MOTIVATION: There is a plethora of measures to evaluate functional similarity (FS) of genes based on their co-expression, protein-protein interactions and sequence similarity. These measures are typically derived from hand-engineered and application-specific metrics to quantify the degree of shared information between two genes using their Gene Ontology (GO) annotations. RESULTS: We introduce deepSimDEF, a deep learning method to automatically learn FS estimation of gene pairs given a set of genes and their GO annotations. deepSimDEF's key novelty is its ability to learn low-dimensional embedding vector representations of GO terms and gene products and then calculate FS using these learned vectors. We show that deepSimDEF can predict the FS of new genes using their annotations: it outperformed all other FS measures by >5-10% on yeast and human reference datasets on protein-protein interactions, gene co-expression and sequence homology tasks. Thus, deepSimDEF offers a powerful and adaptable deep neural architecture that can benefit a wide range of problems in genomics and proteomics, and its architecture is flexible enough to support its extension to any organism. AVAILABILITY AND IMPLEMENTATION: Source code and data are available at https://github.com/ahmadpgh/deepSimDEF. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

OBJECTIVE: In biomedicine, there is a wealth of information hidden in unstructured narratives such as research articles and clinical reports. To exploit these data properly, a word sense disambiguation (WSD) algorithm prevents downstream difficulties in the natural language processing applications pipeline. Supervised WSD algorithms largely outperform un- or semisupervised and knowledge-based methods; however, they train 1 separate classifier for each ambiguous term, necessitating a large number of expert-labeled training data, an unattainable goal in medical informatics. To alleviate this need, a single model that shares statistical strength across all instances and scales well with the vocabulary size is desirable. MATERIALS AND METHODS: Built on recent advances in deep learning, our deepBioWSD model leverages 1 single bidirectional long short-term memory network that makes sense prediction for any ambiguous term. In the model, first, the Unified Medical Language System sense embeddings will be computed using their text definitions; and then, after initializing the network with these embeddings, it will be trained on all (available) training data collectively. This method also considers a novel technique for automatic collection of training data from PubMed to (pre)train the network in an unsupervised manner. RESULTS: We use the MSH WSD dataset to compare WSD algorithms, with macro and micro accuracies employed as evaluation metrics. deepBioWSD outperforms existing models in biomedical text WSD by achieving the state-of-the-art performance of 96.82% for macro accuracy. CONCLUSIONS: Apart from the disambiguation improvement and unsupervised training, deepBioWSD depends on considerably less number of expert-labeled data as it learns the target and the context terms jointly. These merit deepBioWSD to be conveniently deployable in real-time biomedical applications.

MOTIVATION: Measures of protein functional similarity are essential tools for function prediction, evaluation of protein-protein interactions (PPIs) and other applications. Several existing methods perform comparisons between proteins based on the semantic similarity of their GO terms; however, these measures are highly sensitive to modifications in the topological structure of GO, tend to be focused on specific analytical tasks and concentrate on the GO terms themselves rather than considering their textual definitions. RESULTS: We introduce simDEF, an efficient method for measuring semantic similarity of GO terms using their GO definitions, which is based on the Gloss Vector measure commonly used in natural language processing. The simDEF approach builds optimized definition vectors for all relevant GO terms, and expresses the similarity of a pair of proteins as the cosine of the angle between their definition vectors. Relative to existing similarity measures, when validated on a yeast reference database, simDEF improves correlation with sequence homology by up to 50%, shows a correlation improvement >4% with gene expression in the biological process hierarchy of GO and increases PPI predictability by > 2.5% in F1 score for molecular function hierarchy. AVAILABILITY AND IMPLEMENTATION: Datasets, results and source code are available at http://kiwi.cs.dal.ca/Software/simDEF CONTACT: [email protected] or [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

