Detecting drug side effects is crucial in development of drugs. As publicly available biomedical data expands, researchers have devised numerous computational methods for predicting drug-side effect associations (DSAs). Among these, network-based approaches have gained significant attention in the biomedical field. However, the challenge of data scarcity poses a significant hurdle for existing DSAs prediction models. While various data augmentation methods have been created to solve the proble, most rely on random alterations to the original networks, neglecting the causality of DSAs’ existence, thus impacting the predictive performance negatively. In this paper, we introduce a counterfactual inference-based data augmentation method to enhance performance. First,a heterogeneous information network (HIN) is construct by integrating multiple biomedical data sources. We employ community detection on the HIN to preform a counterfactual inference-based method, deriving augmented links and an augmented HIN. Subsequently, we apply a meta-path-based graph neural network to obtain high-quality representations of drugs and side effects, enabling the prediction of DSAs. Our comprehensive experiments confirm the effectiveness of this counterfactual inference-based data augmentation for DSAs prediction.

Drug repositioning is the process of identifying potential associations between approved drugs and diseases (DDAs) to unveil novel therapeutic applications. Unlike traditional drug discovery approaches, a key advantage of drug repositioning lies in its capacity to leverage the existing knowledge and safety profiles of established medications, leading to significant reductions in both the time and costs associated with drug development. While various methods have been proposed to address this challenge using diverse strategies, the conventional approach for training DDAs prediction models typically relies on random sampling of unknown drug-disease pairs to construct negative samples. However, this method may inadvertently introduce unwanted noise or errors by erroneously categorizing some genuine DDAs as negative samples, thereby leaving room for improvement in current methodologies. In this paper, we introduce a novel negative sample selection algorithm for DDAs prediction that explicitly incorporates causal knowledge inherent in DDAs. To accomplish this, we first construct a heterogeneous information network (HIN) that encompasses various biological entities associated with DDAs and their interconnections. Subsequently, we utilize the outcomes of community detection within the HIN as a form of counterfactual inference, resulting in the development of a negative sample selection algorithm based on a thoughtfully designed counterfactual question. By combining the known DDAs (i.e., positive samples) with the newly generated negative samples, we train a prediction model that incorporates a graph learning module to acquire representations of drugs and diseases. Comprehensive experiments confirm the effectiveness of our proposed model for DDAs prediction.

Drug-drug interaction (DDI) pertains to the occurrence where the concomitant use of two or more drugs may lead to interactions in terms of their pharmacokinetic or pharmacodynamic behavior, resulting in unexpected effects. Accurately predicting DDIs holds significant importance in ensuring drug safety. Despite the numerous approaches proposed for DDI prediction, a majority of these methods often overlook the challenge presented by cold-start scenario, consequently limiting their applicability. This paper presents a novel data augmentation approach for the prediction of DDIs in cold-start scenarios. This method leverages counterfactual inference to generate meaningful pseudo samples for drugs with limited prior information. To achieve this, a HIN relevant to DDIs is initially established by amalgamating various associations between drugs and proteins. Subsequently, the identification of drug communities within this HIN is regarded as a form of counterfactual inference treatment, facilitating the generation of counterfactual links for cold-start drugs and thereby augmenting the training dataset. Lastly, we enhance our understanding of drug characteristics through a meta-path-based fusion mechanism, ultimately improving the accuracy of DDIs prediction in cold-start scenarios. We substantiate the effectiveness of our proposed method through an extensive series of experiments.

Online metric learning has been widely exploited for large-scale data classification due to the low computational cost. However, amongst online practical scenarios where the features are evolving (e.g., some features are vanished and some new features are augmented), most metric learning models cannot be successfully applied to these scenarios, although they can tackle the evolving instances efficiently. To address the challenge, we develop a new online Evolving Metric Learning (EML) model for incremental and decremental features, which can handle the instance and feature evolutions simultaneously by incorporating with a smoothed Wasserstein metric distance. Specifically, our model contains two essential stages: a Transforming stage (T-stage) and a Inheriting stage (I-stage). For the T-stage, we propose to extract important information from vanished features while neglecting non-informative knowledge, and forward it into survived features by transforming them into a low-rank discriminative metric space. It further explores the intrinsic low-rank structure of heterogeneous samples to reduce the computation and memory burden especially for highly-dimensional large-scale data. For the I-stage, we inherit the metric performance of survived features from the T-stage and then expand to include the new augmented features. Moreover, a smoothed Wasserstein distance is utilized to characterize the similarity relationships among the heterogeneous and complex samples, since the evolving features are not strictly aligned in the different stages. In addition to tackling the challenges in one-shot case, we also extend our model into multi-shot scenario. After deriving an efficient optimization strategy for both T-stage and I-stage, extensive experiments on several datasets verify the superior performance of our EML model.

