Source-linked AI summary
Deep learning extends de novo protein modelling coverage of genomes using iteratively predicted structural constraints
Joe G Greener, Shaun M Kandathil, David T Jones
TL;DR
Structural annotation of whole genomes is limited by weak covariation evidence in small protein families and by the challenge of turning noisy contacts into models. DMPfold predicts richer distance, hydrogen-bond, and torsion constraints and iteratively refines models, producing accurate structures across benchmark, transmembrane, and Pfam settings. It provides broad, freely available de novo modelling coverage, while longer proteins and some multi-domain cases require caution.
Problem
Small protein families provide limited covariation evidence, restricting whole-genome structural annotation.
Method
DMPfold predicts inter-atomic distance bounds, torsion angles, and hydrogen bonds, then iteratively refines constraints during model generation.
Results
25% (1475/5908) of dark Pfam domains were confidently annotated, and DMPfold produced more accurate CASP12 models than CONFOLD2 and Rosetta.
Takeaways & Limitations
DMPfold supports genome-scale de novo structural annotation with models that can be generated for thousands of targets in days on standard research clusters.
Takeaways & Limitations
Proteins longer than 500 residues should be treated cautiously, and some multi-domain predictions can form difficult-to-distinguish mirror topologies.
Abstract
from arXiv · showhide
The inapplicability of amino acid covariation methods to small protein families has limited their use for structural annotation of whole genomes. Recently, deep learning has shown promise in allowing accurate residue-residue contact prediction even for shallow sequence alignments. Here we introduce DMPfold, which uses deep learning to predict inter-atomic distance bounds, the main chain hydrogen bond network, and torsion angles, which it uses to build models in an iterative fashion. DMPfold produces more accurate models than two popular methods for a test set of CASP12 domains, and works just as well for transmembrane proteins. Applied to all Pfam domains without known structures, confident models for 25% of these so-called dark families were produced in under a week on a small 200 core cluster. DMPfold provides models for 16% of human proteome UniProt entries without structures, generates accurate models with fewer than 100 sequences in some cases, and is freely available.
Introduction
Accurate contact prediction has advanced, but converting incomplete or noisy contacts into reliable de novo protein models remains less developed. DMPfold addresses this gap by predicting richer structural constraints and iteratively refining models, outperforming established methods on CASP12 domains and extending to transmembrane proteins.
- Accurate residue-residue contact prediction has improved through statistical models that separate direct from indirect correlation effects.
- Converting predicted contacts into models remains comparatively underdeveloped because contacts may be incomplete or contain many false positives.
- Fragment-based approaches can require substantial computation, produce variable native-like model yields, and depend on suitable native-like fragments.
- Fragment assembly often fails for complex beta-sheet topologies, especially proteins with high contact order.
- DMPfold predicts inter-atomic distance bounds, torsion angles, and hydrogen bonds, then iteratively refines constraints while building models.
- DMPfold produces more accurate models than CONFOLD2 and Rosetta on CASP12 free-modelling domains and also models transmembrane proteins without task-specific modification.
Results
DMPfold generates accurate protein models from deep-learning-derived distance, torsion, and hydrogen-bond constraints, outperforming contact-based comparisons across CASP12 and Pfam evaluations. It extends modelling to transmembrane proteins, shallow alignments, dark Pfam families, and genome-scale structural coverage, with caution warranted for proteins longer than 500 residues.
- CASP12 evaluation: DMPfold had the best top-1 and top-5 TM-score performance among DMPfold, CONFOLD2, and Rosetta on 22 CASP12 free-modelling domains.It effectively produces a single output, whereas Rosetta’s best model may require selecting from 2,000 generated models.
- Method: DMPfold uses predicted inter-atomic distances, torsion angles, and hydrogen bonds with iterative constraint refinement to build models.Distance constraints are required, while combining all three constraint types gives the best performance.
- Transmembrane proteins: Of 28 transmembrane proteins, 26 had top-model TM-scores of at least 0.5, with a mean TM-score of 0.74 versus 0.60 for FILM3’s final refined models.DMPfold was run without modification and used the same sequence alignments as the FILM3 comparison.
- Shallow alignments: With 50-100 alignment sequences, DMPfold generated a correct-fold model in 22% of cases, rising to 84% with 10^4 or more sequences.Model accuracy was relatively flat across alignment sizes, supporting modelling when comparatively few sequences are available.
- Scope and limitations: DMPfold models domains up to around 600 residues accurately, but users should treat models for proteins longer than 500 residues with caution.The accuracy drop may reflect the 500-residue maximum chain length in the training set, multidomain proteins, or mirror topology effects.
Discussion
DMPfold combines iterative constraint refinement with deep-learning distance prediction to extend de novo modelling toward genome-scale annotation. It produces competitive models and confident structural annotations, while remaining limited by Pfam coverage and unresolved multi-domain topologies.
- Methodological progress: DMPfold combines model generation and constraint refinement iteratively, using predicted distances rather than binary contacts to provide richer modelling information.The method is intended to move beyond identifying folds toward high-quality models useful for applications such as ligand binding.
