An ensemble clustering for mining high-dimensional biological big data

Dewan Md. Farid, Ann Nowé, Bernard ManderickInternational Journal of Design & Nature and Ecodynamics 2016 · DOI

Abstract

Clustering high-dimensional biological big data is an incredibly difficult and challenging task, as the data space is often too large and noisy. Conventional clustering methods can be inefficient because traditional distance measures may be dominated by noise across many dimensions. Furthermore, biological data requires identifying not only clusters of instances (e.g., genes) but also the specific sets of features (conditions) that define those clusters. This paper proposes an ensemble clustering approach incorporating feature selection and grouping. By combining k-means and similarity-based clustering followed by biclustering, the method identifies coherent sub-matrices within genomic data. Applied to 148 unlabeled exome datasets of Brugada syndrome (BrS), the approach successfully discovered previously unknown data patterns, demonstrating high performance in complex biological spaces.

Introduction

The volume of biological data is expanding at an unprecedented rate, often reaching petabyte scales. However, the "curse of dimensionality" poses a significant hurdle: as the number of features increases, the data becomes sparse, and traditional clustering algorithms like k-means struggle. In these high-dimensional spaces, distance metrics often fail because they are influenced more by irrelevant noise than by meaningful biological signals.

Moreover, biological relationships are often local. A group of genes might only show similar expression patterns under a specific subset of conditions. Standard global clustering methods miss these localized "sub-matrices." There is a critical need for frameworks that can simultaneously group instances and identify the relevant feature subsets that manifest those groups.

Methodology

The proposed framework utilizes a multi-stage pipeline designed to transition from coarse global grouping to fine-grained local pattern recognition.

The multi-stage pipeline for processing big biological data, from preprocessing to hidden pattern discovery.

The process consists of four primary stages:

  1. Unsupervised Feature Selection: To mitigate the impact of noise, the "maximum information compression index" is employed to filter out irrelevant features, retaining only those with high information content.
  2. Feature Grouping: The remaining features are categorized into nominal (categorical) and numeric types. This allows the system to apply specialized algorithms suited to each data type.
  3. Ensemble Clustering: A hybrid approach is used to generate robust clusters. Numeric features are processed via k-means clustering, while nominal features are handled using Similarity-Based Clustering (SCM). The results are integrated to form a set of "strong" clusters.
  4. $\delta$-Biclustering: Finally, a biclustering algorithm is applied to each cluster. This step searches for sub-matrices where a subset of instances exhibits highly coherent behavior across a subset of features, guided by mean squared residue scores.

Results

The framework was validated using a real-world genomic dataset consisting of 148 unlabeled exome datasets related to Brugada Syndrome (BrS). The raw data, initially containing 19,687 variants, was filtered down to 2,143 variants across 147 features (a mix of numeric and nominal) after preprocessing.

Comparison between the unclustered genomic data (left) and the five distinct clusters identified by the ensemble method (right).

The ensemble method successfully partitioned the variants into five distinct clusters. Subsequent application of $\delta$-biclustering allowed for the extraction of specific, previously unknown data patterns within these clusters. Qualitative analysis of the distribution plots confirms that the ensemble approach produces much more meaningful partitions than conventional methods, which often fail to separate such high-dimensional data effectively.

Discussion

The primary strength of this approach lies in its "hybrid advantage." By treating numeric and categorical data separately before merging them, the model respects the underlying nature of biological data types. Furthermore, using ensemble clustering as a pre-filter for biclustering significantly improves computational efficiency. Biclustering is notoriously expensive; by "zooming in" on promising clusters first, the search space for coherent sub-matrices is greatly reduced.

The use of mean squared residue scores provides a rigorous mathematical foundation for handling biological variability. It allows the model to identify "perfect" biclusters (score of 0) or "good enough" biclusters that fall under a specific threshold ($\delta$), effectively managing the inherent noise in genomic data.

Conclusion and Future Work

The proposed ensemble clustering framework offers a robust solution for mining high-dimensional biological big data. While successful in discovering novel patterns in Brugada syndrome exomes, several avenues for future research remain. Clinical validation of the identified clusters is necessary to determine their diagnostic or prognostic value. Additionally, moving from "hard" clustering to "soft" clustering—where instances can belong to multiple clusters—would better reflect the multi-functional nature of biological pathways. Finally, automating the selection of the $\delta$ threshold would enhance the tool's usability for researchers across various biological domains.

Summary updated

Abstract

Clustering of high-dimensional biological big data is incredibly difficult and challenging task, as the data space is often too big and too messy. The conventional clustering methods can be inefficient and ineffective on high-dimensional biological big data, because traditional distance measures may be dominated by the noise in many dimensions. An additional challenge in biological big data is that we need to find not only the clusters of instances (genes), but also for each cluster a set of features (conditions) that manifest the cluster. In this paper, we propose an ensemble clustering approach with feature selection and grouping for clustering high-dimensional biological big data. It uses two well-approved clustering methods: (a) k-means clustering and (b) similarity-based clustering. This approach selects the most relevant features in the dataset and grouping them into subset of features to overcome the problems associated with the traditional clustering methods. Also, we applied biclustering on each cluster that generated by ensemble clustering to find the sub-matrices in the biological data by the mean squared residue scores. We have applied the proposed clustering method on unlabeled genomic data (148 Exome datasets) of Brugada syndrome to discover previously unknown data patterns. Experiments verify that the proposed clustering method achieved high performance clustering results on high-dimensional biological big data.
BibTeX
@article{farid2016ensemble,
  author = {Dewan Md. Farid and Ann Nowé and Bernard Manderick},
  title = {{An ensemble clustering for mining high-dimensional biological big data}},
  journal = {International Journal of Design \& Nature and Ecodynamics},
  year = {2016},
  doi = {10.2495/dne-v11-n3-328-337},
}
Figure