基于深度学习与多层次信息融合的药物靶标亲和力预测*

药物研发是非常重要但也十分耗费人力物力的过程。利用计算机辅助预测药物与蛋白质亲和力的方法可以极大地加快药物研发过程。药物靶标亲和力预测的关键在于对药物和蛋白质进行准确详细地信息表征。提出一种基于深度学习与多层次信息融合的药物靶标亲和力的预测模型,试图通过综合药物与蛋白质的多层次信息,来获得更好的预测表现。首先将药物表述成分子图和扩展连接指纹两种形式,分别利用图卷积神经网络模块和全连接层进行学习;其次将蛋白质序列和蛋白质K-mer特征分别输入卷积神经网络模块和全连接层来学习蛋白质潜在特征;随后将4个通道学习到的特征进行融合,再利用全连接层进行预测。在两个基准药物靶标亲和力数据集上验证了所提方法的有效性,并与其他已有模型作对比研究。结果说明提出的模型相比基准模型能得到更好的预测性能,表明提出的综合药物与蛋白质多层次信息的药物靶标亲和力预测策略是有效的。     

A Time-Series-Based Feature Extraction Approach for Prediction of Protein Structural Class

This paper presents a novel feature vector based on physicochemical property of amino acids for prediction protein structural classes. The proposed method is divided into three different stages. First, a discrete time series representation to protein sequences using physicochemical scale is provided. Later on, a wavelet-based time-series technique is proposed for extracting features from mapped amino acid sequence and a fixed length feature vector for classification is constructed. The proposed feature space summarizes the variance information of ten different biological properties of amino acids. Finally, an optimized support vector machine model is constructed for prediction of each protein structural class. The proposed approach is evaluated using leave-one-out cross-validation tests on two standard datasets. Comparison of our result with existing approaches shows that overall accuracy achieved by our approach is better than exiting methods.     

Prediction of protein submitochondria locations based on data fusion of various features of sequences

In this study, the predictors are developed for protein submitochondria locations based on various features of sequences. Information about the submitochondria location for a mitochondria protein can provide much better understanding about its function. We use ten representative models of protein samples such as pseudo amino acid composition, dipeptide composition, functional domain composition, the combining discrete model based on prediction of solvent accessibility and secondary structure elements, the discrete model of pairwise sequence similarity, etc. We construct a predictor based on support vector machines (SVMs) for each representative model. The overall prediction accuracy by the leave-one-out cross validation test obtained by the predictor which is based on the discrete model of pairwise sequence similarity is 1% better than the best computational system that exists for this problem. Moreover, we develop a method based on ordered weighted averaging (OWA) which is one of the fusion data operators. Therefore, OWA is applied on the 11 best SVM-based classifiers that are constructed based on various features of sequence. This method is called Mito-Loc. The overall leave-one-out cross validation accuracy obtained by Mito-Loc is about 95%. This indicates that our proposed approach (Mito-Loc) is superior to the result of the best existing approach which has already been reported.     

Prediction of Deleterious Non-Synonymous SNPs Based on Protein Interaction Network and Hybrid Properties

Non-synonymous SNPs (nsSNPs), also known as Single Amino acid Polymorphisms (SAPs) account for the majority of human inherited diseases. It is important to distinguish the deleterious SAPs from neutral ones. Most traditional computational methods to classify SAPs are based on sequential or structural features. However, these features cannot fully explain the association between a SAP and the observed pathophysiological phenotype. We believe the better rationale for deleterious SAP prediction should be: If a SAP lies in the protein with important functions and it can change the protein sequence and structure severely, it is more likely related to disease. So we established a method to predict deleterious SAPs based on both protein interaction network and traditional hybrid properties. Each SAP is represented by 472 features that include sequential features, structural features and network features. Maximum Relevance Minimum Redundancy (mRMR) method and Incremental Feature Selection (IFS) were applied to obtain the optimal feature set and the prediction model was Nearest Neighbor Algorithm (NNA). In jackknife cross-validation, 83.27% of SAPs were correctly predicted when the optimized 263 features were used. The optimized predictor with 263 features was also tested in an independent dataset and the accuracy was still 80.00%. In contrast, SIFT, a widely used predictor of deleterious SAPs based on sequential features, has a prediction accuracy of 71.05% on the same dataset. In our study, network features were found to be most important for accurate prediction and can significantly improve the prediction performance. Our results suggest that the protein interaction context could provide important clues to help better illustrate SAP''s functional association. This research will facilitate the post genome-wide association studies.     

