Computational Systems Bioinformatics

PREFACE

COMMITTEES

REFEREES

CONTENTS

Metagenomics is an emerging methodology for the direct genomic analysis of a mixed community of uncultured microorganisms. The current analyses of metagenomics data largely rely on the computational tools originally designed for microbial genomics projects. The challenge of assembling metagenomic sequences arises mainly from the short reads and the high species complexity of the community. Alternatively, individual (short) reads will be searched directly against databases of known genes (or proteins) to identify homologous sequences. The latter approach may have low sensitivity and specificity in identifying homologous sequences, which may further bias the subsequent diversity analysis. In this paper, we present a novel approach to metagenomic data analysis, called Metagenomic ORFome Assembly (MetaORFA). The whole computational framework consists of three steps. Each read from a metagenomics project will first be annotated with putative open reading frames (ORFs) that likely encode proteins. Next, the predicted ORFs are assembled into a collection of peptides using an EULER assembly method. Finally, the assembled peptides (i.e., ORFome) are used for database searching of homologs and subsequent diversity analysis. We applied MetaORFA approach to several metagenomics datasets with low coverage short reads. The results show that MetaORFA can produce long peptides even when the sequence coverage of reads is extremely low. Hence, the ORFome assembly significantly increased the sensitivity of homology searching, and may potentially improve the diversity analysis of the metagenomic data. This improvement is especially useful for the metagenomic projects when the genome assembly does not work because of the low sequence coverage.

This paper presents an original Quantum Genetic algorithm for Multiple sequence ALIGNment (QGMALIGN) that combines a genetic algorithm and a quantum algorithm. A quantum probabilistic coding is designed for representing the multiple sequence alignment. A quantum rotation gate as a mutation operator is used to guide the quantum state evolution. Six genetic operators are designed on the coding basis to improve the solution during the evolutionary process. The features of implicit parallelism and state superposition in quantum mechanics and the global search capability of the genetic algorithm are exploited to get efficient computation. A set of well known test cases from BAliBASE2.0 is used as reference to evaluate the efficiency of the QGMALIGN optimization. The QGMALIGN results have been compared with the most popular methods (CLUSTALX, SAGA, DIALIGN, SB_PIMA, and QGMALIGN) results. The QGMALIGN results show that QGMALIGN performs well on the presenting biological data. The addition of genetic operators to the quantum algorithm lowers the cost of overall running time.

With the rapidly expanding field of medical genetics and genetic counseling, genealogy information is becoming increasingly abundant. An important computation on pedigree data is the calculation of identity coefficients, which provide a complete description of the degree of relatedness of a pair of individuals. The areas of application of identity coefficients are numerous and diverse, from genetic counseling to disease tracking, and thus, the computation of identity coefficients merits special attention. However, the computation of identity coefficients is not done directly, but rather as the final step after computing a set of generalized kinship coefficients. In this paper, we first propose a novel Path-Counting Formula for calculating generalized kinship coefficients, which is motivated by Wright's path-counting method for computing the inbreeding coefficient for an individual. We then present an efficient and scalable scheme for calculating generalized kinship coefficients on large pedigrees using NodeCodes, a special encoding scheme for expediting the evaluation of queries on pedigree graph structures. We also perform experiments for evaluating the efficiency of our method, and compare it with the performance of the traditional recursive algorithm for three individuals. Experimental results demonstrate that the resulting scheme is more scalable and efficient than the traditional recursive methods for computing generalized kinship coefficients.

Finding motifs in many sequences is an important problem in computational biology, especially in identification of regulatory motifs in DNA sequences. Let c be a motif sequence. Given a set of sequences, each is planted with a mutated version of c at an unknown position, the motif finding problem is to find these planted motifs and the original c. In this paper, we study the VM model of the planted motif problem, which is proposed by Pevzner and Sze¹. We give a simple Selecting One Voting algorithm and a more powerful Selecting k Voting algorithm. When the length of motif and the number of input sequences are large enough, we prove that the two algorithms can find the unknown motif consensus with high probability. In the proof, we show why a large number of input sequences is so important for finding motifs, which is believed by most researchers. Experimental results on simulated data also support the claim. Selecting k Voting algorithm is powerful, but computational intensive. To speed up the algorithm, we propose a progressive filtering algorithm, which improves the running time significantly and has good accuracy in finding motifs. Our experimental results show that Selecting k Voting algorithm with progressive filtering performs very well in practice and it outperforms some best known algorithms.

