TL;DR
This paper introduces "launch-dCas9," a novel machine learning framework that accurately predicts the impact of guide RNAs (gRNAs) in CRISPR epigenome editing experiments. By integrating sequence and functional genomic features, launch-dCas9 helps researchers design more effective gRNAs for interrogating non-coding regulatory elements, a crucial step for advancing gene editing applications.
What Problem Does This Paper Address?
CRISPR epigenome editing, particularly CRISPRi (interference), is a powerful tool for understanding the function of non-coding regulatory elements like enhancers and promoters. However, designing effective guide RNAs (gRNAs) for these experiments is challenging. Current computational tools primarily focus on predicting the efficiency of DNA cutting in genetic CRISPR applications, not the nuanced effects of epigenomic perturbations. This leaves a gap in predicting gRNA impact on crucial cellular outcomes like cell fitness, gene expression, and even the abundance of the gRNA itself within cells. Researchers need better tools to select gRNAs that are not only functional but also efficient and less toxic, especially when targeting distal regulatory elements that are difficult to study.
Methods
The core of this work is the "launch-dCas9" framework, which employs two distinct machine learning models: a Convolutional Neural Network (CNN) and an XGBoost model.
- Feature Engineering: The models utilize a rich set of input features, including:
- Sequence-based features: The 20-bp gRNA protospacer sequence itself. CNNs use one-hot encoding and various kernel sizes to capture k-mer information, while XGBoost uses a bag-of-words approach with position-dependent and independent mono- and dinucleotide features.
- Functional Annotations: These include thermodynamic properties (e.g., hybridization free energy), epigenetic marks (e.g., H3K4me3, H3K27ac, ATAC-seq), and the "essentiality" of nearby genes.
- Model Architectures:
- CNN: Inspired by previous work on CRISPR-Cas9 efficiency prediction, the CNN uses multiple convolutional layers with varying kernel sizes to extract sequence features, which are then combined with functional annotations before passing through fully connected layers.
- XGBoost: This model uses a combination of engineered sequence features and functional annotations.
- Prediction Tasks: launch-dCas9 is designed to predict three key outcomes:
- Cell Fitness: The impact of gRNA perturbation on cell growth and proliferation.
- Wild-type Abundance: The expected number of gRNAs detected in wild-type cells, which can indicate packaging/transduction efficiency and potential toxicity.
- Gene Expression: The effect of gRNA perturbation on the expression of nearby genes.
- Evaluation: The models were trained and evaluated using a rigorous 5-fold cross-validation strategy, splitting data by chromosome to prevent information leakage.
Datasets
The study leverages large-scale experimental datasets for training and evaluation:
- wgCERES (Whole Genome CRISPR-dCas9 Epigenomic Regulatory Element Screening): This dataset, previously generated by the authors, includes over 1 million gRNAs targeting more than 100,000 putative cis-regulatory elements (CREs) in human K562 leukemia cells. This dataset was primarily used for predicting cell fitness and wild-type abundance.
- scCERES (Single-Cell CRISPR Epigenomic Regulatory Element Screening): This dataset, also from the authors' previous work, involves perturbing 3,029 distinct genomic regions (DHSs) with one gRNA per DHS and measuring differential gene expression in single cells. This was used for predicting gene expression impact.
- MHC Region Datasets: Additional scCERES datasets targeting the Major Histocompatibility Complex (MHC) region across K562, induced pluripotent stem cells (iPSCs), and human iPSC-derived neural progenitor cells (NPCs) were used to assess model generalizability across cell lines.
- Functional Annotation Data: Publicly available datasets from ENCODE and other sources were used for epigenetic marks (H3K4me3, H3K27ac, ATAC-seq) and transcription factor binding. Gene essentiality data was sourced from OGEE v3.
Key Insights and Ideas
- Multifaceted Prediction is Key: Predicting gRNA impact requires considering not just sequence but also functional genomic context. Both sequence and annotation features are complementary and crucial for accurate predictions, especially for cell fitness.
- Feature Importance Varies by Task: The relative importance of sequence versus functional annotation features differs significantly depending on the prediction task. For cell fitness, annotations are critical, while for enhancer regions, sequence information plays a more dominant role.
- Sequence Context Matters for CRISPRi: Unlike standard CRISPR/Cas9, specific nucleotide positions (1 and 3) are particularly predictive for CRISPRi screens, suggesting a unique mechanism of action that relies more on sequence context for targeting.
- Model Generalizability is Promising: launch-dCas9 models demonstrate reasonable transferability across different cell lines, indicating that models trained on one cell type or genome-wide screens can be adapted to optimize gRNA design in others, even with limited experimental data.
- Top gRNAs are Significantly Better: Method-prioritized top gRNAs are substantially more likely to exert significant effects (4.6-fold) compared to other gRNAs targeting the same regulatory region.
Key Results
- High Prediction Accuracy: launch-dCas9 achieved high prediction accuracy, with AUC values up to 0.81 for cell fitness prediction.
- Cell Fitness Prediction:
- For promoter regions, both CNN and XGBoost achieved mean AUCs of ~0.800 when using all features. Ablation studies showed that both sequence and annotation features were important.
- For enhancer regions, CNN and XGBoost achieved mean AUCs of ~0.742/0.730. Here, sequence-only models performed similarly to full models, suggesting sequence information was more critical.
- Wild-type Abundance Prediction: Models achieved strong correlations (Spearman correlation of 0.799/0.689) between predicted and true counts, with CNN outperforming XGBoost due to its ability to capture longer k-mer information.
- Gene Expression Prediction: Models achieved AUCs of ~0.630/0.635. Annotation-only models performed similarly or better than full models, highlighting the importance of functional annotations for this task.
- Improved gRNA Selection: Top predicted gRNAs were significantly more likely to have experimentally validated significant effects (adj. p < 0.05) and larger effect sizes (|log2FC|) compared to other gRNAs within the same genomic region. For example, 36.4-36.6% of top predicted gRNAs had adj. p < 0.05, compared to only 14.5% of other gRNAs.
- Cell Line Generalizability: Models showed reasonable performance across different cell lines in the MHC region, with K562 and iPSCs being easier to predict than NPCs.
Limitations and Future Work
- "Gray Zone" Data: The study acknowledges that many gRNAs fall into a "gray zone" where evidence for significant or insignificant effects is weak. This limits the ability to confidently label outcomes for training.
- Underpowered for Causal Identification: While the study explored the relationship between cell fitness and gene expression, it was underpowered to identify specific causal genes or cascades of expression changes.
- Future Directions:
- Semi-supervised Learning: The authors suggest that semi-supervised models could be explored to leverage the large amount of unlabeled data in the "gray zone."
- Causal Inference: Future experiments could aim to identify causal genes and regulatory pathways linking gRNA perturbations to cell fitness and gene expression.
- Broader Applications: The launch-dCas9 framework is presented as a prototype and could be further developed and applied to a wider range of CRISPR epigenomic editing experiments and cell types.
Summary updated