Feature Selection and the Class Imbalance Problem in Predicting Protein Function from Sequence

When the standard approach to predict protein function by sequence homology fails, other alternative methods can be used that require only the amino acid sequence for predicting function. One such approach uses machine learning to predict protein function directly from amino acid sequence features. However, there are two issues to consider before successful functional prediction can take place: identifying discriminatory features, and overcoming the challenge of a large imbalance in the training data. We show that by applying feature subset selection followed by undersampling of the majority class, significantly better support vector machine (SVM) classifiers are generated compared with standard machine learning approaches. As well as revealing that the features selected could have the potential to advance our understanding of the relationship between sequence and function, we also show that undersampling to produce fully balanced data significantly improves performance. The best discriminating ability is achieved using SVMs together with feature selection and full undersampling; this approach strongly outperforms other competitive learning algorithms. We conclude that this combined approach can generate powerful machine learning classifiers for predicting protein function directly from sequence.     

Predicting subcellular localization of proteins using machine-learned classifiers

MOTIVATION: Identifying the destination or localization of proteins is key to understanding their function and facilitating their purification. A number of existing computational prediction methods are based on sequence analysis. However, these methods are limited in scope, accuracy and most particularly breadth of coverage. Rather than using sequence information alone, we have explored the use of database text annotations from homologs and machine learning to substantially improve the prediction of subcellular location. RESULTS: We have constructed five machine-learning classifiers for predicting subcellular localization of proteins from animals, plants, fungi, Gram-negative bacteria and Gram-positive bacteria, which are 81% accurate for fungi and 92-94% accurate for the other four categories. These are the most accurate subcellular predictors across the widest set of organisms ever published. Our predictors are part of the Proteome Analyst web-service.     

Prediction of functional class of novel plant proteins by a statistical learning method

In plant genomes, the function of a substantial percentage of the putative protein-coding open reading frames (ORFs) is unknown. These ORFs have no significant sequence similarity to known proteins, which complicates the task of functional study of these proteins. Efforts are being made to explore methods that are complementary to, or may be used in combination with, sequence alignment and clustering methods. A web-based protein functional class prediction software, SVMProt, has shown some capability for predicting functional class of distantly related proteins. Here the usefulness of SVMProt for functional study of novel plant proteins is evaluated. To test SVMProt, 49 plant proteins (without a sequence homolog in the Swiss-Prot protein database, not in the SVMProt training set, and with functional indications provided in the literature) were selected from a comprehensive search of MEDLINE abstracts and Swiss-Prot databases in 1999-2004. These represent unique proteins the function of which, at present, cannot be confidently predicted by sequence alignment and clustering methods. The predicted functional class of 31 proteins was consistent, and that of four other proteins was weakly consistent, with published functions. Overall, the functional class of 71.4% of these proteins was consistent, or weakly consistent, with functional indications described in the literature. SVMProt shows a certain level of ability to provide useful hints about the functions of novel plant proteins with no similarity to known proteins.     

机器学习在蛋白质功能预测领域的研究进展

蛋白质是有机生命体内不可或缺的化合物,在生命活动中发挥着多种重要作用,了解蛋白质的功能有助于医学和药物研发等领域的研究。此外,酶在绿色合成中的应用一直备受人们关注,但是由于酶的种类和功能多种多样,获取特定功能酶的成本高昂,限制了其进一步的应用。目前,蛋白质的具体功能主要通过实验表征确定,该方法实验工作繁琐且耗时耗力,同时,随着生物信息学和测序技术的高速发展,已测序得到的蛋白质序列数量远大于功能获得注释的序列数量,高效预测蛋白质功能变得至关重要。随着计算机技术的蓬勃发展,由数据驱动的机器学习方法已成为应对这些挑战的有效解决方案。本文对蛋白质功能及其注释方法以及机器学习的发展历程和操作流程进行了概述,聚焦于机器学习在酶功能预测领域的应用,对未来人工智能辅助蛋白质功能高效研究的发展方向提出了展望。     

