MulTFBS: A Spatial-Temporal Network with Multichannels for Predicting Transcription Factor Binding Sites.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Understanding transcription factor binding specificity is crucial for gene regulation.
• Previous methods for predicting transcription factor binding sites (TFBSs) primarily used DNA structure and one-hot encoding.
• Natural language processing techniques like word embedding have not been fully explored for TFBS prediction.

Purpose of the Study:

• To develop an advanced computational framework for predicting TFBSs.
• To integrate multiple types of DNA sequence features for improved prediction accuracy.
• To leverage deep learning, including natural language processing concepts, for TFBS prediction.

Main Methods:

• Developed a multichanneled deep learning framework, MulTFBS.
• Integrated one-hot encoding, word embedding encoding, and 3D DNA structural features into separate channels.
• Employed a spatial-temporal network combining CNNs, bidirectional LSTMs, and an attention mechanism for feature extraction.

Main Results:

• MulTFBS demonstrated superior performance on 66 universal protein-binding microarray datasets compared to six state-of-the-art methods.
• Achieved an average R² of 0.698 and an average PCC of 0.833 in regression tasks, surpassing the suboptimal CRPTS method.
• Showcased strong performance in TFBS classification tasks using TF ChIP-seq data.

Conclusions:

• The MulTFBS framework effectively integrates diverse DNA sequence features for accurate TFBS prediction.
• The proposed deep learning architecture successfully captures high-level sequence information.
• MulTFBS represents a significant advancement in computational TFBS prediction for both regression and classification.