SPINE-D: accurate prediction of short and long disordered regions by a single neural-network based method

Short and long disordered regions of proteins have different preference for different amino acid residues. Different methods often have to be trained to predict them separately. In this study, we developed a single neural-network-based technique called SPINE-D that makes a three-state prediction first (ordered residues and disordered residues in short and long disordered regions) and reduces it into a two-state prediction afterwards. SPINE-D was tested on various sets composed of different combinations of Disprot annotated proteins and proteins directly from the PDB annotated for disorder by missing coordinates in X-ray determined structures. While disorder annotations are different according to Disprot and X-ray approaches, SPINE-D's prediction accuracy and ability to predict disorder are relatively independent of how the method was trained and what type of annotation was employed but strongly depend on the balance in the relative populations of ordered and disordered residues in short and long disordered regions in the test set. With greater than 85% overall specificity for detecting residues in both short and long disordered regions, the residues in long disordered regions are easier to predict at 81% sensitivity in a balanced test dataset with 56.5% ordered residues but more challenging (at 65% sensitivity) in a test dataset with 90% ordered residues. Compared to eleven other methods, SPINE-D yields the highest area under the curve (AUC), the highest Mathews correlation coefficient for residue-based prediction, and the lowest mean square error in predicting disorder contents of proteins for an independent test set with 329 proteins. In particular, SPINE-D is comparable to a meta predictor in predicting disordered residues in long disordered regions and superior in short disordered regions. SPINE-D participated in CASP 9 blind prediction and is one of the top servers according to the official ranking. In addition, SPINE-D was examined for prediction of functional molecular recognition motifs in several case studies.     

SPINE-D: Accurate Prediction of Short and Long Disordered Regions by a Single Neural-Network Based Method

Abstract

Short and long disordered regions of proteins have different preference for different amino acid residues. Different methods often have to be trained to predict them separately. In this study, we developed a single neural-network-based technique called SPINE-D that makes a three-state prediction first (ordered residues and disordered residues in short and long disordered regions) and reduces it into a two-state prediction afterwards. SPINE-D was tested on various sets composed of different combinations of Disprot annotated proteins and proteins directly from the PDB annotated for disorder by missing coordinates in X-ray determined structures. While disorder annotations are different according to Disprot and X-ray approaches, SPINE-D's prediction accuracy and ability to predict disorder are relatively independent of how the method was trained and what type of annotation was employed but strongly depend on the balance in the relative populations of ordered and disordered residues in short and long disordered regions in the test set. With greater than 85% overall specificity for detecting residues in both short and long disordered regions, the residues in long disordered regions are easier to predict at 81% sensitivity in a balanced test dataset with 56.5% ordered residues but more challenging (at 65% sensitivity) in a test dataset with 90% ordered residues. Compared to eleven other methods, SPINE-D yields the highest area under the curve (AUC), the highest Mathews correlation coefficient for residue-based prediction, and the lowest mean square error in predicting disorder contents of proteins for an independent test set with 329 proteins. In particular, SPINE-D is comparable to a meta predictor in predicting disordered residues in long disordered regions and superior in short disordered regions. SPINE-D participated in CASP 9 blind prediction and is one of the top servers according to the official ranking. In addition, SPINE-D was examined for prediction of functional molecular recognition motifs in several case studies. The server and databases are available at http://sparks.informatics.iupui.edu/.     

Statistical analysis of unstructured amino acid residues in protein structures

We have performed a statistical analysis of unstructured amino acid residues in protein structures available in the databank of protein structures. Data on the occurrence of disordered regions at the ends and in the middle part of protein chains have been obtained: in the regions near the ends (at distance less than 30 residues from the N- or C-terminus), there are 66% of unstructured residues (38% are near the N-terminus and 28% are near the C-terminus), although these terminal regions include only 23% of the amino acid residues. The frequencies of occurrence of unstructured residues have been calculated for each of 20 types in different positions in the protein chain. It has been shown that relative frequencies of occurrence of unstructured residues of 20 types at the termini of protein chains differ from the ones in the middle part of the protein chain; amino acid residues of the same type have different probabilities to be unstructured in the terminal regions and in the middle part of the protein chain. The obtained frequencies of occurrence of unstructured residues in the middle part of the protein chain have been used as a scale for predicting disordered regions from amino acid sequence using the method (FoldUnfold) previously developed by us. This scale of frequencies of occurrence of unstructured residues correlates with the contact scale (previously developed by us and used for the same purpose) at a level of 95%. Testing the new scale on a database of 427 unstructured proteins and 559 completely structured proteins has shown that this scale can be successfully used for the prediction of disordered regions in protein chains.     

