BIOMARKER SIGNATURE IDENTIFICATION IN “OMICS” DATA WITH MULTI-CLASS OUTCOME

Vincenzo Lagani, George Kortas, Ioannis TsamardinosComputational 2013 · DOI

TL;DR

This paper introduces a novel statistical test (MC-CIT) for identifying biomarker signatures in complex "omics" data with multiple outcome categories. Their method, integrated into a constraint-based discovery algorithm, outperforms existing approaches by finding smaller, more accurate predictive sets of molecular features, particularly for multi-class problems.

What Problem Does This Paper Address?

Identifying reliable biomarkers from high-dimensional "omics" data (like gene expression) is crucial for understanding diseases and developing diagnostics. A key challenge is finding the minimal set of molecular features that can accurately predict a given outcome. This becomes significantly harder when the outcome isn't just binary (e.g., healthy vs. diseased) but has multiple classes (e.g., different cancer subtypes or stages). Existing methods often struggle with these multi-class scenarios, either by making simplifying assumptions or by producing overly large and less interpretable biomarker sets. The authors aim to bridge this gap by developing a more robust statistical test for feature selection in these complex settings.

Methods

The core of this paper is the development of a new Multi-Class Conditional Independence Test (MC-CIT). This test is designed to work within constraint-based (CB) local causal discovery algorithms, such as the Max-Min Parent Children (MMPC) algorithm. CB algorithms rely on a series of conditional independence tests (CITs) to determine which variables are relevant for predicting an outcome.

The MC-CIT is novel because it's specifically tailored for situations where you have:

  • Continuous predictors: These are your "omics" measurements (e.g., gene expression levels).
  • Multi-class outcomes: The target variable has more than two categories (e.g., different cancer stages).

Instead of relying on older tests like the Fisher Z test (which assumes linear relationships and normal errors, often violated in "omics" data) or discretizing data for tests like the G² test, MC-CIT uses multinomial logistic regression. It frames the problem as a nested model selection task: it compares a "full" model (predicting the outcome using the predictor and conditioning variables) against an "alternative" model (predicting the outcome using only the conditioning variables). The log-likelihood ratio test is used to determine if the predictor adds significant predictive power, thus assessing conditional independence. For ordinal outcomes, they use an extension called Ordered Logit models.

Datasets

The authors evaluated their MC-CIT method using seven real-world, high-dimensional gene-expression datasets from cancer-related studies. These datasets are publicly available through the Gene Expression Omnibus (GEO) database. The datasets vary in terms of disease, outcome type (categorical or ordinal), number of classes, samples, and variables.

Key Insights and Ideas

  • The limitations of existing tests: The Fisher Z test, while common, makes assumptions (linearity, normality) that are often violated in "omics" data, leading to suboptimal biomarker discovery. Discretizing continuous data for tests like G² can lead to information loss.
  • The power of multinomial logistic regression: By leveraging multinomial logistic regression, MC-CIT can directly model the complex relationships between continuous predictors and multi-class outcomes without making restrictive assumptions.
  • Smaller is better (and more accurate): The study strongly suggests that MC-CIT not only identifies more accurate biomarker signatures but also does so with a significantly smaller number of features compared to traditional methods. This is crucial for practical applications where fewer biomarkers mean lower costs and easier implementation.
  • Constraint-based discovery is effective: The MMPC algorithm, when paired with MC-CIT, proves to be a powerful framework for biomarker signature identification, demonstrating competitive performance against other established methods like Lasso.

Key Results

  • Improved Accuracy: MC-CIT, when used with the MMPC algorithm, consistently achieved comparable or better accuracy than the Fisher Z test across multiple datasets. For some datasets, the difference was statistically significant.
  • More Parsimonious Signatures: A major finding is that MC-CIT identified significantly smaller biomarker signatures. On average, MC-CIT selected substantially fewer features than the G² test and the Lasso method, while often achieving similar or better predictive performance. For example, the global average number of selected features was 5.48 for MC-CIT versus 19.59 for G² and 30.61 for Lasso.
  • Competitive Performance: MC-CIT performed competitively with the Lasso feature selection method, often identifying equally predictive signatures but with fewer features.
  • Robustness: The MC-CIT method showed good stability, with low standard deviations in accuracy across cross-validation folds.

Limitations and Future Work

  • Computational Requirements: The authors acknowledge that they did not optimize the implementation of MC-CIT, so a direct comparison of computational speed with Fisher Z test was not performed.
  • Data Type: The study was limited to gene expression data. While the authors expect similar conclusions for other "omics" data types (RNA-seq, methylation), this remains an area for future validation.
  • Parameter Tuning: The performance of MMPC and Lasso can be sensitive to parameter choices (k-max, alpha, lambda). The authors explored a range of parameters, but further investigation into optimal parameter configurations could be beneficial.
  • Future Directions: The paper suggests that MC-CIT is a valuable tool for biomarker discovery in multi-class "omics" data and encourages its adoption over existing methods like the Fisher Z test. Further validation on diverse datasets and "omics" modalities would strengthen these findings.

Summary updated

Abstract

Biomarker signature identification in "omics" data is a complex challenge that requires specialized feature selection algorithms. The objective of these algorithms is to select the smallest set(s) of molecular quantities that are able to predict a given outcome (target) with maximal predictive performance. This task is even more challenging when the outcome comprises of multiple classes; for example, one may be interested in identifying the genes whose expressions allow discrimination among different types of cancer (nominal outcome) or among different stages of the same cancer, e.g. Stage 1, 2, 3 and 4 of Lung Adenocarcinoma (ordinal outcome). In this work, we consider a particular type of successful feature selection methods, named constraint-based, local causal discovery algorithms. These algorithms depend on performing a series of conditional independence tests. We extend these algorithms for the analysis of problems with continuous predictors and multi-class outcomes, by developing and equipping them with an appropriate conditional independence test procedure for both nominal and ordinal multi-class targets. The test is based on multinomial logistic regression and employs the log-likelihood ratio test for model selection. We present a comparative, experimental evaluation on seven real-world, high-dimensional, gene-expression datasets. Within the scope of our analysis the results indicate that the new conditional independence test allows the identification of smaller and better performing signatures for multi-class outcome datasets, with respect to the current alternatives for performing the independence tests.
BibTeX
@inproceedings{lagani2013biomarker,
  author = {Vincenzo Lagani and George Kortas and Ioannis Tsamardinos},
  title = {{BIOMARKER SIGNATURE IDENTIFICATION IN “OMICS” DATA WITH MULTI-CLASS OUTCOME}},
  booktitle = {Computational},
  year = {2013},
  doi = {10.5936/csbj.201303004},
}
Figure