Problems with the nested granularity of feature domains in bioinformatics: the eXtasy case

Area of Science:

• Biomedical data analysis
• Machine learning in bioinformatics
• Predictive modeling

Background:

• Biomedical data often possesses an implicit hierarchical structure.
• Overlooking this structure can introduce unrecognized bias into predictive models.
• This bias affects model evaluation and results in downstream applications.

Purpose of the Study:

• To identify and detail the bias arising from implicit hierarchical structures in biomedical data.
• To propose a novel, sampling-based solution to mitigate this bias.
• To evaluate the effectiveness of the proposed solution on synthetic and real-world data, specifically within the eXtasy variant prioritization framework.

Main Methods:

• Analysis of bias stemming from heterogeneous feature granularity in hierarchical data.
• Development and application of a simple, sampling-based bias mitigation technique.
• Evaluation using synthetic datasets to explore bias sources and extent.
• Integration and testing within the Random forest classifier core of the eXtasy framework.

Main Results:

• Standard Random forest classifiers and stratified bootstrapping methods are significantly challenged by heterogeneous feature granularity.
• The proposed sampling scheme effectively mitigates the identified bias during classifier training.
• In the eXtasy framework, a Random forest trained with the proposed method shows improved precision over the standard version, without compromising recall.

Conclusions:

• Heterogeneous granularity in biomedical data features presents a significant challenge for standard machine learning classifiers.
• The proposed sampling strategy effectively corrects bias, leading to more reliable predictive models.
• The enhanced Random forest classifier improves performance in variant prioritization, particularly at the top of gene lists, crucial for clinical applications.