Recent years have witnessed the rapidly growing of the amount of semistructured documents in real-world applications. Due to the huge size of the real-world data, how to manage semistructured documents effectively is a big challenge for researchers. As a fundamental task in natural language processing field, document classification is a feasible way to handle the large-scale semistructured documents. However, existing methods fail to explicitly take advantage of the hierarchical semantics in semistructured documents. It's known that the contained semantics is beneficial for understanding the semistructured documents. Considering the hierarchical structure of a given semistructured document, we propose a semistructured document classification framework which explicitly utilizes the semantic hierarchical attention mechanism. More specifically, the hierarchical attention mechanism and graph neural network are employed to model semistructured documents, by which the multilevel semantic relationships and grammatical information are considered. Moreover, we propose an adaptive class cost learning method to treat the issue of data imbalance. Comprehensive experiments are conducted on two real-world data sets, and the results demonstrate that our framework performs better than selected baselines for semistructured document classification.

Weakly-supervised learning has attracted growing research attention on medical lesions segmentation due to significant saving in pixel-level annotation cost. However, 1) most existing methods require effective prior and constraints to explore the intrinsic lesions characterization, which only generates incorrect and rough prediction; 2) they neglect the underlying semantic dependencies among weakly-labeled target enteroscopy diseases and fully-annotated source gastroscope lesions, while forcefully utilizing untransferable dependencies leads to the negative performance. To tackle above issues, we propose a new weakly-supervised lesions transfer framework, which can not only explore transferable domain-invariant knowledge across different datasets, but also prevent the negative transfer of untransferable representations. Specifically, a Wasserstein quantified transferability framework is developed to highlight wide-range transferable contextual dependencies, while neglecting the irrelevant semantic characterizations. Moreover, a novel self-supervised pseudo label generator is designed to equally provide confident pseudo pixel labels for both hard-to-transfer and easy-to-transfer target samples. It inhibits the enormous deviation of false pseudo pixel labels under the self-supervision manner. Afterwards, dynamically-searched feature centroids are aligned to narrow category-wise distribution shift. Comprehensive theoretical analysis and experiments show the superiority of our model on the endoscopic dataset and several public datasets.

Unsupervised domain adaptation has attracted growing research attention on semantic segmentation. However, 1) most existing models cannot be directly applied into lesions transfer of medical images, due to the diverse appearances of same lesion among different datasets; 2) equal attention has been paid into all semantic representations instead of neglecting irrelevant knowledge, which leads to negative transfer of untransferable knowledge. To address these challenges, we develop a new unsupervised semantic transfer model including two complementary modules (i.e., T_D and T_F ) for endoscopic lesions segmentation, which can alternatively determine where and how to explore transferable domain-invariant knowledge between labeled source lesions dataset (e.g., gastroscope) and unlabeled target diseases dataset (e.g., enteroscopy). Specifically, T_D focuses on where to translate transferable visual information of medical lesions via residual transferability-aware bottleneck, while neglecting untransferable visual characterizations. Furthermore, T_F highlights how to augment transferable semantic features of various lesions and automatically ignore untransferable representations, which explores domain-invariant knowledge and in return improves the performance of T_D. To the end, theoretical analysis and extensive experiments on medical endoscopic dataset and several non-medical public datasets well demonstrate the superiority of our proposed model.

Biomedical document classification is a fundamental task in biomedical field. Existing methods do not make full use of the hierarchically semantic structures in biomedical documents which can be utilized to improve the performance of biomedical document classification. In this paper, according to the hierarchical structures in given biomedical documents, we propose two models for biomedical document classification, which are based on the semantically hierarchical attention mechanism. Specifically, we utilize a hierarchical attention mechanism to model biomedical documents, taking into account simultaneously multiple-level semantic relationships in documents. In addition, an adaptive cost sensitive learning method is proposed to address the data imbalance issue. Extensive experiments on two real-world datasets demonstrate the effectiveness of the proposed methods.

The state-of-the-art online learning approaches are only capable of learning the metric for predefined tasks. In this paper, we consider a lifelong learning problem to mimic "human learning," i.e., endowing a new capability to the learned metric for a new task from new online samples and incorporating the previous experiences. Therefore, we propose a new metric learning framework: lifelong metric learning (LML), which only utilizes the data of the new task to train the metric model while preserving the original capabilities. More specifically, the proposed LML maintains a common subspace for all learned metrics, named lifelong dictionary, transfers knowledge from the common subspace to learn each new metric learning task with task-specific idiosyncrasy, and redefines the common subspace over time to maximize performance across all metric tasks. For model optimization, we apply online passive aggressive optimization algorithm to achieve lifelong metric task learning, where the lifelong dictionary and task-specific partition are optimized alternatively and consecutively. Finally, we evaluate our approach by analyzing several multitask metric learning datasets. Extensive experimental results demonstrate effectiveness and efficiency of the proposed framework.