Sequence determinants of protein folding rates: positive correlation between contact energy and contact range indicates selection for fast folding

In comparison with intense investigation of the structural determinants of protein folding rates, the sequence features favoring fast folding have received little attention. Here, we investigate this subject using simple models of protein folding and a statistical analysis of the Protein Data Bank (PDB). The mean-field model by Plotkin and coworkers predicts that the folding rate is accelerated by stronger-than-average interactions at short distance along the sequence. We confirmed this prediction using the Finkelstein model of protein folding, which accounts for realistic features of polymer entropy. We then tested this prediction on the PDB. We found that native interactions are strongest at contact range l = 8. However, since short range contacts tend to be exposed and they are frequently formed in misfolded structures, selection for folding stability tends to make them less attractive, that is, stability and kinetics may have contrasting requirements. Using a recently proposed model, we predicted the relationship between contact range and contact energy based on buriedness and contact frequency. Deviations from this prediction induce a positive correlation between contact range and contact energy, that is, short range contacts are stronger than expected, for 2/3 of the proteins. This correlation increases with the absolute contact order (ACO), as expected if proteins that tend to fold slowly due to large ACO are subject to stronger selection for sequence features favoring fast folding. Our results suggest that the selective pressure for fast folding is detectable only for one third of the proteins in the PDB, in particular those with large contact order.     

Accurate prediction of protein structural classes using functional domains and predicted secondary structure sequences

Protein structural class prediction is one of the challenging problems in bioinformatics. Previous methods directly based on the similarity of amino acid (AA) sequences have been shown to be insufficient for low-similarity protein data-sets. To improve the prediction accuracy for such low-similarity proteins, different methods have been recently proposed that explore the novel feature sets based on predicted secondary structure propensities. In this paper, we focus on protein structural class prediction using combinations of the novel features including secondary structure propensities as well as functional domain (FD) features extracted from the InterPro signature database. Our comprehensive experimental results based on several benchmark data-sets have shown that the integration of new FD features substantially improves the accuracy of structural class prediction for low-similarity proteins as they capture meaningful relationships among AA residues that are far away in protein sequence. The proposed prediction method has also been tested to predict structural classes for partially disordered proteins with the reasonable prediction accuracy, which is a more difficult problem comparing to structural class prediction for commonly used benchmark data-sets and has never been done before to the best of our knowledge. In addition, to avoid overfitting with a large number of features, feature selection is applied to select discriminating features that contribute to achieve high prediction accuracy. The selected features have been shown to achieve stable prediction performance across different benchmark data-sets.     

Accurate prediction of protein structural classes using functional domains and predicted secondary structure sequences

Protein structural class prediction is one of the challenging problems in bioinformatics. Previous methods directly based on the similarity of amino acid (AA) sequences have been shown to be insufficient for low-similarity protein data-sets. To improve the prediction accuracy for such low-similarity proteins, different methods have been recently proposed that explore the novel feature sets based on predicted secondary structure propensities. In this paper, we focus on protein structural class prediction using combinations of the novel features including secondary structure propensities as well as functional domain (FD) features extracted from the InterPro signature database. Our comprehensive experimental results based on several benchmark data-sets have shown that the integration of new FD features substantially improves the accuracy of structural class prediction for low-similarity proteins as they capture meaningful relationships among AA residues that are far away in protein sequence. The proposed prediction method has also been tested to predict structural classes for partially disordered proteins with the reasonable prediction accuracy, which is a more difficult problem comparing to structural class prediction for commonly used benchmark data-sets and has never been done before to the best of our knowledge. In addition, to avoid overfitting with a large number of features, feature selection is applied to select discriminating features that contribute to achieve high prediction accuracy. The selected features have been shown to achieve stable prediction performance across different benchmark data-sets.     

