Graphs effectively model interactions in real-world applications such as social and trade networks, where Graph Neural Networks (GNNs) excel at tasks such as link prediction to enhance user experiences. Despite these benefits, users raise privacy concerns as user data can be exploited to improve GNN performance without consent. Accordingly, various graph unlearning methods have been developed. Prior work shows that comparing models before and after unlearning enables attackers to launch former membership inference attacks (FMIA) on unlearned data. However, the imprint of unlearned data left in the unlearned model itself remains underexplored, and existing membership inference methods mainly exploit overfitting, making them ineffective for identifying unlearned data. To address this, we conducted theoretical analysis and proposed an attack framework targeting unlearned GNNs by learning the distribution patterns of unlearned data to distinguish them from normal test data. Extensive experiments on four real-world datasets and GNN architectures confirm our framework's effectiveness and reveal significant vulnerabilities in current graph unlearning methods.

Pregnancy complications are a leading cause of maternal and neonatal mortality worldwide. Understanding their underlying mechanisms is hindered by dispersed evidence across thousands of studies and complex biological interactions. We present PregBase, a comprehensive knowledge base for pregnancy research comprising automated extraction, validation, and exploration. PregBase was constructed by utilising large language models (LLMs) to extract relationships from literature; 13 LLMs were benchmarked across seven prompting strategies, with ensemble shuffle prompting outperforming single-strategy alternatives. A three-tier validation pipeline combining ontology mapping, graph neural networks, and statistical analysis produced PregKG, a knowledge graph containing 64,087 associations between 13,303 biomedical entities spanning 155 semantic types and 50 vocabularies across 8 relationship types from 8420 articles. Link prediction validated PregBase's inference capability beyond extracted knowledge, recovering established clinical interventions, reconstructing canonical hormonal pathways across maternal-placental-fetal compartments, and identifying novel biomarker candidates for preterm birth, gestational diabetes, and preeclampsia, supported by genetic and expression databases. An interactive web interface (https://pregknowledgebase.com) has been created to enable further exploration via conversational queries. This work provides a scalable foundation for systematic discovery in maternal health research.

The organization of information about genes, genetic variants, and associated diseases from scientific literature texts can facilitate progress in precision medicine. However, the vast scale of this literature demands development of automated strategies for identifying and extracting this information. Generative large language models (LLMs) represent a promising avenue for such automation. We systematically evaluate the performance of LLMs on the extraction of information relating to impacts of genetic variation on disease from the biomedical literature, considering the challenge of genetic and disease named entity recognition in Spanish-language scientific abstracts, and experimenting with a range of instruction strategies over a dataset known as GenoVarDis. We evaluate cross-linguistic prompting, and zero- and few-shot strategies, along with the optional provision of an annotation guideline and variations in the requested output format. A key finding is that the natural language of the prompt had only a limited impact on the model's performance on NER, demonstrating the feasibility of cross-linguistic information extraction. Overall, optimal results were obtained with few-shot prompting. However, we identify that generative LLMs failed to adhere to the instructions provided, leading to the over-generation and fabrication (hallucinations) of entities not appearing in the texts. We find that adding examples to the prompts and providing an overview of the expected output structure reduces hallucinated entities. Lastly, we explore the limitations of the prompting strategies and demonstrate the value of grounding generated outputs in the original texts. Overall, LLMs do not reach the accuracy of task-specific models, but we gain insight into effective strategies for their use.

Graph transformers are a recent advancement in machine learning, offering a new class of neural network models for graph-structured data. The synergy between transformers and graph learning demonstrates strong performance and versatility across various graph-related tasks. This survey provides an in-depth review of recent progress and challenges in graph transformer research. We begin with foundational concepts of graphs and transformers. We then explore design perspectives of graph transformers, focusing on how they integrate graph inductive biases and graph attention mechanisms into the transformer architecture. Furthermore, we propose a taxonomy classifying graph transformers based on depth, scalability, and pretraining strategies, summarizing key principles for effective development of graph transformer models. Beyond technical analysis, we discuss the applications of graph transformer models for node-level, edge-level, and graph-level tasks, exploring their potential in other application scenarios as well. Finally, we identify remaining challenges in the field, such as scalability and efficiency, generalization and robustness, interpretability and explainability, dynamic and complex graphs, as well as data quality and diversity, charting future directions for graph transformer research.