- Benchmark performance: 56% (24/43) of CASP13 FM domains received the correct fold from DMPfold, matching RaptorX-DeepModeller but below AlphaFold's 72% (31/43).The method was freely available and designed for genome-scale use.
- Scope and limitations: Some multi-domain proteins produce a correct fold for one domain but a topological mirror for another, likely because too few inter-domain distance constraints reach the CNS calculation.The authors expect improved distance prediction and constraint assignment to alleviate this problem.
- Scope and limitations: DMPfold's genome-scale reach is constrained because the study models only proteome regions matching Pfam domains; in humans, Pfam annotations cover 50.4% of amino-acid residues.Only 27.9% of human UniProt entries have at least one Pfam annotation.
- Genome-scale utility: 25% (1475/5908) of dark Pfam domains were confidently annotated, and 790 human proteins received high-confidence structural annotation without prior structural information.Across 14 model-organism genomes, the work structurally annotated 8,525 proteins with no prior structural information; the provided Pfam models had an 82.5% likelihood of a correct fold.
Methods
DMPfold uses deep neural networks to predict distance distributions, hydrogen-bond donor/acceptor pairs, and torsion angles for protein model construction. Its architecture increases representational capacity for distance prediction while preserving symmetry in the final distance output.
- Methods: DMPfold predicts inter-residue distance probability distributions, main-chain hydrogen-bond pairs, and torsion angles rather than only binary contacts.The distance predictor outputs a 20-channel histogram for each residue pair, while the hydrogen-bond map represents donor rows and acceptor columns.
- Methods: Hydrogen-bond predictions encode asymmetric donor-acceptor relationships, with donors within 3.5 Å of acceptors.Unlike ordinary contact maps, the hydrogen-bond map is not expected to be symmetric.
- Methods: The distance predictor uses a softmax layer with 20 output channels, each representing the likelihood of a predefined distance range.The output bins cover distances from 3.5 Å through a final bin for distances of 19 Å or greater.
- Methods: The final distance tensor is symmetrized during inference, which also ensembles independent upper- and lower-triangle predictions.Symmetry is required because the underlying distance matrix is symmetric.
- Methods: DMPfold replaces residual-block convolutional layers with maxout layers to increase representation power without increasing depth and GPU-memory demands.Each maxout layer uses four hidden maxout units and avoids a separate nonlinearity such as ReLU.
- Methods: Models were trained with Adam for 75 epochs, using validation cross-entropy to select weights and minibatch gradient accumulation because of GPU-memory limitations.Minibatches contained one protein for forward and backward passes, with gradients accumulated across eight minibatches.
Model generation using CNS
DMPfold converts predicted distances, hydrogen bonds, and torsion angles into pseudo-NOE constraints for CNS model generation. Distance bounds are widened according to prediction uncertainty before models are built.
- Model generation using CNS: CNS generates models from pseudo-NOE constraints derived from DMP distance distributions, hydrogen-bond maps, and torsion angles.The first iteration predicts contact and hydrogen-bond maps, which are converted into upper and lower distance bounds and hydrogen-bond constraints.
- Model generation using CNS: Distance bounds begin with the maximum-likelihood bin and expand into neighboring bins until their total likelihood reaches 0.4.This threshold was found to be optimal, although the overall method was relatively insensitive to threshold changes.
- Model generation using CNS: Less confident distance predictions produce wider bounds, while cases assigned to the unbounded final bin receive no bounds.All upper bounds supplied to CNS are below 19 Å.
Additional constraint types and iterative predictions
DMPfold supplements distance constraints with predicted backbone torsion angles and iteratively refines constraints from generated models. Three iterations were used throughout the reported work.
- Additional constraint types and iterative predictions: DMPfold generates dihedral constraints from predicted main-chain torsion angles using a bidirectional recurrent LSTM output layer.The torsion-angle network embeds each feature-map row in a 256-dimensional vector.
- Additional constraint types and iterative predictions: 18.8° average MAE for φ and 26.7° for ψ was observed on CASP11 free-modelling and hard-template-based targets.ω predictions were not used because the network did not predict rare cis-peptide conformations reliably.
- Additional constraint types and iterative predictions: Generated models are clustered, and the largest cluster’s representative is selected using MODCHECK and MODELLER scores to seed the next iteration.The same distance and hydrogen-bond procedures are then applied with an additional model-derived input feature channel.
- Additional constraint types and iterative predictions: 3 iterations were used throughout the study, although convergence typically required fewer than 5.Parameters were selected using 30 CASP11 free-modelling domains excluded from the training set’s ECOD H-group level.
Enforcement of non-overlap between training and test sets
The study defines train-test separation using structural evolutionary classifications rather than sequence identity alone. ECOD provides the basis for the benchmark split, while sequence-count calculations and fixed alignment resources support comparisons.
- Enforcement of non-overlap between training and test sets: Sequence identity alone is insufficient to exclude evolutionary or structural relationships between protein training and test examples.Many related proteins share less than 20% sequence identity, motivating structural classification for separation.