Predicting functional family of novel enzymes irrespective of sequence similarity: a statistical learning approach

The function of a protein that has no sequence homolog of known function is difficult to assign on the basis of sequence similarity. The same problem may arise for homologous proteins of different functions if one is newly discovered and the other is the only known protein of similar sequence. It is desirable to explore methods that are not based on sequence similarity. One approach is to assign functional family of a protein to provide useful hint about its function. Several groups have employed a statistical learning method, support vector machines (SVMs), for predicting protein functional family directly from sequence irrespective of sequence similarity. These studies showed that SVM prediction accuracy is at a level useful for functional family assignment. But its capability for assignment of distantly related proteins and homologous proteins of different functions has not been critically and adequately assessed. Here SVM is tested for functional family assignment of two groups of enzymes. One consists of 50 enzymes that have no homolog of known function from PSI-BLAST search of protein databases. The other contains eight pairs of homologous enzymes of different families. SVM correctly assigns 72% of the enzymes in the first group and 62% of the enzyme pairs in the second group, suggesting that it is potentially useful for facilitating functional study of novel proteins. A web version of our software, SVMProt, is accessible at http://jing.cz3.nus.edu.sg/cgi-bin/svmprot.cgi.     

Prediction of transporter family from protein sequence by support vector machine approach

Transporters play key roles in cellular transport and metabolic processes, and in facilitating drug delivery and excretion. These proteins are classified into families based on the transporter classification (TC) system. Determination of the TC family of transporters facilitates the study of their cellular and pharmacological functions. Methods for predicting TC family without sequence alignments or clustering are particularly useful for studying novel transporters whose function cannot be determined by sequence similarity. This work explores the use of a machine learning method, support vector machines (SVMs), for predicting the family of transporters from their sequence without the use of sequence similarity. A total of 10,636 transporters in 13 TC subclasses, 1914 transporters in eight TC families, and 168,341 nontransporter proteins are used to train and test the SVM prediction system. Testing results by using a separate set of 4351 transporters and 83,151 nontransporter proteins show that the overall accuracy for predicting members of these TC subclasses and families is 83.4% and 88.0%, respectively, and that of nonmembers is 99.3% and 96.6%, respectively. The accuracies for predicting members and nonmembers of individual TC subclasses are in the range of 70.7-96.1% and 97.6-99.9%, respectively, and those of individual TC families are in the range of 60.6-97.1% and 91.5-99.4%, respectively. A further test by using 26,139 transmembrane proteins outside each of the 13 TC subclasses shows that 90.4-99.6% of these are correctly predicted. Our study suggests that the SVM is potentially useful for facilitating functional study of transporters irrespective of sequence similarity.     

Functional Knowledge Transfer for High-accuracy Prediction of Under-studied Biological Processes

A key challenge in genetics is identifying the functional roles of genes in pathways. Numerous functional genomics techniques (e.g. machine learning) that predict protein function have been developed to address this question. These methods generally build from existing annotations of genes to pathways and thus are often unable to identify additional genes participating in processes that are not already well studied. Many of these processes are well studied in some organism, but not necessarily in an investigator''s organism of interest. Sequence-based search methods (e.g. BLAST) have been used to transfer such annotation information between organisms. We demonstrate that functional genomics can complement traditional sequence similarity to improve the transfer of gene annotations between organisms. Our method transfers annotations only when functionally appropriate as determined by genomic data and can be used with any prediction algorithm to combine transferred gene function knowledge with organism-specific high-throughput data to enable accurate function prediction.We show that diverse state-of-art machine learning algorithms leveraging functional knowledge transfer (FKT) dramatically improve their accuracy in predicting gene-pathway membership, particularly for processes with little experimental knowledge in an organism. We also show that our method compares favorably to annotation transfer by sequence similarity. Next, we deploy FKT with state-of-the-art SVM classifier to predict novel genes to 11,000 biological processes across six diverse organisms and expand the coverage of accurate function predictions to processes that are often ignored because of a dearth of annotated genes in an organism. Finally, we perform in vivo experimental investigation in Danio rerio and confirm the regulatory role of our top predicted novel gene, wnt5b, in leftward cell migration during heart development. FKT is immediately applicable to many bioinformatics techniques and will help biologists systematically integrate prior knowledge from diverse systems to direct targeted experiments in their organism of study.     

