Source-linked AI summary
Distance-based Protein Folding Powered by Deep Learning
Jinbo Xu
TL;DR
Protein folding needs accurate predictions for proteins with few homologs and often relies on time-consuming simulation. This paper predicts whole protein distance matrices with deep learning and directly constructs 3D models from them, achieving strong results on difficult CASP12, CASP13, and CAMEO targets. The approach provides fast distance-based ab initio folding without folding simulation, although integrating distances with fragment-based simulation remains an untested extension that could increase runtime.
Problem
Protein folding requires accurate structural restraints for proteins with few sequence homologs, while contact-based modeling often needs time-consuming folding simulation.
Method
Deep residual networks predict whole inter-atom distance matrices, secondary structure, and torsion angles, which are converted into CNS restraints for direct 3D model construction.
Results
21 of 37 CASP12 hard targets had correct folds, compared with 9 for the authors’ contact-based method and 11 for the best CASP12 human group.
Takeaways & Limitations
Distance-based folding produced more accurate models than contact-based folding and successfully handled several very hard CASP13 and CAMEO targets without folding simulation.
Takeaways & Limitations
Integrating predicted distances into fragment-based folding may improve accuracy but could significantly increase running time.
Abstract
from arXiv · showhide
Contact-assisted protein folding has made very good progress, but two challenges remain. One is accurate contact prediction for proteins lack of many sequence homologs and the other is that time-consuming folding simulation is often needed to predict good 3D models from predicted contacts. We show that protein distance matrix can be predicted well by deep learning and then directly used to construct 3D models without folding simulation at all. Using distance geometry to construct 3D models from our predicted distance matrices, we successfully folded 21 of the 37 CASP12 hard targets with a median family size of 58 effective sequence homologs within 4 hours on a Linux computer of 20 CPUs. In contrast, contacts predicted by direct coupling analysis (DCA) cannot fold any of them in the absence of folding simulation and the best CASP12 group folded 11 of them by integrating predicted contacts into complex, fragment-based folding simulation. The rigorous experimental validation on 15 CASP13 targets show that among the 3 hardest targets of new fold our distance-based folding servers successfully folded 2 large ones with <150 sequence homologs while the other servers failed on all three, and that our ab initio folding server also predicted the best, high-quality 3D model for a large homology modeling target. Further experimental validation in CAMEO shows that our ab initio folding server predicted correct fold for a membrane protein of new fold with 200 residues and 229 sequence homologs while all the other servers failed. These results imply that deep learning offers an efficient and accurate solution for ab initio folding on a personal computer.
Introduction
The paper shifts protein-folding assistance from contact prediction to deep-learning distance prediction, using whole distance matrices to construct 3D models without folding simulation. This approach targets proteins with few homologs and aims to improve accuracy while reducing computational cost.
- De novo protein structure prediction from sequence alone remains a challenging problem with slow progress.
- ResNet improves contact prediction by modeling higher-order residue correlations, global contact-matrix context, and nonlinear relationships across protein families.
- Distance matrices provide finer-grained information and stronger physical constraints than contact matrices, enabling more accurate structure determination.
- The proposed method predicts whole inter-atom distance matrices with deep ResNet and directly constructs 3D models through distance geometry without folding simulation.
- Experiments on CASP12, CASP13, and CAMEO targets showed improved folding of difficult proteins, including targets with few effective sequence homologs.
Datasets and Methods to Compare
The evaluation uses blind or challenging protein-folding datasets and compares distance-based modeling with contact-based approaches under matched feature conditions. Distance prediction uses deep residual networks and multiple atom-pair distance distributions.
- The study evaluates 37 CASP12 hard targets and 41 CAMEO hard targets using distance-based folding methods.
- CASP13 evaluation uses 15 targets with publicly available native structures and compares the authors’ servers with all participating CASP13 servers.
- Matched MSAs and protein features were used for test proteins to ensure fair comparison with earlier results and CASP12 groups.
- Deep 1D and 2D ResNets predict Euclidean inter-atom distance distributions for protein residues.
- Distance prediction uses 25 distance bins rather than the 2 labels used for contact prediction.