Lifelong learning intends to learn new consecutive tasks depending on previously accumulated experiences, i.e., knowledge library. However, the knowledge among different new coming tasks are imbalance. Therefore, in this paper, we try to mimic an effective "human cognition" strategy by actively sorting the importance of new tasks in the process of unknown-to-known and selecting to learn the important tasks with more information preferentially. To achieve this, we consider to assess the importance of the new coming task, i.e., unknown or not, as an outlier detection issue, and design a hierarchical dictionary learning model consisting of two-level task descriptors to sparse reconstruct each task with the l0 norm constraint. The new coming tasks are sorted depending on the sparse reconstruction score in descending order, and the task with high reconstruction score will be permitted to pass, where this mechanism is called as "watchdog." Next, the knowledge library of the lifelong learning framework encode the selected task by transferring previous knowledge, and then can also update itself with knowledge from both previously learned task and current task automatically. For model optimization, the alternating direction method is employed to solve our model and converges to a fixed point. Extensive experiments on both benchmark datasets and our own dataset demonstrate the effectiveness of our proposed model especially in task selection and dictionary learning.

Consider the lifelong machine learning problem where the objective is to learn new consecutive tasks depending on previously accumulated experiences, i.e., knowledge library. In comparison with most state-of-the-arts which adopt knowledge library with prescribed size, in this paper, we propose a new incremental clustered lifelong learning model with two libraries: feature library and model library, called Clustered Lifelong Learning (CL3), in which the feature library maintains a set of learned features common across all the encountered tasks, and the model library is learned by identifying and adding representative models (clusters). When a new task arrives, the original task model can be firstly reconstructed by representative models measured by capped l2-norm distance, i.e., effectively assigning the new task model to multiple representative models under feature library. Based on this assignment knowledge of new task, the objective of our CL3 model is to transfer the knowledge from both feature library and model library to learn the new task. The new task 1) with a higher outlier probability will then be judged as a new representative, and used to refine both feature library and representative models over time; 2) with lower outlier probability will only update the feature library. For the model optimisation, we cast this problem as an alternating direction minimization problem. To this end, the performance of CL3 is evaluated through comparing with most lifelong learning models, even some batch clustered multi-task learning models.

11The corresponding author is Prof. Yang Cong. This work is supported by Nature Science Foundation of China under Grant (61722311, U1613214, 61533015) and CAS-Youth Innovation Promotion Association Scholarship (2012163)Recently, many machine learning problems rely on a valuable tool: metric learning. However, in many applications, large-scale applications embedded in high-dimensional feature space may induce both computation and storage requirements to grow quadratically. In order to tackle these challenges, in this paper, we intend to establish a robust metric learning formulation with the expectation that online metric learning and parallel optimization can solve large-scale and high-dimensional data efficiently, respectively. Specifically, based on the matrix factorization strategy, the first step aims to learn a similarity function in the objective formulation for similarity measurement; in the second step, we derive a variational trace norm to promote low-rankness on the transformation matrix. After converting this variational regularization into its separable form, for the model optimization, we present an parallel block coordinate descent method to learn the optimal metric parameters, which can handle the high-dimensional data in an efficient way. Crucially, our method shares the efficiency and flexibility of block coordinate descent method, and it is also guaranteed to converge to the optimal solution. Finally, we evaluate our approach by analyzing scene categorization dataset with tens of thousands of dimensions, and the experimental results show the effectiveness of our proposed model.

Network clustering is an essential approach to finding latent clusters in real-world networks. As the scale of real-world networks becomes increasingly larger, the existing network clustering algorithms fail to discover meaningful clusters efficiently. In this paper, we propose a framework called AnySCAN, which applies anytime theory to the structural clustering algorithm for networks (SCAN). Moreover, an active learning strategy is proposed to advance the refining procedure in AnySCAN framework. AnySCAN with the active learning strategy is able to find the exactly same clustering result on large-scale networks as the original SCAN in a significantly more efficient manner. Extensive experiments on real-world and synthetic networks demonstrate that our proposed method outperforms existing network clustering approaches.

At SIGMOD 2015, an article was presented with the title “DBSCAN Revisited: Mis-Claim, Un-Fixability, and Approximation” that won the conference’s best paper award. In this technical correspondence, we want to point out some inaccuracies in the way DBSCAN was represented, and why the criticism should have been directed at the assumption about the performance of spatial index structures such as R-trees and not at an algorithm that can use such indexes. We will also discuss the relationship of DBSCAN performance and the indexability of the dataset, and discuss some heuristics for choosing appropriate DBSCAN parameters. Some indicators of bad parameters will be proposed to help guide future users of this algorithm in choosing parameters such as to obtain both meaningful results and good performance. In new experiments, we show that the new SIGMOD 2015 methods do not appear to offer practical benefits if the DBSCAN parameters are well chosen and thus they are primarily of theoretical interest. In conclusion, the original DBSCAN algorithm with effective indexes and reasonably chosen parameter values performs competitively compared to the method proposed by Gan and Tao.

