PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepSF: deep convolutional neural network for mapping protein sequences to folds

Jie Hou, Badri Adhikari, Jianlin Cheng

arXiv:1706.01010v1cs.LGq-bio.BM

TL;DR

Protein fold recognition relies on indirect sequence comparison, while the structures of most proteins remain unknown. DeepSF directly maps protein sequences to the fold space, and the authors report greater fold-recognition accuracy for difficult targets than existing methods.

  • Problem

    Traditional fold-recognition methods indirectly predict folds through sequence comparison and do not directly explain mapping sequences to thousands of folds.

  • Method

    DeepSF uses a deep convolutional neural network to directly classify protein sequences into the entire fold space.

  • Results

    DeepSF is more accurate than widely used methods for recognizing folds of targets with little or no sequence similarity to proteins with known structures.

  • Takeaways & Limitations

    The method directly classifies proteins from sequence space into all 1,195 SCOP 1.75 folds without sequence comparison.

  • Takeaways & Limitations

    Training is challenging for folds represented by five or fewer proteins, limiting accurate classifier training.

Abstract

from arXiv · show

Motivation Protein fold recognition is an important problem in structural bioinformatics. Almost all traditional fold recognition methods use sequence (homology) comparison to indirectly predict the fold of a tar get protein based on the fold of a template protein with known structure, which cannot explain the relationship between sequence and fold. Only a few methods had been developed to classify protein sequences into a small number of folds due to methodological limitations, which are not generally useful in practice. Results We develop a deep 1D-convolution neural network (DeepSF) to directly classify any protein se quence into one of 1195 known folds, which is useful for both fold recognition and the study of se quence-structure relationship. Different from traditional sequence alignment (comparison) based methods, our method automatically extracts fold-related features from a protein sequence of any length and map it to the fold space. We train and test our method on the datasets curated from SCOP1.75, yielding a classification accuracy of 80.4%. On the independent testing dataset curated from SCOP2.06, the classification accuracy is 77.0%. We compare our method with a top profile profile alignment method - HHSearch on hard template-based and template-free modeling targets of CASP9-12 in terms of fold recognition accuracy. The accuracy of our method is 14.5%-29.1% higher than HHSearch on template-free modeling targets and 4.5%-16.7% higher on hard template-based modeling targets for top 1, 5, and 10 predicted folds. The hidden features extracted from sequence by our method is robust against sequence mutation, insertion, deletion and truncation, and can be used for other protein pattern recognition problems such as protein clustering, comparison and ranking.

1 Introduction

Protein fold recognition is needed because most protein structures are unknown, yet sequence-alignment approaches infer folds indirectly and traditional machine learning handles only small fold sets. DeepSF addresses these limitations by directly mapping variable-length protein sequences to all 1,195 SCOP 1.75 folds and extracting features useful for additional protein comparisons.

  • More than 99% of protein structures are unknown, making sequence-based computational fold detection necessary for structure and function prediction.
  • Sequence-alignment methods transfer a template neighbor’s fold to the target and therefore cannot explain the sequence–structure relationship.
  • Traditional machine-learning methods classify only small, pre-selected fold sets and cannot generally predict arbitrary proteins across thousands of folds.
  • DeepSF uses a 1D deep convolutional neural network to classify variable-length protein sequences directly into all 1,195 SCOP 1.75 folds.
  • DeepSF extracts hidden fold-related features without pairwise alignment that can support protein similarity measurement, clustering, and template selection.
  • Across independent SCOP and CASP evaluations, DeepSF was more sensitive than HHSearch and its hidden features were robust to sequence mutations, insertions, deletions, and truncations.

2.1 Datasets

The study uses SCOP-derived datasets spanning 1,195 folds, an independent SCOP 2.06 test set, and CASP template-based and template-free targets. The highly uneven fold distribution makes accurate prediction especially difficult for small folds, while similarity filtering supports independent evaluation.

  • Training and validation datasets: The SCOP 1.75 dataset contains 16,712 proteins across 7 structural classes and 1,195 folds.
  • Training and validation datasets: 69% of folds contain at most 5 proteins, creating a major training challenge, especially for small folds.
  • Training and validation datasets: SCOP 1.75 sequences were split 80/20 into training and validation sets for each fold, with validation filtered by pairwise similarity to training data.
  • Independent SCOP test dataset: The independent SCOP 2.06 test set contains 4,188 domains across 550 folds after removing sequences similar to SCOP 1.75 and filtering remaining pairs to at most 40% similarity.
  • CASP datasets: The CASP dataset covers experiments from CASP9–12 and includes 186 targets: 96 template-free or seemingly template-free and 90 template-based.
  • CASP datasets: CASP target folds were assigned by transferring a SCOP domain’s fold when structural similarity exceeded a TM-score of 0.5.

