Protein secondary structure prediction using modular reciprocal bidirectional recurrent neural networks

Sequence-based prediction of one dimensional structural properties of proteins has been a long-standing subproblem of protein structure prediction. Recently, prediction accuracy has been significantly improved due to the rapid expansion of protein sequence and structure libraries and advances in deep learning techniques, such as residual convolutional networks (ResNets) and Long-Short-Term Memory Cells in Bidirectional Recurrent Neural Networks (LSTM-BRNNs). Here we leverage an ensemble of LSTM-BRNN and ResNet models, together with predicted residue-residue contact maps, to continue the push towards the attainable limit of prediction for 3- and 8-state secondary structure, backbone angles (θ, τ, ϕ and ψ), half-sphere exposure, contact numbers and solvent accessible surface area (ASA). The new method, named SPOT-1D, achieves similar, high performance on a large validation set and test set (≈1000 proteins in each set), suggesting robust performance for unseen data. For the large test set, it achieves 87% and 77% in 3- and 8-state secondary structure prediction and 0.82 and 0.86 in correlation coefficients between predicted and measured ASA and contact numbers, respectively. Comparison to current state-of-the-art techniques reveals substantial improvement in secondary structure and backbone angle prediction. In particular, 44% of 40-residue fragment structures constructed from predicted backbone Cα-based θ and τ angles are less than 6 Å root-mean-squared-distance from their native conformations, nearly 20% better than the next best. The method is expected to be useful for advancing protein structure and function prediction. SPOT-1D and its data is available at: http://sparks-lab.org/. Supplementary data are available at Bioinformatics online.

The accuracy of predicting protein local and global structural properties such as secondary structure and solvent accessible surface area has been stagnant for many years because of the challenge of accounting for non-local interactions between amino acid residues that are close in three-dimensional structural space but far from each other in their sequence positions. All existing machine-learning techniques relied on a sliding window of 10-20 amino acid residues to capture some 'short to intermediate' non-local interactions. Here, we employed Long Short-Term Memory (LSTM) Bidirectional Recurrent Neural Networks (BRNNs) which are capable of capturing long range interactions without using a window. We showed that the application of LSTM-BRNN to the prediction of protein structural properties makes the most significant improvement for residues with the most long-range contacts (|i-j| >19) over a previous window-based, deep-learning method SPIDER2. Capturing long-range interactions allows the accuracy of three-state secondary structure prediction to reach 84% and the correlation coefficient between predicted and actual solvent accessible surface areas to reach 0.80, plus a reduction of 5%, 10%, 5% and 10% in the mean absolute error for backbone ϕ , ψ , θ and τ angles, respectively, from SPIDER2. More significantly, 27% of 182724 40-residue models directly constructed from predicted C α atom-based θ and τ have similar structures to their corresponding native structures (6Å RMSD or less), which is 3% better than models built by ϕ and ψ angles. We expect the method to be useful for assisting protein structure and function prediction. The method is available as a SPIDER3 server and standalone package at http://sparks-lab.org . yaoqi.zhou@griffith.edu.au or yuedong.yang@griffith.edu.au. Supplementary data are available at Bioinformatics online.

Background: Protein secondary structure prediction is an important task in bioinformatics and structural biology. Protein’s structure is the basis for its corresponding function. Experimental methods for determining the tertiary structure of proteins are both costly and time-consuming. Since the tertiary structure of proteins is further formed by secondary structure, leveraging computational approaches for efficient prediction of protein secondary structure is important. Both local and global interactions between amino acids affect the prediction results. Objective: We propose a module aimed at processing sequence profile features for deep feature extraction and constructing a lightweight network to extract fused features. Methods: To enhance the network’s ability to capture both local and global interactions, we propose an efficient method InConTPSS, which integrates convolution operation with different receptive fields and temporal convolutional networks in the inception architecture. Concurrently, InConTPSS takes into account the issue of distribution imbalance across various states of secondary structures and improves the predictive performance of scarce categories. Results: Experimental results on six benchmark datasets (including CASP12, CASP13, CASP14, CB513, TEST2016, and TEST2018) demonstrate our method achieves state-of-the-art performance with a simpler model on both 3-state and 8-state secondary structure prediction. result: Experimental results on benchmark datasets (including CASP12, CASP13, CASP14, CB513, TEST2016 and TEST2018) demonstrate our method achieves state-of-the-art performance with a simpler model on both 3-state and 8-state secondary structure prediction. Conclusion: Through the combination of the convolutional layer and temporal convolutional network, the inception network structure can effectively process the fused features and improve the prediction results. InConTPSS achieves the most advanced performance in protein secondary structure prediction, and the reasonable use of label-distribution-aware margin loss in our method can effectively improve the prediction accuracy of scarce secondary structures.

Protein structure prediction is one of the most important problems in modern computational biology. Protein secondary structure prediction is a key step in prediction of protein tertiary structure. There have emerged many methods based on machine learning techniques, such as neural networks (NN) and support vector machine (SVM) etc., to focus on the prediction of the secondary structures. In this paper, a new method was proposed based on SVM. Different from the existing methods, this method takes into account of the physical-chemical properties and structure properties of amino acids. When tested on the most popular dataset CB513, it achieved a Q(3) accuracy of 0.7844, which illustrates that it is one of the top range methods for protein of secondary structure prediction.