BACKGROUND: The phenome represents a distinct set of information in the human population. It has been explored particularly in its relationship with the genome to identify correlations for diseases. The phenome has been also explored for drug repositioning with efforts focusing on the search space for the most similar candidate drugs. For a comprehensive analysis of the phenome, we assumed that all phenotypes (indications and side effects) were inter-connected with a probabilistic distribution and this characteristic may offer an opportunity to identify new therapeutic indications for a given drug. Correspondingly, we employed Latent Dirichlet Allocation (LDA), which introduces latent variables (topics) to govern the phenome distribution. RESULTS: We developed our model on the phenome information in Side Effect Resource (SIDER). We first developed a LDA model optimized based on its recovery potential through perturbing the drug-phenotype matrix for each of the drug-indication pairs where each drug-indication relationship was switched to "unknown" one at the time and then recovered based on the remaining drug-phenotype pairs. Of the probabilistically significant pairs, 70% was successfully recovered. Next, we applied the model on the whole phenome to narrow down repositioning candidates and suggest alternative indications. We were able to retrieve approved indications of 6 drugs whose indications were not listed in SIDER. For 908 drugs that were present with their indication information, our model suggested alternative treatment options for further investigations. Several of the suggested new uses can be supported with information from the scientific literature. CONCLUSIONS: The results demonstrated that the phenome can be further analyzed by a generative model, which can discover probabilistic associations between drugs and therapeutic uses. In this regard, LDA serves as an enrichment tool to explore new uses of existing drugs by narrowing down the search space.

BACKGROUND: Given the significant impact on public health and drug development, drug safety has been a focal point and research emphasis across multiple disciplines in addition to scientific investigation, including consumer advocates, drug developers and regulators. Such a concern and effort has led numerous databases with drug safety information available in the public domain and the majority of them contain substantial textual data. Text mining offers an opportunity to leverage the hidden knowledge within these textual data for the enhanced understanding of drug safety and thus improving public health. METHODS: In this proof-of-concept study, topic modeling, an unsupervised text mining approach, was performed on the LiverTox database developed by National Institutes of Health (NIH). The LiverTox structured one document per drug that contains multiple sections summarizing clinical information on drug-induced liver injury (DILI). We hypothesized that these documents might contain specific textual patterns that could be used to address key DILI issues. We placed the study on drug-induced acute liver failure (ALF) which was a severe form of DILI with limited treatment options. RESULTS: After topic modeling of the "Hepatotoxicity" sections of the LiverTox across 478 drug documents, we identified a hidden topic relevant to Hy's law that was a widely-accepted rule incriminating drugs with high risk of causing ALF in humans. Using this topic, a total of 127 drugs were further implicated, 77 of which had clear ALF relevant terms in the "Outcome and management" sections of the LiverTox. For the rest of 50 drugs, evidence supporting risk of ALF was found for 42 drugs from other public databases. CONCLUSION: In this case study, the knowledge buried in the textual data was extracted for identification of drugs with potential of causing ALF by applying topic modeling to the LiverTox database. The knowledge further guided identification of drugs with the similar potential and most of them could be verified and confirmed. This study highlights the utility of topic modeling to leverage information within textual drug safety databases, which provides new opportunities in the big data era to assess drug safety.

The big data analytics community has accepted MapReduce as a programming model for processing massive data on distributed systems such as a Hadoop cluster. MapReduce has been evolving to improve its performance. We identified skewed workload among workers in the MapReduce ecosystem. The problem of skewed workload is of serious concern for massive data processing. We tackled the workload balancing issue by introducing a hierarchical MapReduce, or h-MapReduce for short. h-MapReduce identifies a heavy task by a properly defined cost function. The heavy task is divided into child tasks that are distributed among available workers as a new job in MapReduce framework. The invocation of new jobs from a task poses several challenges that are addressed by h-MapReduce. Our experiments on h-MapReduce proved the performance gain over standard MapReduce for data-intensive algorithms. More specifically, the increase of the performance gain is exponential in terms of the size of the networks. In addition to the exponential performance gains, our investigations also found a negative effect of deploying h-MapReduce due to an inappropriate definition of heavy tasks, which provides us a guideline for an effective application of h-MapReduce.

Big data such as complex networks with over millions of vertices and edges is infeasible to process using conventional computation. MapReduce is a programming model that empowers us to analyze big data in a cluster of computers. In this paper we propose a Parallel Structural Clustering Algorithm for big Networks (PSCAN) in MapReduce for the detection of clusters or community structures in big networks such as Twitter. PSCAN is based on the structural clustering algorithm of SCAN, which not only finds cluster accurately, but also identifies vertices playing special roles such as hubs and outliers. An empirical evaluation of PSCAN using both real and synthetic networks demonstrated an outstanding performance in terms of accuracy and running time. We analyzed a Twitter network with over 40 million users and 1.4 billion follower/following relationships by using PSCAN on a Hadoop cluster with 15 computers. The result shows that PSCAN successfully detected interesting communities of people who share common interests.

