SIGMA: self-supervised inference of gene networks via masked auto-encoding

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Gene regulatory network (GRN) inference is crucial for understanding disease pathways.
• Current methods face challenges with accuracy (unsupervised) and data scarcity (supervised).
• Existing models struggle with transferability to different GRN inference tasks.

Purpose of the Study:

• To develop a robust framework for accurate GRN inference.
• To overcome limitations of existing unsupervised and supervised methods.
• To improve the transferability of GRN inference models.

Main Methods:

• Developed SIGMA, a transformer-based framework utilizing self-supervised learning.
• Pretrained the encoder on gene expression profiles by converting pairs into patches and masking some.
• Enabled decoder to reconstruct masked patches, forcing encoder to learn correlation representations without labels.

Main Results:

• SIGMA accurately infers GRNs and demonstrates transferability to other subtypes, reducing label dependency.
• Outperformed state-of-the-art methods on human and mouse datasets.
• Identified novel candidate interactions in breast cancer datasets beyond existing gold-standard networks.

Conclusions:

• SIGMA offers a powerful, label-efficient approach for GRN inference.
• The framework shows promise for discovering new regulatory patterns in complex diseases.
• Further experimental validation of identified candidate interactions is recommended.