DeepPIG: deep neural network architecture with pairwise connected layers and stochastic gates using knockoff...

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Biology

Background:

• Feature selection is crucial for machine learning model performance.
• Knockoff filters control false discovery rates (FDR) but can lack detection power.
• Deep neural networks offer potential for enhanced feature selection.

Purpose of the Study:

• To introduce DeepPIG (Deep neural network with PaIrwise connected layers integrated with stochastic Gates), a novel feature selection model.
• To improve detection power within the knockoff filter framework for enhanced machine learning applications.
• To evaluate DeepPIG's performance against existing methods in synthetic and real-world datasets.

Main Methods:

• Developed DeepPIG, a deep neural network incorporating pairwise connected layers and stochastic gates.
• Utilized the knockoff filter framework for feature selection.
• Compared DeepPIG's detection power and FDR control against DeepPINK, STG, and SHAP on synthetic data.
• Assessed classification performance of DeepPIG-selected features on real-world microbiome and single-cell datasets.

Main Results:

• DeepPIG demonstrated higher detection power than baseline and recent models (DeepPINK, STG, SHAP) on synthetic data, particularly with weak feature signals.
• DeepPIG maintained preselected FDR levels.
• Features selected by DeepPIG led to superior classification performance on real-world cancer prognosis and classification tasks.

Conclusions:

• DeepPIG is a robust and powerful feature selection method.
• The approach is effective even when feature signals are weak.
• DeepPIG offers significant advantages for machine learning tasks requiring precise feature identification.