A deep learning model to identify gene expression level using cobinding transcription factor signals.

Area of Science:

• Computational biology
• Genomics
• Systems biology

Background:

• Gene expression is regulated by transcription factors (TFs) binding in complex combinations.
• Systematically inferring TF cooperative binding effects on gene activity remains challenging.

Purpose of the Study:

• To quantitatively analyze TF correlations and interaction networks in relation to gene expression.
• To develop a deep learning model for predicting gene expression levels based on TF binding patterns.

Main Methods:

• Quantitative analysis of TF correlations and interaction networks in GM12878 and K562 cell lines.
• Identification of six TF modules associated with gene expression per cell line.
• Construction and application of a convolutional neural network (TFCNN) model for gene expression prediction.

Main Results:

• The TFCNN model achieved high prediction performance, with AUC values up to 0.976.
• TFCNN outperformed Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) models in predicting gene expression.
• Abundant TF binding primarily drives gene expression, while cooperative TF interactions have subtle effects, and regulation is nonlinear.

Conclusions:

• The TFCNN model effectively captures combinatorial TF interactions for accurate gene expression prediction.
• Understanding TF combinations is crucial for deciphering gene expression regulatory mechanisms.
• This approach provides insights into the complex, nonlinear regulation of gene expression by TFs.