Availability: The software is available upon request.

The emergence of high-throughput technologies leads to abundant protein-protein interaction (PPI) data and microarray gene expression profiles, and provides a great opportunity for the identification of novel protein complexes using computational methods. Although it has been demonstrated in the literature that methods using protein-protein interaction data alone can successfully predict a large number of protein complexes, the incorporation of gene expression profiles could help refine the putative complexes and hence improve the accuracy of the computational methods.

By combining protein-protein interaction data and microarray gene expression profiles, we propose a novel Graph Fragmentation Algorithm (GFA) for protein complex identification. Adapted from a classical max-flow algorithm for finding the (weighted) densest subgraphs, GFA first finds large (weighted) dense subgraphs in a protein-protein interaction network and then breaks each such subgraph into fragments iteratively by weighting its nodes appropriately in terms of their corresponding log fold changes in the microarray data, until the fragment subgraphs are suffciently small. Our extensive tests on three widely used protein-protein interaction datasets and comparisons with the latest methods for protein complex identification demonstrate the superior performance of our method in terms of accuracy, effciency, and capability in predicting novel protein complexes. Given the high specificity (or precision) that our method has achieved, we conjecture that our prediction results imply more than 200 novel protein complexes.

Along with the wide application of mass spectrometry in proteomics, more and more mass spectrometry data are becoming publicly available. Several public mass spectrometry data repositories have been built on the Internet. However, most of these repositories are devoid of effective searching methods. In this paper we describe a new mass spectrometry data library, and a novel method to efficiently index and search in the library for spectra that are similar to a query spectrum. A public online server have been set up and demonstrated outstanding speed and scalability of our methods.

Together with the mass spectrometry library, our searching method can improve the protein identification confidence by comparing a spectrum with the ones that are already characterized in the database. The searching method can also be used alone to cluster the similar spectra in a mass spectrometry dataset together, in order to to improve the speed and accuracy of the protein identification or quantification.

Almost every cellular process requires the interactions of pairs or larger complexes of proteins. High throughput protein-protein interaction (PPI) data have been generated using techniques such as the yeast two-hybrid systems, mass spectrometry method, and many more. Such data provide us with a new perspective to predict protein functions and to generate protein-protein interaction networks, and many recent algorithms have been developed for this purpose. However, PPI data generated using high throughput techniques contain a large number of false positives. In this paper, we have proposed a novel method to evaluate the support for PPI data based on gene ontology information. If the semantic similarity between genes is computed using gene ontology information and using Resnik's formula, then our results show that we can model the PPI data as a mixture model predicated on the assumption that true protein-protein interactions will have higher support than the false positives in the data. Thus semantic similarity between genes serves as a metric of support for PPI data. Taking it one step further, new function prediction approaches are also being proposed with the help of the proposed metric of the support for the PPI data. These new function prediction approaches outperform their conventional counterparts. New evaluation methods are also proposed.

Mass Spectrometry (MS) is increasingly being used to discover disease related proteomic patterns. The peak detection step is one of most important steps in the typical analysis of MS data. Recently, many new algorithms have been proposed to increase true position rate with low false position rate in peak detection. Most of them follow two approaches: one is denoising approach and the other one is decomposing approach. In the previous studies, the decomposition of MS data method shows more potential than the first one. In this paper, we propose a new method named GaborLocal which can detect more true peaks with a very low false position rate. The Gaussian local maxima is employed for peak detection, because it is robust to noise in signals. Moreover, the maximum rank of peaks is defined at the first time to identify peaks instead of using the signal-to-noise ratio and the Gabor filter is used to decompose the raw MS signal. We perform the proposed method on the real SELDI-TOF spectrum with known polypeptide positions. The experimental results demonstrate our method outperforms other common used methods in the receiver operating characteristic (ROC) curve.

Site-directed mutagenesis affects protein stability in a manner dependent on the local structural environment of the mutated residue; e.g., a hydrophobic to polar substitution would behave differently in the core vs. on the surface of the protein. Thus site-directed mutagenesis followed by stability measurement enables evaluation of and selection among predicted structure models, based on consistency between predicted and experimental stability changes (∆∆G° values). This paper develops a method for planning a set of individual site-directed mutations for protein structure model selection, so as to minimize the Bayes error, i.e., the probability of choosing the wrong model. While in general it is hard to calculate exactly the multi-dimensional Bayes error defined by a set of mutations, we leverage the structure of “∆∆G° space” to develop tight upper and lower bounds. We further develop a lower bound on the Bayes error of any plan that uses a fixed number of mutations from a set of candidates. We use this bound in a branch-and-bound planning algorithm to find optimal and near-optimal plans. We demonstrate the significance and effectiveness of this approach in planning mutations for elucidating the structure of the pTfa chaperone protein from bacteriophage lambda.

