NIMEFI: gene regulatory network inference using multiple ensemble feature importance algorithms.

Area of Science:

• Computational Systems Biology
• Bioinformatics
• Genomics

Background:

• Gene regulatory network (GRN) topology inference from high-throughput omics data remains a significant challenge in computational systems biology.
• Community-driven challenges like DREAM4 and DREAM5 have benchmarked network inference methods using gene expression data.

Purpose of the Study:

• To generalize the GENIE3 algorithm's regression decomposition strategy to various feature importance methods.
• To evaluate the performance of ensemble variants of different regression algorithms for network inference.
• To introduce and assess NIMEFI, a novel approach combining multiple ensemble feature importance algorithms.

Main Methods:

• Generalized the GENIE3 regression decomposition approach to include support vector regression, elastic net, random forest regression, and symbolic regression.
• Developed a subsampling strategy to create ensemble feature importance algorithms from any feature selection method.
• Implemented and evaluated NIMEFI, which averages predictions from multiple ensemble algorithms.

Main Results:

• Ensemble variants of regression methods significantly outperformed individual methods in network inference.
• The NIMEFI approach demonstrated superior performance compared to individual methods across general cases.
• Specific network structures may still favor single, optimized inference algorithms.

Conclusions:

• Ensemble learning is crucial for robust gene regulatory network inference.
• NIMEFI offers a powerful strategy for improving network inference accuracy by integrating multiple ensemble algorithms.
• The developed NIMEFI approach and its implementation are publicly available.