Improving protein disorder prediction by deep bidirectional long short-term memory recurrent neural networks

Detecting Proline and Non-Proline Cis Isomers in Protein Structures from Sequences Using Deep Residual Ensemble Learning

It has been long established that cis conformations of amino acid residues play many biologically important roles despite their rare occurrence in protein structure. Because of this rarity, few methods have been developed for predicting cis isomers from protein sequences, most of which are based on outdated datasets and lack the means for independent testing. In this work, using a database of >10000 high-resolution protein structures, we update the statistics of cis isomers and develop a sequence-based prediction technique using an ensemble of residual convolutional and long short-term memory bidirectional recurrent neural networks that allow learning from the whole protein sequence. We show that ensembling eight neural network models yields maximum Matthews correlation coefficient values of approximately 0.35 for cis-Pro isomers and 0.1 for cis-nonPro residues. The method should be useful for prioritizing functionally important residues in cis isomers for experimental validations and improving the sampling of rare protein conformations for ab initio protein structure prediction.

IDP⁻CRF: Intrinsically Disordered Protein/Region Identification Based on Conditional Random Fields

Accurate prediction of intrinsically disordered proteins/regions is one of the most important tasks in bioinformatics, and some computational predictors have been proposed to solve this problem. How to efficiently incorporate the sequence-order effect is critical for constructing an accurate predictor because disordered region distributions show global sequence patterns. In order to capture these sequence patterns, several sequence labelling models have been applied to this field, such as conditional random fields (CRFs). However, these methods suffer from certain disadvantages. In this study, we proposed a new computational predictor called IDP⁻CRF, which is trained on an updated benchmark dataset based on the MobiDB database and the DisProt database, and incorporates more comprehensive sequence-based features, including PSSMs (position-specific scoring matrices), kmer, predicted secondary structures, and relative solvent accessibilities. Experimental results on the benchmark dataset and two independent datasets show that IDP⁻CRF outperforms 25 existing state-of-the-art methods in this field, demonstrating that IDP⁻CRF is a very useful tool for identifying IDPs/IDRs (intrinsically disordered proteins/regions). We anticipate that IDP⁻CRF will facilitate the development of protein sequence analysis.

De novo profile generation based on sequence context specificity with the long short-term memory network

Background

Long short-term memory (LSTM) is one of the most attractive deep learning methods to learn time series or contexts of input data. Increasing studies, including biological sequence analyses in bioinformatics, utilize this architecture. Amino acid sequence profiles are widely used for bioinformatics studies, such as sequence similarity searches, multiple alignments, and evolutionary analyses. Currently, many biological sequences are becoming available, and the rapidly increasing amount of sequence data emphasizes the importance of scalable generators of amino acid sequence profiles.

Results

We employed the LSTM network and developed a novel profile generator to construct profiles without any assumptions, except for input sequence context. Our method could generate better profiles than existing de novo profile generators, including CSBuild and RPS-BLAST, on the basis of profile-sequence similarity search performance with linear calculation costs against input sequence size. In addition, we analyzed the effects of the memory power of LSTM and found that LSTM had high potential power to detect long-range interactions between amino acids, as in the case of beta-strand formation, which has been a difficult problem in protein bioinformatics using sequence information.

Conclusion

We demonstrated the importance of sequence context and the feasibility of LSTM on biological sequence analyses. Our results demonstrated the effectiveness of memories in LSTM and showed that our de novo profile generator, SPBuild, achieved higher performance than that of existing methods for profile prediction of beta-strands, where long-range interactions of amino acids are important and are known to be difficult for the existing window-based prediction methods. Our findings will be useful for the development of other prediction methods related to biological sequences by machine learning methods.

Electronic supplementary material

The online version of this article (10.1186/s12859-018-2284-1) contains supplementary material, which is available to authorized users.

Predicting human protein function with multi-task deep neural networks

Machine learning methods for protein function prediction are urgently needed, especially now that a substantial fraction of known sequences remains unannotated despite the extensive use of functional assignments based on sequence similarity. One major bottleneck supervised learning faces in protein function prediction is the structured, multi-label nature of the problem, because biological roles are represented by lists of terms from hierarchically organised controlled vocabularies such as the Gene Ontology. In this work, we build on recent developments in the area of deep learning and investigate the usefulness of multi-task deep neural networks (MTDNN), which consist of upstream shared layers upon which are stacked in parallel as many independent modules (additional hidden layers with their own output units) as the number of output GO terms (the tasks). MTDNN learns individual tasks partially using shared representations and partially from task-specific characteristics. When no close homologues with experimentally validated functions can be identified, MTDNN gives more accurate predictions than baseline methods based on annotation frequencies in public databases or homology transfers. More importantly, the results show that MTDNN binary classification accuracy is higher than alternative machine learning-based methods that do not exploit commonalities and differences among prediction tasks. Interestingly, compared with a single-task predictor, the performance improvement is not linearly correlated with the number of tasks in MTDNN, but medium size models provide more improvement in our case. One of advantages of MTDNN is that given a set of features, there is no requirement for MTDNN to have a bootstrap feature selection procedure as what traditional machine learning algorithms do. Overall, the results indicate that the proposed MTDNN algorithm improves the performance of protein function prediction. On the other hand, there is still large room for deep learning techniques to further enhance prediction ability.

Sixty-five years of the long march in protein secondary structure prediction: the final stretch?

