Predicting transcriptional activation domain function using Graph Neural Networks.

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Predicting transcriptional activation domains (TADs) is challenging due to sequence diversity and disordered structures.
• Existing methods like logistic regression and CNNs struggle with complex patterns and structural features.
• TADs' structural properties offer potential for improved prediction accuracy.

Purpose of the Study:

• To develop a more accurate method for predicting the functionality of transcriptional activation domains.
• To explore the utility of graph neural networks (GNNs) in analyzing TAD structural features.
• To identify key sequence and structural properties that determine TAD functionality.

Main Methods:

• Utilized graph neural networks (GNNs) with residues or atoms as nodes to model TAD structures.
• Experimented with various feature combinations, including amino acid type, position, physicochemical properties, and secondary structure.
• Developed a logistic regression model to analyze feature importance and compare performance.

Main Results:

• The residue-level GNN model, incorporating amino acid type, position, physicochemical properties, and secondary structure, achieved 97.9% accuracy, 71% F1 score, and 97.1% AUROC.
• This GNN model outperformed existing literature methods on the tested dataset.
• Logistic regression identified amino acid frequency as a dominant feature, with acidic and aromatic residues indicating functionality, while basic residues suggested non-functionality.

Conclusions:

• Graph neural networks represent a powerful approach for predicting TAD functionality by capturing complex structural relationships.
• Specific residue properties, such as acidity and aromaticity, are critical determinants of TAD function.
• The findings highlight the importance of integrating structural information for enhanced prediction of regulatory elements.