SPREd: a simulation-supervised neural network tool for gene regulatory network reconstruction.

Area of Science:

• Computational Biology
• Systems Biology
• Machine Learning

Background:

• Gene regulatory network (GRN) reconstruction from expression data is a critical challenge in systems biology.
• Current methods often rely on predicting gene expression from transcription factors (TFs) or use feature importance, which can be indirect.
• These approaches face limitations with high TF co-expression and small sample sizes.

Purpose of the Study:

• To develop a novel machine learning paradigm for direct GRN edge prediction.
• To introduce SPREd, a simulation-supervised neural network designed for accurate GRN inference.
• To evaluate SPREd's performance against state-of-the-art methods on synthetic and real biological data.

Main Methods:

• SPREd utilizes a neural network trained on synthetic data generated by a biophysics-inspired simulation.
• The model takes expression relationships (correlation, mutual information) between genes and TFs as input.
• It outputs binary labels indicating TF-gene regulatory relationships, directly inferring GRN edges.

Main Results:

• SPREd demonstrates superior performance compared to GENIE3, ENNET, PORTIA, and TIGRESS on synthetic datasets, especially with high TF co-expression.
• The method shows significant robustness to datasets with a limited number of conditions (columns).
• SPREd achieves comparable or better results than existing tools on gold-standard yeast GRN datasets.

Conclusions:

• SPREd offers a new, accurate, and efficient approach to GRN reconstruction.
• Its ability to handle complex regulatory relationships and limited data makes it a valuable tool.
• This work represents a step towards integrating biophysical principles into machine learning for gene regulation studies.