Multiple multi-sample testing under arbitrary covariance dependency.

Area of Science:

• Biostatistics
• Bioinformatics
• Computational Biology

Background:

• High-throughput biomedical devices generate large-scale, high-dimensional datasets.
• Extracting meaningful features from thousands of variables presents a significant analytical challenge.
• Existing methods struggle with simultaneous analysis of multiple features under complex dependencies.

Purpose of the Study:

• To develop a robust statistical procedure for evaluating associations between categorical response variables and multiple features.
• To establish a framework for large-scale multiple testing accommodating arbitrary correlation dependencies.
• To provide a practical method for feature selection in high-dimensional biomedical data analysis.

Main Methods:

• Utilized marginal multinomial regressions for individual feature analysis.
• Employed multiple marginal models to establish asymptotic joint normality of regression coefficients.
• Estimated the limiting covariance matrix for coefficients across all marginal models.
• Approximated the realized false discovery proportion for p-value thresholding.

Main Results:

• The proposed approach effectively balances true and false discoveries.
• Demonstrated a practical application using hyperspectral imaging data from MALDI instruments.
• Successfully applied the method to classify cancer subtypes based on spectral data.

Conclusions:

• The developed framework offers a significant advancement in analyzing complex, high-dimensional biomedical data.
• The method provides a sensible trade-off between identifying true associations and controlling false positives.
• This approach holds promise for enhancing clinical diagnosis, particularly in cancer research using MALDI imaging.