Source-linked AI summary
Computational Protein Design with Deep Learning Neural Networks
Jingxue Wang, Huali Cao, John Z. H. Zhang, Yifei Qi
TL;DR
Designing proteins for specified structures and functions remains difficult despite the growth of solved structures and accumulated fold information. This study uses deep-learning neural networks and structural features to predict amino-acid probabilities at individual residues, achieving 38.3% accuracy and improving sequence identity in Rosetta designs and over SPIN. The authors conclude that these predictions can support computational protein design, while residue prediction remains limited by structural ambiguity and amino-acid tolerability.
Problem
Designing proteins for specified structures and functions remains challenging, despite rapidly increasing structure data and saturation of unique folds.
Method
The study trains deep-learning neural networks on structural features from target residues and their three-dimensional neighbors to predict probabilities for 20 amino-acid types.
Results
38.3% accuracy was achieved on the SI90 dataset with 15 neighboring residues, while network outputs improved Rosetta sequence identity and produced ~3% higher identity than SPIN.
Takeaways & Limitations
Using predicted amino-acid probabilities as residue-type restraints can improve computational designs, although the optimal restraint set is system dependent.
Takeaways & Limitations
Prediction accuracy is limited by structural ambiguity and the average number of amino acids tolerated at each residue, with weaker performance on hydrophobic residues.
Abstract
from arXiv · showhide
Computational protein design has a wide variety of applications. Despite its remarkable success, designing a protein for a given structure and function is still a challenging task. On the other hand, the number of solved protein structures is rapidly increasing while the number of unique protein folds has reached a steady number, suggesting more structural information is being accumulated on each fold. Deep learning neural network is a powerful method to learn such big data set and has shown superior performance in many machine learning fields. In this study, we applied the deep learning neural network approach to computational protein design for predicting the probability of 20 natural amino acids on each residue in a protein. A large set of protein structures was collected and a multi-layer neural network was constructed. A number of structural properties were extracted as input features and the best network achieved an accuracy of 38.3%. Using the network output as residue type restraints was able to improve the average sequence identity in designing three natural proteins using Rosetta. Moreover, the predictions from our network show ~3% higher sequence identity than a previous method. Results from this study may benefit further development of computational protein design methods.
Introduction
Computational protein design can generate sequences for target structures, but designing proteins with specified structures and functions remains challenging. The paper motivates deep-learning methods by the growing structural database and presents a neural-network approach using local structural context to improve residue prediction and design.
- Introduction: Computational protein design generates sequences likely to fold into a desired backbone structure for experimental verification.Its applications include designed folds, enzymes, vaccines, antibodies, assemblies, ligand-binding proteins, and membrane proteins.
- Introduction: Approximately 132,000 PDB structures and 10,000 new structures yearly were available by July 2017, while unique folds were saturating.This pattern suggests that additional structural data are accumulating for existing folds.
- Introduction: Deep-learning neural networks require large datasets and had already been applied to diverse biological and chemical prediction tasks.Examples include protein-ligand scoring, interaction prediction, secondary-structure prediction, contact-map prediction, and toxicity prediction.
- Introduction: SPIN predicted protein sequence profiles from backbone structures with 30.3% sequence identity on a 500-protein test set.Its inputs included backbone dihedrals, a fragment-derived sequence profile, and a rotamer-based energy profile.
- Introduction: The study applies deep-learning neural networks with new structural features, architecture, and a larger structure dataset to predict residue identities using local three-dimensional neighborhoods.It compares input setups and investigates whether network outputs can improve protein design.
Results
The study predicts residue identities from local structural features using a shared-parameter neural architecture, then evaluates accuracy, top-K enrichment, and design applications. The network improves sequence identity in Rosetta designs and exceeds SPIN on a 50-protein comparison.
- Network architecture: The network predicts 20 residue-type probabilities by combining target-residue features with those of 10–30 neighbors through shared residue-probability and weighting subnetworks.The weighted outputs are concatenated and passed through fully connected layers to a final 20-dimensional softmax.
- Overall accuracy: Including 15 neighboring residues produced the best accuracy, whereas fewer neighbors underrepresented the target environment and more introduced input noise.The authors note that distance-based neighbor selection would require a flexible input size, which they leave for future work.
- Amino-acid-specific accuracy: Pro achieved 92.1% recall and 62.7% precision on SI90N15, while lower-performing amino acids generally had lower abundance in the training set.The authors attribute Pro’s performance to its conformational rigidity and explain that Gly has similarly distinctive structural properties.
- Top-K accuracy: Top-2, top-3, top-5, and top-10 accuracies reached 54.3%, 64.0%, 76.3%, and 91.7%, respectively, on SI90N15.These results indicate that native residues were enriched among the highest-probability predictions.
- Comparison with SPIN: The network achieved approximately 3% higher sequence identity than SPIN on the 50-protein comparison, despite only a nonsignificant accuracy improvement from nearly tripling the training samples.On de novo designed Top7, top-1 identity was 0.250 for SPIN versus 0.283, 0.304, and 0.402 for SI30N15, SI50N15, and SI90N15.
Discussions
The network achieves 38.3% accuracy, but performance is constrained by structural ambiguity, residue-specific difficulty, and incomplete environmental representation. Its predictions can nevertheless narrow design choices, while integration with scoring functions remains important.
- Performance and limitations: 38.3% accuracy was achieved with a 90% sequence-identity cutoff and 15 neighboring residues, but the authors attribute the limit partly to protein structural degeneracy.Proteins with similar structures can have low sequence identity, and many mutations may be tolerated at individual positions.
- Performance and limitations: Top-5 and top-10 accuracies reach 76.3% and 91.7%, respectively, but hydrophobic-residue predictions are less satisfactory than Gly and Pro.The authors suggest incorporating solvation energy into molecular-mechanics scoring functions as a future improvement.
- Environmental representation: Selecting the N closest residues can omit contacting residues or add noncontacting residues, especially for exposed termini.With N=20 and N=30, 3.8% and 1.1% of contacting residues, respectively, are excluded; distance cutoffs could address this but require flexible input sizes.
- Application to design: Using neural-network outputs as residue-type restraints improved sequence identity in Rosetta designs, although the optimal restraint size depends on the protein system.Because residue predictions are independent, real designs should also account for neighboring-residue identities with molecular-mechanics or statistical scoring functions.
Methods
The study trains neural networks on filtered protein structures using residue-level structural features and weighted optimization, then evaluates Rosetta designs with and without neural-network residue restraints.
- Datasets: Training structures were selected from PDB using X-ray crystallography, resolution better than 2 Å, chain length longer than 50, and exclusion of DNA/RNA molecules.Datasets used 30%, 50%, and 90% sequence-identity cutoffs to examine homology effects.
- Datasets: The resulting datasets contained 10,173, 14,064, and 17,607 structures at 30%, 50%, and 90% sequence identity, respectively.Biological assemblies were used, and non-protein residues were removed.
- Input features: Features encoded backbone dihedrals, backbone SASA, secondary structure, neighbor distance, and the neighbor-to-target direction vector.Central-residue and neighbor-residue features were calculated separately for the neural-network inputs.
- Neural-network training: The network used ReLU activations, categorical cross entropy, stochastic gradient descent, and abundance-based sample weighting before outputting probabilities for 20 amino acids.Weights were defined as W_i = N_max/N_i to emphasize underrepresented residue types.
- Rosetta evaluation: Rosetta fixbb designs used the talaris2014 score, with 500 designs per protein performed with and without residue-type restraints.Restraints were incorporated through the “-resfile” option.