The International Conference on Intelligent Biology and Medicine (ICIBM) 2016: from big data to big analytical tools

Zhandong Liu, W. Jim Zheng, Genevera I. Allen, Yin Liu, Jianhua Ruan, Zhongming ZhaoBMC Bioinformatics 2017 · DOI

TL;DR

This paper summarizes research presented at the International Conference on Intelligent Biology and Medicine (ICIBM) 2016, highlighting nine selected articles that showcase advancements in applying big data analytics and machine learning to biological and medical challenges. The featured work spans areas like cancer genomics, drug discovery, and systems biology, demonstrating novel computational tools and insights derived from large-scale biological datasets.

What Problem Does This Paper Address?

The rapid growth of biological data, particularly from genomics and high-throughput experiments, presents a significant challenge for traditional analytical methods. Researchers are faced with the need to develop sophisticated computational tools and analytical frameworks to extract meaningful insights from these massive datasets. This includes identifying patterns related to disease mechanisms, predicting patient outcomes, discovering new therapeutic targets, and understanding complex biological systems. The ICIBM 2016 conference aimed to bridge the gap between computational scientists and experimental biologists, fostering the development and application of advanced analytical tools for big biological data.

Methods

The paper highlights a diverse range of computational methods presented at ICIBM 2016. These include:

  • Unsupervised Deep Learning: A novel deep learning method was developed to identify low-dimensional representations of cancer gene expression data, aiming to uncover latent biological mechanisms and patient survival patterns.
  • Text Mining and Machine Learning: Algorithms were employed to mine medical abstracts, specifically identifying genes crucial for cancer cell survival by analyzing large corpora of scientific literature.
  • Short Tandem Repeat (STR) Profiling: A new tool, STRScan, was developed to identify and quantify STRs from whole-genome sequencing data, improving upon existing methods.
  • Splice Junction Identification: A novel algorithm, Read-Split-Fly (RSF), was created to identify non-canonically spliced regions across the genome.
  • Copy Number Variation (CNV) Estimation: An extension of the PennCNV tool was developed to estimate CNVs from whole-genome sequencing data by processing BAM files.
  • Differential Expression and Polyadenylation Analysis: An updated BioConductor package, XBSeq2, was introduced to improve the speed and accuracy of testing for differential gene expression and differential alternative polyadenylation.
  • Statistical Methods for Genetic Analysis: An expectation least squares (ELS) algorithm and binomial analysis of three-point gametes (BAT) were proposed for estimating gamete frequencies and recombination fractions.
  • Biomedical Entity Normalization: A novel convolutional neural network (CNN) architecture was developed for linking entity mentions in text to standard knowledge bases.
  • Database Development: A user-friendly database, ScaPD, was created to curate and present information on scaffold proteins and their associated signaling pathways.

Datasets

The paper mentions the use of several datasets for evaluation and demonstration:

  • Cancer Gene Expression Data: Used for unsupervised deep learning to find low-dimensional representations.
  • 1000 Genomes Project Data and Venter's Genome: Used to test the STRScan algorithm for STR identification.
  • 70 ENCODE Samples: Used to evaluate the RSF algorithm for identifying non-canonical spliced regions.
  • High-quality sample NA12878 from NIST database: Used to test the PennCNV-Seq method for CNV estimation.
  • Ten simulated artificial samples with CNVs: Also used to test PennCNV-Seq.
  • RNA-seq data: Used for testing differential expression (DE) and differential alternative polyadenylation (APA) with XBSeq2.
  • F2 dominant and codominant genotype data: Used to evaluate the ELS and BAT methods.
  • ShARe/CLEF and NCBI datasets: Used to evaluate the CNN-based biomedical entity normalization.
  • Scaffold protein data: Curated and computationally predicted data form the basis of the ScaPD database.

Key Insights and Ideas

  • Deep learning can uncover hidden biological patterns: The deep learning approach demonstrated that hidden layer representations can encode crucial information for classifying cancer subtypes and predicting patient survival, suggesting a powerful way to understand complex biological relationships.
  • Text mining can identify novel therapeutic targets: Mining scientific literature can effectively pinpoint genes essential for cancer cell survival, potentially revealing new targets for drug development that might be missed by other methods.
  • Improved tools for genomic analysis are crucial: The development of tools like STRScan and PennCNV-Seq highlights the ongoing need for more accurate and efficient methods to analyze large-scale genomic data, such as STRs and CNVs.
  • Non-canonical splicing is an important area of research: The identification of non-canonical spliced regions using RSF suggests that this phenomenon is more prevalent than previously thought and may play a significant role in disease.
  • Comprehensive databases are vital for systems biology: The creation of databases like ScaPD, which consolidate information on scaffold proteins and signaling pathways, is essential for advancing our understanding of complex biological networks.

Key Results

  • The unsupervised deep learning method provided novel insights into the mechanisms of tumorigenesis and patient survival by encoding relevant information in hidden layer representations.
  • Text mining algorithms identified genes that are highly involved in survival pathways and, importantly, several genes not previously implicated as essential oncogenes.
  • STRScan demonstrated a 20% increase in identification of STRs compared to existing approaches and was implemented in C with an open-source license.
  • Preliminary results using RSF on 70 ENCODE samples indicated that 5' splicing sites with U12-type signatures are more frequent than U2-type in non-canonical junctions.
  • The CNN-based biomedical entity normalization achieved accuracies of 90.30% and 86.10% on ShARe/CLEF and NCBI datasets, respectively, significantly improving upon existing baseline systems.
  • ScaPD currently contains 273 scaffold proteins and 1118 associated signaling pathways, providing a valuable resource for researchers.

Limitations and Future Work

  • Limitations of Deep Learning: While powerful, the paper notes that understanding the biological relationships encoded in hidden layer representations could lead to novel insights, implying that direct interpretability of these layers is an ongoing area of exploration.
  • Future Directions for Text Mining: The identified genes from text mining could serve as novel targets for treating complex diseases and cancers, suggesting a clear path for future experimental validation and therapeutic development.
  • Future of STR Analysis: The STRScan algorithm's open-source nature and implementation in C suggest it is ready for broader adoption and further development by the research community.
  • Future of Splice Junction Analysis: The RSF algorithm addresses a "gap in knowledge" regarding undiscovered spliced sequences, indicating that further research will likely uncover more such sequences and their functional roles.
  • Future of CNV Calling: PennCNV-Seq's integration into existing pipelines suggests its utility for more accurate CNV calling in genomic regions.
  • Future of RNA-seq Analysis: XBSeq2's improvements in speed and capacity for differential APA testing point towards more efficient and comprehensive RNA-seq data analysis.
  • Future of Genetic Analysis: The ELS algorithm's extensibility to search for accurate estimation of hidden variables suggests its potential application beyond genetic mapping.
  • Future of Entity Normalization: The high accuracies achieved by the CNN model suggest its potential for widespread use in biomedical text processing and knowledge extraction.
  • Future of Scaffold Protein Databases: ScaPD is positioned as a valuable resource for understanding how individual signaling pathways are regulated, implying future research will leverage this database for deeper mechanistic studies.

Summary updated

BibTeX
@article{liu2017international,
  author = {Zhandong Liu and W. Jim Zheng and Genevera I. Allen and Yin Liu and Jianhua Ruan and Zhongming Zhao},
  title = {{The International Conference on Intelligent Biology and Medicine (ICIBM) 2016: from big data to big analytical tools}},
  journal = {BMC Bioinformatics},
  year = {2017},
  doi = {10.1186/s12859-017-1797-3},
}
Figure