A Protein Interaction Information-based Generative Model for Enhancing Gene Clustering

Pratik Dutta, Sriparna Saha, Sanket Pai, Aviral KumarScientific 2020 · DOI

TL;DR

This paper introduces a novel generative model for gene clustering that leverages protein-protein interaction (PPI) data and weak supervision sources (multi-objective optimization and Gene Ontology) to improve clustering accuracy. By integrating PPI information as a latent variable, the model enhances its ability to infer probabilistic gene labels, outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Accurately grouping genes with similar functionalities is crucial for understanding cellular mechanisms, diagnosing diseases, and developing new treatments. Gene clustering is a primary method for this, but traditional approaches often struggle with the complexity and noise inherent in biological data. Ensemble clustering, which combines multiple clustering solutions, offers a promising avenue. However, creating high-quality, consensus clusters often requires extensive labeled data, which is expensive and time-consuming to obtain due to the need for expert knowledge. This paper addresses the challenge of improving gene clustering accuracy by developing a generative model that can effectively utilize readily available but imperfect "weak supervision" sources, thereby reducing reliance on costly ground truth labels.

Methods

The core of this work is a generative model, inspired by the Snorkel framework, that integrates multiple weak supervision sources to infer probabilistic gene labels. The novelty lies in incorporating protein-protein interaction (PPI) information as a latent variable within this generative model.

Here's a breakdown of the approach:

  • Weak Supervision Sources:
    • Multi-Objective Optimization (MOO)-based Clustering: This approach uses algorithms like NSGA-II to simultaneously optimize multiple objective functions (Fuzzy Partition Coefficient, PBM index, DB index) on gene expression data. The resulting non-dominated solutions (Pareto front) are treated as weak labels.
  • Gene Ontology (GO)-based Solutions: Leveraging the Gene Ontology Consortium (GOC), the largest knowledge base of gene functions, the authors generate three weak supervision sets based on molecular function (MF), biological process (BP), and cellular component (CC). These are derived from enrichment analysis using tools like PANTHER. While these are considered "near ground truth," they don't label all genes, making them imperfect weak sources.
  • Generative Model (Modified Snorkel): The proposed model takes these weak supervision labels as input. The key innovation is the integration of PPI information. This information is used to:
  • Weight MOO-based solutions: PPI data helps assign appropriate weights to different MOO-based solutions, reflecting their accuracy.
  • Enhance accuracy parameter: PPI information directly influences the accuracy parameter of the generative model, particularly for MOO-based solutions.
  • Model dependencies: The generative model is structured as a factor graph that captures dependencies between different weak supervision sources, including the PPI information.
  • Ensemble Approach: The model is built incrementally, starting with MOO-based solutions (M1), then adding PPI information (M2), and finally incorporating GO-based solutions (M3). This incremental approach allows for analysis of the performance gains at each stage.

Datasets

The study evaluates the proposed method on three real-life NCBI's Gene Expression Omnibus (GEO) datasets:

  • B-CLL chronic lymphocytic leukemia: Contains gene expression data from B-CLL samples.
  • ILD Interstitial lung disease: Gene expression data related to interstitial lung disease.
  • Prostate dataset: Gene expression data from prostate cancer studies.

The paper notes that these datasets were preprocessed to filter for differentially expressed (DE) genes using statistical tests and bootstrapped p-values. The preprocessed datasets are available in a supporting online repository.

Key Insights and Ideas

  • PPI as a Latent Variable is Powerful: Integrating protein-protein interaction data directly into the generative model as a latent variable significantly boosts its ability to infer accurate probabilistic gene labels. This highlights the importance of biological network information for understanding gene function.
  • Weak Supervision is Sufficient: The model demonstrates that combining multiple imperfect weak supervision sources (MOO clustering, GO annotations) can yield high-quality clustering results without relying on fully labeled data. This is a significant advantage in biological research where labeled data is scarce.
  • Incremental Integration Yields Performance Gains: Building the model by progressively adding weak supervision sources (MOO -> MOO+PPI -> MOO+PPI+GO) clearly shows how each component contributes to improved clustering performance, validating the design choices.
  • GO's "Near Ground Truth" Value: While not perfect, GO-based solutions, derived from curated biological knowledge, act as strong anchors for the generative model, significantly improving biological relevance compared to models relying solely on MOO-based solutions.
  • Generative Models as Ensemble Tools: The paper reinforces the utility of generative models as sophisticated ensemble methods capable of handling complex dependencies and noise in weak supervision signals.

Key Results

The proposed generative model (M3: MO + PPI + GO + GM) consistently outperformed other state-of-the-art techniques across three datasets (B-CLL, ILD, Prostrate) when evaluated using biological relevance metrics:

  • Biological Homogeneity Index (BHI): M3 achieved notable improvements. For example, on the BCLL dataset, M3 showed a 50.42% improvement over a MOO-based ensemble technique. On the ILD dataset, it achieved a 11.24% improvement.
  • Biological Stability Index (BSI): M3 achieved very high BSI scores, such as 0.994 for BCLL, 0.941 for ILD, and 0.945 for Prostrate.
  • Silhouette Index: M3 also showed competitive performance on the Silhouette index, indicating good cluster separation.

The paper also includes a statistical analysis (Welch's t-test) showing significant improvements in BHI for the proposed technique. The authors also performed biological enrichment analysis on the obtained gene clusters, confirming that genes within a cluster are more engaged in similar biological mechanisms compared to the remaining genes.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that GO-based solutions do not classify all genes, leaving some unmapped. This means the model cannot leverage GO information for every gene.
  • Relying solely on GO-based solutions could lead to overfitting due to their limited number.
  • The performance of the generative model is dependent on the quality of the input weak supervision solutions.
  • Future Work:
    • The authors plan to apply this ensemble method to other biomedical functions where real class labels are unavailable.
  • They aim to develop an enhanced version of the ensemble method by modifying the generative model's variables to improve performance further.
  • The paper suggests exploring other protein interaction databases and GO enrichment analysis tools.

Summary updated

Abstract

In the field of computational bioinformatics, identifying a set of genes which are responsible for a particular cellular mechanism, is very much essential for tasks such as medical diagnosis or disease gene identification. Accurately grouping (clustering) the genes is one of the important tasks in understanding the functionalities of the disease genes. In this regard, ensemble clustering becomes a promising approach to combine different clustering solutions to generate almost accurate gene partitioning. Recently, researchers have used generative model as a smart ensemble method to produce the right consensus solution. In the current paper, we develop a protein-protein interaction-based generative model that can efficiently perform a gene clustering. Utilizing protein interaction information as the generative model's latent variable enables enhance the generative model's efficiency in inferring final probabilistic labels. The proposed generative model utilizes different weak supervision sources rather utilizing any ground truth information. For weak supervision sources, we use a multi-objective optimization based clustering technique together with the world's largest gene ontology based knowledge-base named Gene Ontology Consortium(GOC). These weakly supervised labels are supplied to a generative model that eventually assigns all genes to probabilistic labels. The comparative study with respect to silhouette score, Biological Homogeneity Index (BHI) and Biological Stability Index (BSI) proves that the proposed generative model outperforms than other state-of-the-art techniques.
BibTeX
@inproceedings{dutta2020protein,
  author = {Pratik Dutta and Sriparna Saha and Sanket Pai and Aviral Kumar},
  title = {{A Protein Interaction Information-based Generative Model for Enhancing Gene Clustering}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-57437-5},
}
Figure