Enabling full-length evolutionary profiles based deep convolutional neural network for predicting DNA-binding...

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Science

Background:

• Sequence-based DNA-binding protein (DBP) prediction is a significant challenge.
• Traditional methods using sliding windows on Position-Specific Substitution Matrices (PSSMs) struggle with variable protein lengths.
• Current PSSM summary features may not capture the full predictive potential of PSSMs.

Purpose of the Study:

• To evaluate if zero-vector padded PSSMs (pPSSMs) improve DBP prediction compared to PSSM summary features.
• To compare the performance of Multilayer Perceptron (MLP) and Deep Convolutional Neural Network (CNN) models using different feature sets.
• To determine the optimal feature representation and model architecture for accurate DBP prediction.

Main Methods:

• Utilized zero-vector padding to create fixed-size pPSSMs from variable-length PSSMs.
• Employed Multilayer Perceptron (MLP) and Deep Convolutional Neural Network (CNN) architectures.
• Compared performance using PSSM summary features versus pPSSM features on diverse datasets.

Main Results:

• pPSSMs outperformed summary features on diverse PDB-derived datasets, indicating reduced redundancy.
• A consensus approach combining summary features and pPSSMs yielded superior results.
• CNN models consistently outperformed corresponding MLP models across all evaluated feature sets.
• Model performance was dependent on the chosen input features, with CNNs achieving the highest accuracy.

Conclusions:

• Scaled PSSMs (pPSSMs) offer enhanced predictive power for DNA-binding proteins, especially on diverse datasets.
• Convolutional Neural Networks (CNNs) represent a more effective architecture than MLPs for this task.
• Combining different feature representations can further improve prediction accuracy, highlighting the importance of feature engineering and model selection.