Functional annotation of proteomic sequences based on consensus of sequence and structural analysis

To maximise the assignment of function of the proteins encoded by a genome and to aid the search for novel drug targets, there is an emerging need for sensitive methods of predicting protein function on a genome-wide basis. GeneAtlas is an automated, high-throughput pipeline for the prediction of protein structure and function using sequence similarity detection, homology modelling and fold recognition methods. GeneAtlas is described in detail here. To test GeneAtlas, a 'virtual' genome was used, a subset of PDB structures from the SCOP database, in which the functional relationships are known. GeneAtlas detects additional relationships by building 3D models in comparison with the sequence searching method PSI-BLAST. Functionally related proteins with sequence identity below the twilight zone can be recognised correctly.     

Reduced alphabet motif methodology for GPCR annotation

Identification and Classification of G-protein coupled receptors (GPCRs) using protein sequences is an important computational challenge, given that experimental screening of thousands of ligands is an expensive proposition. There are two distinct but complementary approaches to GPCR classification --machine learning and sequence motif analysis. Machine learning methodologies typically suffer from problems of class imbalance and lack of multi-class classification. Many sequence motif methods, meanwhile, are too dependent on the similarity of the primary sequence alignments. It is desirable to have a motif discovery and application methodology that is not strongly dependent on primary sequence similarity. It should also overcome limitations of machine learning. We propose and evaluate the effectiveness of a simple methodology that uses a reduced protein functional alphabet representation, where similar functional residues have similar symbols. Regular expression motifs can then be obtained by ClustalW based multiple sequence alignment, using an identity matrix. Since evolutionary matrices like BLOSUM, PAM are not used, this method can be useful for any set of sequences that do not necessarily share a common ancestry. Reduced alphabet motifs can accurately classify known GPCR proteins and the results are comparable to PRINTS and PROSITE. For well known GPCR proteins from SWISSPROT, there were no false negatives and only a few false positives. This methodology covers most currently known classes of GPCRs, even if there are very few representative sequences. It also predicts more than one class for certain sequences, thus overcoming the limitation of machine learning methods. We also annotated, 695 orphan receptors, and 121 were identified as belonging to Family A. A simple JavaScript based web interface has been developed to predict GPCR families and subfamilies (www.insilico-consulting.com/gpcrmotif.html).     

Prediction of functional class of novel bacterial proteins without the use of sequence similarity by a statistical learning method

A substantial percentage of the putative protein-encoding open reading frames (ORFs) in bacterial genomes have no homolog of known function, and their function cannot be confidently assigned on the basis of sequence similarity. Methods not based on sequence similarity are needed and being developed. One method, SVMProt (http://jing.cz3.nus.edu.sg/cgi-bin/svmprot.cgi), predicts protein functional family irrespective of sequence similarity (Nucleic Acids Res. 2003;31:3692-3697). While it has been tested on a large number of proteins, its capability for non-homologous proteins has so far been evaluated for a relatively small number of proteins, and additional tests are needed to more fully assess SVMProt. In this work, 90 novel bacterial proteins (non-homologous to known proteins) are used to evaluate the capability of SVMProt. These proteins are such that none of their homologs are in the Swiss-Prot database, their functions not clearly described in the literature, and they themselves and their homologs are not included in the training sets of SVMProt. They represent proteins whose function cannot be confidently predicted by sequence similarity methods at present. The predicted functional class of 76.7% of each of these proteins shows various levels of consistency with the literature-described function, compared to the overall accuracy of 87% for the SVMProt functional class assignment of 34,582 proteins that have at least one homolog of known function. Our study suggests that SVMProt is capable of assigning functional class for novel bacterial proteins at a level not too much lower than that of sequence alignment methods for homologous proteins.     