- Summing probabilities for distances ≤8Å also yields contact predictions that are slightly more accurate than those from a directly contact-trained model.
- Five atom-pair types are predicted and used together for folding, outperforming use of predicted Cβ-Cβ distances alone on average.
Predict secondary structure and torsion angles by 1D deep residual network
A 1D deep ResNet predicts secondary structure and backbone torsion-angle distributions from sequence-derived features. The angle model outputs residue-specific distribution parameters rather than discrete angle classes.
- A 19-layer 1D deep ResNet predicts three-state secondary structure and backbone torsion angles φ and ψ for each residue.
- The model uses HHblits-derived PSSM features and a 20×L binary representation of the primary sequence.
- Torsion angles are predicted as continuous distributions rather than classification bins.
- For each residue, the angle model predicts a mean, two variances, and a correlation parameter using maximum-likelihood training.
Folding by predicted distance, secondary structure and torsion angles
The folding pipeline converts predicted distances, secondary structure, and torsion angles into CNS restraints, then generates 3D models by distance geometry. Multiple decoys are produced because CNS solutions are not globally optimal.
- Predicted distances, secondary structure, and torsion angles are converted into CNS restraints for 3D model construction.
- CNS uses distance geometry to build models from distance restraints without folding simulation.
- For each atom type, the method selects 7L residue pairs with the highest predicted probability of distance <15Å and estimates their distances.
- Each protein receives 200 CNS decoys, and the five models with the fewest distance-restraint violations are retained.
- Meff quantifies the number of non-redundant sequence homologs in an MSA and serves as a difficulty measure for contact prediction.
Performance metrics
Performance is evaluated with TMscore-based 3D-model quality and five distance-prediction metrics, while CASP12 results compare distance-based folding with contact-based and leading CASP12 methods.
- TMscore measures similarity to the native structure from 0 to 1, with TMscore≥0.5 indicating a correct fold.
- Distance accuracy is assessed using absolute error, relative error, precision, recall, and F1 for restrained atom pairs.Only pairs with predicted distance≤15Å are used as restraints and evaluated.
- Average TMscore reaches 0.466 for top-1 and 0.476 for best-of-top-5 models on 37 CASP12 hard targets, exceeding contact-based and top CASP12 methods.The distance-based method uses neither template fragments nor CASP12 server predictions.
- 21 of 37 CASP12 targets have correct folds with distance-based folding, compared with 9 for contact-based folding and 11 for the best CASP12 human group.MetaPSICOV contacts produced 2 correct folds, while CCMpred contacts produced none.
- 8 of 21 targets with Meff≤100 are folded correctly, including 5 with TMscore>0.6, whereas top CASP12 groups achieve TMscore>0.6 on at most 2 targets.The distance-based model quality is correlated with Meff, with a good chance of a correct fold when Meff>55.
- At a distance error of 8Å, predicted 3D-model quality is equivalent to a random model.3D-modeling quality shows a strong correlation with distance-prediction error.
Results on the 41 CAMEO hard targets
On 41 CAMEO hard targets, distance-based ab initio folding outperforms contact-based folding in model quality and correct-fold counts, while performance remains related to sequence-homolog availability.
- Average TMscore is 0.551 for top-1 and 0.577 for best-of-top-5 distance-based models, about 10% above contact-based scores of 0.504 and 0.524.The distance-based and contact-based methods predict correct folds for 30 and 23 targets, respectively.
- Distance-based model quality is correlated with Meff, yet one target with Meff=1 is folded correctly.
- When Meff>55, there is a very good chance that average Cβ-Cβ distance-prediction error is less than 4Å.Distance-prediction error is inversely proportional to the logarithm of Meff.
Rigorous experimental validation in CASP13
CASP13 validation found that the authors’ servers performed strongly on several hard targets, including correct folds for two large, hardest targets with fewer than 150 effective sequence homologs. RaptorX-Contact also achieved the best contact accuracy across the six hard targets, while RaptorX-DeepModeller had the best overall TMscore across all 15 evaluated targets.