We have developed a feedback algorithm for protein structure alignment between two protein backbones. A web portal implementing this method has been constructed and is freely available for use at http://fpsa.cs.uno.edu/ with a mirror site at http://fpsa.cs.panam.edu/FPSA/. We compare our algorithm with three other, commonly used methods: CE, DaliLite and SSM. The results show that in most cases our algorithm outputs a larger number of aligned positions when the (C_α) RMSD is comparable. Also, in many cases where the number of aligned positions is larger or comparable, our learning method is able to achieve a smaller (C_α) RMSD than the other methods tested. This trend of larger number of aligned positions and smaller (C_α) RMSD is observed more frequently in cases where the similarity between protein structures is weak.

Identifying B-cell epitopes play an important role in vaccine design, immunodiagnostic tests, and antibody production. Therefore, computational tools for reliably predicting B-cell epitopes are highly desirable. We explore two machine learning approaches for predicting fiexible length linear B-cell epitopes. The first approach utilizes four sequence kernels for determining a similarity score between any arbitrary pair of variable length sequences. The second approach utilizes four different methods of mapping a variable length sequence into a fixed length feature vector. Based on our empirical comparisons, we propose FBCPred, a novel method for predicting flexible length linear B-cell epitopes using the subsequence kernel. Our results demonstrate that FBCPred significantly outperforms all other classifiers evaluated in this study. An implementation of FBCPred and the datasets used in this study are publicly available through our linear B-cell epitope prediction server, BCPREDS, at: http://ailab.cs.iastate.edu/bcpreds/.

Establishing structural or functional relationship between sequences, for instance to infer the structural class of an unannotated protein, is a key task in biological sequence analysis. Recent computational methods such as profile and neighborhood mismatch kernels have shown very promising results for protein sequence classification, at the cost of high computational complexity. In this study we address the multi-class sequence classification problems using a class of string-based kernels, the sparse spatial sample kernels (SSSK), that are both biologically motivated and efficient to compute. The proposed methods can work with very large databases of protein sequences and show substantial improvements in computing time over the existing methods. Application of the SSSK to the multi-class protein prediction problems (fold recognition and remote homology detection) yields significantly better performance than existing state-of-theart algorithms.

Detecting non-coding RNAs (ncRNAs) in genomic DNA is an important part of annotation. However, the most widely used tool for modeling ncRNA families, the covariance model (CM), incurs a high computational cost when used for search. This cost can be reduced by using a filter to exclude sequence that is unlikely to contain the ncRNA of interest, applying the CM only where it is likely to match strongly. Despite recent advances, designing an efficient filter that can detect nearly all ncRNA instances while excluding most irrelevant sequences remains challenging.

This work proposes a systematic procedure to convert a CM for an ncRNA family to a secondary structure profile (SSP), which augments a conservation profile with secondary structure information but can still be efficiently scanned against long sequences. We use dynamic programming to estimate an SSP's sensitivity and FP rate, yielding an efficient, fully automated filter design algorithm. Our experiments demonstrate that designed SSP filters can achieve significant speedup over unfiltered CM search while maintaining high sensitivity for various ncRNA families, including those with and without strong sequence conservation. For highly structured ncRNA families, including secondary structure conservation yields better performance than using primary sequence conservation alone.

There is an increasing number of proteins with known structure but unknown function. Determining their function would have a significant impact on understanding diseases and designing new therapeutics. However, experimental protein function determination is expensive and very time-consuming. Computational methods can facilitate function determination by identifying proteins that have high structural and chemical similarity. Our focus is on methods that determine binding site similarity. Although several such methods exist, it still remains a challenging problem to quickly find all functionally-related matches for structural motifs in large data sets with high specificity. In this context, a structural motif is a set of 3D points annotated with physicochemical information that characterize a molecular function. We propose a new method called LabelHash that creates hash tables of n-tuples of residues for a set of targets. Using these hash tables, we can quickly look up partial matches to a motif and expand those matches to complete matches. We show that by applying only very mild geometric constraints we can find statistically significant matches with extremely high specificity in very large data sets and for very general structural motifs. We demonstrate that our method requires a reasonable amount of storage when employing a simple geometric filter and further improves on the specificity of our previous work while maintaining very high sensitivity. Our algorithm is evaluated on 20 homolog classes and a non-redundant version of the Protein Data Bank as our background data set. We use cluster analysis to analyze why certain classes of homologs are more difficult to classify than others. The LabelHash algorithm is implemented on a web server at http://kavrakilab.org/labelhash/.

