NLCD: A method to discover nonlinear causal relations among genes

Area of Science:

• Genomics
• Systems Biology
• Bioinformatics

Background:

• Distinguishing correlation from causation is crucial in biology, especially with observational data.
• Existing causal discovery methods often assume linear relationships, limiting their application to complex biological systems.
• Genomics data frequently exhibit nonlinear interactions that are missed by linear models.

Purpose of the Study:

• To develop a novel method for discovering nonlinear causal relationships in genomics data.
• To extend existing linear causal discovery techniques to handle nonlinearities.
• To improve the accuracy and scope of causal gene network inference.

Main Methods:

• Developed NonLinear Causal Discovery (NLCD) using nonlinear regression and conditional feature importance.
• Extended statistical tests from the linear Causal Inference Test (CIT) framework.
• Benchmarked NLCD against CIT, Findr, and MRPC using simulated and real genomic datasets.

Main Results:

• NLCD achieved comparable performance to state-of-the-art methods for linear relations (AUPRC=0.94).
• NLCD significantly outperformed existing methods in detecting nonlinear gene relations (AUPRC=0.76 vs. 0.56-0.73).
• NLCD identified known causal gene interactions in yeast and revealed novel causal pairs (IRF1 → PSME1, HLA-C → HLA-T) in human muscle tissue.

Conclusions:

• NLCD effectively discovers nonlinear causal relationships in genomics data.
• The method advances the ability to infer complex causal gene networks.
• NLCD offers a promising approach for understanding gene regulation in human tissues.