CITRUS

CITRUS infers patient-specific transcription factor activities and transcriptional regulatory programs from somatic alterations and chromatin-informed multi-omic data in cancer using a partially interpretable neural network with a self-attention mechanism.

Key Features:

• Partially Interpretable Neural Network Model: Integrates genomic, transcriptomic, and epigenomic data to model how somatic alterations affect transcriptional regulation.
• Self-Attention Mechanism: Employs a self-attention module to capture the contextual impact of somatic alterations on gene expression.
• Explicit Transcription Factor Representation: Includes hidden nodes that explicitly represent TF states and enable prediction of TF activities.
• TF–Target Linking via Binding Motifs in Open Chromatin: Learns connections between TFs and target genes based on binding motifs located in open chromatin regions.
• Patient-Specific TF Activity Prediction: Predicts TF activity profiles at the individual tumor level, enabling identification of TFs associated with somatic alterations.
• Multi-cancer Application on TCGA: Applied to data from 17 cancer types profiled by The Cancer Genome Atlas (TCGA) to analyze diverse tumor profiles.
• Multi-omic Integration to Model Somatic Alteration Influence: Integrates multiple omic layers to infer the influence of somatic alterations on downstream gene expression.

Scientific Applications:

• Precision oncology: Predicts patient-specific TF activities to identify regulatory drivers linked to somatic alterations for individualized interpretation.
• Transcriptional program characterization: Characterizes variation in transcriptional programs between and within tumor types.
• Mechanistic interpretation of gene dysregulation: Links somatic alterations to downstream TF activity and gene expression changes to explain tumor-specific dysregulation.
• Prioritization for therapeutic strategies: Identifies candidate TFs and regulatory programs that can inform personalized therapeutic hypotheses.

Methodology:

Uses a partially interpretable neural network with hidden nodes representing TFs and a self-attention mechanism, integrating genomic, transcriptomic, and epigenomic (open chromatin) data and using TF binding motifs in open chromatin regions to learn TF–target gene connections and infer TF activity from somatic alterations.