Spark-IDPP: high-throughput and scalable prediction of intrinsically disordered protein regions with Spark clusters on the Cloud

Determination of 3D protein structures became an important part of protein bioinformatics, since the knowledge of 3D protein structures allows to draw conclusions about molecular mechanisms of cellular biochemical reactions or particular diseases, and it supports drug design. However, as was gradually observed from 1990s, not all known proteins have stable (ordered) native 3D structure. Some proteins have shorter or longer segments that indicate instability or flexibility, which are called intrinsically disordered regions (IDRs). Such proteins are usually known as intrinsically disordered proteins (IDPs), and they may consist of one or several IDRs, or, in extreme case, can be completely unstructured (intrinsically unstructured proteins, IUPs). Intrinsically disordered proteins play many important roles in cells of living organisms, and on the basis of various studies carried out by scientists on whole proteomes, it is estimated that the percentage of IDPs in mammals is very large [16]. Determination of 3D structures of intrinsically disordered proteins with traditional methods, like the X-ray crystallography or Nuclear Magnetic Resonance (NMR), is difficult, since, e.g., the lack of electron density in crystal structures (which is marked in PDB files describing protein macromolecules as REMARK465 record). For this reason, IDP predictors have become playing an important role in determination of unstructured regions. IDP prediction became an important part of the computational protein structure determination. It supports studies of protein structural features, functional analysis of proteins, and investigations of the relationships between IDRs and the occurrence of particular diseases. As strictly computational tools, IDP predictors are able to predict protein disorder from pure amino acid sequence. However, taking into account the number of protein amino acid sequences in the UniProtKB/Swiss-Prot database, it is essential to provide efficient and effective tools for IDP prediction. These tools should be able to scale the computational procedure in order to accommodate the growing volume of DNA, and consequently, protein sequences.

With the growing volume of biological data describing various aspects of living forms, increasing demand for fast data analysis (velocity), and a variety of data formats, structural bioinformatics has been facing the Big Data challenge. This requires changing the way how we cope with the biological data and redevelopment of existing tools for data analysis and processing, thus directing us to Big Data ecosystems. Apache Hadoop with MapReduce processing model and Apache Spark are both popular big data frameworks. However, since the Spark allows for running most computations in memory, without the necessity to store intermediate results in a file system, it provides better performance for many applications [88]. Therefore, in order to meet the requirements for fast IDP predictions, we decided to use Apache Spark for the development of the efficient IDP meta-predictor described in the paper.

Apache Spark [88] is a platform for large-scale data processing and general-purpose computing. Spark was generally created to run computations on computer clusters allowing to distribute computations over the cluster nodes. Spark divides the computational job into a number of tasks and executes the tasks on Worker nodes of the cluster within processes called executors. In addition to performing tasks, executors are responsible for keeping data in memory or disk storage across tasks.

Operational data that must be processed on Spark cluster are delivered as resilient distributed dataset (RDD). RDD [87] is a fault-tolerant collection of elements that Spark uses to operate on in parallel. RDDs can be created by parallelizing an already existing collection through a transformation or by referencing a data set in an external storage system, such as Hadoop Distributed File System (HDFS). Both methods are used in our solution presented in the paper. Data delivered in the form of RDDs are divided into a number of partitions. Spark then adjusts the number of tasks sent for execution to the number of partitions the RDD collection is cut into. RDDs allow to store processing results in memory in any stage of data processing, e.g., between various transformations. Transformations allow to perform various operations on data and create new data sets (RDDs) from existing ones.

Spark can be deployed in several modes, including private clusters and clusters located on public clouds. The latter solution enables flexible scaling the Spark cluster on-demand in order to accommodate data growth, since the Cloud allows users to access and use configurable computing resources (e.g., servers, storage, applications, networks) as a service [49] without the necessity to build entire infrastructure that supports the resources on premises. The Spark cluster can be hosted on resources provisioned by a cloud provider. This allows to raise the computing power and was applied in our solution presented in the paper. We used Microsoft Azure cloud platform in order to quickly scale the system performing predictions of IDPs on large scale.