- Contact prediction: 64.0% was RaptorX-Contact’s accuracy for top L/2 long-range contacts on six hard targets, exceeding the other listed servers.Its top L/2 medium-range contact accuracy was 55.0%.
- Hard-target modeling: All servers failed to predict correct folds for T0953s1, T0960, and T0963, whose best TMscores were below 0.4, 0.3, and 0.3, respectively.The latter two are multi-domain proteins, making domain-level evaluation more appropriate.
- Hard-target modeling: 0.5632 was RaptorX-DeepModeller’s top TMscore for T0950, the only correct fold among the listed servers at the 0.5 cutoff.T0950 is an α protein of 353 residues with Meff=148.
- Hard-target modeling: Only RaptorX-Contact generated a correct fold for T0953s2, while all other listed servers failed.T0958 also received a correct RaptorX-Contact fold with TMscore=0.5294 despite having only Meff=18.
- Overall CASP13 outcome: 2 large, hardest targets were folded correctly by the authors’ group with fewer than 150 effective sequence homologs, while the other servers got none correct.Across all 15 targets, RaptorX-Contact had the best contact accuracy and RaptorX-DeepModeller the best overall TMscore.
Rigorous experimental validation in CAMEO
In a blind CAMEO test, the distance-based ab initio server correctly folded a 200-residue membrane protein with a new fold and 229 effective sequence homologs. Its best model had TMscore=0.68 and RMSD=5.65Å, whereas the other servers failed to predict a correct fold.
- CAMEO result: TMscore=0.68 and RMSD=5.65Å were achieved by the best model for CAMEO target 2018-11-03_00000053_1.The target was a 200-residue membrane protein with a new fold and 229 effective sequence homologs.
- CAMEO result: The other servers failed to predict a correct fold for this membrane-protein target.Figure 3 ranks submitted models by decreasing TMscore and shows superimposition of the predicted and native structures.
Running time
The distance-based ab initio pipeline avoids folding simulation and completes folding in minutes to a few hours on 20 CPUs, enabling rapid processing of benchmark targets.
- Runtime: 10 minutes to a few hours are required for the third folding step on a Linux computer with 20 CPUs.The pipeline generates alignments and features, predicts angles and distances on a GPU, then folds with CNS restraints.
- Runtime: ~13 hours were needed for 41 CAMEO targets and ~4 hours for 37 CASP12 targets when the third step ran in parallel.The first step takes minutes for most targets and the second takes seconds on a GPU card.
Conclusion and Discussion
The study concludes that deep-learning-predicted distances can support accurate, fast folding without simulation, while discussing restraint choices, distance discretization, and possible future integration with fragment-based simulation.
- Conclusion: Distance-based folding produced better 3D models than contact-based folding without folding simulation.The method avoids fragment assembly and complex energy functions and generated correct folds for many hard targets.
- Conclusion: 10 minutes to a few hours were needed to generate 200 decoys on a Linux computer with 20 CPUs.The authors state that ab initio folding may therefore be performed on a laptop equipped with a GPU card.
- Discussion: ~0.008 was the average TMscore decrease when predicted torsion angles were omitted.The study did not evaluate the 1D secondary-structure and torsion-angle models in detail; torsion angles may reduce mirror images.
- Discussion: All 5 atom-pair types were used in the reported folding results, although Cβ-Cβ alone was most useful for evaluating Cα conformation.Using all five types may reduce noise and improve side-chain packing.
- Discussion: 25 and 52 distance bins performed similarly on average, and both outperformed 12 bins.The authors also considered predicting real-valued distance matrices using a log-normal assumption.
- Future work: Integrating predicted distances into fragment-based folding simulation such as Rosetta may improve accuracy but could significantly increase running time.The current ResNet predicts inter-atom distances from sequence-derived features, after which CNS constructs 3D models.
Appendix
Figure S1 depicts the deep network architecture used to predict protein distance matrices, combining sequential and pairwise processing pathways.
- A 1D deep residual neural network transforms sequential features such as sequence profiles and predicted secondary structure.
- A 2D deep dilated residual neural network transforms pairwise features.
- The middle pathway converts convolved sequential features into pairwise features and combines them with the original pairwise features.