LINA: A Linearizing Neural Network Architecture for Accurate First-Order and Second-Order Interpretations.

Area of Science:

• Computational biology
• Machine learning
• Genomics

Background:

• Neural networks excel in prediction but lack interpretability.
• Identifying salient features and interactions is vital for biomedical applications, especially predictive genomics.
• Current interpretation methods struggle with both feature importance and interaction identification.

Purpose of the Study:

• To develop a novel method, Linearizing Neural Network Architecture (LINA), for comprehensive model interpretation.
• To provide both first-order (feature importance) and second-order (feature interaction) interpretations.
• To validate LINA's performance against existing methods in synthetic and real-world datasets.

Main Methods:

• LINA combines a deep inner attention neural network with a linearized intermediate representation.
• It offers instance-wise and model-wise interpretations at both first and second orders.
• Performance was evaluated using synthetic datasets and a predictive genomics application.

Main Results:

• LINA demonstrated superior Spearman correlation for first-order interpretation compared to DeepLIFT, LIME, Grad*Input, and L2X on synthetic data.
• LINA achieved higher precision in identifying ground-truth feature interactions (second-order interpretation) than NID and GEH.
• In predictive genomics, LINA identified more significant single nucleotide polymorphisms (SNPs) and SNP interactions than other methods at controlled false discovery rates.

Conclusions:

• LINA provides accurate and versatile model interpretation for neural networks.
• The method effectively addresses the challenge of identifying salient features and interactions.
• LINA enhances the deployability of neural networks in interpretable biomedical applications, particularly genomics.