Secondary structure predictions are increasingly becoming the workhorse for several methods aiming at predicting protein structure and function. Here we use ensembles of bidirectional recurrent neural network architectures, PSI-BLAST-derived profiles, and a large nonredundant training set to derive two new predictors: (a) the second version of the SSpro program for secondary structure classification into three categories and (b) the first version of the SSpro8 program for secondary structure classification into the eight classes produced by the DSSP program. We describe the results of three different test sets on which SSpro achieved a sustained performance of about 78% correct prediction. We report confusion matrices, compare PSI-BLAST to BLAST-derived profiles, and assess the corresponding performance improvements. SSpro and SSpro8 are implemented as web servers, available together with other structural feature predictors at: http://promoter.ics.uci.edu/BRNN-PRED/.

The secondary structure of a protein is critical for establishing a link between the protein primary and tertiary structures. For this reason, it is important to design methods for accurate protein secondary structure prediction. Most of the existing computational techniques for protein structural and functional prediction are based on machine learning with shallow frameworks. Different deep learning architectures have already been applied to tackle protein secondary structure prediction problem. In this study, deep learning based models, i.e., convolutional neural network and long short-term memory for protein secondary structure prediction were proposed. The input to proposed models is amino acid sequences which were derived from CulledPDB dataset. Hyperparameter tuning with cross validation was employed to attain best parameters for the proposed models. The proposed models enables effective processing of amino acids and attain approximately 87.05% and 87.47% Q3 accuracy of protein secondary structure prediction for convolutional neural network and long short-term memory models, respectively.

Combining Prediction of Protein Aggregation Propensities with Prediction of Other One-Dimensional Properties

Protein secondary structure predictions and amino acid long range contact map predictions from primary sequence of proteins have been explored to aid in modelling protein tertiary structures. In order to evaluate the usefulness of secondary structure and 3D-residue contact prediction methods to model protein structures we have used the known Q3 (alpha-helix,beta-strands and irregular turns/loops) secondary structure information, along with residue-residue contact information as restraints for MODELLER. We present here results of our modelling studies on 30 best resolved single domain protein structures of varied lengths. The results shows that it is very difficult to obtain useful models even with 100% accurate secondary structure predictions and accurate residue contact predictions for up to 30% of residues in a sequence. The best models that we obtained for proteins of lengths 37, 70, 118, 136 and 193 amino acid residues are of RMSDs 4.17, 5.27, 9.12, 7.89 and 9.69,respectively. The results show that one can obtain better models for the proteins which have high percent of alpha-helix content. This analysis further shows that MODELLER restrain optimization program can be useful only if we have truly homologous structure(s) as a template where it derives numerous restraints, almost identical to the templates used. This analysis also clearly indicates that even if we satisfy several true residue-residue contact distances, up to 30%of their sequence length with fully known secondary structural information, we end up predicting model structures much distant from their corresponding native structures.

Accurate prediction of protein secondary structure is essential for accurate sequence alignment, three-dimensional structure modeling, and function prediction. The accuracy of ab initio secondary structure prediction from sequence, however, has only increased from around 77 to 80% over the past decade. Here, we developed a multistep neural-network algorithm by coupling secondary structure prediction with prediction of solvent accessibility and backbone torsion angles in an iterative manner. Our method called SPINE X was applied to a dataset of 2640 proteins (25% sequence identity cutoff) previously built for the first version of SPINE and achieved a 82.0% accuracy based on 10-fold cross validation (Q(3)). Surpassing 81% accuracy by SPINE X is further confirmed by employing an independently built test dataset of 1833 protein chains, a recently built dataset of 1975 proteins and 117 CASP 9 targets (critical assessment of structure prediction techniques) with an accuracy of 81.3%, 82.3% and 81.8%, respectively. The prediction accuracy is further improved to 83.8% for the dataset of 2640 proteins if the DSSP assignment used above is replaced by a more consistent consensus secondary structure assignment method. Comparison to the popular PSIPRED and CASP-winning structure-prediction techniques is made. SPINE X predicts number of helices and sheets correctly for 21.0% of 1833 proteins, compared to 17.6% by PSIPRED. It further shows that SPINE X consistently makes more accurate prediction in helical residues (6%) without over prediction while PSIPRED makes more accurate prediction in coil residues (3-5%) and over predicts them by 7%. SPINE X Server and its training/test datasets are available at http://sparks.informatics.iupui.edu/