In this study we propose a new method to classify sentiments in messages posted on online forums. Traditionally, sentiment classification relies on analysis of emotionally-charged words and discourse units found in the classified text. In coherent online discussions, however, messages’ non-lexical meta-information can be sufficient to achieve reliable classification results. Our empirical evidence is obtained through multi-class classification of messages posted on a medical forum. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Feature selection is based on the notion that redundant and/or irrelevant variables bring no additional information about the data classes and can be considered noise for the predictor. As a result, the total feature set of a dataset could be minimized to only few features containing maximum discrimination information about the class. Classification accuracy is used as the evaluation measure in guiding the feature selection process. At the same time, such measure does not take into account the privacy of the resulting dataset. In this work, we incorporate privacy considerations into the very evaluation measure that is used to evaluate and select feature subsets. We consider privacy “during” the feature selection process and as such introduce a two-dimensional measure in automatic feature selection that takes into account both objectives of privacy and efficacy (e.g. accuracy) simultaneously and provides the data user with the flexibility of trading-off one for another.

A large amount of digital information collected and stored in databases creates new opportunities for knowledge discovery and data mining. The datasets, however, may contain personally identifiable information that needs to be protected. With high dimensionality of many large datasets, dimensionality reduction such as feature selection becomes indispensible. In this work, we aim at incorporating privacy into the very process of feature selection and as such, propose a privacy-aware filter-based feature selection method (PF-IFR). Our method enables data custodians to define a trade-off measure for controlling the amount of privacy and efficacy using filter-based feature selection techniques.

Abstract The user-generated Web content has been intensively analyzed in Information Extraction and Natural Language Processing research. Web-posted reviews of consumer goods are studied to find customer opinions about the products. We hypothesize that nonemotionally charged descriptions can be applied to predict those opinions. The descriptions may include indicators of product size (tall), commonplace (some), frequency of happening (often), and reviewer certainty (maybe). We first construct patterns of how the descriptions are used in consumer-written texts and then represent individual reviews through these patterns. We propose a semantic hierarchy that organizes individual words into opinion types. We run machine learning algorithms on five data sets of user-written product reviews: four are used in classification experiments, another one for regression and classification. The obtained results support the use of non-emotional descriptions in opinion learning.

OBJECTIVE: There has been a consistent concern about the inadvertent disclosure of personal information through peer-to-peer file sharing applications, such as Limewire and Morpheus. Examples of personal health and financial information being exposed have been published. We wanted to estimate the extent to which personal health information (PHI) is being disclosed in this way, and compare that to the extent of disclosure of personal financial information (PFI). DESIGN: After careful review and approval of our protocol by our institutional research ethics board, files were downloaded from peer-to-peer file sharing networks and manually analyzed for the presence of PHI and PFI. The geographic region of the IP addresses was determined, and classified as either USA or Canada. MEASUREMENT: We estimated the proportion of files that contain personal health and financial information for each region. We also estimated the proportion of search terms that return files with personal health and financial information. We ascertained and discuss the ethical issues related to this study. RESULTS: Approximately 0.4% of Canadian IP addresses had PHI, as did 0.5% of US IP addresses. There was more disclosure of financial information, at 1.7% of Canadian IP addresses and 4.7% of US IP addresses. An analysis of search terms used in these file sharing networks showed that a small percentage of the terms would return PHI and PFI files (ie, there are people successfully searching for PFI and PHI on the peer-to-peer file sharing networks). CONCLUSION: There is a real risk of inadvertent disclosure of PHI through peer-to-peer file sharing networks, although the risk is not as large as for PFI. Anyone keeping PHI on their computers should avoid installing file sharing applications on their computers, or if they have to use such tools, actively manage the risks of inadvertent disclosure of their, their family's, their clients', or patients' PHI.

We present a learning algorithm for decision lists which allows features that are constructed from the data and allows a trade-off between accuracy and complexity. We provide bounds on the generalization error of this learning algorithm in terms of the number of errors and the size of the classifier it finds on the training data. We also compare its performance on some natural data sets with the set covering machine and the support vector machine. Furthermore, we show that the proposed bounds on the generalization error provide effective guides for model selection.

We introduce a new learning algorithm for decision lists to allow features that are constructed from the data and to allow a trade- ofi between accuracy and complexity. We bound its generalization error in terms of the number of errors and the size of the classifler it flnds on the training data. We also compare its performance on some natural data sets with the set covering machine and the support vector machine.