TL;DR
This paper introduces a novel semi-supervised approach to identify differentially expressed genes, particularly useful when sample labels are incomplete or unavailable. By extending the Optimal Discovery Procedure (ODP) to incorporate latent variable models, they demonstrate improved gene discovery performance, especially in scenarios with limited labeled data, by effectively leveraging unlabeled samples.
What Problem Does This Paper Address?
Identifying genes that show significant differences in expression levels across different conditions (differential gene discovery) is a cornerstone of gene expression analysis. However, real-world biological studies, especially clinical ones, often face challenges in obtaining complete and accurate condition labels for all samples. This can be due to physical, financial, or time constraints. Traditional supervised methods, which require all samples to be labeled, are thus not always applicable. While unsupervised methods can be used when no labels are available, they might miss subtle signals. This paper tackles the gap of efficiently identifying differential genes when only a subset of samples is labeled, a scenario termed "semi-supervised learning."
Methods
The core of this work is an extension of the Optimal Discovery Procedure (ODP) framework, originally designed for multiple hypothesis testing. The authors propose a latent variable model, specifically a mixture model, to represent gene expression. This model allows for the incorporation of "hidden variables" that can represent the underlying condition of a sample, even if that condition is not explicitly labeled.
Key methodological contributions include:
- Latent Variable Model: A mixture of normal distributions is used to model gene expression, where a hidden variable indicates whether a gene is "on" or "off" in a given sample.
- Unsupervised and Semi-supervised Scores:
- Unsupervised: This score is derived when no sample labels are available at all.
- Semi-supervised: This score is derived when some samples are labeled, and others are not. It modifies the prior probability of the hidden variable to incorporate the observed labels.
- ODP Extension: The ODP framework is adapted to these latent variable models. A crucial aspect is the exploration of two ways to handle hidden variables within ODP:
- Implicit Sharing: Hidden variables implicitly affect the alternative model through shared parameters.
- Explicit Sharing: Estimated hidden variables are explicitly shared across multiple tests. The paper argues that explicit sharing improves detectability.
- Likelihood Ratio Scores: The paper also discusses and compares standard likelihood ratio scores (LR-U for unsupervised, LR-S for supervised) as baselines.
Datasets
The paper utilizes both artificial and real-world datasets for evaluation:
- Artificial Data: A synthetic dataset was generated with 6,400 genes and 16 samples, divided into four groups (two inactive, two active). This dataset was designed to mimic different noise levels and expression patterns, allowing for controlled evaluation of the proposed methods.
- Real Data:
- Prostate Data Set: Consists of 52 prostate tumor samples and 50 normal samples, with expression data from Affymetrix U95Av2 arrays.
- Colon Data Set: Includes 40 colorectal tumor samples and 22 normal samples, measured using Affymetrix Hum6000 arrays.
For real data experiments, sample labels were intentionally masked to simulate semi-supervised scenarios with varying numbers of labeled samples.
Key Insights and Ideas
- Leveraging Unlabeled Data is Powerful: The most significant insight is that incorporating unlabeled samples in a semi-supervised manner can drastically enhance the stability and accuracy of gene rankings, especially when labeled data is scarce.
- Explicit Sharing of Hidden Variables Improves Performance: The study demonstrates that explicitly sharing the estimation of hidden variables across multiple genes within the ODP framework leads to better detection of truly active genes.
- ODP's Adaptability: The ODP framework, originally for supervised settings, proves to be highly adaptable to latent variable models, enabling its application in unsupervised and semi-supervised scenarios.
- Semi-supervised vs. Supervised Trade-offs: While semi-supervised methods (like ODP-SS) show clear advantages with limited labels, their performance can approach that of fully supervised methods (like ODP-S) when a sufficient number of labels are available.
- Unsupervised Scores as Pre-processing: Unsupervised scores can be valuable for initial filtering of genes, especially when the underlying biological differences are not strictly aligned with the primary labels of interest.
Key Results
- Artificial Data: ODP-U (unsupervised, explicit sharing) outperformed other unsupervised methods (LR-U, ODPP-U) in sensitivity and specificity, particularly in scenarios where genes shared a common biology or pattern.
- Semi-supervised Artificial Data: ODP-SS (semi-supervised, using all samples) achieved the best ROC curves, demonstrating superior performance compared to supervised methods (SAM, LR-S, ODP-S) when some labels were missing.
- Real Data (Prostate): ODP-SS consistently showed better concordance (Spearman's rank correlation and AUC metrics) than SAM and ODP-S, especially when the number of labeled samples was small. ODP-SS at zero labeled samples (equivalent to ODP-U) also performed moderately well.
- Real Data (Colon): While ODP-SS showed some improvement, the differences were less pronounced than in the Prostate dataset. Interestingly, ODP-SS performed worse than ODP-U in some Colon data scenarios with few labels, suggesting potential sensitivity to labeling errors or specific data characteristics.
Limitations and Future Work
- Labeling Errors: The paper acknowledges that labeling errors in real datasets can impact the performance of semi-supervised methods, and this warrants further investigation.
- FDR Estimation: The study focused on gene ranking and did not estimate the False Discovery Rate (FDR) for unsupervised cases, as permutation of class labels is not possible. This remains an area for future work.
- Model Complexity: The paper used a simple normal mixture model. Future work could explore other latent variable models (e.g., t-distribution, Cox's proportional hazards model) within the ODP framework to potentially improve gene ranking.
- Computational Cost: ODP has a computational cost of O(N^2), where N is the number of genes. While not explicitly detailed as a limitation, this could be a consideration for very large datasets.
Summary updated