TL;DR
This paper introduces ORVAL, a novel web platform designed to predict and explore combinations of genetic variants that may cause diseases, moving beyond the traditional focus on single-variant causes. By integrating machine learning with network visualization, ORVAL aims to uncover the complex genetic underpinnings of oligogenic diseases, which are influenced by multiple genetic factors.
What Problem Does This Paper Address?
The advent of Next-Generation Sequencing (NGS) has generated a massive amount of genetic data, revolutionizing our understanding of Mendelian diseases (caused by a single gene variant). However, many complex human diseases are not explained by single gene defects. Instead, they are thought to arise from the combined effects of variants in multiple genes, a phenomenon known as oligogenicity. Identifying these "variant combinations" is a significant challenge because current bioinformatics tools primarily focus on individual variants. This gap leaves a substantial portion of human diseases without adequate diagnostic and research tools, hindering our ability to understand their complex genetic basis and develop targeted therapies.
Methods
ORVAL employs a multi-stage pipeline to address this challenge:
- Variant Submission and Filtering: Users can submit genetic variants via VCF files or manual input. The platform allows for filtering based on Minor Allele Frequency (MAF), genomic/exonic positions, and synonymous effects, ensuring that the input variants align with those used for training the predictive models. A gene panel can also be specified to narrow down the analysis.
- Annotation: Submitted variants are annotated using public bioinformatics resources, including Ensembl, UniProt, dbNSFP, and others. Key annotations include CADD scores (for variant deleteriousness), protein sequences, gene recessiveness/haploinsufficiency probabilities, Gene Damage Index (GDI), and Biological Distance (based on protein-protein interactions).
- Variant Combination Prediction: ORVAL generates all possible bi-, tri-, and tetra-allelic combinations of variants within gene pairs. It then uses a machine learning model called VarCoPP (Variant Combination Pathogenicity Predictor) to predict the pathogenicity of these combinations. VarCoPP provides two scores: a Support Score (SS) and a Classification Score (CS), along with confidence levels (99% and 95%).
- Digenic Effect (DE) Prediction: For combinations predicted as potentially pathogenic, a DE predictor determines the likelihood of them being a "True Digenic," "Monogenic + Modifier," or "Dual Molecular Diagnosis" case.
- Network Construction and Visualization: The platform aggregates predictions at the gene-pair level to build an "Oligogenic Network." This network visualizes candidate pathogenic gene pairs, with edges colored by pathogenicity scores. Users can explore this network, filter genes by centrality or pathogenicity, and identify "Oligogenic Modules" (connected components of genes).
- Exploration Tools: ORVAL provides interactive visualizations, including:
- S-plots: To visualize digenic combinations based on SS and CS.
- Protein-Protein Interaction (PPI) Networks: To explore physical connections between proteins within an oligogenic module.
- Pathway Mappings (Tree-maps): To identify biological pathways enriched by the genes in a selected module.
- Feature Contribution Plots: To understand how VarCoPP arrived at its predictions for specific combinations.
Datasets
The paper mentions the use of several publicly available datasets for annotation and training:
- ExAC database: For Minor Allele Frequency (MAF) filtering.
- Ensembl GRCh37/hg19 genome version database: For variant annotation.
- UniProt: For protein sequences.
- dbNSFP: For gene recessiveness and haploinsufficiency probabilities.
- Digenic Diseases Database (DIDA): This database is mentioned as a crucial resource for identifying and classifying digenic effects, and it served as a basis for developing the VarCoPP predictor. The paper notes that DIDA is curated from scientific literature.
- Reactome: For pathway ontology information.
- ComPPI database: For protein-protein interaction data.
No new benchmark datasets were introduced in this paper; rather, ORVAL integrates existing resources.
Key Insights and Ideas
- Shifting Focus from Single Variants to Combinations: The core idea is that many diseases are not caused by one "bad" gene but by the interplay of several genetic factors. ORVAL provides a framework to explore these complex interactions.
- Network-Based Exploration for Biological Context: Simply predicting variant combinations isn't enough. ORVAL's strength lies in visualizing these combinations within biological networks (gene-gene, protein-protein, pathways), allowing researchers to understand the potential biological mechanisms.
- Machine Learning for Complex Genetic Models: Traditional methods struggle with oligogenic diseases. ORVAL leverages machine learning (VarCoPP) to predict the pathogenicity of variant combinations, a task that is computationally intensive and complex.
- Interactive Exploration for Hypothesis Generation: The platform is designed for interactive exploration, enabling clinicians and researchers to formulate new hypotheses about the genetic causes of complex diseases.
Key Results
The paper presents ORVAL as a functional platform with several key capabilities:
- Successful Integration of Multiple Resources: ORVAL effectively integrates variant annotation, machine learning prediction, and network visualization tools into a single, user-friendly web platform.
- Demonstrated Support for Oligogenic Hypotheses: The paper provides case studies illustrating how ORVAL can support hypotheses for patients with mild hypertrophic cardiomyopathy and congenital long QT syndrome, identifying candidate variant combinations and their potential biological relevance. For example, in one case, ORVAL confirmed suspicions for the gene pair KCNQ1 and KCNH2, which had the highest median pathogenicity score.
- Interactive Network Visualization: The platform generates interactive oligogenic networks, PPI networks, and pathway tree-maps, allowing for detailed exploration of candidate oligogenic signatures.
- Publication-Ready Output: Tables can be downloaded in TSV format, networks in GraphML, and S-plots in PNG, facilitating further analysis and reporting.
Limitations and Future Work
The authors acknowledge several limitations and outline future directions:
- Interpretation Requires Manual Effort: While ORVAL provides rich visualizations, interpreting the results still requires significant manual inspection by experts.
- Integration of Trio Data: The plan is to integrate parent (trio) variant data to better assess inheritance patterns.
- Additional Filtering Options: Further pre-filtering options for variants will be explored.
- Incorporating Phenotypic Information: Integrating patient phenotypic information and its relation to other phenotypes could provide more context.
- Expanding Network Resources: Exploring and integrating other network-related resources, such as disease networks, is planned.
- Improving Predictive Methods: The authors aim to improve or integrate new predictive methods for oligogenic analysis.
- Patient Cohort Analysis: A big data pipeline is being developed to support patient cohort analyses, which are not currently supported.
- DIDA Data Quality: The predictive quality is dependent on the DIDA database, which can be noisy. Efforts are underway to expand and improve DIDA, potentially introducing community rating mechanisms.
Summary updated