ChimerDB 4.0: an updated and expanded database of fusion genes

Ye Eun Jang, Insu Jang, Sunkyu Kim, Subin Cho, Daehan Kim, Keonwoo Kim, Jaewon Kim, Jimin Hwang, Sangok Kim, Jaesang Kim, Jaewoo Kang, Byungwook Lee, Sanghyuk Lee

TL;DR

ChimerDB 4.0 is a significantly updated and expanded database for fusion genes, crucial players in cancer development and potential therapeutic targets. This latest version integrates more deep sequencing data, leverages advanced deep learning for literature mining, and introduces new visualization tools to help researchers assess the functional significance of these genetic anomalies.

What Problem Does This Paper Address?

Fusion genes, formed when two genes are abnormally joined, are increasingly recognized as important biomarkers and therapeutic targets in various cancers. Identifying these fusion events from the vast amounts of genomic and transcriptomic data being generated is a significant challenge. Furthermore, understanding the functional implications of these fusions and their prevalence across different cancer types requires robust, curated databases and effective analytical tools. Previous versions of ChimerDB have addressed this by integrating data from sequencing analysis (ChimerSeq), text mining of publications (ChimerPub), and manual curation (ChimerKB). However, the rapid growth of sequencing data and the need for more sophisticated text analysis and functional assessment tools necessitated this update.

Methods

ChimerDB 4.0 builds upon its tripartite structure (ChimerKB, ChimerPub, ChimerSeq) with several key methodological advancements:

  • ChimerSeq: This module was enhanced by re-analyzing the entire TCGA (The Cancer Genome Atlas) dataset, comprising 10,565 tumor samples across 33 cancer types. The analysis employed multiple high-performance fusion detection tools (STAR-Fusion, FusionScan, TopHat-Fusion, and others) to identify fusion candidates. A "ChimerSeq-Plus" group was created, containing fusion genes predicted by at least two of these pipelines, to ensure higher reliability.
  • ChimerPub: A novel "deep learning"-based text-mining model, built on BioBERT, was developed to identify fusion gene mentions in PubMed abstracts. This model is capable of extracting fusion relations even when gene names are not explicitly delimited by symbols like '-' or '/'. The training data for this model was derived from existing ChimerKB entries. Extensive manual curation of full-text articles was performed to validate findings and remove false positives.
  • ChimerKB: This knowledge base was updated to incorporate newly identified fusion genes from the enhanced ChimerPub module, particularly those supported by experimental evidence.
  • Functional Significance Tools: New features were implemented to aid in assessing the functional significance of fusion genes. These include a fusion structure viewer (displaying domains), gene expression plots comparing fusion-positive and negative patients, and a STRING network viewer for protein-protein interactions.

Datasets

  • TCGA (The Cancer Genome Atlas): This was the primary dataset for ChimerSeq, with 10,565 tumor samples and 1,144 normal samples analyzed.
  • PubMed Abstracts: Approximately 30 million PubMed abstracts were mined for fusion gene information using the new deep learning model.
  • Public Resources: Data from other public resources like TumorFusions, TCGA FAWG, and ChiTaRS were integrated into ChimerSeq.
  • Human Kinome Database, ONGene, and TSGene 2.0: These were used for functional annotation of fusion genes.

Key Insights and Ideas

  • Deep Learning Revolutionizes Literature Mining: The application of a deep learning model (BioBERT) to PubMed abstracts significantly improved the ability to identify fusion gene mentions, even in complex sentence structures, leading to the discovery of many new potential fusion genes.
  • Manual Curation is Still King: Despite advanced computational methods, extensive manual curation of full-text articles was crucial for validating fusion genes, removing false positives (e.g., distinguishing gene fusions from gene-gene interactions), and ensuring high data quality.
  • Functional Context is Key: Simply identifying fusion genes is not enough. The integration of tools to visualize fusion structures, gene expression patterns, and protein interaction networks provides critical context for understanding their biological roles and potential as therapeutic targets.
  • TCGA Data is a Goldmine: The comprehensive re-analysis of the TCGA dataset, using multiple robust pipelines, has yielded a massive number of fusion gene candidates, significantly expanding the ChimerSeq module.

Key Results

  • ChimerDB 4.0 now contains 67,610 fusion gene pairs.
  • ChimerSeq includes 65,945 fusion candidates, with 21,106 identified by multiple prediction programs (ChimerSeq-Plus).
  • ChimerPub has been updated to include 1,257 fusion genes, with 777 supported by experimental evidence (ChimerPub-Plus).
  • ChimerKB now contains 1,597 fusion genes with publication support and experimental evidence.
  • The new deep learning model for ChimerPub achieved high precision (0.901) and recall (0.909) for identifying fusion genes from abstracts, even without explicit delimiters.
  • The database now offers enhanced visualization tools, including a fusion structure viewer, gene expression plots, and a STRING network viewer.

Limitations and Future Work

  • False Positives in Text Mining: While the deep learning model is powerful, text mining inevitably generates false positives, necessitating rigorous manual curation. The paper notes that even with extensive curation, some entries might still require further refinement.
  • Novel Fusion Classification: Transcripts not included in ChimerKB and ChimerPub were classified as "novel fusion." Further validation and characterization of these novel fusions would be beneficial.
  • Future Directions: The authors suggest that mining recurrent fusion genes from ChimerSeq can be a promising strategy for identifying novel cancer biomarkers. The continued expansion of the database with new sequencing data and literature will be ongoing. The development of more sophisticated tools to assess functional significance and predict therapeutic responses remains an important area for future work.

Summary updated

BibTeX
@article{jang2020chimerdb,
  author = {Ye Eun Jang and Insu Jang and Sunkyu Kim and Subin Cho and Daehan Kim and Keonwoo Kim and Jaewon Kim and Jimin Hwang and Sangok Kim and Jaesang Kim and Jaewoo Kang and Byungwook Lee and Sanghyuk Lee},
  title = {{ChimerDB 4.0: an updated and expanded database of fusion genes}},
  journal = {Nucleic Acids Res.},
  year = {2020},
  doi = {10.1093/nar/gkz1013},
}
Figure