Gene function prediction using labeled and unlabeled data

Xing‐Ming Zhao, Yong Wang, Luonan Chen, Kazuyuki AiharaBMC 2008 · DOI

TL;DR

Predicting gene functions is crucial for understanding biology, but a major hurdle is the lack of negative examples (genes without a specific function). This paper introduces AGPS, a novel method that intelligently generates these missing negative samples from unlabeled data, leading to more accurate gene function predictions in yeast.

What Problem Does This Paper Address?

In the post-genomic era, a key goal is to decipher the functions of genes. Machine learning offers powerful tools for this, but most methods require both positive (genes known to have a function) and negative (genes known not to have a function) examples for training. The challenge in gene function prediction is that biological databases primarily provide positive annotations. It's difficult to definitively say a gene lacks a function, as genes can have multiple roles, and current annotations might be incomplete. Simply treating all unannotated genes as negative leads to imbalanced datasets and can degrade classifier performance. This paper tackles this "negative sample problem" head-on.

Methods

The core innovation of this paper is the Annotating Genes with Positive Samples (AGPS) algorithm. AGPS addresses the lack of negative samples through a multi-stage process:

  1. Functional Linkage Graph Construction: Multiple data sources (protein-protein interactions, gene expression profiles, and protein complexes) are integrated into a unified graph representing functional relationships between genes.
  2. Dimensionality Reduction with SVD: Singular Value Decomposition (SVD) is applied to this graph to reduce dimensionality, remove noise, and capture the dominant underlying structure. This is crucial for handling the high-dimensional nature of biological data.
  3. Intelligent Negative Sample Generation: This is the heart of AGPS. Instead of relying on external negative labels, AGPS iteratively learns from the available positive and unlabeled data.
  4. It starts by using a one-class SVM to draw an initial boundary around positive samples, identifying potential negative candidates from the unlabeled pool.
  5. Then, it employs a two-class SVM to refine and expand this negative set. This iterative process aims to find a representative subset of negative samples that best distinguishes positive from unlabeled data, avoiding the pitfalls of using all unannotated genes.
  6. Classification: Once representative negative samples are defined, AGPS uses a standard two-class SVM (or similar classifier) to predict the functions of unknown genes.

Datasets

The study utilizes several well-established yeast datasets:

  • Functional Annotation Data: Obtained from the Gene Ontology (GO) and MIPS Functional Catalogue (FunCat 2.0) databases. The paper focuses on 13 general functional classes.
  • Protein Interaction Data: Sourced from the BioGRID database.
  • Gene Expression Profiles: Downloaded from the Stanford Microarray Database (SMD).
  • Protein Complex Data: From the MIPS database.

The paper also defines a "test set" of 386 yeast genes that were newly annotated between March 2004 and 2006, providing a valuable benchmark for evaluating prediction accuracy on previously unknown genes.

Key Insights and Ideas

  • The "Negative Sample Problem" is a Major Bottleneck: The paper clearly articulates why the absence of reliable negative examples is a significant challenge in gene function prediction and how traditional approaches fail.
  • Intelligent Negative Sample Generation is Key: AGPS's strength lies in its ability to learn what constitutes a good negative sample from the available data, rather than making arbitrary assumptions.
  • Balancing Data Improves Performance: The results highlight that imbalanced datasets (many more negative than positive examples) significantly degrade classifier performance, reinforcing the importance of AGPS's approach to creating balanced training sets.
  • Integration of Multiple Data Sources is Powerful: Combining protein interactions, expression data, and complex information into a functional linkage graph provides a richer representation for prediction.
  • SVD is Essential for Handling Complex Data: Using SVD for dimensionality reduction is critical for making the integrated data computationally tractable and improving the signal-to-noise ratio.

Key Results

  • Improved Recall: AGPS demonstrated a higher recall rate compared to other methods, meaning it was better at identifying genes that truly possess a specific function among the unknown genes. This is crucial for biologists who are primarily interested in discovering gene functions.
  • Competitive Overall Performance: AGPS achieved competitive or superior performance across various metrics (precision, recall, F1 score, ROC score) when compared to established methods like PSOL, one-class SVMs, and two-class SVMs (both balanced and imbalanced).
  • Validation Against GO Annotations: When predicting functions for previously unannotated genes, AGPS successfully recovered GO annotations for approximately 71% of the genes that had matching GO terms.
  • Class-by-Class Superiority: Figures and tables show that AGPS consistently outperformed other methods on a class-by-class basis, particularly in recovering unknown genes.

Limitations and Future Work

  • Functional Class Independence: The current AGPS implementation treats each functional class independently. The authors acknowledge that biological functions are often correlated, and considering these correlations could further improve performance.
  • Generalizability: While AGPS was applied to yeast, the authors suggest it could be extended to other organisms and bioinformatics tasks.
  • Data Source Expansion: The method can readily incorporate other data types like protein subcellular localization and protein domains, which could enhance prediction accuracy.
  • Correlation Among Functions: Future work will focus on incorporating the correlation among biological functions into the prediction model.

Summary updated

Abstract

We proposed a new method for defining negative samples in gene function prediction. Experimental results on yeast genes show that AGPS yields good performances on both training and test sets. In addition, the overlapping between prediction results and GO annotations on unknown genes also demonstrates the effectiveness of the proposed method.
BibTeX
@inproceedings{zhao2008gene,
  author = {Xing‐Ming Zhao and Yong Wang and Luonan Chen and Kazuyuki Aihara},
  title = {{Gene function prediction using labeled and unlabeled data}},
  booktitle = {BMC},
  year = {2008},
  doi = {10.1186/1471-2105-9-57},
}
Figure