GRAM

GRAM (GeneRAlized Model) is a software tool that predicts the impact of non-coding genetic variants on gene expression in a cell-specific manner. The tool integrates multiple genomic predictors to provide a well-defined and transferable measure of a variant's effect on the expression of its associated gene.

Key features of GRAM:

1. Feature engineering: GRAM uses LASSO, a regularized linear model, to identify the most predictive features. Transcription factor (TF) binding, especially for hub TFs in the regulatory network, was the most informative, while evolutionary conservation had minimal contribution.

2. Integration of SELEX features and expression profiles: GRAM combines a universal regulatory score based on in vitro SELEX TF binding data with cell-specific expression profiles, making it easily adaptable to different cell types.

3. Benchmarking: GRAM achieved AUROC scores of 0.72 in GM12878 and 0.66 in a multi-cell line dataset when tested on large-scale MPRA datasets.

4. Validation: Luciferase assays in MCF7 and K562 cell lines were used to evaluate GRAM's performance on targeted regions, highlighting the importance of carefully defining the prediction target.

5. Practical application: GRAM can fine-map causal variants within larger linkage-disequilibrium blocks associated with a phenotype of interest, such as eQTLs.