TL;DR
This paper introduces CoRNN, a deep learning model that accurately predicts genome A/B compartments using histone modification data, bypassing the need for expensive Hi-C experiments. The model achieves high accuracy and identifies key histone marks like H3K27ac and H3K36me3 as crucial predictors, offering a more accessible way to study genome organization.
What Problem Does This Paper Address?
The 3D organization of DNA within the cell nucleus is fundamental to gene regulation and overall cellular function. A key aspect of this organization is the segregation of the genome into "A" (active) and "B" (inactive) compartments. These compartments are typically identified using Hi-C, a technique that maps physical interactions between DNA segments. While Hi-C provides valuable insights into A/B compartment status and its cell-type specific variations, it is prohibitively expensive and technically challenging to generate for a wide range of cell and tissue types. This limitation hinders comprehensive studies of genome organization across diverse biological contexts. The paper aims to address this by developing a method to predict A/B compartments using more readily available data, specifically histone modification profiles.
Methods
The core of this work is the CoRNN (Compartment prediction using Recurrent Neural Network) model. CoRNN is designed as an end-to-end deep learning framework that leverages histone modification data to predict A/B compartments.
- Recurrent Neural Networks (GRUs): The model employs Gated Recurrent Units (GRUs), a type of recurrent neural network known for its ability to capture sequential dependencies. This is crucial for processing the linear sequence of histone modification signals along the genome. GRUs process the histone modification signals for each 100 kbp genomic region sequentially, allowing them to learn patterns and dependencies across consecutive genomic bins.
- Input Representation: The input to CoRNN for a given genomic region is a matrix representing the average ChIP-seq signal for six key histone modifications (H3K4me1, H3K27ac, H3K4me3, H3K36me3, H3K27me3, and H3K9me3) across 100 bins of 1000 bp each. Additionally, the model takes the mean compartment value from the training cell lines as a supplementary input.
- Fully Connected Network (FCN): The output from the GRU layers is concatenated with the mean compartment value and fed into a two-layer fully connected network. This network generates the final prediction for the A/B compartment.
- Training and Prediction: The model is trained in a cross-cell type setting, where it learns from five cell lines to predict the compartment status of a sixth, held-out cell line. The A/B compartment labels are derived from Hi-C data, with positive eigenvector values assigned to A compartments and negative values to B compartments. These are then binarized for classification.
Datasets
The study utilized the following datasets:
- Training and Validation: Hi-C and histone modification ChIP-seq data from six human cell lines: NHEK, IMR90, HMEC, GM12878, K562, and HUVEC.
- Independent Test Set: Histone modification and Hi-C data from human muscle and colon tissue samples (accession code ENCDO845WKR from the ENCODE portal).
Key Insights and Ideas
- Histone Modifications as Proxies for 3D Genome Structure: The study demonstrates that histone modification profiles, which are much easier to obtain than Hi-C data, contain sufficient information to predict A/B genome compartments. This significantly lowers the barrier to studying genome organization.
- Key Predictive Histone Marks: Through perturbation analysis, H3K27ac and H3K36me3 were identified as the most predictive histone marks for A/B compartment classification. This suggests specific roles for these marks in defining active and gene-body regions, respectively, which are strongly correlated with compartment status.
- Ambiguous Regions and Cell-Type Specificity: The model's mispredictions often highlight regions with ambiguous compartment status or those that are highly cell-type specific. These regions may represent bivalent enhancers or exhibit variability within cell populations, suggesting limitations of a simple two-state compartment model.
- Generalizability Beyond Cell Lines: CoRNN's ability to predict compartments in independent tissue samples, even when these samples were not part of the training set, underscores its robustness and potential for broad application.
Key Results
- High Prediction Accuracy: CoRNN achieved an average Area Under the ROC Curve (AUROC) of 90.9% across six cell lines in a cross-cell type prediction setting.
- Outperforming Baselines: CoRNN consistently outperformed the "mean baseline" (average compartment value across training cell lines) for five out of six cell lines, demonstrating its ability to capture cell-type specific signals. It also outperformed other baseline models like Random Forest and Logistic Regression.
- Predictive Power of Specific Marks: Perturbation analysis showed that H3K27ac and H3K36me3 were the most important histone marks, with models using only these two marks still achieving accurate predictions.
- Functional Relevance of Predictions: Regions correctly predicted by CoRNN but missed by the mean baseline showed a high overlap (99%) with candidate cis-regulatory elements (cCREs), suggesting that CoRNN's predictions capture functionally important genomic regions.
- Tissue Sample Performance: CoRNN achieved AUROC scores of 0.72 for muscle and 0.77 for colon tissue, demonstrating its generalizability to independent tissue samples, albeit with slightly lower scores compared to cell lines, potentially due to tissue heterogeneity.
Limitations and Future Work
- Ambiguous Compartments: The paper acknowledges that CoRNN struggles with regions exhibiting ambiguous compartment status, often characterized by weak Hi-C eigenvector signals or variability between biological replicates. These regions may also be influenced by the resolution of the Hi-C data (100 kb).
- Tissue Heterogeneity: The slightly lower performance on tissue samples compared to cell lines is attributed to the inherent heterogeneity of tissues.
- Future Directions: The authors envision using CoRNN to assign A/B compartment labels for the entire ENCODE catalog of cell lines and tissues, especially by leveraging imputed ChIP-seq signals. They also suggest that sub-compartment calling could provide additional information for difficult-to-predict regions. The availability of the software and trained model is expected to facilitate further research in this area.
Summary updated