Length-dependent prediction of protein intrinsic disorder

Background  

Due to the functional importance of intrinsically disordered proteins or protein regions, prediction of intrinsic protein disorder from amino acid sequence has become an area of active research as witnessed in the 6th experiment on Critical Assessment of Techniques for Protein Structure Prediction (CASP6). Since the initial work by Romero et al. (Identifying disordered regions in proteins from amino acid sequences, IEEE Int. Conf. Neural Netw., 1997), our group has developed several predictors optimized for long disordered regions (>30 residues) with prediction accuracy exceeding 85%. However, these predictors are less successful on short disordered regions (≤30 residues). A probable cause is a length-dependent amino acid compositions and sequence properties of disordered regions.     

Predicting disordered regions in proteins based on decision trees of reduced amino acid composition.

Intrinsically unstructured proteins (IUPs) are proteins lacking a fixed three dimensional structure or containing long disordered regions. IUPs play an important role in biology and disease. Identifying disordered regions in protein sequences can provide useful information on protein structure and function, and can assist high-throughput protein structure determination. In this paper we present a system for predicting disordered regions in proteins based on decision trees and reduced amino acid composition. Concise rules based on biochemical properties of amino acid side chains are generated for prediction. Coarser information extracted from the composition of amino acids can not only improve the prediction accuracy but also increase the learning efficiency. In cross-validation tests, with four groups of reduced amino acid composition, our system can achieve a recall of 80% at a 13% false positive rate for predicting disordered regions, and the overall accuracy can reach 83.4%. This prediction accuracy is comparable to most, and better than some, existing predictors. Advantages of our approach are high prediction accuracy for long disordered regions and efficiency for large-scale sequence analysis. Our software is freely available for academic use upon request.     

PONDR-FIT: A meta-predictor of intrinsically disordered amino acids

Protein intrinsic disorder is becoming increasingly recognized in proteomics research. While lacking structure, many regions of disorder have been associated with biological function. There are many different experimental methods for characterizing intrinsically disordered proteins and regions; nevertheless, the prediction of intrinsic disorder from amino acid sequence remains a useful strategy especially for many large-scale proteomic investigations. Here we introduced a consensus artificial neural network (ANN) prediction method, which was developed by combining the outputs of several individual disorder predictors. By eight-fold cross-validation, this meta-predictor, called PONDR-FIT, was found to improve the prediction accuracy over a range of 3 to 20% with an average of 11% compared to the single predictors, depending on the datasets being used. Analysis of the errors shows that the worst accuracy still occurs for short disordered regions with less than ten residues, as well as for the residues close to order/disorder boundaries. Increased understanding of the underlying mechanism by which such meta-predictors give improved predictions will likely promote the further development of protein disorder predictors. Access to PONDR-FIT is available at www.disprot.org.     

iPTREE-STAB: interpretable decision tree based method for predicting protein stability changes upon mutations

We have developed a web server, iPTREE-STAB for discriminating the stability of proteins (stabilizing or destabilizing) and predicting their stability changes (delta deltaG) upon single amino acid substitutions from amino acid sequence. The discrimination and prediction are mainly based on decision tree coupled with adaptive boosting algorithm, and classification and regression tree, respectively, using three neighboring residues of the mutant site along N- and C-terminals. Our method showed an accuracy of 82% for discriminating the stabilizing and destabilizing mutants, and a correlation of 0.70 for predicting protein stability changes upon mutations. AVAILABILITY: http://bioinformatics.myweb.hinet.net/iptree.htm. SUPPLEMENTARY INFORMATION: Dataset and other details are given.     