Substantial percent of global Internet users are now actively use Twitter. In recent times, Twitter has been experiencing explosive growth, attracting celebrities consequently a growing mass of user coverage. Insights of such a social network aid researchers in understanding behavioral dynamics of the society. Though there have been attempts to study social networks, they did not scale to process social networks on the scale of Twitter user-follower network. In this paper we uncovered some of the essential properties of the complete Twitter user-follower network. The properties include degree distribution, connectivity, strength of following relationships, clustering coefficient. Our investigations showed that the Twitter user-follower network follows power-law degree distribution. We found Twitter being a connected network. The strength of the relationships among users is distributed nearly uniform on the scale of 0.0 to 1.0. Nearly 90% of the users possess '0' clustering coefficient, which refers to the least possibility to find communities in the network. In addition to the listed properties, this study found communities of users with high clustering coefficient despite many users with low clustering coefficient. A sample of the communities is validated manually for accuracy. The validation proved that the communities are representing users of similar interests. The communities found from this work yields to friend recommendations, target based advertisements, etc.

BACKGROUND: High Content Screening (HCS) has become an important tool for toxicity assessment, partly due to its advantage of handling multiple measurements simultaneously. This approach has provided insight and contributed to the understanding of systems biology at cellular level. To fully realize this potential, the simultaneously measured multiple endpoints from a live cell should be considered in a probabilistic relationship to assess the cell's condition to response stress from a treatment, which poses a great challenge to extract hidden knowledge and relationships from these measurements. METHOD: In this work, we applied a text mining method of Latent Dirichlet Allocation (LDA) to analyze cellular endpoints from in vitro HCS assays and related to the findings to in vivo histopathological observations. We measured multiple HCS assay endpoints for 122 drugs. Since LDA requires the data to be represented in document-term format, we first converted the continuous value of the measurements to the word frequency that can processed by the text mining tool. For each of the drugs, we generated a document for each of the 4 time points. Thus, we ended with 488 documents (drug-hour) each having different values for the 10 endpoints which are treated as words. We extracted three topics using LDA and examined these to identify diagnostic topics for 45 common drugs located in vivo experiments from the Japanese Toxicogenomics Project (TGP) observing their necrosis findings at 6 and 24 hours after treatment. RESULTS: We found that assay endpoints assigned to particular topics were in concordance with the histopathology observed. Drugs showing necrosis at 6 hour were linked to severe damage events such as Steatosis, DNA Fragmentation, Mitochondrial Potential, and Lysosome Mass. DNA Damage and Apoptosis were associated with drugs causing necrosis at 24 hours, suggesting an interplay of the two pathways in these drugs. Drugs with no sign of necrosis we related to the Cell Loss and Nuclear Size assays, which is suggestive of hepatocyte regeneration. CONCLUSIONS: The evidence from this study suggests that topic modeling with LDA can enable us to interpret relationships of endpoints of in vitro assays along with an in vivo histological finding, necrosis. Effectiveness of this approach may add substantially to our understanding of systems biology.

BACKGROUND: Drug repositioning offers an opportunity to revitalize the slowing drug discovery pipeline by finding new uses for currently existing drugs. Our hypothesis is that drugs sharing similar side effect profiles are likely to be effective for the same disease, and thus repositioning opportunities can be identified by finding drug pairs with similar side effects documented in U.S. Food and Drug Administration (FDA) approved drug labels. The safety information in the drug labels is usually obtained in the clinical trial and augmented with the observations in the post-market use of the drug. Therefore, our drug repositioning approach can take the advantage of more comprehensive safety information comparing with conventional de novo approach. METHOD: A probabilistic topic model was constructed based on the terms in the Medical Dictionary for Regulatory Activities (MedDRA) that appeared in the Boxed Warning, Warnings and Precautions, and Adverse Reactions sections of the labels of 870 drugs. Fifty-two unique topics, each containing a set of terms, were identified by using topic modeling. The resulting probabilistic topic associations were used to measure the distance (similarity) between drugs. The success of the proposed model was evaluated by comparing a drug and its nearest neighbor (i.e., a drug pair) for common indications found in the Indications and Usage Section of the drug labels. RESULTS: Given a drug with more than three indications, the model yielded a 75% recall, meaning 75% of drug pairs shared one or more common indications. This is significantly higher than the 22% recall rate achieved by random selection. Additionally, the recall rate grows rapidly as the number of drug indications increases and reaches 84% for drugs with 11 indications. The analysis also demonstrated that 65 drugs with a Boxed Warning, which indicates significant risk of serious and possibly life-threatening adverse effects, might be replaced with safer alternatives that do not have a Boxed Warning. In addition, we identified two therapeutic groups of drugs (Musculo-skeletal system and Anti-infective for systemic use) where over 80% of the drugs have a potential replacement with high significance. CONCLUSION: Topic modeling can be a powerful tool for the identification of repositioning opportunities by examining the adverse event terms in FDA approved drug labels. The proposed framework not only suggests drugs that can be repurposed, but also provides insight into the safety of repositioned drugs.

