Modeling in-vivo protein-DNA binding by combining multiple-instance learning with a hybrid deep neural network

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Modeling in-vivo protein-DNA binding is crucial for understanding gene regulation but remains computationally challenging.
• Existing deep learning methods often use fully supervised learning and overlook sequence dependencies or weakly supervised information.

Purpose of the Study:

• To develop a weakly supervised framework for more accurate in-vivo protein-DNA binding modeling.
• To address limitations of existing methods by incorporating multiple-instance learning and advanced sequence encoding.

Main Methods:

• A hybrid deep neural network combining convolutional and recurrent neural networks was developed.
• K-mer encoding was used to transform DNA sequences into image-like inputs, capturing nucleotide dependencies.
• Multiple-instance learning with a sliding window approach segmented sequences into instances.
• The Noisy-and method was employed to integrate predictions from individual instances.

Main Results:

• The proposed framework demonstrated superior performance on in-vivo datasets compared to existing methods.
• K-mer encoding proved effective in modeling dependencies among nucleotides.
• Integrating recurrent layers further enhanced the framework's predictive performance.

Conclusions:

• The developed weakly supervised framework offers a significant advancement in modeling in-vivo protein-DNA binding.
• The combination of k-mer encoding, hybrid deep networks, and multiple-instance learning provides a robust approach for computational biology challenges.