Source-linked AI summary
Accurate De Novo Prediction of Protein Contact Map by Ultra-Deep Learning Model
Sheng Wang, Siqi Sun, Zhen Li, Renyu Zhang, Jinbo Xu
TL;DR
The paper targets further improvement of supervised protein contact prediction. It presents a very deep residual neural network and reports dramatically improved contact-prediction results, including performance on membrane proteins.
Problem
Existing supervised learning methods for contact prediction motivate further improvement, while techniques effective for image labeling may also work for contact prediction.
Method
The method uses a very deep residual neural network for contact prediction, incorporating co-evolution information, pairwise contact, and distance-potential features.
Results
The authors report dramatically improved contact-prediction results over existing methods, with almost the same accuracy on membrane proteins despite using no membrane proteins for training.
Takeaways & Limitations
The method predicts all contacts of a protein simultaneously, a design that allows the authors to model long-range inter-contact relationships.
Takeaways & Limitations
The authors indicate that contact-prediction accuracy may be further improved by enlarging the training set.
Abstract
from arXiv · showhide
Recently exciting progress has been made on protein contact prediction, but the predicted contacts for proteins without many sequence homologs is still of low quality and not very useful for de novo structure prediction. This paper presents a new deep learning method that predicts contacts by integrating both evolutionary coupling (EC) and sequence conservation information through an ultra-deep neural network formed by two deep residual networks. This deep neural network allows us to model very complex sequence-contact relationship as well as long-range inter-contact correlation. Our method greatly outperforms existing contact prediction methods and leads to much more accurate contact-assisted protein folding. Tested on three datasets of 579 proteins, the average top L long-range prediction accuracy obtained our method, the representative EC method CCMpred and the CASP11 winner MetaPSICOV is 0.47, 0.21 and 0.30, respectively; the average top L/10 long-range accuracy of our method, CCMpred and MetaPSICOV is 0.77, 0.47 and 0.59, respectively. Ab initio folding using our predicted contacts as restraints can yield correct folds (i.e., TMscore>0.6) for 203 test proteins, while that using MetaPSICOV- and CCMpred-predicted contacts can do so for only 79 and 62 proteins, respectively. Further, our contact-assisted models have much better quality than template-based models. Using our predicted contacts as restraints, we can (ab initio) fold 208 of the 398 membrane proteins with TMscore>0.5. By contrast, when the training proteins of our method are used as templates, homology modeling can only do so for 10 of them. One interesting finding is that even if we do not train our prediction models with any membrane proteins, our method works very well on membrane protein prediction. Finally, in recent blind CAMEO benchmark our method successfully folded 5 test proteins with a novel fold.
Introduction
Protein contact prediction remains difficult for proteins with few sequence homologs, motivating a very deep residual network designed to capture complex sequence–contact relationships and contact correlations. The method improves prediction accuracy and contact-assisted folding relative to existing approaches.
- DCA methods usually need many sequence homologs, while existing supervised methods remain limited by shallow architectures.These limitations motivate a deeper learning architecture for proteins with sparse homologous-sequence information.
- Contact prediction differs from image labeling because contact maps cannot be resized, inputs combine sequential and pairwise features, and contacts comprise less than 2% of labels.These properties create distinct modeling challenges despite the analogy to image labeling.
- The paper introduces a very deep residual neural network for protein contact prediction.The model is motivated partly by residual networks’ strong performance in computer vision and treats contact prediction as related to pixel-level labeling.
- The network is intended to capture complex sequence–contact relationships and high-order contact correlations.It is trained on proteins with solved structures and evaluated on public CASP, CAMEO, and membrane-protein datasets.
- The experiments report much better contact-prediction accuracy and more accurate contact-assisted folding than existing methods.The method is also presented as potentially useful for protein–protein and protein–RNA interfacial-contact prediction.
Results
Across contact-prediction benchmarks, the method substantially outperforms DCA and MetaPSICOV baselines, including proteins with few homologs, and produces more accurate contact-assisted structures. Its predicted contacts also outperform template-based modeling, especially for membrane proteins and proteins lacking close templates.
- Contact prediction accuracy: The method outperforms all tested DCA methods and MetaPSICOV by a large margin across four test sets and multiple prediction depths.Supervised learning methods also greatly outperform pure DCA methods.
- Contact prediction accuracy: On CASP11, CAMEO, and membrane proteins, long-range top L accuracy exceeds CCMpred by 0.32, 0.27, and 0.26, respectively.The corresponding advantages over MetaPSICOV are 0.20, 0.17, and 0.17.
