FilterDCA: Interpretable supervised contact prediction using inter-domain coevolution.

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• Predicting protein 3D structure and complexes from sequence is crucial in computational biology.
• Deep learning and coevolutionary analysis show progress but have limitations for complex prediction and interpretability.
• Current methods struggle with large training sets for multi-protein complexes.

Purpose of the Study:

• Introduce FilterDCA, a simpler, interpretable supervised predictor for inter-domain and inter-protein contacts.
• Improve prediction accuracy for protein complex assembly using sequence information.
• Provide a transparent alternative to deep learning for contact prediction.

Main Methods:

• Integrate averaged contact patterns with coevolutionary scores from Direct Coupling Analysis (DCA).
• Utilize inherent contact patterns in protein contact maps derived from secondary structure.
• Develop a supervised learning approach for predicting residue-residue contacts.

Main Results:

• FilterDCA demonstrates improved performance over standard coevolutionary analysis.
• The method remains fully transparent and interpretable, unlike deep learning models.
• Successfully predicts inter-domain and inter-protein contacts.

Conclusions:

• FilterDCA offers a more interpretable and applicable approach for predicting protein complex structures.
• The integration of contact patterns and coevolutionary data enhances prediction accuracy.
• This method advances the field of computational biology for protein structure prediction.