Fast and accurate gene regulatory network inference by normalized least squares regression

Area of Science:

• Systems Biology
• Bioinformatics
• Computational Biology

Background:

• Accurate gene regulatory network (GRN) inference is crucial for systems biology.
• Existing methods like Least Squares Cut-Off (LSCO) struggle with extreme values, leading to false positives (hyper-connected genes).
• Computational methods must be fast and accurate to handle large modern experimental datasets.

Purpose of the Study:

• To develop a novel GRN inference method, LSCON, that addresses the limitations of LSCO.
• To improve the accuracy and reduce false positives in GRN inference, particularly in the presence of extreme data values.
• To provide a computationally efficient method for GRN inference.

Main Methods:

• Developed Least Squares Cut-Off with Normalization (LSCON), an extension of the LSCO algorithm.
• LSCON incorporates regularization through normalization to mitigate the impact of extreme values.
• Benchmarked LSCON against Genie3, LASSO, LSCO, and Ridge regression using accuracy, speed, and hyper-connectivity metrics.

Main Results:

• LSCON effectively reduces false positives by avoiding hyper-connected genes.
• LSCON demonstrates superior or comparable accuracy to LASSO, the previous best method, especially on datasets with extreme values.
• LSCON is an order of magnitude faster than LASSO due to its reliance on fast least squares regression.

Conclusions:

• LSCON is a highly accurate and efficient method for gene regulatory network inference.
• The normalization-based regularization in LSCON successfully tackles the issue of extreme values and hyper-connected genes.
• LSCON offers a significant advancement for systems biology research requiring robust GRN analysis.