DeepMR

DeepMR estimates causal relationships between genomic marks learned by multi-task deep learning models by applying Mendelian randomization to in silico mutagenesis-derived perturbation effects.

Key Features:

• Causal Estimation: Estimates whether multi-task deep learning (DL) models capture causal relationships between genomic marks.
• Mendelian randomization with in silico mutagenesis: Combines Mendelian randomization principles with in silico mutagenesis to derive causal effect estimates from model-predicted sequence perturbations.
• Local and global estimates: Produces locus-specific (local) and global estimates of assumed linear causal relationships among genomic features.
• Simulation validation: Validated in simulations that recover pairwise causal relations between transcription factors (TFs), producing accurate and unbiased estimates of true global causal effects under controlled conditions.
• Confounding sensitivity: Coverage and performance can decline in the presence of sequence-dependent confounding.

Scientific Applications:

• BPNet transcription factor relationships: Applied to BPNet to estimate global relationships among TFs involved in cellular reprogramming, validating previously hypothesized TF interactions and suggesting novel potential interactions.

Methodology:

Apply in silico mutagenesis to multi-task DL models and use Mendelian randomization on the resulting perturbation effects to compute local and global causal effect estimates; validate estimates using simulations of pairwise TF causal relations.