DeepSELEX: inferring DNA-binding preferences from HT-SELEX data using multi-class CNNs.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Transcription factor (TF) DNA-binding is crucial for gene regulation, necessitating accurate models for predicting binding sites.
• High-throughput SELEX (Systematic Evolution of Ligands by Exponential Enrichment) generates extensive data on TF-DNA interactions.
• Existing computational methods inadequately leverage the full potential of high-throughput SELEX data and advanced techniques like deep neural networks.

Purpose of the Study:

• To develop a novel computational method, DeepSELEX, for inferring intrinsic DNA-binding preferences of TFs.
• To utilize deep neural networks to exploit the richness of high-throughput SELEX data for improved binding inference.
• To enhance the accuracy of predicting TF binding sites in DNA sequences.

Main Methods:

• Development of DeepSELEX, a deep neural network-based algorithm.
• Training DeepSELEX on high-throughput SELEX data, learning from sequence changes across experimental cycles.
• Comparative analysis against existing methods for DNA-binding inference.

Main Results:

• DeepSELEX demonstrates superior performance in predicting TF DNA-binding preferences from high-throughput SELEX data compared to extant methods.
• The algorithm achieves state-of-the-art accuracy for in vivo binding prediction.
• Analysis of DeepSELEX model parameters reveals the learning of biologically relevant features, offering insights into TF binding mechanisms.

Conclusions:

• DeepSELEX represents a significant advancement in inferring TF DNA-binding preferences from high-throughput SELEX data.
• The method effectively utilizes deep learning to extract complex binding information from experimental data.
• DeepSELEX provides a powerful tool for understanding gene regulation and TF function.