FCNN

FCNN measures conditional relatedness between genes by integrating co-expression and prior-knowledge similarities to predict gene-gene relationships under specific conditions.

Key Features:

• Input integration: Uses co-expression and prior-knowledge similarity matrices as model inputs.
• Neural architecture: Implements a convolutional neural network (CNN) architecture with a fully connected first layer to capture complex mapping relations between inputs and conditional relatedness.
• Mapping capability: Produces conditional relatedness scores between gene pairs reflecting context-specific relationships.
• Validation strategy: Employs grid-search 10-fold cross-validation for model selection and performance assessment.
• Performance benchmarks: Reported average accuracy improvements of 3.0% on validation and 2.7% on test sets compared to existing methods, with additional verification improvements of 1.8% on GeneFriends and 7.6% on DIP.
• Tested datasets: Evaluated on COXPRESdb, KEGG, TRRUST, the benchmark dataset by Xiao-Yong et al., and verified on GeneFriends and DIP.

Scientific Applications:

• Gene–gene interaction inference: Identifies gene-gene interactions by estimating conditional relatedness under specific conditions.
• Conditional relatedness quantification: Quantifies context-dependent relationships between gene pairs using integrated similarity inputs.
• Cancer gene network construction: Supports construction of cancer gene networks by providing condition-specific relatedness scores for network inference.
• Detection of relationships missed by classical methods: Captures intricate gene relationships that are often missed by classical machine learning and co-expression-only approaches.

Methodology:

Integrates co-expression and prior-knowledge similarities as inputs into a CNN-based model with a fully connected first layer, and evaluates performance via grid-search 10-fold cross-validation on COXPRESdb, KEGG, TRRUST and the Xiao-Yong benchmark, with verification on GeneFriends and DIP.