High-throughput structure determination based on solution Nuclear Magnetic Resonance (NMR) spectroscopy plays an important role in structural genomics. One of the main bottlenecks in NMR structure determination is the interpretation of NMR data to obtain a sufficient number of accurate distance restraints by assigning nuclear Overhauser effect (NOE) spectral peaks to pairs of protons. The difficulty in automated NOE assignment mainly lies in the ambiguities arising both from the resonance degeneracy of chemical shifts and from the uncertainty due to experimental errors in NOE peak positions. In this paper we present a novel NOE assignment algorithm, called HAusdorff-based NOE Assignment (HANA), that starts with a high-resolution protein backbone computed using only two residual dipolar couplings (RDCs) per residue^{37, 39}, employs a Hausdorff-based pattern matching technique to deduce similarity between experimental and back-computed NOE spectra for each rotamer from a statistically diverse library, and drives the selection of optimal position-specific rotamers for filtering ambiguous NOE assignments. Our algorithm runs in time O(tn³ + tn log t), where t is the maximum number of rotamers per residue and n is the size of the protein. Application of our algorithm on biological NMR data for three proteins, namely, human ubiquitin, the zinc finger domain of the human DNA Y-polymerase Eta (pol η) and the human Set2-Rpb1 interacting domain (hSRI) demonstrates that our algorithm overcomes spectral noise to achieve more than 90% assignment accuracy. Additionally, the final structures calculated using our automated NOE assignments have backbone RMSD < 1.7 Å and all-heavy-atom RMSD < 2.5 Å from reference structures that were determined either by X-ray crystallography or traditional NMR approaches. These results show that our NOE assignment algorithm can be successfully applied to protein NMR spectra to obtain high-quality structures.

Structural similarity between proteins gives us insights on the evolutionary relationship between proteins which have low sequence similarity. In this paper, we present a novel approach called STSA for non-sequential pair-wise structural alignment. Starting from an initial alignment, our approach iterates over a two-step process, a superposition step and an alignment step, until convergence. Given two superposed structures, we propose a novel greedy algorithm to construct both sequential and non-sequential alignments. The quality of STSA alignments is evident in the high agreement it has with the reference alignments in the challenging-to-align RPIC set. Moreover, on a dataset of 4410 protein pairs selected from the CATH database, STSA has a high sensitivity and high specificity values and is competitive with state-of-the-art alignment methods and gives longer alignments with lower rmsd. The STSA software along with the data sets will be made available on line at http://www.cs.rpi.edu/~zaki/software/STSA.

Solvent accessibility is an important structural feature for a protein. We propose a new method for solvent accessibility prediction that uses known structure and sequence information more efficiently. We first estimate the relative solvent accessibility of the query protein using fuzzy mean operator from the solvent accessibilities of known structure fragments that have similar sequences to the query protein. We then integrate the estimated solvent accessibility and the position specific scoring matrix of the query protein using a neural network. We tested our method on a large data set consisting of 3386 non-redundant proteins. The comparison with other methods show slightly improved prediction accuracies with our method. The resulting system does need not be re-trained when new data is available. We incorporated our method into the MUPRED system, which is available as a web server at http://digbio.missouri.edu/mupred.

With some simplifications, computational protein folding can be understood as an optimization problem of a potential energy function on a variable space consisting of all conformation for a given protein molecule. It is well known that realistic energy potentials are very “rough” functions, when expressed in the standard variables, and the folding trajectories can be easily trapped in multiple local minima. We have integrated our variation of Parallel Tempering optimization into the protein folding program Rosetta in order to improve its capability to overcome energy barriers and estimate how such improvement will influence the quality of the folded protein domains. Here we report that (1) Parallel Tempering Rosetta (PTR) is significantly better in the exploration of protein structures than previous implementations of the program; (2) systematic improvements are observed across a large benchmark set in the parameters that are normally followed to estimate robustness of the folding; (3) these improvements are most dramatic in the subset of the shortest domains, where high-quality structures have been obtained for >75% of all tested sequences. Further analysis of the results will improve our understanding of protein conformational space and lead to new improvements in the protein folding methodology, while the current PTR implementation should be very efficient for short (up to ~80 a.a.) protein domains and therefore may find practical application in system biology studies.

