Knowledge Discovery Using Big Data in Biomedical Systems [Guest Editorial]

Sarath Chandra Janga, Dongxiao Zhu, Jake Y. Chen, Mohammed J. ZakiIEEE/ACM 2015 · DOI · OpenAlex

Abstract

This special section presents research from the 13th International Workshop on Data Mining in Bioinformatics (BIOKDD'14), held in conjunction with the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining on August 24, 2014, in New York, NY. The workshop brought together international researchers in data mining, systems biology, and bioinformatics to address the challenges of extracting knowledge from massive biomedical datasets and modeling complex cellular networks.

Introduction: The Biomedical Data Deluge

The rapid advancement of high-throughput technologies has created a "data deluge" in molecular and cellular biology. While we have gained deep insights into individual processes like metabolism or transcription, the next great challenge is integration. Biological systems do not work in isolation; they form complex, integrated networks. This editorial addresses the need for computational and statistical tools—specifically from data mining and machine learning—that can model the topological and dynamic properties of these unified cellular networks. The goal is to move beyond data collection toward building predictive models that can inform the development of novel therapeutics and diagnostics in post-genomic medicine.

Methodological Innovations

The editorial highlights three distinct methodological contributions from the BIOKDD’14 workshop designed to handle large and disparate biological datasets:

1. Parallelized Protein Structure Comparison

To solve the Contact Map Overlap (CMO) problem—a computationally expensive method for comparing protein structures—researchers proposed using approximate 2D-pattern matching. By breaking down large contact maps into smaller, matched pairs, the problem can be processed in parallel using fast heuristic algorithms. This approach achieved significantly better running times and, in some cases, better structural overlap for specific protein folds compared to traditional methods.

2. Flexible Plaid Models (BiP Algorithm)

Traditional clustering often assumes a gene belongs to only one group. The "BiP" algorithm uses "plaid models," which allow for an additive composition of contributions. This means a gene’s expression can be modeled as the sum of multiple overlapping biological processes occurring simultaneously. This approach successfully unraveled non-trivial functional interactions between biological processes that were previously obscured by simpler clustering models.

3. Ranked Neighborhood Comparison (RaNC)

RaNC is an intuitive, unsupervised method for structure detection. Instead of forcing data into strictly separated clusters, RaNC orders objects based on similarity and mutual overlap of nearest neighbors. This preserves a network structure where objects (like genes) can belong to multiple functional groups. The method demonstrated high robustness against experimental and measurement noise, effectively isolating outliers as "singletons" to prevent them from distorting the underlying data structure.

Key Insights and Results

  • The "Plaid" Reality of Biology: A significant shift in perspective is the move away from "hard" clustering. In biology, a single gene often performs multiple roles; plaid models and neighborhood ranking acknowledge that biological data is inherently overlapping and "fuzzy."
  • Parallelization via Pattern Matching: By treating protein structure comparison as a 2D pattern matching problem rather than a monolithic optimization task, researchers show that even NP-hard problems in bioinformatics can be made tractable through smart decomposition.
  • Robustness to Noise: The RaNC method offers a shift in perspective on noise. Instead of requiring aggressive pre-processing to remove outliers, the algorithm naturally isolates them, preventing them from distorting the underlying structure of the "clean" data.

Limitations and Future Work

While these advancements are significant, the "open challenge" remains the full integration of diverse cellular programs into a single, quantitative model.

  • Predictive Accuracy: There is a continuing need to move from descriptive models (what is happening) to accurate predictive models (what will happen) for complex diseases.
  • Clinical Translation: Exploiting these networks for therapeutic benefit requires further bridging the gap between data mining research and clinical application.
  • Scalability: As datasets continue to grow, the need for even more efficient parallelization and heuristic approaches remains a primary concern for the field.

Summary updated

Abstract

The papers in this special section were presented at the 13th International Workshop on Data Mining in Bioinformatics (BIOKDD'14) was organized in conjunction with the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining that was held on August 24, 2014 in New York, NY. It brought together international researchers in the interacting disciplines of data mining, systems biology, and bioinformatics at the Bloomberg Headquarters venue. The goal of this workshop is to encourage Knowledge Discovery and Data mining (KDD) researchers to take on the numerous challenges that Bioinformatics offers.
BibTeX
@inproceedings{janga2015knowledge,
  author = {Sarath Chandra Janga and Dongxiao Zhu and Jake Y. Chen and Mohammed J. Zaki},
  title = {{Knowledge Discovery Using Big Data in Biomedical Systems [Guest Editorial]}},
  booktitle = {IEEE/ACM},
  year = {2015},
  doi = {10.1109/tcbb.2015.2454551},
}

Loading related content...

Figure