Avoiding pitfalls in L1-regularised inference of gene networks

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Statistical regularisation methods, including LASSO and L1 regularised regression, are widely employed for gene regulatory network (GRN) modeling.
• While theoretically capable of inferring correct network structures, these methods often introduce errors, omitting existing links or including spurious ones in practice.

Purpose of the Study:

• To investigate the performance of L1 regularisation methods in GRN inference, particularly with ill-conditioned gene expression data.
• To identify conditions under which L1 methods fail and to propose alternative strategies for accurate network reconstruction.

Main Methods:

• Analysis of L1 regularisation performance on gene expression data with varying condition numbers.
• Comparison of L1 regularisation with least-squares regression (with parameter thresholding) and robust network inference (intersection of non-rejectable models).

Main Results:

• L1 regularisation methods produce suboptimal GRN models when gene expression data matrices are ill-conditioned (high condition number), despite sufficient underlying information.
• Accurate network structures can be recovered from informative data using least-squares regression or robust network inference, even when L1 methods fail.

Conclusions:

• The condition number of the gene expression data matrix is a critical factor influencing the success of L1 regularised GRN inference.
• Researchers should assess data condition numbers and consider alternative methods like robust network inference or thresholded least-squares regression to avoid pitfalls associated with L1 regularisation, especially with commonly encountered ill-conditioned experimental datasets.