STRUCTURED GAUSSIAN PROCESSES FOR UNCERTAINTY-AWARE CLASSIFICATION OF HIGH-DIMENSIONAL, SMALL-SAMPLED OMICS DATA

High-dimensional, small-sample omics datasets, such as microbiome profiles, present significant challenges for disease classification due to nonlinear interactions and severe class imbalance. Traditional classifiers often rely solely on feature abundance, ignoring the underlying biological interaction networks. This paper investigates whether incorporating topological context via graph-encoded pathways into a Gaussian process (GP) framework can improve classification performance and provide more reliable uncertainty estimates.

The authors propose a hybrid kernel for GP classification that combines standard abundance-based features with a graph-based similarity measure. This graph kernel is constructed by propagating node-level signals through a random walk on a sample-specific interaction network, resulting in a fixed-dimensional embedding that captures multi-hop structural relationships. To address class imbalance, the authors evaluate several strategies, including SMOTE, a custom adaptive oversampling technique (AdaLoRAS), and confusion-matrix-based threshold calibration to adjust predictive probabilities for test-time class priors.

The hybrid GP framework demonstrates competitive performance across three microbiome datasets. The most significant improvements were observed in the Sinha cohort, where graph-augmented models substantially outperformed unstructured baselines, suggesting that topological information is particularly valuable when disease is linked to coordinated community-level dysbiosis. The framework also provides well-calibrated predictive uncertainty, which is essential for distinguishing between confident predictions and ambiguous samples in clinical settings.

This work provides a principled Bayesian approach to omics classification that leverages existing biological knowledge. By moving beyond simple abundance-based models, the proposed framework offers a more robust way to handle the complexities of microbiome data, enabling researchers to better identify disease-associated signatures while maintaining a clear understanding of model confidence.