Protein secondary structure prediction began in 1951 when Pauling and Corey predicted helical and sheet conformations for protein polypeptide backbone even before the first protein structure was determined. Sixty-five years later, powerful new methods breathe new life into this field. The highest three-state accuracy without relying on structure templates is now at 82-84%, a number unthinkable just a few years ago. These improvements came from increasingly larger databases of protein sequences and structures for training, the use of template secondary structure information and more powerful deep learning techniques. As we are approaching to the theoretical limit of three-state prediction (88-90%), alternative to secondary structure prediction (prediction of backbone torsion angles and Cα-atom-based angles and torsion angles) not only has more room for further improvement but also allows direct prediction of three-dimensional fragment structures with constantly improved accuracy. About 20% of all 40-residue fragments in a database of 1199 non-redundant proteins have <6 Å root-mean-squared distance from the native conformations by SPIDER2. More powerful deep learning methods with improved capability of capturing long-range interactions begin to emerge as the next generation of techniques for secondary structure prediction. The time has come to finish off the final stretch of the long march towards protein secondary structure prediction.

Grid-based prediction of torsion angle probabilities of protein backbone and its application to discrimination of protein intrinsic disorder regions and selection of model structures

BACKGROUND

Protein structure can be described by backbone torsion angles: rotational angles about the N-Cα bond (φ) and the Cα-C bond (ψ) or the angle between Cαi-1-Cαi-Cαi + 1 (θ) and the rotational angle about the Cαi-Cαi + 1 bond (τ). Thus, their accurate prediction is useful for structure prediction and model refinement. Early methods predicted torsion angles in a few discrete bins whereas most recent methods have focused on prediction of angles in real, continuous values. Real value prediction, however, is unable to provide the information on probabilities of predicted angles.

RESULTS

Here, we propose to predict angles in fine grids of 5° by using deep learning neural networks. We found that this grid-based technique can yield 2-6% higher accuracy in predicting angles in the same 5° bin than existing prediction techniques compared. We further demonstrate the usefulness of predicted probabilities at given angle bins in discrimination of intrinsically disorder regions and in selection of protein models.

CONCLUSIONS

The proposed method may be useful for characterizing protein structure and disorder. The method is available at http://sparks-lab.org/server/SPIDER2/ as a part of SPIDER2 package.

Structural analysis of human glycoprotein butyrylcholinesterase using atomistic molecular dynamics: The importance of glycosylation site ASN241

Human butyrylcholinesterase (BChE) is a glycoprotein capable of bioscavenging toxic compounds such as organophosphorus (OP) nerve agents. For commercial production of BChE, it is practical to synthesize BChE in non–human expression systems, such as plants or animals. However, the glycosylation profile in these systems is significantly different from the human glycosylation profile, which could result in changes in BChE’s structure and function. From our investigation, we found that the glycan attached to ASN₂₄₁ is both structurally and functionally important due to its close proximity to the BChE tetramerization domain and the active site gorge. To investigate the effects of populating glycosylation site ASN₂₄₁, monomeric human BChE glycoforms were simulated with and without site ASN₂₄₁ glycosylated. Our simulations indicate that the structure and function of human BChE are significantly affected by the absence of glycan 241.

Protein remote homology detection based on bidirectional long short-term memory

BACKGROUND

Protein remote homology detection plays a vital role in studies of protein structures and functions. Almost all of the traditional machine leaning methods require fixed length features to represent the protein sequences. However, it is never an easy task to extract the discriminative features with limited knowledge of proteins. On the other hand, deep learning technique has demonstrated its advantage in automatically learning representations. It is worthwhile to explore the applications of deep learning techniques to the protein remote homology detection.

RESULTS

In this study, we employ the Bidirectional Long Short-Term Memory (BLSTM) to learn effective features from pseudo proteins, also propose a predictor called ProDec-BLSTM: it includes input layer, bidirectional LSTM, time distributed dense layer and output layer. This neural network can automatically extract the discriminative features by using bidirectional LSTM and the time distributed dense layer.

CONCLUSION

Experimental results on a widely-used benchmark dataset show that ProDec-BLSTM outperforms other related methods in terms of both the mean ROC and mean ROC50 scores. This promising result shows that ProDec-BLSTM is a useful tool for protein remote homology detection. Furthermore, the hidden patterns learnt by ProDec-BLSTM can be interpreted and visualized, and therefore, additional useful information can be obtained.

DisBind: A database of classified functional binding sites in disordered and structured regions of intrinsically disordered proteins

Background

Intrinsically unstructured or disordered proteins function via interacting with other molecules. Annotation of these binding sites is the first step for mapping functional impact of genetic variants in coding regions of human and other genomes, considering that a significant portion of eukaryotic genomes code for intrinsically disordered regions in proteins.

Results

DisBind (available at http://biophy.dzu.edu.cn/DisBind) is a collection of experimentally supported binding sites in intrinsically disordered proteins and proteins with both structured and disordered regions. There are a total of 226 IDPs with functional site annotations. These IDPs contain 465 structured regions (ORs) and 428 IDRs according to annotation by DisProt. The database contains a total of 4232 binding residues (from UniProt and PDB structures) in which 2836 residues are in ORs and 1396 in IDRs. These binding sites are classified according to their interacting partners including proteins, RNA, DNA, metal ions and others with 2984, 258, 383, 350, and 262 annotated binding sites, respectively. Each entry contains site-specific annotations (structured regions, intrinsically disordered regions, and functional binding regions) that are experimentally supported according to PDB structures or annotations from UniProt.

Conclusion

The searchable DisBind provides a reliable data resource for functional classification of intrinsically disordered proteins at the residue level.