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 factor (TF) expression, indirectly inferring regulatory relationships.
• These approaches can struggle with complex biological systems, particularly with high TF co-expression and limited experimental conditions.

Approach:

• We developed SPREd, a simulation-supervised neural network that directly predicts GRN edges.
• SPREd utilizes expression relationships between genes and TFs, and between TF pairs, as input.
• The model is trained on synthetic data generated by a biophysics-inspired simulation incorporating diverse GRN architectures and regulatory dynamics.

Key Points:

• SPREd demonstrates superior performance compared to state-of-the-art tools like GENIE3, ENNET, PORTIA, and TIGRESS on synthetic datasets with high TF co-expression.
• The approach exhibits robustness to datasets with a limited number of conditions, a common limitation of existing GRN inference methods.
• Evaluation on yeast GRN benchmarks shows SPREd achieving comparable or better results than existing methods, highlighting its accuracy and speed.

Conclusions:

• SPREd offers a novel paradigm for GRN inference by directly predicting regulatory edges.
• Its performance on both synthetic and real biological data underscores its potential for advancing systems biology research.
• This work represents a significant step towards integrating biophysical principles into machine learning-based GRN reconstruction.