蛋白质序列的混合特征值对折叠速率的影响

鉴于蛋白质折叠速率预测对研究其蛋白质功能的重要性,许多的科研工作者都开始对影响蛋白质折叠速率的因素进行研究。各种预测参数和方法被提出。利用蛋白质编码序列的不同特征参数,不同的二级结构及不同的折叠类的蛋白质对折叠速率的不同影响,我们选取蛋白质编码序列的新的特征值,即选取蛋白质序列的LZ复杂度,等电点等特征值。然后把这些特征值与20种氨基酸的属性αc、Cα、K0、Pβ、Ra、ΔASA、PI、ΔGhD、Nm、LZ、Mu、El融合,建立多元线性回归模型,并利用回归模型计算了13个全α类蛋白质、18个全β类蛋白质、13个混合类蛋白质和39个未分类蛋白质的ln（kf）与预测值之间的相关系数分别达到0.89、0.93、0.98、0.86。在Jack-knife方法的验证下发现在不同的结构中混合特征值与相应折叠速率有很好的相关性。结果表明,在蛋白质折叠过程中,蛋白质序列的LZ复杂度、等电点等特征值可能影响蛋白质的折叠速率及其结构。     

Topology prediction of Brucella abortus Omp2b and Omp2a porins after critical assessment of transmembrane beta strands prediction by several secondary structure prediction methods

In order to propose a reliable model for Brucella porin topology, several structure prediction methods were evaluated in their ability to predict porin topology. Four porins of known structure were selected as test-cases and their secondary structure delineated. The specificity and sensitivity of 11 methods were separately evaluated. Our critical assessment shows that some secondary structure prediction methods (PHD, Dsc, Sopma) originally designed to predict globular protein structure are useful on porin topology prediction. The overall best prediction is obtained by combining these three "generalist" methods with a transmembrane beta strand prediction technique. This "consensus" method was applied to Brucella porins Omp2b and Omp2a, sharing no sequence homology with any other porin. The predicted topology is a 16-stranded antiparallel beta barrel with Omp2a showing a higher number of negatively charged residue in the exposed loops than Omp2b. Experiments are in progress to validate the proposed topology and the functional hypotheses. The ability of the proposed consensus method to predict topology of complex outer membrane protein is briefly discussed.     

Improved recognition of native-like protein structures using a combination of sequence-dependent and sequence-independent features of proteins.

We describe the development of a scoring function based on the decomposition P(structure/sequence) proportional to P(sequence/structure) *P(structure), which outperforms previous scoring functions in correctly identifying native-like protein structures in large ensembles of compact decoys. The first term captures sequence-dependent features of protein structures, such as the burial of hydrophobic residues in the core, the second term, universal sequence-independent features, such as the assembly of beta-strands into beta-sheets. The efficacies of a wide variety of sequence-dependent and sequence-independent features of protein structures for recognizing native-like structures were systematically evaluated using ensembles of approximately 30,000 compact conformations with fixed secondary structure for each of 17 small protein domains. The best results were obtained using a core scoring function with P(sequence/structure) parameterized similarly to our previous work (Simons et al., J Mol Biol 1997;268:209-225] and P(structure) focused on secondary structure packing preferences; while several additional features had some discriminatory power on their own, they did not provide any additional discriminatory power when combined with the core scoring function. Our results, on both the training set and the independent decoy set of Park and Levitt (J Mol Biol 1996;258:367-392), suggest that this scoring function should contribute to the prediction of tertiary structure from knowledge of sequence and secondary structure.     