Flavors of protein disorder

Intrinsically disordered proteins are characterized by long regions lacking 3-D structure in their native states, yet they have been so far associated with 28 distinguishable functions. Previous studies showed that protein predictors trained on disorder from one type of protein often achieve poor accuracy on disorder of proteins of a different type, thus indicating significant differences in sequence properties among disordered proteins. Important biological problems are identifying different types, or flavors, of disorder and examining their relationships with protein function. Innovative use of computational methods is needed in addressing these problems due to relative scarcity of experimental data and background knowledge related to protein disorder. We developed an algorithm that partitions protein disorder into flavors based on competition among increasing numbers of predictors, with prediction accuracy determining both the number of distinct predictors and the partitioning of the individual proteins. Using 145 variously characterized proteins with long (>30 amino acids) disordered regions, 3 flavors, called V, C, and S, were identified by this approach, with the V subset containing 52 segments and 7743 residues, C containing 39 segments and 3402 residues, and S containing 54 segments and 5752 residues. The V, C, and S flavors were distinguishable by amino acid compositions, sequence locations, and biological function. For the sequences in SwissProt and 28 genomes, their protein functions exhibit correlations with the commonness and usage of different disorder flavors, suggesting different flavor-function sets across these protein groups. Overall, the results herein support the flavor-function approach as a useful complement to structural genomics as a means for automatically assigning possible functions to sequences.     

Disorder predictors also predict backbone dynamics for a family of disordered proteins

Several algorithms have been developed that use amino acid sequences to predict whether or not a protein or a region of a protein is disordered. These algorithms make accurate predictions for disordered regions that are 30 amino acids or longer, but it is unclear whether the predictions can be directly related to the backbone dynamics of individual amino acid residues. The nuclear Overhauser effect between the amide nitrogen and hydrogen (NHNOE) provides an unambiguous measure of backbone dynamics at single residue resolution and is an excellent tool for characterizing the dynamic behavior of disordered proteins. In this report, we show that the NHNOE values for several members of a family of disordered proteins are highly correlated with the output from three popular algorithms used to predict disordered regions from amino acid sequence. This is the first test between an experimental measure of residue specific backbone dynamics and disorder predictions. The results suggest that some disorder predictors can accurately estimate the backbone dynamics of individual amino acids in a long disordered region.     

Protein flexibility and intrinsic disorder

Comparisons were made among four categories of protein flexibility: (1) low-B-factor ordered regions, (2) high-B-factor ordered regions, (3) short disordered regions, and (4) long disordered regions. Amino acid compositions of the four categories were found to be significantly different from each other, with high-B-factor ordered and short disordered regions being the most similar pair. The high-B-factor (flexible) ordered regions are characterized by a higher average flexibility index, higher average hydrophilicity, higher average absolute net charge, and higher total charge than disordered regions. The low-B-factor regions are significantly enriched in hydrophobic residues and depleted in the total number of charged residues compared to the other three categories. We examined the predictability of the high-B-factor regions and developed a predictor that discriminates between regions of low and high B-factors. This predictor achieved an accuracy of 70% and a correlation of 0.43 with experimental data, outperforming the 64% accuracy and 0.32 correlation of predictors based solely on flexibility indices. To further clarify the differences between short disordered regions and ordered regions, a predictor of short disordered regions was developed. Its relatively high accuracy of 81% indicates considerable differences between ordered and disordered regions. The distinctive amino acid biases of high-B-factor ordered regions, short disordered regions, and long disordered regions indicate that the sequence determinants for these flexibility categories differ from one another, whereas the significantly-greater-than-chance predictability of these categories from sequence suggest that flexible ordered regions, short disorder, and long disorder are, to a significant degree, encoded at the primary structure level.     

Reduced amino acid alphabet is sufficient to accurately recognize intrinsically disordered protein

Intrinsically disordered proteins are an important class of proteins with unique functions and properties. Here, we have applied a support vector machine (SVM) trained on naturally occurring disordered and ordered proteins to examine the contribution of various parameters (vectors) to recognizing proteins that contain disordered regions. We find that a SVM that incorporates only amino acid composition has a recognition accuracy of 87+/-2%. This result suggests that composition alone is sufficient to accurately recognize disorder. Interestingly, SVMs using reduced sets of amino acids based on chemical similarity preserve high recognition accuracy. A set as small as four retains an accuracy of 84+/-2%; this suggests that general physicochemical properties rather than specific amino acids are important factors contributing to protein disorder.     

