TL;DR
This paper introduces "peaBrain," a novel two-stage deep learning framework that predicts how genetic variations in non-coding DNA affect gene expression. peaBrain significantly improves our ability to interpret the functional consequences of genetic variants, particularly those found in non-coding regions, which are often challenging to decipher.
What Problem Does This Paper Address?
Genome-wide association studies (GWAS) have identified thousands of genetic loci associated with complex traits and diseases. However, a major challenge is that most of these associated variants are located in non-coding regions of the genome. Understanding how these non-coding variants influence gene expression and, consequently, disease risk, is crucial for translating GWAS findings into functional insights and potential therapeutic targets. Existing methods struggle to accurately predict the impact of these non-coding variants, especially when considering tissue-specific effects and individual genetic variations.
Methods
The authors developed a two-stage computational framework called "peaBrain."
- Stage 1: This stage focuses on predicting the mean tissue-specific abundance of genes. It uses convolutional neural networks (CNNs) to model the transcriptional machinery of a tissue, taking DNA sequences and optionally epigenetic annotations as input. This stage aims to capture the general regulatory landscape of gene expression in a given tissue.
- Stage 2: This stage incorporates the transcriptomic consequences of individual genotype variation. It builds upon Stage 1 by using individual-level genetic data to predict gene expression on a subject-by-subject basis. This allows for a more personalized prediction of how specific genetic variations affect gene expression.
Key methodological aspects include:
- Convolutional Neural Networks (CNNs): Used for their ability to learn complex patterns from sequence data.
- Two-stage modeling: Separating the prediction of general gene abundance from the impact of individual genetic variation.
- Tissue-specific modeling: Allowing the framework to learn distinct regulatory mechanisms across different tissues.
- Non-coding impact scores: Generating scores that quantify the functional impact of genomic positions on gene expression.
- Gene embeddings: Creating compressed representations of genes that capture their regulatory context.
Datasets
The study utilized several large-scale datasets:
- GTEx (Genotype-Tissue Expression) dataset: This was the primary dataset for training and evaluating Stage 1 models, providing gene expression and genotype data across numerous human tissues.
- Geuvadis dataset: Used for validating Stage 2 predictions, particularly for eQTL analysis in lymphoblastoid cell lines (LCLs).
- COSMIC (Catalogue of Somatic Mutations in Cancer): Used to assess the predictive capacity of non-coding metrics for identifying cancer-associated somatic mutations.
- Ensembl, UCSC, and other public databases: Used for genomic annotations, reference genomes, and downloading various functional genomics data (e.g., epigenomic marks, transcription factor binding sites).
- Roadmap Epigenomics Project: Provided consolidated epigenomes for tissue-specific annotations.
Key Insights and Ideas
- Non-coding regions hold significant regulatory power: The study demonstrates that a substantial portion of the variation in gene expression across tissues is explained by sequences within the 4kbps core promoter region. This highlights the critical role of non-coding DNA in gene regulation.
- Tissue-specificity is paramount: Incorporating tissue-specific epigenetic annotations significantly improves the prediction of gene expression, underscoring that gene regulation is highly context-dependent.
- Deep learning excels where linear models falter: peaBrain, a deep learning framework, significantly outperforms traditional regularized linear models in predicting gene expression and the consequences of individual genetic variation.
- Predicting functional impact beyond correlation: peaBrain's impact scores are derived directly from expression prediction, making them independent of external annotations and providing a more direct measure of functional consequence.
- Identifying functional tissues for complex traits: Tissue-specific peaBrain scores can pinpoint the most relevant tissues for complex traits, outperforming methods relying solely on eQTL-GWAS co-localization.
- Unlocking functional architecture inaccessible to current assays: Stage 2 of peaBrain can identify putatively functional variants that are missed by high-throughput experimental methods, suggesting it can reveal regulatory mechanisms not easily captured by current experimental techniques.
Key Results
- Stage 1: peaBrain models (class-C, with tissue-specific annotations) captured over 50% of the variance in mean gene abundance across most GTEx tissues. Even DNA-only models (class-A) captured nearly a fifth of the variance.
- Stage 2: peaBrain significantly outperformed regularized linear models in predicting individual genotype variation's consequences. For example, in LCLs, peaBrain captured cis-heritability for 113 genes, compared to only 28 genes for linear models.
- Non-coding impact scores: peaBrain's non-tissue-specific impact scores were significantly predictive of cancer-associated somatic mutations and allele-specific transcription factor binding, outperforming CADD and Eigen in certain tasks.
- Tissue identification for complex traits: peaBrain framework identified more relevant functional tissues for traits like LDL, HDL, total cholesterol, and triglycerides compared to the RTC (eQTL-based) method. For instance, liver was ranked 3rd by peaBrain but 32nd by the RTC method for HDL.
- Functional variant identification: peaBrain predictions were significantly enriched in DNase accessible sites and transcriptionally active regions, and depleted from repressed regions, demonstrating its ability to identify functionally relevant variants.
Limitations and Future Work
- "Black-box" nature: The authors acknowledge that peaBrain, as a deep learning model, is a "black-box" approach. While they implemented strategies to mitigate bias (dropout, out-of-sample testing, comparisons), fully elucidating the scientific rationale behind its predictions remains challenging.
- Computational cost: Training peaBrain models, especially for Stage 2 with 1Mbps input sequences, can be computationally intensive, requiring significant resources.
- Future Work: The authors suggest that the gene embeddings generated by peaBrain can be used for downstream analyses like network-based studies, potentially discovering co-regulation and co-expression patterns. They also highlight the importance of the invariant genomic context and distance to the TSS for interpreting non-coding variation effects. Further exploration of the model's ability to predict more complex traits is also implied.
Summary updated