- Enforcement of non-overlap between training and test sets: ECOD was used to define train-test splits for all benchmarks because structural relationships can persist below 20% sequence identity.The authors identify ECOD, CATH, and SCOPe as structural classification databases suitable for this purpose.
- Enforcement of non-overlap between training and test sets: Neff was calculated by clustering each multiple sequence alignment with CD-HIT at 62% sequence identity and counting the resulting clusters.Comparisons with the Baker group used their corresponding effective-sequence-count measure, denoted Nf.
- Enforcement of non-overlap between training and test sets: Input alignments used HHblits with the February 2016 UniProt20 HMM library, restricting inputs to sequences available at the start of CASP12.DMPfold distance predictions were converted to contacts by summing likelihoods for bins below 8 Å.
Running DMPfold on transmembrane targets
The study evaluates DMPfold across Pfam families and a validation set designed to exclude training-related structural overlap. It also uses a defined subset of families for de novo modelling.
- FILM3 comparison: FILM3 alignments matched those used in the original study, enabling direct comparison with its reported results.None of the FILM3 targets were homologous to training examples at the ECOD T-group level.
- De novo modelling set: 5,214 families with target sequences 50–800 residues were taken forward for de novo modelling.The starting set contained 5,908 families after excluding likely templates.
- Validation set: The Pfam validation set comprised families with available structures that were not used for DMPfold training.Structures were excluded when they belonged to the same ECOD T-group as training structures.
Estimating Model Accuracy
DMPfold estimates model accuracy with a neural network using sequence, alignment, and distance-prediction features. At a predicted TM-score threshold of 0.5, it identifies models with correct folds with high precision and recall.
- Accuracy prediction: The TM-score predictor uses sequence length, effective alignment depth, and summed and average distance-histogram likelihoods.These likelihoods come from the first iteration of distance-histogram prediction.
- Evaluation: 100 random training, validation, and test splits were used to train the accuracy-estimation network.The network was trained on 3-D models from the Pfam validation set.
- Evaluation: 82.5% mean precision and 82.2% recall were achieved for recognizing correct folds at a predicted TM-score threshold of 0.5.Correct folds were defined as models with observed TM-score > 0.5.
Data availability
DMPfold’s trained neural-network models and Pfam 3-D models are openly available through its GitHub repository. The implementation and documentation are also provided there.
- Released resources: Trained neural-network models and Pfam 3-D models are available from the DMPfold GitHub repository.The repository URL is https://github.com/psipred/DMPfold.
- Implementation: DMPfold’s deep-learning components are implemented in PyTorch, with source code and documentation available online.The source code and documentation are provided at the same GitHub repository.
Deep learning extends de novo protein modelling coverage of genomes
DMPfold combines deep-learning-derived constraints with iterative modelling to expand structural coverage beyond existing homology annotations. Its evaluation includes runtime, prediction accuracy, constraint ablations, proteome coverage, and post-release validation resources.
- Running DMPfold: Each CASP12 FM domain was evaluated using total runtime, input-generation time, and model-generation time on one Intel Xeon processor.The input stage mainly involved alignment generation and PSICOV, while model generation mainly involved CNS.
- Estimating accuracy: The Pearson correlation between real and predicted TM-align scores on the Pfam validation set was 0.733.Cross-validation reported predictions for hold-out proteins.
- Constraint evaluation: The ablation study compares CASP12 FM results when different combinations of distance, torsion, and hydrogen-bond constraints are used.Results are reported for the top model only and ordered by descending mean TM-score.
- Proteome coverage: Supplementary proteome data report residues covered by Pfam annotations and those additionally covered by homology models or DMPfold predictions.The comparison is made across several model proteomes.
- Proteome coverage: Supplementary UniProt data distinguish entries gaining high-confidence DMPfold models according to whether they already have direct PDB or template coverage.Only active entries with at least one Pfam annotation were considered.
- Further validation: A supplementary validation set comprises Pfam families released in 2019 after the PDB template search for this study.This set provides a further validation of DMPfold models.
- Input features: Single-residue features are striped horizontally and vertically into L x L maps, occupying twice as many input channels as residue-pair features.This construction converts single-residue features into 2-D feature maps for the neural network.
- Input features: The input-feature table lists sequence profiles, Shannon entropy, sequence bounds, feature dimensionality, and channel occupancy.Sequence profiles occupy 42 channels, while Shannon entropy occupies 2 channels in the listed examples.
FILM3 DMPfold
The FILM3 analysis evaluates DMPfold models using sequence-alignment inputs, iterative predictors, and structural clustering. The workflow includes distance, hydrogen-bond, and torsion-angle predictions and reports improvement over iterations.
- The FILM3 workflow reports improvement over iterations and tracks the number of sequences in each alignment.
- DMPfold uses covariance data and contact predictions as inputs for transmembrane-protein modelling.
- Distance, hydrogen-bond, and torsion-angle predictors provide the structural constraints used by the workflow.
- Iterative distance and hydrogen-bond predictors refine the modelling process across successive stages.
- DMPfold models are clustered to select final structures from the generated candidates.