Identification of RNA-binding sites in proteins by integrating various sequence information

RNA–protein interactions play a pivotal role in various biological processes, such as mRNA processing, protein synthesis, assembly, and function of ribosome. In this work, we have introduced a computational method for predicting RNA-binding sites in proteins based on support vector machines by using a variety of features from amino acid sequence information including position-specific scoring matrix (PSSM) profiles, physicochemical properties and predicted solvent accessibility. Considering the influence of the surrounding residues of an amino acid and the dependency effect from the neighboring amino acids, a sliding window and a smoothing window are used to encode the PSSM profiles. The outer fivefold cross-validation method is evaluated on the data set of 77 RNA-binding proteins (RBP77). It achieves an overall accuracy of 88.66% with the Matthew’s correlation coefficient (MCC) of 0.69. Furthermore, an independent data set of 39 RNA-binding proteins (RBP39) is employed to further evaluate the performance and achieves an overall accuracy of 82.36% with the MCC of 0.44. The result shows that our method has good generalization abilities in predicting RNA-binding sites for novel proteins. Compared with other previous methods, our method performs well on the same data set. The prediction results suggest that the used features are effective in predicting RNA-binding sites in proteins. The code and all data sets used in this article are freely available at .     

Sequence fingerprints distinguish erroneous from correct predictions of intrinsically disordered protein regions

More than 60 prediction methods for intrinsically disordered proteins (IDPs) have been developed over the years, many of which are accessible on the World Wide Web. Nearly, all of these predictors give balanced accuracies in the ~65%–~80% range. Since predictors are not perfect, further studies are required to uncover the role of amino acid residues in native IDP as compared to predicted IDP regions. In the present work, we make use of sequences of 100% predicted IDP regions, false positive disorder predictions, and experimentally determined IDP regions to distinguish the characteristics of native versus predicted IDP regions. A higher occurrence of asparagine is observed in sequences of native IDP regions but not in sequences of false positive predictions of IDP regions. The occurrences of certain combinations of amino acids at the pentapeptide level provide a distinguishing feature in the IDPs with respect to globular proteins. The distinguishing features presented in this paper provide insights into the sequence fingerprints of amino acid residues in experimentally determined as compared to predicted IDP regions. These observations and additional work along these lines should enable the development of improvements in the accuracy of disorder prediction algorithm.     

Comparing and combining predictors of mostly disordered proteins

Intrinsically disordered proteins and regions carry out varied and vital cellular functions. Proteins with disordered regions are especially common in eukaryotic cells, with a subset of these proteins being mostly disordered, e.g., with more disordered than ordered residues. Two distinct methods have been previously described for using amino acid sequences to predict which proteins are likely to be mostly disordered. These methods are based on the net charge-hydropathy distribution and disorder prediction score distribution. Each of these methods is reexamined, and the prediction results are compared herein. A new prediction method based on consensus is described. Application of the consensus method to whole genomes reveals that approximately 4.5% of Yersinia pestis, 5% of Escherichia coli K12, 6% of Archaeoglobus fulgidus, 8% of Methanobacterium thermoautotrophicum, 23% of Arabidopsis thaliana, and 28% of Mus musculus proteins are mostly disordered. The unexpectedly high frequency of mostly disordered proteins in eukaryotes has important implications both for large-scale, high-throughput projects and also for focused experiments aimed at determination of protein structure and function.     

Conservation of intrinsic disorder in protein domains and families: I. A database of conserved predicted disordered regions

Many protein regions have been shown to be intrinsically disordered, lacking unique structure under physiological conditions. These intrinsically disordered regions are not only very common in proteomes, but also crucial to the function of many proteins, especially those involved in signaling, recognition, and regulation. The goal of this work was to identify the prevalence, characteristics, and functions of conserved disordered regions within protein domains and families. A database was created to store the amino acid sequences of nearly one million proteins and their domain matches from the InterPro database, a resource integrating eight different protein family and domain databases. Disorder prediction was performed on these protein sequences. Regions of sequence corresponding to domains were aligned using a multiple sequence alignment tool. From this initial information, regions of conserved predicted disorder were found within the domains. The methodology for this search consisted of finding regions of consecutive positions in the multiple sequence alignments in which a 90% or more of the sequences were predicted to be disordered. This procedure was constrained to find such regions of conserved disorder prediction that were at least 20 amino acids in length. The results of this work included 3,653 regions of conserved disorder prediction, found within 2,898 distinct InterPro entries. Most regions of conserved predicted disorder detected were short, with less than 10% of those found exceeding 30 residues in length.     