BACKGROUND: The Food and Drug Administration (FDA) approved drug labels contain a broad array of information, ranging from adverse drug reactions (ADRs) to drug efficacy, risk-benefit consideration, and more. However, the labeling language used to describe these information is free text often containing ambiguous semantic descriptions, which poses a great challenge in retrieving useful information from the labeling text in a consistent and accurate fashion for comparative analysis across drugs. Consequently, this task has largely relied on the manual reading of the full text by experts, which is time consuming and labor intensive. METHOD: In this study, a novel text mining method with unsupervised learning in nature, called topic modeling, was applied to the drug labeling with a goal of discovering "topics" that group drugs with similar safety concerns and/or therapeutic uses together. A total of 794 FDA-approved drug labels were used in this study. First, the three labeling sections (i.e., Boxed Warning, Warnings and Precautions, Adverse Reactions) of each drug label were processed by the Medical Dictionary for Regulatory Activities (MedDRA) to convert the free text of each label to the standard ADR terms. Next, the topic modeling approach with latent Dirichlet allocation (LDA) was applied to generate 100 topics, each associated with a set of drugs grouped together based on the probability analysis. Lastly, the efficacy of the topic modeling was evaluated based on known information about the therapeutic uses and safety data of drugs. RESULTS: The results demonstrate that drugs grouped by topics are associated with the same safety concerns and/or therapeutic uses with statistical significance (P<0.05). The identified topics have distinct context that can be directly linked to specific adverse events (e.g., liver injury or kidney injury) or therapeutic application (e.g., antiinfectives for systemic use). We were also able to identify potential adverse events that might arise from specific medications via topics. CONCLUSIONS: The successful application of topic modeling on the FDA drug labeling demonstrates its potential utility as a hypothesis generation means to infer hidden relationships of concepts such as, in this study, drug safety and therapeutic use in the study of biomedical documents.

BACKGROUND: Protein-protein interactions (PPIs) are a critical component for many underlying biological processes. A PPI network can provide insight into the mechanisms of these processes, as well as the relationships among different proteins and toxicants that are potentially involved in the processes. There are many PPI databases publicly available, each with a specific focus. The challenge is how to effectively combine their contents to generate a robust and biologically relevant PPI network. METHODS: In this study, seven public PPI databases, BioGRID, DIP, HPRD, IntAct, MINT, REACTOME, and SPIKE, were used to explore a powerful approach to combine multiple PPI databases for an integrated PPI network. We developed a novel method called k-votes to create seven different integrated networks by using values of k ranging from 1-7. Functional modules were mined by using SCAN, a Structural Clustering Algorithm for Networks. Overall module qualities were evaluated for each integrated network using the following statistical and biological measures: (1) modularity, (2) similarity-based modularity, (3) clustering score, and (4) enrichment. RESULTS: Each integrated human PPI network was constructed based on the number of votes (k) for a particular interaction from the committee of the original seven PPI databases. The performance of functional modules obtained by SCAN from each integrated network was evaluated. The optimal value for k was determined by the functional module analysis. Our results demonstrate that the k-votes method outperforms the traditional union approach in terms of both statistical significance and biological meaning. The best network is achieved at k = 2, which is composed of interactions that are confirmed in at least two PPI databases. In contrast, the traditional union approach yields an integrated network that consists of all interactions of seven PPI databases, which might be subject to high false positives. CONCLUSIONS: We determined that the k-votes method for constructing a robust PPI network by integrating multiple public databases outperforms previously reported approaches and that a value of k=2 provides the best results. The developed strategies for combining databases show promise in the advancement of network construction and modeling.

Similarity breeds connections, the principle of homophily, has been well studied in existing sociology literature. %Several studies have observed this phenomena by conducting surveys on human subjects. These studies have concluded that new ties are formed between similar individuals. This phenomenon has been used to explain several socio-psychological concepts such as segregation, community development, social mobility, etc. However, due to the nature of these studies and limitations because of involvement of human subjects, conclusions from these studies are not easily extensible in online social media. %Social media, which is becoming the infinite space for interactions, has exceeded all the expectations in terms of growth, for reasons beyond human mind. New ties are formed in social media just like the way they emerge in real-world. However, given the differences between real-world and online social media, do the same factors that govern the construction of new ties in real-world also govern the construction of new ties in social media? In other words, does homophily exist in social media? In this article, we study this extremely significant question. We propose a systematic approach by studying two online social media sites, BlogCatalog and Last.fm and report our findings along with some interesting observations.

Many systems in sciences, engineering and nature can be modeled as networks. Examples include the Internet, WWW and social networks. Finding hidden structures is important for making sense of complex networked data. In this paper we present a new network clustering method that can find clusters in an agglomerative fashion using structural similarity of vertices in the given network. Experiments conducted on real datasets demonstrate promising performance of the new method.

