GRAM

GRAM predicts cell-type specific effects of non-coding genetic variants on gene expression by integrating transcription factor binding information and regulatory scores to identify expression-modulating variants.

Key Features:

• Predictive Focus: Predicts the expression-modulating effect of non-coding variants on their associated genes with transferability across cell types.
• Feature Engineering: Uses LASSO (Least Absolute Shrinkage and Selection Operator) regularized linear regression to identify predictive features, with transcription factor (TF) binding identified as highly predictive, especially for TFs that are network hubs, while evolutionary conservation contributes minimally.
• Data Sources: Integrates TF binding information inferred from SELEX (Systematic Evolution of Ligands by Exponential Enrichment) experiments and compares effectiveness to in vivo ChIP-Seq data.
• Model Integration: Combines a universal regulatory score with cell-type specific modifiers derived from expression profiles to produce context-specific predictions.
• Performance Evaluation: Benchmarked on large-scale MPRA (Massively Parallel Reporter Assay) datasets with reported AUROC of 0.72 in GM12878 and 0.66 across multiple cell lines, and additionally evaluated with luciferase assays in MCF7 and K562.
• Practical Applications: Provides a computational pipeline to facilitate fine-mapping of causal variants within linkage-disequilibrium blocks and identification of variants modulating expression, applicable to eQTL and phenotype-associated loci.

Scientific Applications:

• Fine-mapping causal variants: Prioritizes candidate causal non-coding variants within linkage-disequilibrium blocks for eQTL and phenotype-associated loci.
• Regulatory mechanism inference: Infers likely transcription factor–mediated regulatory effects of non-coding variants to link variants to changes in gene expression and disease-associated regulatory mechanisms.

Methodology:

Feature selection via LASSO; integration of SELEX-derived TF binding data (with comparisons to ChIP-Seq); construction of a universal regulatory score combined with cell-type specific modifiers from expression profiles; benchmarking using AUROC on MPRA datasets and targeted luciferase assay evaluations.