In order to improve the prediction of protein-ligand binding sites through homology modeling, we incorporate knowledge of the binding residues into the modeling framework. Residues are identified as binding or nonbinding based on their true labels as well as labels predicted from structure and sequence. The sequence predictions were made using a support vector machine framework which employs a sophisticated window-based kernel. Binding labels are used with a very sensitive sequence alignment method to align the target and template. Relevant parameters governing the alignment process are searched for optimal values. Based on our results, homology models of the binding site can be improved if a priori knowledge of the binding residues is available. For target-template pairs with low sequence identity and high structural diversity our sequence-based prediction method provided sufficient information to realize this improvement.

With the advent of high-throughput gene perturbation screens (e.g., RNAi assays, genome-wide deletion mutants), modeling the complex relationship between genes and phenotypes has become a paramount problem. One broad class of methods uses ‘guilt by association’ methods to impute phenotypes to genes based on the interactions between the given gene and other genes with known phenotypes. But these methods are inadequate for genes that have no cataloged interactions but which nevertheless are known to result in important phenotypes. In this paper, we present an approach to first model relationships between phenotypes using the notion of ‘relative importance’ and subsequently use these derived relationships to make phenotype predictions. Besides improved accuracy on S. cerevisiae deletion mutants and C. elegans knock-down datasets, we show how our approach sheds insight into relations between phenotypes.

Pathways show how different biochemical entities interact with each other to perform vital functions for the survival of organisms. Similarities between pathways indicate functional similarities that are difficult to identify by comparing the individual entities that make up those pathways. When interacting entities are of single type, the problem of identifying similarities reduces to graph isomorphism problem. However, for pathways with varying types of entities, such as metabolic pathways, alignment problem is more challenging. Existing methods, often, address the metabolic pathway alignment problem by ignoring all the entities except for one type. This kind of abstraction reduces the relevance of the alignment significantly as it causes losses in the information content. In this paper, we develop a method to solve the pairwise alignment problem for metabolic pathways. One distinguishing feature of our method is that it aligns reactions, compounds and enzymes without abstraction of pathways. We pursue the intuition that both pairwise similarities of entities (homology) and their organization (topology) are crucial for metabolic pathway alignment. In our algorithm, we account for both by creating an eigenvalue problem for each entity type. We enforce the consistency by considering the reachability sets of the aligned entities. Our experiments show that, our method finds biologically and statistically significant alignments in the order of seconds for pathways with ∼ 100 entities.

The recent explosion in the number of clinical studies involving microarray data calls for novel computational methods for their dissection. Human protein interaction networks are rapidly growing and can assist in the extraction of functional modules from microarray data. We describe a novel methodology for extraction of connected network modules with coherent gene expression patterns that are correlated with a specific clinical parameter. Our approach suits both numerical (e.g., age or tumor size) and logical parameters (e.g., gender or mutation status). We demonstrate the method on a large breast cancer dataset, where we identify biologically-relevant modules related to nine clinical parameters including patient age, tumor size, and metastasis-free survival. Our method is capable of detecting disease-relevant pathways that could not be found using other methods. Our results support some previous hypotheses regarding the molecular pathways underlying diversity of breast tumors and suggest novel ones.

We improve on guided genome halving algorithms so that several thousand gene sets, each containing two paralogs in the descendant T of the doubling event and their single ortholog from an undoubled reference genome R, can be analyzed to reconstruct the ancestor A of T at the time of doubling. At the same time, large numbers of defective gene sets, either missing one paralog from T or missing their ortholog in R, may be incorporated into the analysis in a consistent way. We apply this genomic rearrangement distance-based approach to the recently sequenced poplar (Populus trichocarpa) and grapevine (Vitis vinifera) genomes, as T and R respectively.