BACKGROUND: Biological systems can be modeled as complex network systems with many interactions between the components. These interactions give rise to the function and behavior of that system. For example, the protein-protein interaction network is the physical basis of multiple cellular functions. One goal of emerging systems biology is to analyze very large complex biological networks such as protein-protein interaction networks, metabolic networks, and regulatory networks to identify functional modules and assign functions to certain components of the system. Network modules do not occur by chance, so identification of modules is likely to capture the biologically meaningful interactions in large-scale PPI data. Unfortunately, existing computer-based clustering methods developed to find those modules are either not so accurate or too slow. RESULTS: We devised a new methodology called SCAN (Structural Clustering Algorithm for Networks) that can efficiently find clusters or functional modules in complex biological networks as well as hubs and outliers. More specifically, we demonstrated that we can find functional modules in complex networks and classify nodes into various roles based on their structures. In this study, we showed the effectiveness of our methodology using the budding yeast (Saccharomyces cerevisiae) protein-protein interaction network. To validate our clustering results, we compared our clusters with the known functions of each protein. Our predicted functional modules achieved very high purity comparing with state-of-the-art approaches. Additionally the theoretical and empirical analysis demonstrated a linear running-time of the algorithm, which is the fastest approach for networks. CONCLUSION: We compare our algorithm with well-known modularity based clustering algorithm CNM. We successfully detect functional groups that are annotated with putative GO terms. Top-10 clusters with minimum p-value theoretically prove that newly proposed algorithm partitions network more accurately then CNM. Furthermore, manual interpretations of functional groups found by SCAN show superior performance over CNM.

Squamous cell carcinoma is the most common type of head and neck cancer affecting about 30,000 Americans each year [1]. Diagnosis of tumor is backed by histopathologic examination of excised tissue in which lesion is speckled. Computer vision systems have yet to contribute significantly to the investigation of tumor areas in terms of histological slide analysis. Recently the improvements in imaging techniques led to the discovery of virtual histological slides. Virtual slides are of sufficiently high quality to generate immense interest within the research community. We describe a novel method to tackle automatic delineation of head and neck squamous cell carcinoma problem in virtual histological slides. A density-based clustering algorithm improved in this study plays a key role in the determination of the proliferative cell nuclei. The experimental results on high-resolution head and neck slides show that the proposed algorithm performed well, obtaining an average of 96% accuracy.

Network clustering (or graph partitioning) is an important task for the discovery of underlying structures in networks. Many algorithms find clusters by maximizing the number of intra-cluster edges. While such algorithms find useful and interesting structures, they tend to fail to identify and isolate two kinds of vertices that play special roles - vertices that bridge clusters (hubs) and vertices that are marginally connected to clusters (outliers). Identifying hubs is useful for applications such as viral marketing and epidemiology since hubs are responsible for spreading ideas or disease. In contrast, outliers have little or no influence, and may be isolated as noise in the data. In this paper, we proposed a novel algorithm called SCAN (Structural Clustering Algorithm for Networks), which detects clusters, hubs and outliers in networks. It clusters vertices based on a structural similarity measure. The algorithm is fast and efficient, visiting each vertex only once. An empirical evaluation of the method using both synthetic and real datasets demonstrates superior performance over other methods such as the modularity-based algorithms.

BACKGROUND: Histopathology, which is one of the most important routines of all laboratory procedures used in pathology, is decisive for the diagnosis of cancer. Experienced histopathologists review the histological slides acquired from biopsy specimen in order to outline malignant areas. Recently, improvements in imaging technologies in terms of histological image analysis led to the discovery of virtual histological slides. In this technique, a computerized microscope scans a glass slide and generates virtual slides at a resolution of 0.25 mum/pixel. As the recognition of intrinsic cancer areas is time consuming and error prone, in this study we develop a novel method to tackle automatic squamous cell carcinoma of the head and neck detection problem in high-resolution, wholly-scanned histopathological slides. RESULTS: A density-based clustering algorithm improved for this study plays a key role in the determination of the corrupted cell nuclei. Using the Support Vector Machines (SVMs) Classifier, experimental results on seven head and neck slides show that the proposed algorithm performs well, obtaining an average of 96% classification accuracy. CONCLUSION: Recent advances in imaging technology enable us to investigate cancer tissue at cellular level. In this study we focus on wholly-scanned histopathological slides of head and neck tissues. In the context of computer-aided diagnosis, delineation of malignant regions is achieved using a powerful classification algorithm, which heavily depends on the features extracted by aid of a newly proposed cell nuclei clustering technique. The preliminary experimental results demonstrate a high accuracy of the proposed method.

Memory-based collaborative filtering (CF) has been studied extensively in the literature and has proven to be successful in various types of personalized recommender systems. In this paper, we develop a probabilistic framework for memory-based CF (PMCF). While this framework has clear links with classical memory-based CF, it allows us to find principled solutions to known problems of CF-based recommender systems. In particular, we show that a probabilistic active learning method can be used to actively query the user, thereby solving the "new user problem." Furthermore, the probabilistic framework allows us to reduce the computational cost of memory-based CF by working on a carefully selected subset of user profiles, while retaining high accuracy. We report experimental results based on two real-world data sets, which demonstrate that our proposed PMCF framework allows an accurate and efficient prediction of user preferences.