CaMELS: In silico prediction of calmodulin binding proteins and their binding sites

Due to Ca²⁺‐dependent binding and the sequence diversity of Calmodulin (CaM) binding proteins, identifying CaM interactions and binding sites in the wet‐lab is tedious and costly. Therefore, computational methods for this purpose are crucial to the design of such wet‐lab experiments. We present an algorithm suite called CaMELS (CalModulin intEraction Learning System) for predicting proteins that interact with CaM as well as their binding sites using sequence information alone. CaMELS offers state of the art accuracy for both CaM interaction and binding site prediction and can aid biologists in studying CaM binding proteins. For CaM interaction prediction, CaMELS uses protein sequence features coupled with a large‐margin classifier. CaMELS models the binding site prediction problem using multiple instance machine learning with a custom optimization algorithm which allows more effective learning over imprecisely annotated CaM‐binding sites during training. CaMELS has been extensively benchmarked using a variety of data sets, mutagenic studies, proteome‐wide Gene Ontology enrichment analyses and protein structures. Our experiments indicate that CaMELS outperforms simple motif‐based search and other existing methods for interaction and binding site prediction. We have also found that the whole sequence of a protein, rather than just its binding site, is important for predicting its interaction with CaM. Using the machine learning model in CaMELS, we have identified important features of protein sequences for CaM interaction prediction as well as characteristic amino acid sub‐sequences and their relative position for identifying CaM binding sites. Python code for training and evaluating CaMELS together with a webserver implementation is available at the URL: http://faculty.pieas.edu.pk/fayyaz/software.html#camels .     

Integrating structure-based machine learning and co-evolution to investigate specificity in plant sesquiterpene synthases

Sesquiterpene synthases (STSs) catalyze the formation of a large class of plant volatiles called sesquiterpenes. While thousands of putative STS sequences from diverse plant species are available, only a small number of them have been functionally characterized. Sequence identity-based screening for desired enzymes, often used in biotechnological applications, is difficult to apply here as STS sequence similarity is strongly affected by species. This calls for more sophisticated computational methods for functionality prediction. We investigate the specificity of precursor cation formation in these elusive enzymes. By inspecting multi-product STSs, we demonstrate that STSs have a strong selectivity towards one precursor cation. We use a machine learning approach combining sequence and structure information to accurately predict precursor cation specificity for STSs across all plant species. We combine this with a co-evolutionary analysis on the wealth of uncharacterized putative STS sequences, to pinpoint residues and distant functional contacts influencing cation formation and reaction pathway selection. These structural factors can be used to predict and engineer enzymes with specific functions, as we demonstrate by predicting and characterizing two novel STSs from Citrus bergamia.     

DNA-binding protein prediction using plant specific support vector machines: validation and application of a new genome annotation tool

There are currently 151 plants with draft genomes available but levels of functional annotation for putative protein products are low. Therefore, accurate computational predictions are essential to annotate genomes in the first instance, and to provide focus for the more costly and time consuming functional assays that follow. DNA-binding proteins are an important class of proteins that require annotation, but current computational methods are not applicable for genome wide predictions in plant species. Here, we explore the use of species and lineage specific models for the prediction of DNA-binding proteins in plants. We show that a species specific support vector machine model based on Arabidopsis sequence data is more accurate (accuracy 81%) than a generic model (74%), and based on this we develop a plant specific model for predicting DNA-binding proteins. We apply this model to the tomato proteome and demonstrate its ability to perform accurate high-throughput prediction of DNA-binding proteins. In doing so, we have annotated 36 currently uncharacterised proteins by assigning a putative DNA-binding function. Our model is publically available and we propose it be used in combination with existing tools to help increase annotation levels of DNA-binding proteins encoded in plant genomes.     