2.2 Input feature generation and label assignment

DeepSF represents each variable-length protein sequence with residue-level sequence, profile, secondary-structure, and solvent-accessibility features, then assigns an SCOP 1.75 fold label. These features encode each protein as L × 45 numbers for classification into 1,195 fold categories.

  • Input feature generation: Four input feature types represent sequence, profile, predicted secondary structure, and predicted solvent accessibility.Each residue is encoded using amino-acid identity, PSSM values, secondary-structure classes, and solvent-accessibility states.
  • Input feature generation: Each residue is represented by a 45-number vector, so a protein of length L is encoded by L × 45 numbers.
  • Input feature generation: Sequence features use a 20-dimension one-hot amino-acid vector at every residue position.
  • Input feature generation: PSSM features are generated by PSI-BLAST against the nr90 database and provide 20 position-specific values per residue.
  • Label assignment: Each protein receives a predefined SCOP 1.75 fold index from 0 to 1,194 as its class label.

2.3 Deep convolutional neural network for fold classification

DeepSF uses a 1D convolutional network to transform variable-length protein features into fixed-length hidden representations and predict probabilities over 1,195 folds. Convolutions use multiple window sizes, followed by K-max pooling and fully connected classification layers.

  • Architecture: The network contains 15 layers, including 10 convolutional layers, K-max pooling, flattening, a fully connected hidden layer, and an output layer.
  • Architecture: The input layer accepts L × 45 positional features for proteins of variable length L.
  • Convolutional layers: Each convolution filter produces L hidden features through convolution, batch normalization, and ReLU transformation.
  • Convolutional layers: Two window sizes, 6 and 10, are used in the convolutional layers and were selected as close to average beta-sheet and alpha-helix lengths.
  • Pooling and classification: K-max pooling extracts the 30 highest activations from each feature map and combines both window sizes into a 10 × 30 × 2 feature vector.
  • Pooling and classification: The pooled vector feeds a 500-node hidden layer and then 1,195 softmax output nodes predicting fold probabilities.

2.4 Model training and validation

DeepSF training accommodates variable sequence lengths by grouping proteins into length-based bins and padding within each mini-batch. Model selection uses validation accuracy across multiple sequence-similarity levels before independent testing.

  • Training procedure: Proteins up to 1,419 residues are grouped into fixed-length bins so sequences within a mini-batch have similar lengths.
  • Training procedure: Sequences shorter than the maximum length in their bin receive zero-padding.
  • Training procedure: Mini-batches are trained for 100 epochs, while proteins in each bin receive fewer epochs to reduce over-training on a specific bin.
  • Model selection: Different bin sizes and model parameters are evaluated to select a suitable training configuration.
  • Model selection: The final model is selected by the best average validation accuracy across datasets with 95%, 70%, 40%, and 25% sequence similarity to training data.

2.5 Model evaluation and benchmarking

DeepSF is evaluated on independent SCOP 2.06 and CASP datasets, with performance analyzed across imbalanced fold-size groups and compared against majority assignment, HHSearch, and PSI-BLAST. CASP evaluation reports top-1, top-5, and top-10 fold predictions.

  • Test datasets: Independent testing uses SCOP 2.06 and CASP datasets.
  • Evaluation design: Because fold sizes are highly unbalanced, the 1,195 folds are divided into small, medium, and large groups by protein count.
  • Evaluation design: DeepSF is evaluated on all test proteins and separately within each fold-size category.
  • Baselines: Benchmarking compares DeepSF with majority assignment, HHSearch, and PSI-BLAST.

2.6 Hidden fold-related feature extraction and template ranking

DeepSF’s hidden-layer outputs provide fold-related sequence features that bridge protein sequence and fold spaces, enabling protein comparison and template ranking.