The application range of memory-based collaborative filtering (CF) is limited due to CF's high memory consumption and long runtime. The approach presented in this paper removes redundant and inconsistent instances (users) from the data. This paper aims to distinguish informative instances (users) from large raw user preference database and thus alleviate the memory and runtime cost of the widely used memory-based collaborative filtering (CF) algorithm. Our work shows that a satisfactory accuracy can be achieved by using only a small portion of the original data set, thereby alleviating the storage and runtime cost of the CF algorithm. In our approach, we consider instance selection as the problem of selecting informative data that increase the We begin by discussing the instance selection problem in a general sense that is to increase a posteriori probability of the optimal model by selecting informative data. We evaluate the empirical performance of our approach PF on two real-world data sets and attain very promisingpositive experimental results. The dData size and the prediction time are significantly reduced, while the prediction accuracy is on a par with almost the same as the results achieved by using the complete database.

Text clustering methods can be used to structure large sets of text or hypertext documents. The well-known methods of text clustering, however, do not really address the special problems of text clustering: very high dimensionality of the data, very large size of the databases and understandability of the cluster description. In this paper, we introduce a novel approach which uses frequent item (term) sets for text clustering. Such frequent sets can be efficiently discovered using algorithms for association rule mining. To cluster based on frequent term sets, we measure the mutual overlap of frequent sets with respect to the sets of supporting documents. We present two algorithms for frequent term-based text clustering, FTC which creates flat clusterings and HFTC for hierarchical clustering. An experimental evaluation on classical text documents as well as on web documents demonstrates that the proposed algorithms obtain clusterings of comparable quality significantly more efficiently than state-of-the- art text clustering algorithms. Furthermore, our methods provide an understandable description of the discovered clusters by their frequent term sets.

Collaborative filtering (CF) has become an important data mining technique to make personalized recommendations for books, web pages or movies, etc. One popular algorithm is the memory-based collaborative filtering, which predicts a user's preference based on his or her similarity to other users (instances) in the database. However, the tremendous growth of users and the large number of products, memory-based CF algorithms results in the problem of deciding the right instances to use during prediction, in order to reduce executive cost and excessive storage, and possibly to improve the generalization accuracy by avoiding noise and overfitting. In this paper, we focus our work on a typical user preference database that contains many missing values, and propose four novel instance reduction techniques called TURF1-TURF4 as a preprocessing step to improve the efficiency and accuracy of the memory-based CF algorithm. The key idea is to generate prediction from a carefully selected set of relevant instances. We evaluate the techniques on the well-known EachMovie data set. Our experiments showed that the proposed algorithms not just dramatically speed up the prediction, but also improved the accuracy.

Collaborative filtering uses a database about consumers' preferences to make personal product recommendations and is achieving widespread success in both E-Commerce and Information Filtering Applications nowadays. However, the traditional collaborative filtering algorithms do not scale well to the ever-growing number of consumers. The quality of the recommendation also needs to be improved in order to gain more trust from the consumers. In this paper, we present a novel method to improve the scalability and the accuracy of the collaborative filtering algorithm. We introduce an information theoretic approach to measure the relevance of a consumer (instance) for predicting the preference for the given product (target concept). The proposed method reduces the training data set by selecting only highly relevant instances. Our experimental evaluation on the well-known EachMovie data set shows that our method doesn't only significantly speed up the prediction, but also results in a better accuracy.

Precision agriculture is a new approach to farming in which environmental characteristics at a sub-field level are used to guide crop production decisions. Instead of applying management actions and production inputs uniformly across entire fields, they are varied to match site-specific needs. A first step in this process is to define spatial regions having similar characteristics and to build local regression models describing the relationship between field characteristics and yield. From these yield prediction models, one can then determine optimum production input levels. Discovery of "similar" regions in fields is done by applying the DBSCAN clustering algorithm on data from more than one field, ignoring spatial attributes and the corresponding yield values. The experimental results on real life agriculture data show observable improvements in prediction accuracy, although there are many unresolved issues in applying the proposed method in practice.

The problem of detecting clusters of points belonging to a spatial point process arises in many applications. In this paper, we introduce the new clustering algorithm DBCLASD (Distribution-Based Clustering of LArge Spatial Databases) to discover clusters of this type. The results of experiments demonstrate that DBCLASD, contrary to partitioning algorithms such as CLARANS (Clustering Large Applications based on RANdomized Search), discovers clusters of arbitrary shape. Furthermore, DBCLASD does not require any input parameters, in contrast to the clustering algorithm DBSCAN (Density-Based Spatial Clustering of Applications with Noise) requiring two input parameters, which may be difficult to provide for large databases. In terms of efficiency, DBCLASD is between CLARANS and DBSCAN, close to DBSCAN. Thus, the efficiency of DBCLASD on large spatial databases is very attractive when considering its nonparametric nature and its good quality for clusters of arbitrary shape.