Prediction of RNA-binding proteins from primary sequence by a support vector machine approach

Elucidation of the interaction of proteins with different molecules is of significance in the understanding of cellular processes. Computational methods have been developed for the prediction of protein-protein interactions. But insufficient attention has been paid to the prediction of protein-RNA interactions, which play central roles in regulating gene expression and certain RNA-mediated enzymatic processes. This work explored the use of a machine learning method, support vector machines (SVM), for the prediction of RNA-binding proteins directly from their primary sequence. Based on the knowledge of known RNA-binding and non-RNA-binding proteins, an SVM system was trained to recognize RNA-binding proteins. A total of 4011 RNA-binding and 9781 non-RNA-binding proteins was used to train and test the SVM classification system, and an independent set of 447 RNA-binding and 4881 non-RNA-binding proteins was used to evaluate the classification accuracy. Testing results using this independent evaluation set show a prediction accuracy of 94.1%, 79.3%, and 94.1% for rRNA-, mRNA-, and tRNA-binding proteins, and 98.7%, 96.5%, and 99.9% for non-rRNA-, non-mRNA-, and non-tRNA-binding proteins, respectively. The SVM classification system was further tested on a small class of snRNA-binding proteins with only 60 available sequences. The prediction accuracy is 40.0% and 99.9% for snRNA-binding and non-snRNA-binding proteins, indicating a need for a sufficient number of proteins to train SVM. The SVM classification systems trained in this work were added to our Web-based protein functional classification software SVMProt, at http://jing.cz3.nus.edu.sg/cgi-bin/svmprot.cgi. Our study suggests the potential of SVM as a useful tool for facilitating the prediction of protein-RNA interactions.     

Identification of functional subclasses in the DJ-1 superfamily proteins

Genomics has posed the challenge of determination of protein function from sequence and/or 3-D structure. Functional assignment from sequence relationships can be misleading, and structural similarity does not necessarily imply functional similarity. Proteins in the DJ-1 family, many of which are of unknown function, are examples of proteins with both sequence and fold similarity that span multiple functional classes. THEMATICS (theoretical microscopic titration curves), an electrostatics-based computational approach to functional site prediction, is used to sort proteins in the DJ-1 family into different functional classes. Active site residues are predicted for the eight distinct DJ-1 proteins with available 3-D structures. Placement of the predicted residues onto a structural alignment for six of these proteins reveals three distinct types of active sites. Each type overlaps only partially with the others, with only one residue in common across all six sets of predicted residues. Human DJ-1 and YajL from Escherichia coli have very similar predicted active sites and belong to the same probable functional group. Protease I, a known cysteine protease from Pyrococcus horikoshii, and PfpI/YhbO from E. coli, a hypothetical protein of unknown function, belong to a separate class. THEMATICS predicts a set of residues that is typical of a cysteine protease for Protease I; the prediction for PfpI/YhbO bears some similarity. YDR533Cp from Saccharomyces cerevisiae, of unknown function, and the known chaperone Hsp31 from E. coli constitute a third group with nearly identical predicted active sites. While the first four proteins have predicted active sites at dimer interfaces, YDR533Cp and Hsp31 both have predicted sites contained within each subunit. Although YDR533Cp and Hsp31 form different dimers with different orientations between the subunits, the predicted active sites are superimposable within the monomer structures. Thus, the three predicted functional classes form four different types of quaternary structures. The computational prediction of the functional sites for protein structures of unknown function provides valuable clues for functional classification.     

