De Novo Protein Structure Prediction by Model Quality Assessment Dynamic Feedback Mechanism Using Deep Learning

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in protein science

Background:

• Accurate de novo protein structure prediction is crucial but challenging, especially without homologous templates or strong evolutionary data.
• Current end-to-end methods like AlphaFold2 are accurate but lack transparency and flexibility for external evaluation.
• Integrating model quality assessment (MQA) into prediction pipelines offers a potential avenue for iterative accuracy improvement.

Purpose of the Study:

• To investigate the integration of MQA as a closed-loop feedback mechanism for iterative de novo protein structure prediction.
• To develop and evaluate DGMFold, a novel method employing a feedback loop between geometric constraint prediction, structural simulation, and quality evaluation.
• To assess DGMFold's performance against state-of-the-art methods, including AlphaFold2 and RoseTTAFold, on benchmark and challenging protein targets.

Main Methods:

• DGMFold utilizes a three-component feedback loop: GeomNet for geometric constraint prediction from MSAs, a structural simulation module, and EmaNet for model quality evaluation.
• GeomNet predicts inter-residue geometric constraints using an improved residual neural network, guiding structure folding.
• EmaNet estimates structure accuracy (distance deviation, lDDT) and feeds this information back to GeomNet for iterative refinement.

Main Results:

• The closed-loop feedback mechanism in DGMFold significantly enhances prediction performance.
• DGMFold demonstrated superior accuracy compared to trRosetta and RaptorX on benchmark and CASP14 FM targets.
• DGMFold achieved higher accuracy than AlphaFold2 and RoseTTAFold on specific subsets of human proteins where these methods had lower TM-scores.

Conclusions:

• DGMFold's iterative, feedback-driven approach effectively improves de novo protein structure prediction accuracy.
• The integration of MQA within a closed-loop system represents a promising strategy for advancing protein structure prediction.
• DGMFold offers a competitive alternative, particularly for challenging protein targets, showcasing the benefits of integrating predictive and evaluative components.