Semi-supervised data-integrated feature importance enhances performance and interpretability of biological...

Area of Science:

• Computational Biology
• Machine Learning
• Bioinformatics

Background:

• Standard model performance metrics do not guarantee alignment with human knowledge, limiting real-world applicability.
• Integrating prior knowledge into machine learning models is crucial for enhancing relevance and interpretability.

Purpose of the Study:

• To introduce Semi-Supervised Data-Integrated Feature Importance (DIFI), a novel method for integrating a priori knowledge into model feature weighting.
• To improve the alignment between model feature importance and human knowledge.

Main Methods:

• DIFI numerically integrates external knowledge, represented as a sparse knowledge map, into the model's feature weighting process.
• A loss function incorporates the similarity between the knowledge map and the model's feature map to guide feature weighting.
• The method was tested on two biological tasks: cancer type prediction and enzyme/non-enzyme classification.

Main Results:

• DIFI improved neural network performance in both cancer type prediction from gene expression data and enzyme classification from protein sequences.
• The method resulted in feature weightings that are interpretable and aligned with biological knowledge (cancer biomarkers and catalytic residues).
• DIFI demonstrated enhanced model alignment and interpretability.

Conclusions:

• DIFI is an effective method for injecting domain knowledge into machine learning models, leading to improved performance and interpretability.
• The approach successfully aligns model feature importance with human expertise in biological applications.
• Code and models are publicly available for reproducibility and further research.