Accurate and efficient prediction of disordered regions in protein structures on the basis of pure amino acid sequence is one of the challenges in computational biology and structural genomics. Existing methods for IDP prediction rely on the analysis of various features of proteins, e.g., protein amino acid composition or physical-chemical characteristics of particular amino acids. Some of the methods are grounded in statistical observations, other use machine learning algorithms. For example, GlobPlot [40] is a simple prediction method that identifies regions of globularity and disorder within protein sequences on the basis of propensities of particular amino acids to be in globular or non-globular states. GlobPlot calculates a running sum of the propensity for amino acids to be in an ordered or disordered state, and uses the sum to classify regions of the given amino acid sequence to a particular class. IUPred [15] is also footed on the physical foundations and statistical observations of inter-residue interactions in proteins structures. To discriminate between ordered and disordered class, the IUPred uses statistical interaction potentials. On the other hand, DisEMBL [39] uses trained Artificial Neural Networks (ANNs) to predict protein disorders. There are three variants of the method: (1) Coils used to predict classic loops and coils as defined by DSSP [29], (2) Hot loops for prediction of flexible loops with a high degree of mobility determined by temperature factors (B-factors), (3) Remark 465 that predicts missing coordinates in X-ray structures, as defined by REMARK465 records in the PDB files describing macromolecular structures of proteins. DISpro [10] uses evolutionary information in the form of protein profiles, predicted secondary structure and relative solvent accessibility, and ensembles of 1D-recursive neural networks. RONN [85] classifies disordered regions on the basis of observed homology between protein sequences. The homology is expressed in calculated alignment scores, while comparing given protein sequences with a series of amino acid sequences of known folding state (ordered, disordered, or a mixture of both). Obtained alignment scores are then used in the prediction process performed by suitably trained regional order neural network. SPINE-D [89] classifies residues into three classes, as structurally ordered, disordered, and semi-ordered. The method also uses a trained artificial neural network. Several methods, including DISOPRED2 [78], Poodle-s [66], Poodle-l [21], PrDOS [26], and Spritz [73], use SVM-based classifiers to predict disordered regions on the basis of various features extracted from protein sequences. For example, Poodle-s, Poodle-l, and PrDOS perform prediction on the basis of position-specific scoring matrices (PSSMs) generated by PSI-BLAST [1], while Spritz uses amino acid frequencies in disordered regions. PSSMs with respect to physicochemical properties of amino acids are also used in iPDA server [68] and its underlying DisPSSMP predictor.

Recent works in this area, including Xue et al. [83] and Kozlowski and Bujnicki [33], show that meta-prediction with the use of ensembles composed of subsets of described basic predictors may improve the quality of prediction results. The method presented in our paper also works on the basis of ensemble of basic predictors. However, so far, none of the prediction or meta-prediction methods was designed to perform large-scale identification of disordered regions and to deal with large volumes of protein sequences.

Meanwhile, there are real-life problems, in which massive parallelization of computations on Apache Hadoop or Spark, and the use of scalable environments, like the Cloud, brought significant improvements in performance of data processing and analysis. Big data challenge was observed and solved in various works devoted to intelligent transport and smart cities [11, 19, 42, 43, 74, 75, 84], water monitoring [12, 22, 90], social networks analysis [13, 14, 77], multimedia processing [72, 82], internet of things (IoT) [9], social media monitoring [50], Life sciences [3, 31, 32, 44, 58, 69] and disease data analysis [6, 45, 81], telecommunication [27], and finance [2], to mention just a few. Many hot issues in various sub-fields of bioinformatics were also solved with the use of Big Data ecosystems and Cloud computing, e.g., mapping next-generation sequence data to the human genome and other reference genomes, for use in a variety of biological analyzes including SNP discovery, genotyping and personal genomics [65], sequence analysis and assembly [17, 30, 34, 35, 47, 62], multiple alignments of DNA and RNA sequences [86, 91], codon analysis with local MapReduce aggregations [63], NGS data analysis [8], phylogeny [24, 48], proteomics [37], analysis of protein-ligand binding sites [23], and others. Regarding the analysis of 3D protein structures, it is worth mentioning several works, including Hazelhurst et al. [20] and Małysiak-Mrozek et al. [46] devoted to exploration of various atomic interactions within protein structures, works of Che-Lun Hung and Yaw-Ling Lin [25], and Mrozek et al. [51, 53, 55‐57], devoted to comparison and alignment of 3D protein structures, and cloud-based system for 3D protein structure modeling presented in [54]. However, none of the mentioned works was focused on prediction of disordered regions.