Kinship analysis using genetic data is important for many biological applications, including many in conservation biology. Wide availability of microsatellites has boosted studies in wild populations that rely on the knowledge of kinship, particularly sibling relationships (sibship). While there exist many methods for reconstructing sibling relationships, almost none account for errors and mutations in microsatellite data, which are prevalent and affect the quality of reconstruction. We present an error-tolerant method for reconstructing sibling relationships based on the concept of consensus methods. We test our approach on both real and simulated data, with both pre-existing and introduced errors. Our method is highly accurate on almost all simulations, giving over 90% accuracy in most cases. Ours is the first method designed to tolerate errors while making no assumptions about the population or the sampling.

We study the problem of merging genetic maps, when the individual genetic maps are given as directed acyclic graphs. The problem is to build a consensus map, which includes and is consistent with all (or, the vast majority of) the markers in the individual maps. When markers in the input maps have ordering conflicts, the resulting consensus map will contain cycles. We formulate the problem of resolving cycles in a combinatorial optimization framework, which in turn is expressed as an integer linear program. A faster approximation algorithm is proposed, and an additional speed-up heuristic is developed. According to an extensive set of experimental results, our tool is consistently better than JOINMAP, both in terms of accuracy and running time.

We study the haplotype inference problem from pedigree data under the zero recombination assumption, which is well supported by real data for tightly linked markers (i.e., single nucleotide polymorphisms (SNPs)) over a relatively large chromosome segment. We solve the problem in a rigorous mathematical manner by formulating genotype constraints as a linear system of inheritance variables. We then utilize disjoint-set structures to encode connectivity information among individuals, to detect constraints from genotypes, and to check consistency of constraints. On a tree pedigree without missing data, our algorithm can output a general solution as well as the number of total specific solutions in a nearly linear time O(mn · α(n)), where m is the number of loci, n is the number of individuals and α is the inverse Ackermann function⁴, which is a further improvement over existing ones^{3, 8, 12, 15}. We also extend the idea to looped pedigrees and pedigrees with missing data by considering existing (partial) constraints on inheritance variables. The algorithm has been implemented in C ++ and will be incorporated into our PedPhase package⁸. Experimental results show that it can correctly identify all 0-recombinant solutions with great efficiency. Comparisons with other two popular algorithms show that the proposed algorithm achieves 10 to 10⁵-fold improvements over a variety of parameter settings. The experimental study also provides empirical evidences on the complexity bounds suggested by theoretical analysis.

Biological and pharmaceutical research relies heavily on microscopically imaging cell populations for understanding their structure and function. Much work has been done on automated analysis of biological images, but image analysis tools are generally focused only on extracting quantitative information for validating a particular hypothesis. Images contain much more information than is normally required for testing individual hypotheses. The lack of symbolic knowledge representation schemes for representing semantic image information and the absence of knowledge mining tools are the biggest obstacles in utilizing the full information content of these images. In this paper we first present a graph-based scheme for integrated representation of semantic biological knowledge contained in cellular images acquired in spatial, spectral, and temporal dimensions. We then present a spatio-temporal knowledge mining framework for extracting non-trivial and previously unknown association rules from image data sets. This mechanism can change the role of biological imaging from a tool used to validate hypotheses to one used for automatically generating new hypotheses. Results for an apoptosis screen are also presented.

We present two heuristics for speeding up a time series alignment algorithm that is related to dynamic time warping (DTW). In previous work, we developed our multisegment alignment algorithm to answer similarity queries for toxicogenomic time-series data. Our multisegment algorithm returns more accurate alignments than DTW at the cost of time complexity; the multisegment algorithm is O(n⁵) whereas DTW is O(n²). The first heuristic we present speeds up our algorithm by a constant factor by restricting alignments to a cone shape in alignment space. The second heuristic restricts the alignments considered to those near one returned by a DTW-like method. This heuristic adjusts the time complexity to O(n³). Importantly, neither heuristic results in a loss in accuracy.

In this paper we introduce a novel graph classification algorithm and demonstrate its efficacy in drug design. In our method, we use graphs to model chemical structures and apply a wavelet analysis of graphs to create features capturing graph local topology. We design a novel graph kernel function to utilize the created feature to build predictive models for chemicals. We call the new graph kernel a graph wavelet-alignment kernel.

We have evaluated the efficacy of the wavelet-alignment kernel using a set of chemical structure-activity prediction benchmarks. Our results indicate that the use of the kernel function yields performance profiles comparable to, and sometimes exceeding that of the existing state-of-the-art chemical classification approaches. In addition, our results also show that the use of wavelet functions significantly decreases the computational costs for graph kernel computation with more than 10 fold speed up.

AUTHOR INDEX