GeneSilico protein structure prediction meta-server

Rigorous assessments of protein structure prediction have demonstrated that fold recognition methods can identify remote similarities between proteins when standard sequence search methods fail. It has been shown that the accuracy of predictions is improved when refined multiple sequence alignments are used instead of single sequences and if different methods are combined to generate a consensus model. There are several meta-servers available that integrate protein structure predictions performed by various methods, but they do not allow for submission of user-defined multiple sequence alignments and they seldom offer confidentiality of the results. We developed a novel WWW gateway for protein structure prediction, which combines the useful features of other meta-servers available, but with much greater flexibility of the input. The user may submit an amino acid sequence or a multiple sequence alignment to a set of methods for primary, secondary and tertiary structure prediction. Fold-recognition results (target-template alignments) are converted into full-atom 3D models and the quality of these models is uniformly assessed. A consensus between different FR methods is also inferred. The results are conveniently presented on-line on a single web page over a secure, password-protected connection. The GeneSilico protein structure prediction meta-server is freely available for academic users at http://genesilico.pl/meta.     

Prediction of cystine connectivity using SVM

One of the major contributors to protein structures is the formation of disulphide bonds between selected pairs of cysteines at oxidized state. Prediction of such disulphide bridges from sequence is challenging given that the possible combination of cysteine pairs as the number of cysteines increases in a protein. Here, we describe a SVM (support vector machine) model for the prediction of cystine connectivity in a protein sequence with and without a priori knowledge on their bonding state. We make use of a new encoding scheme based on physico-chemical properties and statistical features (probability of occurrence of each amino acid residue in different secondary structure states along with PSI-blast profiles). We evaluate our method in SPX (an extended dataset of SP39 (swiss-prot 39) and SP41 (swiss-prot 41) with known disulphide information from PDB) dataset and compare our results with the recursive neural network model described for the same dataset.     

A Mycoplasma hyorhinis protein with sequence similarities to nucleotide-binding enzymes.

We have determined the nucleotide (nt) and deduced amino acid (aa) sequence of a unique 115-kDa Mycoplasma hyorhinis protein (P115) with an N-terminal region containing a highly conserved consensus sequence characteristics of nt-binding domains of several ATPase and GTPase enzymes. However, P115 lacked additional conserved features characteristic of some classes of nt-binding proteins. Based on the hydropathy profile of the deduced aa sequence, the absence of a leader peptide, its exclusive partitioning into the hydrophilic phase during Triton X-114 phase fractionation of M. hyorhinis, and immunofluorescence analysis indicating no surface-exposed domains, it was concluded that P115 is a cytoplasmic protein lacking intrinsic membrane interaction. M. hyorhinis P115 appears to be a species-specific protein, since it was not detected in any other mycoplasmal or bacterial species examined with specific antibody or genomic probes. Since genetic systems for direct mutational analysis are currently unavailable in this organism, sequence analysis provides critical information in establishing the possible function of this protein. Moreover, the nt sequence encoding P115 reported here supports a previously proposed model, based on synthesis of P115-related proteins in Escherichia coli, suggesting that multiple polypeptide products can be generated from mycoplasma genes by promiscuous translation initiation in this heterologous expression system.     

Prediction of protein structural class for the twilight zone sequences

Structural class characterizes the overall folding type of a protein or its domain. This paper develops an accurate method for in silico prediction of structural classes from low homology (twilight zone) protein sequences. The proposed LLSC-PRED method applies linear logistic regression classifier and a custom-designed, feature-based sequence representation to provide predictions. The main advantages of the LLSC-PRED are the comprehensive representation that includes 58 features describing composition and physicochemical properties of the sequences and transparency of the prediction model. The representation also includes predicted secondary structure content, thus for the first time exploring synergy between these two related predictions. Based on tests performed with a large set of 1673 twilight zone domains, the LLSC-PRED's prediction accuracy, which equals over 62%, is shown to be better than accuracy of over a dozen recently published competing in silico methods and similar to accuracy of other, non-transparent classifiers that use the proposed representation.     

Protein subcellular localization based on PSI-BLAST and machine learning

Subcellular location is an important functional annotation of proteins. An automatic, reliable and efficient prediction system for protein subcellular localization is necessary for large-scale genome analysis. This paper describes a protein subcellular localization method which extracts features from protein profiles rather than from amino acid sequences. The protein profile represents a protein family, discards part of the sequence information that is not conserved throughout the family and therefore is more sensitive than the amino acid sequence. The amino acid compositions of whole profile and the N-terminus of the profile are extracted, respectively, to train and test the probabilistic neural network classifiers. On two benchmark datasets, the overall accuracies of the proposed method reach 89.1% and 68.9%, respectively. The prediction results show that the proposed method perform better than those methods based on amino acid sequences. The prediction results of the proposed method are also compared with Subloc on two redundance-reduced datasets.     

