A novel deep learning identifier for promoters and their strength using heterogeneous features

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Promoters are crucial DNA regions regulating gene transcription.
• Genetic variations in promoters are linked to diseases like cancer and cardiovascular conditions.
• Experimental promoter identification is resource-intensive, necessitating computational solutions.

Purpose of the Study:

• To develop a computational tool for accurate promoter identification and strength classification.
• To overcome the limitations of experimental methods in promoter characterization.
• To aid in drug discovery by enabling large-scale promoter analysis.

Main Methods:

• Designed PROCABLES, a bi-layer deep-learning predictor.
• Utilized five distinct features: word2vec, k-spaced nucleotide pairs, trinucleotide propensity, trinucleotide composition, and electron-ion interaction pseudopotentials.
• Integrated a convolutional neural network (CNN) with bidirectional long-short-term memory (LSTM) in a stacked framework.

Main Results:

• PROCABLES achieved high accuracy: 0.971 for promoter identification and 0.920 for strength classification.
• Achieved MCC scores of 0.940 and 0.840 for the respective layers.
• Outperformed existing computational predictors in identifying promoters and their types.

Conclusions:

• PROCABLES offers a powerful and efficient computational approach for promoter recognition.
• The tool can significantly aid in large-scale promoter analysis and understanding genetic disease mechanisms.
• This work provides valuable insights for broader DNA-related computational challenges.