SPOT-Disorder2: Improved Protein Intrinsic Disorder Prediction by Ensembled Deep Learning

Intrinsically disordered or unstructured proteins (or regions in proteins) have been found to be important in a wide range of biological functions and implicated in many diseases. Due to the high cost and low efficiency of experimental determination of intrinsic disorder and the exponential increase of unannotated protein sequences, developing complementary computational prediction methods has been an active area of research for several decades. Here, we employed an ensemble of deep Squeeze-and-Excitation residual inception and long short-term memory (LSTM) networks for predicting protein intrinsic disorder with input from evolutionary information and predicted one-dimensional structural properties. The method, called SPOT-Disorder2, offers substantial and consistent improvement not only over our previous technique based on LSTM networks alone, but also over other state-of-the-art techniques in three independent tests with different ratios of disordered to ordered amino acid residues, and for sequences with either rich or limited evolutionary information. More importantly, semi-disordered regions predicted in SPOT-Disorder2 are more accurate in identifying molecular recognition features (MoRFs) than methods directly designed for MoRFs prediction. SPOT-Disorder2 is available as a web server and as a standalone program at https://sparks-lab.org/server/spot-disorder2/.     

The importance of intrinsic disorder for protein phosphorylation

Reversible protein phosphorylation provides a major regulatory mechanism in eukaryotic cells. Due to the high variability of amino acid residues flanking a relatively limited number of experimentally identified phosphorylation sites, reliable prediction of such sites still remains an important issue. Here we report the development of a new web-based tool for the prediction of protein phosphorylation sites, DISPHOS (DISorder-enhanced PHOSphorylation predictor, http://www.ist.temple. edu/DISPHOS). We observed that amino acid compositions, sequence complexity, hydrophobicity, charge and other sequence attributes of regions adjacent to phosphorylation sites are very similar to those of intrinsically disordered protein regions. Thus, DISPHOS uses position-specific amino acid frequencies and disorder information to improve the discrimination between phosphorylation and non-phosphorylation sites. Based on the estimates of phosphorylation rates in various protein categories, the outputs of DISPHOS are adjusted in order to reduce the total number of misclassified residues. When tested on an equal number of phosphorylated and non-phosphorylated residues, the accuracy of DISPHOS reaches 76% for serine, 81% for threonine and 83% for tyrosine. The significant enrichment in disorder-promoting residues surrounding phosphorylation sites together with the results obtained by applying DISPHOS to various protein functional classes and proteomes, provide strong support for the hypothesis that protein phosphorylation predominantly occurs within intrinsically disordered protein regions.     

Intrinsic disorder in the Protein Data Bank

The Protein Data Bank (PDB) is the preeminent source of protein structural information. PDB contains over 32,500 experimentally determined 3-D structures solved using X-ray crystallography or nuclear magnetic resonance spectroscopy. Intrinsically disordered regions fail to form a fixed 3-D structure under physiological conditions. In this study, we compare the amino-acid sequences of proteins whose structures are determined by X-ray crystallography with the corresponding sequences from the Swiss-Prot database. The analyzed dataset includes 16,370 structures, which represent 18,101 PDB chains and 5,434 different proteins from 910 different organisms (2,793 eukaryotic, 2,109 bacterial, 288 viral, and 244 archaeal). In this dataset, on average, each Swiss-Prot protein is represented by 7 PDB chains with 76% of the crystallized regions being represented by more than one structure. Intriguingly, the complete sequences of only approximately 7% of proteins are observed in the corresponding PDB structures, and only approximately 25% of the total dataset have >95% of their lengths observed in the corresponding PDB structures. This suggests that the vast majority of PDB proteins is shorter than their corresponding Swiss-Prot sequences and/or contain numerous residues, which are not observed in maps of electron density. To determine the prevalence of disordered regions in PDB, the residues in the Swiss-Prot sequences were grouped into four general categories, "Observed" (which correspond to structured regions), "Not observed" (regions with missing electron density, potentially disordered), "Uncharacterized," and "Ambiguous," depending on their appearance in the corresponding PDB entries. This non-redundant set of residues can be viewed as a 'fragment' or empirical domain database that contains a set of experimentally determined structured regions or domains and a set of experimentally verified disordered regions or domains. We studied the propensities and properties of residues in these four categories and analyzed their relations to the predictions of disorder using several algorithms. "Non-observed," "Ambiguous," and "Uncharacterized" regions were shown to possess the amino acid compositional biases typical of intrinsically disordered proteins. The application of four different disorder predictors (PONDR(R) VL-XT, VL3-BA, VSL1P, and IUPred) revealed that the vast majority of residues in the "Observed" dataset are ordered, and that the "Not observed" regions are mostly disordered. The "Uncharacterized" regions possess some tendency toward order, whereas the predictions for the short "Ambiguous" regions are really ambiguous. Long "Ambiguous" regions (>70 amino acid residues) are mostly predicted to be ordered, suggesting that they are likely to be "wobbly" domains. Overall, we showed that completely ordered proteins are not highly abundant in PDB and many PDB sequences have disordered regions. In fact, in the analyzed dataset approximately 10% of the PDB proteins contain regions of consecutive missing or ambiguous residues longer than 30 amino-acids and approximately 40% of the proteins possess short regions (> or =10 and < 30 amino-acid long) of missing and ambiguous residues.     

