TL;DR
This paper introduces OncoVar, a comprehensive database and analysis platform designed to identify and prioritize cancer-driving mutations, genes, and pathways. By integrating multiple bioinformatics algorithms and known driver data from large-scale cancer sequencing projects like TCGA and ICGC, OncoVar aims to provide researchers and clinicians with a powerful tool for understanding cancer mechanisms and identifying potential therapeutic targets.
What Problem Does This Paper Address?
Cancer is driven by accumulated genetic mutations, but distinguishing the truly "driver" mutations that cause cancer from the numerous "passenger" mutations that are merely along for the ride is a significant challenge. This is further complicated by the fact that many somatic mutations have neutral effects, making it difficult to pinpoint the critical genetic alterations. While numerous computational tools exist to predict driver mutations and genes, these often lack consistency and can produce false positives. Furthermore, there's a need for a centralized, user-friendly platform that integrates data from major cancer genomics initiatives (TCGA, ICGC) to facilitate the systematic identification and interpretation of these critical cancer drivers.
Methods
The OncoVar platform employs a multi-pronged approach:
- AI-Driver Method: The core of their mutation identification is a novel machine learning method called "AI-Driver." This method integrates 23 pathogenicity prediction algorithms, leveraging features like evolutionary conservation, structural constraints, and physicochemical attributes to predict the "driverness" of missense mutations. Mutations with a high AI-Driver score (≥0.95) and observed in at least two patients were classified as driver mutations.
- Driver Gene Prioritization: To identify driver genes, OncoVar combines predictions from established methods like MutSigCV and MutPanning. They also incorporate gene-level covariates such as the pLI score (probability of loss-of-function intolerance) and the maximum AI-Driver score of mutations within a gene. A Fisher's method is used to combine p-values from MutSigCV and MutPanning, followed by Benjamini-Hochberg correction. Genes with a corrected p-value < 0.05 are designated as "AI-DriverGenes."
- Consensus-Based Classification: A five-tiered consensus-scoring system is used to classify driver genes. This system combines predictions from 16 different driver sets (including established databases and their own AI-DriverGene predictions) using an improved Borda approach. Genes are then graded from non-pathogenic to pathogenic, with only Level 4 (pathogenic) genes considered true driver genes.
- Pathway Identification: Driver genes are used to identify oncogenic pathways through a hypergeometric test, with a Benjamini-Hochberg FDR threshold of 0.05.
- Data Integration and Visualization: OncoVar integrates data from TCGA and ICGC, providing four distinct views: 'Mutation', 'Gene', 'Pathway', and 'Cancer.' This allows researchers to explore relationships between these entities. The platform also includes tools for batch mutation annotation, cancer-specific pathway enrichment, and genome build conversion.
Datasets
- TCGA (The Cancer Genome Atlas): Data from 10,769 exomes across 33 cancer types were reanalyzed. This included harmonized results from the Multi-Center Mutation-Calling in Multiple Cancers (MC3) network.
- ICGC (International Cancer Genome Consortium): Data from 1,942 genomes across 18 cancer types were curated.
- gnomAD (Genome Aggregation Database): Over 270 million variants from gnomAD v2.1 were used to assess allele frequencies and gene-level loss-of-function metrics (pLI and LOEUF). Specifically, 14,967,411 mutations from 125,748 control individuals were aggregated.
- OncoKB: Used for identifying actionable driver genes.
- Other Databases: Various other databases and prediction tools (e.g., CGC, MutPanning, SIFT, PolyPhen2) were integrated or used for benchmarking.
Key Insights and Ideas
- The Power of Integration: The paper strongly emphasizes that combining algorithmic predictions with prior knowledge (like known driver genes and pathways) is crucial for accurate driver identification. Their AI-Driver method and consensus scoring system exemplify this.
- Actionable Targets for Therapy: OncoVar not only identifies driver mutations and genes but also highlights potentially "actionable" alterations, linking them to known drugs and suggesting opportunities for combination therapies. This bridges the gap between basic research and clinical application.
- Pan-Cancer vs. Cancer-Specific Drivers: The platform allows for the exploration of both common driver genes across many cancer types and those specific to individual cancers, revealing the diverse genetic landscapes of different malignancies.
- Novel Driver Discovery: Beyond cataloging known drivers, OncoVar identified novel driver genes and pathways, expanding our understanding of cancer etiology.
Key Results
- Driver Mutations: Identified 20,162 cancer driver mutations.
- Driver Genes: Identified 814 driver genes (713 from TCGA and 686 from ICGC pan-cancer cohorts, with additional cancer-specific drivers).
- Pathways: Identified 2,360 pathogenic pathways.
- Actionable Genes: Identified 30 actionable driver genes in 29 TCGA cancer types.
- Combination Therapy Potential: Revealed 148 potential combination therapies across 19 cancer types, with specific examples for COAD and UCEC.
- High Performance: The AI-Driver method demonstrated a superior and stable performance with high Area Under the ROC Curve (AUC) values (e.g., 0.972 for TCGA).
- Website Functionality: OncoVar provides a user-friendly interface for browsing, searching, and downloading this extensive data.
Limitations and Future Work
- Focus on Coding Regions: The current database primarily focuses on missense driver mutations affecting protein-coding genes. The authors acknowledge that copy-number alterations (CNAs), rearrangements, and non-coding somatic mutations (e.g., in regulatory regions) also play significant roles in cancer and are not comprehensively included.
- Continuous Updates: The authors plan to regularly update the database with new data from other cancer projects and literature, indicating an ongoing development process.
- In-silico Prescription: The integration of drug-gene interactions aims to improve "in-silico prescription" for actionable cancer drivers, suggesting future work in refining personalized treatment recommendations.
- Broader Scope: Future research could expand to include other types of genetic alterations and non-coding regions to provide an even more complete picture of cancer drivers.
Summary updated