MultiCapsNet

MultiCapsNet performs integrative, interpretable classification of heterogeneous biological datasets by applying Capsule Network–based architectures (CapsNet, scCapsNet) to generate importance scores for data sources such as variant calls, single-cell RNA sequencing (scRNA-seq), protein-protein interactions (PPI), and protein-DNA interactions (PDI).

Key Features:

• Integration of Diverse Data Types: Integrates variant calls, scRNA-seq, PPI, and PDI across heterogeneous formats and lengths.
• Interpretable Classification: Generates direct importance scores for data sources to quantify each source's contribution to predictions without additional post hoc steps.
• Modular Input Handling: Processes modular features and PPI cluster nodes that are challenging for conventional interpretable methods such as linear regression or random forest.

Scientific Applications:

• Variant Call Classification: Classifies labeled variant call datasets with performance comparable to traditional neural networks while providing direct importance scores per data source.
• Single-Cell RNA Sequencing Analysis: Integrates scRNA-seq with PPI and PDI to classify cell types with accuracy comparable to neural networks and random forest and to identify roles of transcription factors (TFs) and PPI cluster nodes.
• Comparison with Existing Methods: Identified cell-type relevant TFs that were consistent with SCENIC in comparative analyses.

Methodology:

Built on Capsule Networks (CapsNet) and scCapsNet, the architecture captures spatial hierarchies in data and produces direct importance scores for input data sources to enable interpretable classification.