Bagging with CTD- A Novel Signature for the Hierarchical Prediction of Secreted Protein Trafficking in Eukaryotes

Protein trafficking or protein sorting in eukaryotes is a complicated process and is carried out based on the information contaified in the protein. Many methods reported prediction of the subcellular location of proteins from sequence information. However, most of these prediction methods use a flat structure or parallel architecture to perform prediction. In this work, we introduce ensemble classifiers with features that are extracted directly from full length protein sequences to predict locations in the protein-sorting pathway hierarchically. Sequence driven features, sequence mapped features and sequence autocorrelation features were tested with ensemble learners and their performances were compared. When evaluated by independent data testing, ensemble based-bagging algorithms with sequence feature composition, transition and distribution （CTD） successfully classified two datasets with accuracies greater than 90%. We compared our results with similar published methods, and our method equally performed with the others at two levels in the secreted pathway. This study shows that the feature CTD extracted from protein sequences is effective in capturing biological features among compartments in secreted pathways.     

A novel numerical model for protein sequences analysis based on spherical coordinates and multiple physicochemical properties of amino acids

How to characterize short protein sequences to make an effective connection to their functions is an unsolved problem. Here we propose to map the physicochemical properties of each amino acid onto unit spheres so that each protein sequence can be represented quantitatively. We demonstrate the usefulness of this representation by applying it to the prediction of cell penetrating peptides. We show that its combination with traditional composition features yields the best performance across different datasets, among several methods compared. For the convenience of users, a web server has been established for automatic calculations of the proposed features at http://biophy.dzu.edu.cn/SNumD/ .     

Predicting protein subcellular localization by pseudo amino acid composition with a segment-weighted and features-combined approach

Information of protein subcellular location plays an important role in molecular cell biology. Prediction of the subcellular location of proteins will help to understand their functions and interactions. In this paper, a different mode of pseudo amino acid composition was proposed to represent protein samples for predicting their subcellular localization via the following procedures: based on the optimal splice site of each protein sequence, we divided a sequence into sorting signal part and mature protein part, and extracted sequence features from each part separately. Then, the combined features were fed into the SVM classifier to perform the prediction. By the jackknife test on a benchmark dataset in which none of proteins included has more than 90% pairwise sequence identity to any other, the overall accuracies achieved by the method are 94.5% and 90.3% for prokaryotic and eukaryotic proteins, respectively. The results indicate that the prediction quality by our method is quite satisfactory. It is anticipated that the current method may serve as an alternative approach to the existing prediction methods.     

Cysteine separations profiles on protein sequences infer disulfide connectivity

MOTIVATION: Disulfide bonds play an important role in protein folding. A precise prediction of disulfide connectivity can strongly reduce the conformational search space and increase the accuracy in protein structure prediction. Conventional disulfide connectivity predictions use sequence information, and prediction accuracy is limited. Here, by using an alternative scheme with global information for disulfide connectivity prediction, higher performance is obtained with respect to other approaches. RESULT: Cysteine separation profiles have been used to predict the disulfide connectivity of proteins. The separations among oxidized cysteine residues on a protein sequence have been encoded into vectors named cysteine separation profiles (CSPs). Through comparisons of their CSPs, the disulfide connectivity of a test protein is inferred from a non-redundant template set. For non-redundant proteins in SwissProt 39 (SP39) sharing less than 30% sequence identity, the prediction accuracy of a fourfold cross-validation is 49%. The prediction accuracy of disulfide connectivity for proteins in SwissProt 43 (SP43) is even higher (53%). The relationship between the similarity of CSPs and the prediction accuracy is also discussed. The method proposed in this work is relatively simple and can generate higher accuracies compared to conventional methods. It may be also combined with other algorithms for further improvements in protein structure prediction. AVAILABILITY: The program and datasets are available from the authors upon request. CONTACT: cykao@csie.ntu.edu.tw.