  • The 14th-layer outputs are used as hidden fold-related features, called SF-Features.These features bridge protein sequence space and protein fold space.
  • SF-Features extracted from protein sequences can assess protein similarity and rank template proteins for a target.For each target, DeepSF compares its features with proteins in predicted folds and ranks templates by feature distance.
  • Four distance or similarity metrics are evaluated for comparing fold-related features.The metrics are used to calculate distances between proteins’ feature vectors for clustering and template ranking.
  • Hierarchical clustering evaluates the metrics by repeatedly clustering sampled proteins from five folds using their fold-related feature vectors.The process samples five folds, uses up to 100 proteins, repeats 1,000 times, and compares clustering accuracy.

3 Results

DeepSF achieves strong fold-classification accuracy across SCOP and CASP evaluations, outperforming alignment-based baselines and supporting protein comparison and structure prediction.

  • 3.1 Training and validation on SCOP 1.75 dataset: Bin size 15 converges fastest and gives the highest training and validation accuracy among the tested bin sizes.The evaluated bin sizes are 500, 200, 50, 30, and 15.
  • 3.1 Training and validation on SCOP 1.75 dataset: 80.4% top-1 and 93.7% top-5 accuracy are achieved on the highest-similarity SCOP 1.75 validation set.Accuracy declines with lower sequence similarity, reaching 66.9% top-1 and 87.6% top-5 at 25% similarity.
  • 3.1 Training and validation on SCOP 1.75 dataset: 76.9% top-1 and 92.2% top-5 accuracy are obtained on the independent SCOP 2.06 dataset, versus 4.3% and 16.0% for majority assignment.Top-1 accuracy is 82.9%, 72.1%, and 61.0% on Large, Medium, and Small folds, respectively.
  • 3.2 Performance on CASP dataset: DeepSF outperforms HHSearch, PSI-BLAST, and majority assignment on both template-based and template-free CASP targets.The comparison covers 90 template-based and 96 template-free proteins.
  • 3.2 Performance on CASP dataset: 14.5%, 20.8%, and 29.1% higher accuracy than HHSearch are achieved for top-1, top-5, and top-10 predictions on hardest template-free targets.DeepSF’s corresponding accuracies are 26.0%, 57.3%, and 72.9%.
  • 3.4 Fold-classification assisted protein structure prediction: DeepSF achieves a higher average TMscore than HHSearch for the evaluated structure-prediction targets.The average TMscores are 0.27 for DeepSF and 0.25 for HHSearch, with Wilcoxon paired-test p-value 0.019.
  • DeepSF features remain more similar to original-sequence features than to random-sequence features after mutations, insertions, and deletions.The Wilcoxon rank-sum test reports p < 2.2e-16.

4 Conclusion

DeepSF directly maps variable-length protein sequences to the entire SCOP 1.75 fold space without sequence comparison. Its learned features are robust to sequence perturbations and support additional protein-analysis tasks.

  • DeepSF directly classifies variable-length protein sequences into all 1,195 folds defined in SCOP 1.75.
  • The method maps sequences to fold space without pairwise sequence comparison and automatically extracts fold-related hidden features through deep convolution.
  • Unlike earlier machine-learning methods, DeepSF does not rely on fixed-size windows or human expertise for feature extraction.
  • The extracted features are robust against sequence perturbations and can support protein comparison, clustering, template ranking, and structure prediction.
  • DeepSF was more accurate than widely used profile-profile alignment methods for recognizing folds of target proteins with little or no sequence similarity to known-structure proteins.
  • The direct sequence-to-fold approach complements traditional sequence-alignment methods and provides a new way to study sequence-structure relationships.

Supplementary Figures

The supplementary figures describe dataset composition, protein-length coverage, batch construction, training behavior, feature clustering, and sequence-perturbation experiments.

  • The SCOP 1.75 training dataset groups 1,195 folds into Small, Medium, and Large categories based on the number of proteins per fold.
  • Proteins shorter than 1,419 residues collectively contain all 1,195 folds in the SCOP 1.75 dataset.
  • The SCOP 1.75 proteins with sequence identity at most 95% range from 9 to 1,419 residues, with the figure also showing domain counts by protein length.
  • Mini-batch distributions are shown for bin sizes of 500, 200, 50, 30, and 15, with length intervals on the x axis and protein counts on the y axis.
  • Training-dataset classification accuracy is plotted against training epochs for five different bin sizes.
  • Hierarchical clustering of SF-features groups proteins from five folds according to similar hidden-feature values.
  • For d1ipaa2, the perturbation set contains 45 deletion, 41 insertion, and 20 mutation sequences, compared through KL-D divergence with 500 random sequences.
Loading 1706.01010v1…