A deep learning framework for improving long-range residue-residue contact prediction using a hierarchical strategy
Dapeng Xiong1,2, Jianyang Zeng2,3, Haipeng Gong1,2
1MOE Key Laboratory of Bioinformatics, School of Life Sciences.
Bioinformatics (Oxford, England)
|May 5, 2017
Summary
DeepConPred improves protein structure prediction by accurately identifying long-range residue contacts, even for proteins with limited sequence data. This computational tool enhances accuracy for challenging targets, aiding in protein structure determination.
Area of Science:
- Computational Biology
- Structural Bioinformatics
- Machine Learning in Biology
Background:
- Accurate prediction of residue-residue contacts is crucial for protein structure prediction.
- Long-range contact information significantly simplifies conformational sampling.
- Current methods struggle with proteins having limited homologous sequences (hard targets).
Purpose of the Study:
- To develop a computational program, DeepConPred, for improved prediction of long-range residue contacts.
- To enhance the accuracy of contact prediction for proteins with limited sequence information.
Main Methods:
- Utilized a pipeline of two novel deep-learning methods: DeepCCon and DeepRCon.
- Incorporated a contact refinement step.
- Trained models using coevolutionary information from limited homologous sequences for robustness on hard targets.
Main Results:
- DeepConPred demonstrated improved prediction accuracy for long-range residue contacts.
- Achieved high percentages of correct predictions for CASP10/CASP11 targets across different top prediction thresholds (e.g., 70.00% for top 5).
- Ranked among the best methods in independent tests for CASP targets.
Conclusions:
- DeepConPred offers a robust and effective approach for predicting long-range residue contacts, particularly for challenging protein targets.
- The framework's ability to utilize limited sequence data makes it valuable for proteins lacking extensive homologous sequences.