In this paper, we present an ensemble method for prediction of intrinsically disordered proteins implemented in our IDP meta-predictor (IDPP, Intrinsically Disordered Proteins meta-Predictor, Sect. 2.1). We also show the implementation of the method on the Spark cluster, called Spark-IDPP (Sect. 2.4). Spark-IDPP allows for large-scale prediction of intrinsically disordered proteins or intrinsically disordered regions of protein structures on the basis of amino acid sequences. The IDPP classifies disordered regions on the basis of consensus of votes cast by component, basic predictors (Sect. 2.2). In the paper, we examine four consensus modes in terms of the effectiveness of prediction showing that the use of ensemble methods leads to the increased effectiveness (Sect. 3.4). The method is implemented on the Apache Spark, hence Spark-IDPP, in order to provide significantly better performance for large volumes of sequence data that are processed and analyzed in protein bioinformatics. In this paper, we also present results of performance tests on scaling out the Spark cluster on the Microsoft Azure cloud and increasing the degree of parallelism with the intention of improving efficiency of performed predictions (Sect. 3.5).

IDPP meta-predictor performs prediction of intrinsically disordered proteins or disordered regions for the given protein amino acid sequences. General architecture of the IDPP meta-predictor is presented in Fig. 2. It consists of the following modules:

The Component Predictors Module (CPM) is responsible for prediction of disordered regions. Prediction can be performed for many protein sequences provided in the FASTA format [41] at the input of the CPM. A sample input sequence in the FASTA format is presented in Listing 1. First line of each entry in the FASTA file consists of descriptive information, e.g., identifiers of the protein sequence in various public repositories. Next lines consist of amino acid sequence of the protein.

The Component Predictors Module contains several basic predictors for intrinsically disordered regions. Each of the component predictors accepts an amino acid sequence at the input and performs independent predictions of intrinsically disorder regions on the basis of the input sequence. The CPM module of the IDPP meta-predictor implements the following basic prediction algorithms:Each of the component predictors generates its results containing the list of regions which are supposed to be disordered. These results contain the information on the probability that each amino acid belongs to the disordered region or not, binary classification of belonging to the disordered region (1—belongs, 0—does not belong), and lists of disordered regions expressed as ranges of amino acid positions.

The Consensus Module (CoM) aggregates results from component predictors on the basis of consensus. There are four consensus approaches (also called consensus modes) implemented and tested in the Consensus Module: two operating on binary classification from component predictors and two operating on float values of the returned probability. All will be described in details in Sect. 2.2. In the next phase, the CoM performs filtering of the consensus results with the use of fuzzy smoothing filter in order to remove probable outliers. After filtering out single disordered positions, the Consensus Module finally classifies particular amino acids in the protein sequence as belonging to a disorder region or not, on the basis of aggregated probabilities or binary classification results. The final classification is performed for the given cutoff threshold obtained experimentally.

The Analysis Module (AM) allows to assess the quality of classification and to find potential cutoff thresholds for component predictors. The AM allows to analyze results of the prediction process and compare them to a ground truth (actual disordered regions that were discovered experimentally or collected manually from literature, stored in the DisProt database [59]). Results of the analyzes are provided in the form of True Positive, True Negative, False Positive, and False Negative rates.

Results of the prediction process are returned as text files, saved on hard disk drive, or on the standard output of the IDPP meta-predictor. For example, a result of a prediction of disordered regions for a single amino acid sequence of protein Cyclin-dependent kinase inhibitor 1B from Homo sapiens (accession number: P46527, entry name: CDN1B_HUMAN in the UniProtKB/Swiss-Prot database) is shown in Listing 2. First line of the result contains the information that allows to identify the protein. The second line contains the information on disordered regions identified in the sequence, expressed as ranges of amino acid positions.

IDPP meta-predictor consists of several component predictors that classify amino acids as belonging to a disordered region or not. These predictors work on the basis of various algorithms, and they have various prediction effectiveness. However, we assumed that the aggregation of their votes may give improvements in the prediction effectiveness. In our IDPP meta-predictor, we applied the following four consensus modes—two operating on results of binary classification from seven component predictors (RONN, IUPred in two variants, GlobProt, and DisEMBL in three variants) and two operating on float values of the returned probabilities of belonging to a disordered region (as presented in Fig. 3):

The Simple Binary (SB) consensus mode makes use of only binary scores returned by all component predictors (0/1 classification), and votes of all component predictors are equally important while aggregating component decisions (equal weights). In this consensus mode, the IDPP meta-predictor evaluates votes of component predictors and aggregates ranks for each ith residue in the protein amino acid sequence according to the following formula:where J is the set of basic component predictors (RONN, GlobProt, IUPred short, IUPred long, DisEMBL Hotloops, DisEMBL Remark465, DisEMBL Coils), \(Score^{Bin}_{i,j}\) is the binary score returned by the jth component predictor for the ith residue (1—belongs to a disordered region, 0—does not belong), and \(w_j=1.0\) are weights of importance of particular component predictors (\(w_j=0.0\), if we want to eliminate the jth predictor from voting).

