deepManReg

deepManReg applies deep manifold-regularized learning to integrate heterogeneous multi-modal biological data and predict phenotypes by aligning modality-specific feature manifolds into a unified latent space.

Key Features:

• Cross-Modal Manifold Learning and Alignment: Uses deep neural networks to learn underlying manifolds of multi-modal data and align features from different modalities into a unified latent space while preserving global consistency and local smoothness.
• Deep Neural Network-based Representation: Learns nonlinear, higher-order relationships across modalities through deep network representations of modality-specific feature manifolds.
• Cross-Modal Manifolds as Feature Graphs: Constructs cross-modal manifolds represented as feature graphs that capture inter-feature relationships across modalities.
• Graph-based Regularization of Classifiers: Uses the constructed feature graphs to regularize classifiers for phenotype prediction, improving predictive performance.
• Feature Prioritization: Prioritizes significant multi-modal features and interactions relevant to specific phenotypes through manifold-based regularization.
• Support for Single-cell Multi-modal Data: Applicable to single-cell multi-modal datasets, including Patch-seq datasets combining transcriptomics and electrophysiology from neuronal cells in the mouse brain.

Scientific Applications:

• Cellular Phenotype Prediction: Predicts cellular phenotypes from integrated multi-modal single-cell data such as Patch-seq.
• Feature and Interaction Discovery: Identifies key genes and electrophysiological features and their multi-modal interactions associated with phenotypes.
• Multi-modal Phenotype Analysis: Enables phenotype prediction and interpretation across diverse biological contexts involving heterogeneous data modalities.

Methodology:

Deep neural networks learn modality-specific manifolds and align them into a unified latent space that preserves global consistency and local smoothness; cross-modal manifolds are constructed as feature graphs and used to regularize classifiers and prioritize multi-modal features for phenotype prediction.