning, and genomic data integration:

• Large-Scale Tiling Screens:

  • Designed a primary library targeting 1,000 bp upstream and downstream of the MANE Select TSS for 201 essential and 198 non-essential genes (108,574 sgRNAs).
  • Screened these in A549 and HCT116 cell lines using different dCas9-KRAB fusion architectures (Kox1 vs. Zim3; N-terminal vs. C-terminal; direct fusion vs. nanobody-mediated recruitment).
  • Conducted a secondary tiling screen on a subset of 60 genes to validate secondary systems (including Kox1-MeCP2 and nanobody recruitment).

• Feature Analysis & Model Development:

  • Integrated primary screen data with previously published tiling datasets (Nunez et al., Gilbert et al.).
  • Evaluated predictive features: sgRNA sequence (using Rule Set 3 Sequence), distance from TSS, and chromatin accessibility (DHS, ATAC-seq, H3K4me3, H3K27ac).
  • Developed Rule Set 3 Interference (RS3i), a machine learning model (XGBoost) to predict on-target efficacy, identifying the top three features: TSS offset, RS3 Sequence score, and ATAC peak overlap.

• Off-Target Characterization:

  • Analyzed "promiscuous" guides (those causing depletion of non-essential genes) to identify sequence drivers.
  • Discovered that Seed Score (the count of "GG" dinucleotides within the 12bp PAM-proximal seed region) is a strong predictor of CRISPRi off-target activity, distinct from CRISPRko profiles.

• Library Design (Katsano):

  • Designed a new genome-wide library targeting 20,106 genes.
  • Target Selection: Combined Ensembl Canonical transcripts with the Jaganathan set (high-confidence FANTOM5 CAGE-seq supported TSSs) to capture alternative promoters.
  • Guide Selection: Prioritized guides with high RS3i scores, Seed Scores < 3, and minimal CFD off-target sites. Selected 3 guides per TSS.

3. Key Contributions

• Optimized dCas9 Architecture: Demonstrated that N-terminal fusions of KRAB domains to dCas9 generally outperform C-terminal fusions. Specifically, Zim3-dCas9 (N-terminal) showed the highest repression efficacy.
• RS3i Model: Created a CRISPRi-specific on-target prediction model that integrates sequence features, TSS proximity, and ATAC-seq chromatin accessibility. It outperforms previous CRISPRko-trained models for interference tasks.
• Seed-Driven Off-Target Mechanism: Identified that guides with 3 or more "GG" motifs in the 12bp seed sequence are highly prone to off-target repression in CRISPRi. This provides a simple, computationally cheap filter to reduce false positives.
• Katsano Library: Released a new, compact, high-performance genome-wide CRISPRi library that addresses alternative TSS usage and incorporates the new design rules.

4. Key Results

• System Performance:
  • Zim3 and Kox1 domains showed comparable performance, but N-terminal fusions consistently yielded stronger depletion of essential genes than C-terminal fusions.
  • Nanobody-mediated recruitment (ALFA-Nb) was shown to be a viable alternative to direct fusion, achieving comparable knockdown (93% for CD47).
• Predictive Features:
  • Guides within 0–75 bp downstream of the TSS showed the highest activity.
  • ATAC-seq and histone ChIP-seq (H3K4me3, H3K27ac) data were significantly more predictive of guide efficacy than DNase hypersensitivity (DHS) data, likely due to higher sensitivity and better peak calling.
  • The RS3i model achieved high correlation between predicted and experimental activity; 81.6% of guides with a predicted score > 1.2 were highly active (z-score > 2).
• Off-Target Findings:
  • Guides with a Seed Score ≥ 3 captured 52.7% of promiscuous guides but represented only 9% of non-promiscuous guides. Excluding these significantly reduces off-target effects without sacrificing library coverage.
  • The CRISPRko off-target model (CRISPick Aggregate CFD) failed to predict CRISPRi promiscuity, confirming the need for CRISPRi-specific filters.
• Katsano Library Performance:
  • Coverage: Targets 20,106 genes with 62,404 unique sgRNAs (avg. 3 guides/gene), significantly smaller than previous libraries but with higher predicted efficacy.
  • Efficacy: In viability screens (A375 and K562), Katsano recovered 63 fewer false negatives than the Dolcetto library at a 5% FDR, despite using half the guides per target.
  • Specificity: Showed significantly lower off-target depletion of non-essential genes compared to Dolcetto and hCRISPRiv2.
  • Alternative TSSs: Successfully identified essential genes (e.g., CHEK1, SAP18) that were missed by libraries targeting only MANE Select TSSs, proving the value of including alternative promoter data.

5. Significance

This work establishes a new gold standard for CRISPRi library design. By decoupling CRISPRi optimization from CRISPRko assumptions, the authors provide:

1. A Robust Model (RS3i): A tool for accurately predicting guide efficacy based on chromatin state and sequence.
2. A Simple Filter (Seed Score): A computationally trivial method to eliminate a major source of false positives in CRISPRi screens.
3. An Optimized Resource (Katsano): A compact, high-sensitivity library that enables more cost-effective and accurate functional genomics screens.
4. Architectural Guidance: Clear evidence that N-terminal Zim3 fusions are the optimal configuration for dCas9-mediated repression.

These advancements allow researchers to perform CRISPRi screens with higher confidence, lower false discovery rates, and the ability to detect dependencies driven by alternative transcriptional isoforms.