Likewise the Simple Binary consensus mode, the Weighted Binary (WB) consensus mode uses only binary scores returned by all component predictors (0/1 classification), but votes of all component predictors have different weights while aggregating component decisions. Weights of importance for particular component predictors (\(w_j\) in Eq. 1) are calculated according to the following formula:where: \(W_{disorder}\) equals the fraction of disordered residues, \(W_{order}\) equals the fraction of ordered residues, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives. The \(S_{W,j}\) weighted score rewards a correct disorder prediction higher than a correct order prediction [28]. This is done to avoid over-prediction of an ordered state due the fact that ordered regions are more common in known proteins. Values of the \(S_{W,j}\) weighted score for particular component predictors of the IDPP were obtained experimentally (see Table 3 in Sect. 3.4).

In both binary consensus modes, particular amino acids in the protein sequence (i) are then pre-classified, whether they belong to a disordered region or not:where \(preScore^{IDPP-Bin}_i\) is the score aggregated in one of the binary consensus modes (Eq. 1), and \(\lambda _0\) is a qualification threshold determined experimentally to 0.5.

The Simple Float (SF) consensus mode makes use of probabilities returned by all component predictors for each residue of the amino acid chain. Votes of all component predictors are equally important while aggregating component decisions (equal weights, like in the Simple Binary consensus mode). In this consensus mode, the IDPP meta-predictor evaluates votes of component predictors and aggregates probabilities for each ith residue in the protein amino acid sequence according to the following formula:where J is the set of basic component predictors, \(Prob_{i,j}\) is the probability returned by the jth component predictor for the ith residue that the residue belongs to a disordered region, and \(w_j=1.0\) are weights of importance of particular component predictors (\(w_j=0.0\), if we want to eliminate the jth predictor from voting).

Similarly, the Weighted Float (WF) consensus mode works on the basis of probabilities that a residue belongs to a disordered region, calculated by component predictors. However, while aggregating component decisions, votes of all component predictors are weighted according to the \(S_{W,j}\) score (Eq. 2).

In contrast, to binary consensus modes, results of float consensus modes are not pre-classified, but fuzzy filtering in the next phase is performed directly on aggregated probabilities \(preScore^{IDPP-Flo}_i\).

Disordered regions are usually formed by many successive residues in the protein chain, rather than single amino acids. Therefore, single amino acids or short segments predicted as disordered regions separated by short segments of ordered regions should be eliminated from the final result. To this purpose, we implemented fuzzy smoothing filter that discards such small disordered “islands”. The filter runs through the sequence of classes (0/1) or probabilities each residue in protein sequence was assigned in the Consensus Module, and replaces each entry by a new probability value. The new probability value for the ith residue is calculated on the basis of classes \(preClass^{IDPP-Bin}_i\) (for binary consensus modes):or scores \(preScore^{IDPP-Flo}_i\) (for float consensus modes):for neighboring residues. The pattern of neighboring residues is called the window, and slides residue by residue, over the entire sequence of amino acids. In Eqs. 5 and 6 the window size is equal to 5. For the ith residue, the 5-residue window consists of the residues located at absolute positions \(i-2,i-1,i,i+1,i+2\) in the protein amino acid chain. The k iterates through neighboring residues (points to absolute positions) in the sliding window with respect to the ith residue being processed, l transforms the absolute position of the kth residue to the position in the sliding window (\(l=1..5\)), and \(\mu (l)\) is the weight for the lth residue in the 5-residue sliding window. The \(\mu (l)\) weight is calculated as it is presented in Fig. 4. Such a fuzzy set defined by a membership function allows assigning lower weights for the first two residues in the sliding window. In consequence, this weakens the influence of the first two elements (elements preceding the current ith element in a sequence), which is important when switching between classes. If the following elements \(i+1\) and \(i+2\) are of the same class as the ith element, the influence of the preceding two elements \(i-1\) and \(i-2\), especially if one of them is of a different IDP class than the ith element, decreases with the distance from the current element. In other words, this fuzzy set models the uncertainty at the border of regions belonging to two different classes.

