Deep mendelian randomization: Investigating the causal knowledge of genomic deep learning models

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Deep learning (DL) models excel at predicting genomic marks from DNA sequence.
• However, understanding if these models capture causal relationships between genomic marks remains a challenge.

Purpose of the Study:

• To introduce Deep Mendelian Randomization (DeepMR), a novel method for estimating causal relationships between genomic marks as learned by DL models.
• To assess the accuracy and limitations of DeepMR in identifying causal effects.

Main Methods:

• DeepMR combines Mendelian randomization principles with in silico mutagenesis.
• It generates both local (locus-specific) and global estimates of linear causal relationships between genomic marks.
• The method was tested using simulations of transcription factor (TF) interactions.

Main Results:

• DeepMR provided accurate and unbiased estimates of global causal effects in simulations.
• The method's accuracy decreased with sequence-dependent confounding, highlighting a limitation.
• Application to BPNet, a DL model for genomic data, validated known TF relationships and suggested novel ones.

Conclusions:

• DeepMR is a powerful tool for inferring causal relationships from genomic DL models.
• The findings support the utility of DeepMR in both validating existing biological hypotheses and generating new ones for TF interactions.