- Effect of homolog number: The method remains superior when sequence-homolog information is limited, with especially strong long-range accuracy for ln(Meff)≤7.Its advantage persists even when Meff is very large, suggesting use of information beyond pairwise co-evolution.
- Contact-assisted folding: Among 579 test proteins, the best of five contact-assisted models has TMscore>0.6 for 203 proteins, compared with 79 for MetaPSICOV and 62 for CCMpred.This result directly links improved contact prediction to more frequent correct folds.
- Blind CAMEO evaluation: In blind CAMEO evaluation, the server produced TMscore>0.5 for 28 of 41 targets and TMscore>0.6 for 16, and successfully folded five novel-fold targets.Its best-of-five average TMscore was 0.535, compared with 0.316 for CCMpred and 0.392 for MetaPSICOV.
Conclusion and Discussion
The paper presents an ultra-deep residual-network method that predicts all protein contacts simultaneously, substantially outperforming existing methods and improving contact-assisted folding. Its performance generalizes to membrane proteins and novel-fold targets, while larger training sets, deeper models, and distance prediction remain future directions.
- Method and novelty: The method combines deep residual networks with simultaneous whole-protein contact prediction to model sequence-contact relationships and high-order residue correlations.Later stages can correct earlier prediction errors, while deeper architectures capture information from a larger context.
- Contact-prediction performance: The method substantially outperforms CCMpred, Evfold, PSICOV, and MetaPSICOV in blind contact-prediction tests.The authors attribute the advantage partly to learning from thousands of protein families and explicitly modeling high-order contact patterns.
- Structure modeling: Contact restraints enable ab initio models with correct folds for many test proteins, and the resulting contact-assisted models outperform template-based models.The authors report better model quality than homology models built from training proteins.
- Generalization: The model trained only on non-membrane proteins performs well on membrane proteins, with nearly the same accuracy as models trained with membrane proteins.This suggests that sequence-structure relationships learned from non-membrane proteins generalize to membrane-protein contact prediction.
- Blind evaluation: The blind CAMEO test folded several targets with a new fold, including 5f5pH.The paper also reports a server that folded four targets with a new fold plus one target released later.
- Limitations and future work: The authors identify insufficient GPU memory as the current limit on extending the model beyond 100 layers.They plan multi-GPU training to test models with hundreds of layers.
Method
The method uses two residual neural networks to transform sequential and pairwise protein features, then predicts an entire contact map jointly. Its ultra-deep architecture captures long-range dependencies between input features, contacts, and residue pairs.
- Depth and receptive field: Stacking many residual blocks enables long-range interdependency modeling despite small convolution windows.The method captures dependencies between input features and contacts as well as between different residue pairs.
- Depth and receptive field: The 1D residual network has fixed depth 6, while the 2D network depth is varied experimentally.Performance is reported as good with approximately 60 hidden neurons per position and 60 convolution layers in the second network.
- Training: The network accepts variable-length proteins and uses minibatches grouped by similar sequence lengths, with zero padding when needed.This accommodates proteins of different lengths during training.
- Architecture: Two multi-layer residual neural networks form the contact-prediction model.Residual connections pass both linear and nonlinear information from input to output.
- Contact-map prediction: The model predicts contacts for the whole protein simultaneously rather than separately for individual residue pairs.This design models long-range contact and high-order residue correlations.
- Feature transformation: The first network processes sequential features, while the second models pairwise features such as coevolution.The sequential representation is converted into pairwise features before contact-map prediction.
3D model construction by contact-assisted folding
The paper converts predicted contacts and secondary-structure predictions into structural restraints, then uses CNS to generate and rank three-dimensional models. Template-based models are constructed separately using HHblits, HHsearch, and MODELLER for comparison.
- Contact-assisted folding: Predicted secondary structure is converted into distance, angle, and hydrogen-bond restraints before folding.The conversion uses RaptorX-Property predictions and a Confold script.
- Contact-assisted folding: The top 2L predicted contacts are converted into distance restraints assuming residue separations of 3.5Å to 8.0Å.Short-, medium-, and long-range contacts are all eligible for selection.
- Model generation: CNS generates twenty three-dimensional models without force fields, and the five lowest-NOE-score models are selected.The NOE score reflects violations of the input restraints.
- Model generation: For comparisons using CCMpred or MetaPSICOV contacts, the same predicted secondary structure is supplied to support a fair evaluation.All contact-assisted models are evaluated against their input constraints using NOE scores.
- Template-based comparison: Template-based models are built by searching the training-protein library with HHblits and HHsearch, followed by MODELLER construction from the top five templates.The procedure uses an HMM for template searching.