Synchrotron-radiation vacuum-ultraviolet circular dichroism (VUVCD) spectroscopy can significantly improve the predictive accuracy of the contents and segment numbers of protein secondary structures by extending the short-wavelength limit of the spectra. In the present study, we combined VUVCD spectra down to 160 nm with neural-network (NN) method to improve the sequence-based prediction of protein secondary structures. The secondary structures of 30 target proteins (test set) were assigned into alpha-helices, beta-strands, and others by the DSSP program based on their X-ray crystal structures. Combining the alpha-helix and beta-strand contents estimated from the VUVCD spectra of the target proteins improved the overall sequence-based predictive accuracy Q(3) for three secondary-structure components from 59.5 to 60.7%. Incorporating the position-specific scoring matrix in the NN method improved the predictive accuracy from 70.9 to 72.1% when combining the secondary-structure contents, to 72.5% when combining the numbers of segments, and finally to 74.9% when filtering the VUVCD data. Improvement in the sequence-based prediction of secondary structures was also apparent in two other indices of the overall performance: the correlation coefficient (C) and the segment overlap value (SOV). These results suggest that VUVCD data could enhance the predictive accuracy to over 80% when combined with the currently best sequence-prediction algorithms, greatly expanding the applicability of VUVCD spectroscopy to protein structural biology.

Protein structure prediction has always been an important research area in biochemistry. In particular, the prediction of protein secondary structure has been a well-studied research topic. The experimental methods currently used to determine protein structure are accurate, yet costly both in terms of equipment and time. Despite the recent breakthrough of combining multiple sequence alignment information and artificial intelligence algorithms to predict protein secondary structure, the Q3 accuracy of various computational prediction methods rarely has exceeded 75%. In this paper, a newly developed rule-based data-mining approach called RT-RICO (Relaxed Threshold Rule Induction from Coverings) is presented. This method identifies dependencies between amino acids in a protein sequence and generates rules that can be used to predict secondary structure. RT-RICO achieved a Q3 score of 81.75% on the standard test dataset RS126 and a Q3 score of 79.19% on the standard test dataset CB396, an improvement over comparable computational methods.

In order to study the three dimensional structure (3D) and functionalities of proteins, the secondary structure of proteins is imperative. To date, Convolutional Neural Network (CNN) and Long Short Term Memory network (LSTM) are two networks that have been used triumphantly in the study of protein sequences. In this paper, two models for the prediction of eight-state accuracy (Q8 accuracy) for protein secondary structure are proposed, when provided with the amino acid sequence. The first model is the Concatenated Convolutional1D model (CCN1D) wherein the Convolutional 1D networks are concatenated and this model obtains only the short range information between residues. CCN1D does not obtain complete information between the amino acids present within the protein. In order to obtain long range dependency among amino acid residues, Bidirectional LSTM (BLSTM) is used. Thus in order to get accurate results, the second model known as Convolution-BLSTM (C-BLSTM) is proposed. It is a cascaded network of CNN and BLSTM. The evolutionary profiles of proteins are used as the input to both models. The Q8 Accuracy of CCN1D and C-BLSTM is 71.73% at 0.38 Dropout rate and 72.47% at Dropout rate 0.3 respectively for the CullPDB5926 dataset.

The accurate prediction of secondary structures of proteins (SSPs) is a critical challenge in molecular biology and structural bioinformatics. Despite recent advancements, this task remains complex and demands further exploration. This study presents a novel approach to SSP prediction using atom-centric substructural multilevel neighborhoods of atoms (MNA) descriptors for protein molecular fragments. A dataset comprising over 335,000 SSPs, annotated by the Dictionary of Secondary Structure in Proteins (DSSP) software from 37,000 proteins, was constructed from Protein Data Bank (PDB) records with a resolution of 2 Å or better. Protein fragments were converted into structural formulae using the RDKit Python package and stored in SD files using the MOL V3000 format. Classification sequence–structure–property relationships (SSPR) models were developed with varying levels of MNA descriptors and a Bayesian algorithm implemented in MultiPASS software. The average prediction accuracy (AUC) for eight SSP types, calculated via leave-one-out cross-validation, was 0.902. For independent test sets (ASTRAL and CB513 datasets), the best SSPR models achieved AUC, Q3, and Q8 values of 0.860, 77.32%, 70.92% and 0.889, 78.78%, 74.74%, respectively. Based on the created models, a freely available web application MNA-PSS-Pred was developed.

Predicting protein secondary structure using a mixed-modal SVM method in a compound pyramid model

This article briefly describes our program Jamsek written in FORTRAN for an ICL 2950/10 computer. Jamsek combines statistical and stereochemical rules most frequently encountered in literature to predict protein secondary structure from its sequence, into a single algorithm. The composite algorithm does not work better than the best existing single algorithms of Garnier et al. (J. Mol. Biol., 120, 97-120, 1978) or Lim (J. Mol. Biol., 88, 873-894, 1974) if percentage of residues with a correctly predicted secondary structure is taken as a criterion. However, it is fairly reliable in predicting the total amount of alpha-helices and beta-sheets in proteins, the secondary structure of highly ordered proteins or their parts and identification of long alpha-helices. It surpasses the previous algorithms by providing a possibility to make a notion about confidence of the prediction of the particular secondary structure elements thanks to the simultaneous availability of four independent predictions of the secondary structure and other relevant data (hydrophobic profile and helical wheel representation). The main body of this article is devoted to a demonstration that output data of Jamsek can simply be used for the prediction of protein topological class, identification of globular proteins containing hydrophobic alpha-helices and, as an auxiliary means, to distinguish between protein coding and non-coding nucleotide sequences.