Analyzing effect of quadruple multiple sequence alignments on deep learning based protein inter-residue distance...

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Protein 3D structure prediction accuracy has improved due to deep learning methods.
• Multiple sequence alignments (MSAs) are crucial for predicting inter-residue contacts and distances.

Purpose of the Study:

• To develop a novel deep learning approach, AttentiveDist, for enhanced protein 3D structure prediction.
• To improve co-evolutionary information by utilizing MSAs generated with different E-values.

Main Methods:

• Implemented an attention layer in a deep neural network to weigh the importance of different MSA features.
• Integrated multiple MSAs with varying E-value cutoffs into a single model.
• Incorporated bond angle predictions as an additional task.

Main Results:

• Combining MSAs with different E-values outperformed single E-value MSA predictions.
• The attention layer further improved prediction performance.
• Adding bond angle predictions led to additional gains in distance prediction accuracy.

Conclusions:

• AttentiveDist effectively leverages diverse co-evolutionary information for superior protein distance predictions.
• Improved distance predictions translate to enhanced protein tertiary structure modeling.
• The attention mechanism and multi-task learning contribute to the model's performance gains.