Predicting protein inter-residue contacts using composite likelihood maximization and deep learning.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Accurate protein tertiary structure prediction relies on precise identification of inter-residue contacts.
• Co-evolutionary analysis is a key method for inferring these contacts.
• Markov Random Field (MRF) models offer a framework for contact prediction but face a trade-off between computational efficiency and accuracy.

Purpose of the Study:

• To introduce a novel method, clmDCA, that addresses the accuracy-efficiency dilemma in MRF-based protein contact prediction.
• To demonstrate that composite likelihood provides a superior approximation to the true likelihood function compared to pseudo-likelihood.
• To validate the effectiveness of clmDCA in predicting protein inter-residue contacts accurately and efficiently.

Main Methods:

• Developed clmDCA, a method utilizing composite likelihood (product of conditional probabilities of all residue pairs) for MRF parameter estimation.
• Compared clmDCA against existing MRF-based methods using pseudo-likelihood on benchmark datasets (PSICOV, CASP-11).
• Integrated deep learning techniques with clmDCA for further refinement of contact predictions.

Main Results:

• clmDCA significantly outperformed existing MRF-based approaches in prediction accuracy on benchmark datasets.
• The accuracy of clmDCA predictions was further enhanced when combined with deep learning refinement.
• Successfully applied clmDCA-predicted contacts to accurately construct tertiary protein structures.

Conclusions:

• Composite likelihood maximization is an efficient algorithm for estimating MRF parameters.
• This approach substantially improves the accuracy of protein inter-residue contact prediction.
• clmDCA offers a robust and accurate method for advancing protein structure prediction.