With the position-specific weight function as defined in Fig. 4, we can notice that the following condition always holds:For positions in the sliding window that go beyond the beginning or end of the protein sequence, we assume that the IDPP meta-predictor assigns the class of 1, i.e., disordered, since residues located at each end of proteins sequences are, on average, more frequently disordered than residues located in the middle of the protein chain.

Values of the new probability calculated for each residue in the protein sequence are then compared to the experimentally determined cutoff threshold \(\lambda\) (default value of \(\lambda\) is 0.5). The ith residue is classified as disordered, if the probability is greater or equal to the \(\lambda\) threshold:After this step, each residue is finally classified as ordered or disordered, and the Consensus Module returns a list of disordered regions accompanied by the sequence of disorder probabilities, for each protein sequence provided on the input of the IDPP meta-predictor. An example of the filtering applied on a sequence of classes produced in binary classification for a sample protein from the DisProt database is shown in Fig. 5.

Spark-IDP meta-predictor (Spark-IDPP) performs predictions of disordered regions on large scale by parallelizing computations on the Apache Spark cluster located on Microsoft Azure cloud. The Spark-IDPP was created and tested on Spark 1.6.1 working on Linux platform within Microsoft Azure HDInsight service HDI 3.4.

The Spark cluster used for most of the performed experiments was established on the Microsoft Azure public cloud as the HDInsight service (HDI 3.4) hosted on D13v2-sized virtual machines (VMs) with Linux operating system. Virtual machines in the Dv2-series are intended for applications that demand faster CPUs, better temporary storage performance, or have higher memory demands. The whole series is based on the 2.4 GHz Intel Xeon E5-2673 v3 (Haswell) processor, and with the Intel Turbo Boost Technology 2.0, can go up to 3.1 GHz. D13v2-sized virtual machines used by Worker nodes of the Spark cluster had 8 virtual CPUs, 56GB RAM, and 400GB of temporal storage space on SSD drives. Such a prepared Spark cluster was used to run parallel procedures of the Spark-IDPP meta-predictor and carry out assumed performance experiments.

During our experiments we used two databases of protein sequences—one for testing effectiveness of the Spark-IDPP meta-predictor and other basic predictors, and another data set for testing efficiency of the Spark-IDPP. While testing effectiveness of the Spark-IDPP meta-predictor, we used the DisProt data set [59, 67]. The data set that we used contained 1539 disordered regions located in 694 proteins. These regions were discovered with the use of experimental methods, and evidences of disorder were manually collected from the literature. The DisProt data set provided a “ground truth” while testing effectiveness of all primary predictors and the Spark-IDPP and allowed us to calculate weighted scores \(S_W\) for basic predictors in all consensus modes.

The second database, UniProtKB [71], was used to test efficiency of the Spark-IDPP. This database provides millions of protein amino acid sequences and was used due to its large size. In our experiments we used various subsets of the database.

Our experiments involved testing the quality of predictions of disordered regions and efficiency of the proposed meta-predictor on the Spark cluster. The quality of predictions was tested for four consensus modes in order to select the best one. Then, we tested the performance of the meta-predictor working on the Spark cluster for the selected consensus mode. While testing performance, we first experimentally established how to divide data into data chunks and what number and size of data chunks should be used in order to minimize the execution time for the current cluster configuration (the number of nodes). The most efficient division of data into data chunks was used in the following experiments. Afterward, we changed the cluster size for the same amount of input data in order to verify scalability of the solution. And finally, we verified efficiency and speedup achieved by the system built on the 32-node Spark cluster (cluster size was constant) for the growing volume of data.

Effectiveness of the prediction performed with the use of the Spark-IDPP meta-predictor was verified in a series of tests and compared to the effectiveness of component predictors. All component predictors and the Spark-IDPP meta-predictor were tested on the DisProt data set ver. 6.02. This data set contains protein sequences together with the information on the experimentally identified disordered regions. We had to exclude three protein sequences from the data set due to errors generated by particular component predictors: DP00642 and DP00651 (errors generated by the DisEMBL), and DP00195 (not fulfilling the conditions of RONN).

We conducted a broad series of tests in order to verify performance of the proposed Spark-IDPP method. In those tests we investigated the execution time and speedup achieved for various configurations of the Spark cluster and various data load.