Intrinsic Disorder in the Protein Data Bank

Abstract

The Protein Data Bank (PDB) is the preeminent source of protein structural information. PDB contains over 32,500 experimentally determined 3-D structures solved using X-ray crystallography or nuclear magnetic resonance spectroscopy. Intrinsically disordered regions fail to form a fixed 3-D structure under physiological conditions. In this study, we compare the amino-acid sequences of proteins whose structures are determined by X-ray crystallography with the corresponding sequences from the Swiss-Prot database. The analyzed dataset includes 16,370 structures, which represent 18,101 PDB chains and 5,434 different proteins from 910 different organisms (2,793 eukaryotic, 2,109 bacterial, 288 viral, and 244 archaeal). In this dataset, on average, each Swiss-Prot protein is represented by 7 PDB chains with 76% of the crystallized regions being represented by more than one structure. Intriguingly, the complete sequences of only ~7% of proteins are observed in the corresponding PDB structures, and only ~25% of the total dataset have >95% of their lengths observed in the corresponding PDB structures. This suggests that the vast majority of PDB proteins is shorter than their corresponding Swiss-Prot sequences and/or contain numerous residues, which are not observed in maps of electron density. To determine the prevalence of disordered regions in PDB, the residues in the Swiss-Prot sequences were grouped into four general categories, “Observed” (which correspond to structured regions), “Not observed” (regions with missing electron density, potentially disordered), “Uncharacterized,” and “Ambiguous,” depending on their appearance in the corresponding PDB entries. This non-redundant set of residues can be viewed as a ‘fragment’ or empirical domain database that contains a set of experimentally determined structured regions or domains and a set of experimentally verified disordered regions or domains. We studied the propensities and properties of residues in these four categories and analyzed their relations to the predictions of disorder using several algorithms. “Non-observed,” “Ambiguous,” and “Uncharacterized” regions were shown to possess the amino acid compositional biases typical of intrinsically disordered proteins. The application of four different disorder predictors (PONDR® VL-XT, VL3-BA, VSL1P, and IUPred) revealed that the vast majority of residues in the “Observed” dataset are ordered, and that the “Not observed” regions are mostly disordered. The “Uncharacterized” regions possess some tendency toward order, whereas the predictions for the short “Ambiguous” regions are really ambiguous. Long “Ambiguous” regions (>70 amino acid residues) are mostly predicted to be ordered, suggesting that they are likely to be “wobbly” domains.

Overall, we showed that completely ordered proteins are not highly abundant in PDB and many PDB sequences have disordered regions. In fact, in the analyzed dataset ~10% of the PDB proteins contain regions of consecutive missing or ambiguous residues longer than 30 amino-acids and ~40% of the proteins possess short regions (≥10 and <30 amino-acid long) of missing and ambiguous residues.     

Human-readable rule generator for integrating amino scid sequence information and stability of mutant proteins

Most of the bioinformatics tools developed for predicting mutant protein stability appear as a black box and the relationship between amino acid sequence/structure and stability is hidden to the users. We have addressed this problem and developed a human-readable rule generator for integrating the knowledge of amino acid sequence and experimental stability change upon single mutation. Using information about the original residue, substituted residue, and three neighboring residues, classification rules have been generated to discriminate the stabilizing and destabilizing mutants and explore the basis for experimental data. These rules are human readable, and hence, the method enhances the synergy between expert knowledge and computational system. Furthermore, the performance of the rules has been assessed on a nonredundant data set of 1,859 mutants and we obtained an accuracy of 80 percent using cross validation. The results showed that the method could be effectively used as a tool for both knowledge discovery and predicting mutant protein stability. We have developed a Web for classification rule generator and it is freely available at http://bioinformatics.myweb.hinet.net/irobot.htm.