Chromatin accessibility prediction via convolutional long short-term memory networks with k-mer embedding

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Experimental chromatin accessibility assays are costly and time-intensive.
• Existing computational methods rely on handcrafted k-mer features or convolutional neural networks.
• A comprehensive framework integrating k-mer co-occurrence with deep learning is needed.

Purpose of the Study:

• To develop a computational approach for predicting open chromatin regions from DNA sequences.
• To integrate k-mer co-occurrence information with deep learning for improved chromatin accessibility prediction.
• To provide a robust and high-performing model for genomic studies.

Main Methods:

• Developed a convolutional Long Short-Term Memory (LSTM) network with k-mer embedding.
• Pre-trained k-mer embedding vectors using an unsupervised representation learning approach based on co-occurrence matrices.
• Constructed a deep learning architecture with embedding, convolutional, and Bidirectional LSTM (BLSTM) layers for feature learning and classification.

Main Results:

• The proposed method consistently outperforms baseline methods in chromatin accessibility prediction.
• K-mer embedding effectively enhances model performance.
• Convolutional and BLSTM layers contribute significantly to the model's efficacy.
• The model demonstrates robustness to hyper-parameter variations.

Conclusions:

• The developed deep learning framework offers a powerful tool for chromatin accessibility prediction.
• The integration of k-mer co-occurrence and deep learning advances genomic research.
• This approach can enhance the understanding of chromatin accessibility mechanisms.