A biologically informed hierarchical feature engineering and learning framework for genetic disorder prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Genetics

Background:

• Genetic disorder prediction is challenging due to complex inheritance, symptom interdependencies, and high-dimensional data.
• Traditional machine learning methods often fail to leverage the inherent biological structure and hierarchical nature of genetic disorders.
• Existing approaches lack effective utilization of biological context and hierarchical relationships.

Purpose of the Study:

• To develop a hierarchical, biology-informed framework for accurate genetic disorder prediction.
• To enhance feature engineering and selection by incorporating biological knowledge and explainability.
• To improve the classification performance by employing a two-stage hierarchical strategy.

Main Methods:

• A Genetically-Aware Feature Engineering Framework (GAFE-FE) was developed to create biologically relevant data representations.
• A Consensus Attention-SHAP-Frequency Feature Selection (CASF-FS) approach was used for robust and interpretable predictor selection.
• A Hierarchical Two-Stage Classification strategy was implemented, predicting subclasses first, then final disorders.

Main Results:

• The framework achieved 89.36% accuracy in Stage-1 (subclass prediction) and 86.92% in Stage-2 (disorder prediction) with GAFE-FE and XGBoost.
• Incorporating CASF-FS improved accuracy to 95.57%.
• The overall hierarchical two-stage framework demonstrated superior performance with 95.1% accuracy in Stage-1 and 98.7% in Stage-2.

Conclusions:

• The proposed hierarchical, biology-informed framework significantly improves genetic disorder prediction accuracy.
• GAFE-FE and CASF-FS effectively address challenges related to feature engineering and selection in genetic data.
• The hierarchical two-stage classification strategy provides a robust and accurate approach for complex genetic disorder diagnosis.