Enhancing Transcription Factor Prediction via Domain Knowledge Integration With Logic Tensor Networks

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence

Background:

• Transcription factors (TFs) are crucial for gene expression regulation.
• Current TF prediction methods lack accuracy or interpretability, especially when biological knowledge is limited.
• Deep learning models often require extensive data and struggle to incorporate domain-specific insights.

Purpose of the Study:

• To introduce LTN-TFpredict, a novel neurosymbolic framework for enhanced TF prediction.
• To improve both the accuracy and interpretability of TF prediction models.
• To integrate deep learning with symbolic reasoning and biological domain knowledge.

Main Methods:

• Utilized Logic Tensor Networks (LTNs) integrated with deep learning.
• Employed pre-trained protein language models for sequence embeddings.
• Incorporated logical constraints from five key TF motifs (zinc fingers, leucine zippers, etc.) to guide learning.

Main Results:

• LTN-TFpredict achieved state-of-the-art accuracy in TF prediction.
• The model consistently outperformed traditional and deep learning methods, including CNNs and transformers.
• Demonstrated improved biological validity and logical compliance with known TF characteristics.

Conclusions:

• LTN-TFpredict offers a robust, interpretable, and biologically grounded solution for TF prediction.
• The neurosymbolic approach effectively bridges deep learning and symbolic reasoning in computational biology.
• This framework advances the application of AI for understanding gene regulatory mechanisms.