Machine learning approaches for prediction of linear B-cell epitopes on proteins

Identification and characterization of antigenic determinants on proteins has received considerable attention utilizing both, experimental as well as computational methods. For computational routines mostly structural as well as physicochemical parameters have been utilized for predicting the antigenic propensity of protein sites. However, the performance of computational routines has been low when compared to experimental alternatives. Here we describe the construction of machine learning based classifiers to enhance the prediction quality for identifying linear B-cell epitopes on proteins. Our approach combines several parameters previously associated with antigenicity, and includes novel parameters based on frequencies of amino acids and amino acid neighborhood propensities. We utilized machine learning algorithms for deriving antigenicity classification functions assigning antigenic propensities to each amino acid of a given protein sequence. We compared the prediction quality of the novel classifiers with respect to established routines for epitope scoring, and tested prediction accuracy on experimental data available for HIV proteins. The major finding is that machine learning classifiers clearly outperform the reference classification systems on the HIV epitope validation set.     

A top-down approach to classify enzyme functional classes and sub-classes using random forest

Advancements in sequencing technologies have witnessed an exponential rise in the number of newly found enzymes. Enzymes are proteins that catalyze bio-chemical reactions and play an important role in metabolic pathways. Commonly, function of such enzymes is determined by experiments that can be time consuming and costly. Hence, a need for a computing method is felt that can distinguish protein enzyme sequences from those of non-enzymes and reliably predict the function of the former. To address this problem, approaches that cluster enzymes based on their sequence and structural similarity have been presented. But, these approaches are known to fail for proteins that perform the same function and are dissimilar in their sequence and structure. In this article, we present a supervised machine learning model to predict the function class and sub-class of enzymes based on a set of 73 sequence-derived features. The functional classes are as defined by International Union of Biochemistry and Molecular Biology. Using an efficient data mining algorithm called random forest, we construct a top-down three layer model where the top layer classifies a query protein sequence as an enzyme or non-enzyme, the second layer predicts the main function class and bottom layer further predicts the sub-function class. The model reported overall classification accuracy of 94.87% for the first level, 87.7% for the second, and 84.25% for the bottom level. Our results compare very well with existing methods, and in many cases report better performance. Using feature selection methods, we have shown the biological relevance of a few of the top rank attributes.     

Prediction of Saccharomyces cerevisiae protein functional class from functional domain composition

MOTIVATION: A key goal of genomics is to assign function to genes, especially for orphan sequences. RESULTS: We compared the clustered functional domains in the SBASE database to each protein sequence using BLASTP. This representation for a protein is a vector, where each of the non-zero entries in the vector indicates a significant match between the sequence of interest and the SBASE domain. The machine learning methods nearest neighbour algorithm (NNA) and support vector machines are used for predicting protein functional classes from this information. We find that the best results are found using the SBASE-A database and the NNA, namely 72% accuracy for 79% coverage. We tested an assigning function based on searching for InterPro sequence motifs and by taking the most significant BLAST match within the dataset. We applied the functional domain composition method to predict the functional class of 2018 currently unclassified yeast open reading frames. AVAILABILITY: A program for the prediction method, that uses NNA called Functional Class Prediction based on Functional Domains (FCPFD) is available and can be obtained by contacting Y.D.Cai at y.cai@umist.ac.uk     

MiRANN: a reliable approach for improved classification of precursor microRNA using Artificial Neural Network model

MicroRNA (miRNA) is a special class of short noncoding RNA that serves pivotal function of regulating gene expression. The computational prediction of new miRNA candidates involves various methods such as learning methods and methods using expression data. This article has proposed a reliable model - miRANN which is a supervised machine learning approach. MiRANN used known pre-miRNAs as positive set and a novel negative set from human CDS regions. The number of known miRNAs is now huge and diversified that could cover almost all characteristics of unknown miRNAs which increases the quality of the result (99.9% accuracy, 99.8% sensitivity, 100% specificity) and provides a more reliable prediction. MiRANN performs better than other state-of-the-art approaches and declares to be the most potential tool to predict novel miRNAs. We have also tested our result using a previous negative set. MiRANN, opens new ground using ANN for predicting pre-miRNAs with a promise of better performance.