Artificial intelligence (AI) is a multidisciplinary scientific field that uses machines to solve real-world problems and predict outcomes. Despite the current enthusiasm about AI's potential as a clinical support tool, there is also a growing awareness and concern about the potentially harmful effects of AI. Because AI will likely impact expert-based decision-making in medicine, it is critical to consider the issues that AI raises in medical research. This paper outlines the AI guidelines of the Australian Epilepsy Project. This large-scale platform aims to democratise specialist care in epilepsy and use AI for clinical decision support based on prospective multimodal datasets (MRI, genetic, clinical, and cognitive data) from thousands of people with epilepsy. As AI develops rapidly, we focus on key areas of medical AI identified in the literature, including Trust, Responsibility and Safety. We believe AI is changing medicine, and we believe it is imperative to advance and update our AI guidelines adaptably while preparing for an era of augmented-intelligence-based medicine.

Clinical predictive analysis is a crucial task with numerous applications and has been extensively studied using machine learning approaches. Clinical notes, a vital data source, have been employed to develop natural language processing (NLP) models for risk prediction in healthcare with robust performance. However, clinical notes vary considerably in text composition-written by diverse healthcare providers for different purposes-and the impact of these variations on NLP modeling is also underexplored. It also remains uncertain whether the recent Large Language Models (LLMs) with instruction-following capabilities can effectively handle the risk prediction task out-of-the-box, especially when using routinely collected clinical notes instead of polished text. We address these two important research questions in the context of in-hospital mortality prediction within the critical care setting. Specifically, we propose a supervised hierarchical network with note-specific modules to account for variations across different note categories, and provide a detailed comparison with strong supervised baselines and LLMs. We benchmark 34 instruction-following LLMs based on zero-shot, few-shot, and chain-of-thought prompting with diverse prompt templates. Our results demonstrate that the note-specific network delivers improved risk prediction performance compared to established supervised baselines from both measurement-based and text-based modeling. In contrast, LLMs consistently underperform on this critical task, despite their remarkable performances in other domains. This highlights important limitations and raises caution regarding the use of LLMs for risk assessment in the critical setting. Additionally, we show that the proposed model can be leveraged to select informative clinical notes to enhance the training of other models.

Aparna Elangovan, Jiayuan He, Yuan Li, Karin Verspoor. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Thinh Hung Truong, Yulia Otmakhova, Karin Verspoor, Trevor Cohn, Timothy Baldwin. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

MOTIVATION: Biological background knowledge plays an important role in the manual quality assurance (QA) of biological database records. One such QA task is the detection of inconsistencies in literature-based Gene Ontology Annotation (GOA). This manual verification ensures the accuracy of the GO annotations based on a comprehensive review of the literature used as evidence, Gene Ontology (GO) terms, and annotated genes in GOA records. While automatic approaches for the detection of semantic inconsistencies in GOA have been developed, they operate within predetermined contexts, lacking the ability to leverage broader evidence, especially relevant domain-specific background knowledge. This paper investigates various types of background knowledge that could improve the detection of prevalent inconsistencies in GOA. In addition, the paper proposes several approaches to integrate background knowledge into the automatic GOA inconsistency detection process. RESULTS: We have extended a previously developed GOA inconsistency dataset with several kinds of GOA-related background knowledge, including GeneRIF statements, biological concepts mentioned within evidence texts, GO hierarchy and existing GO annotations of the specific gene. We have proposed several effective approaches to integrate background knowledge as part of the automatic GOA inconsistency detection process. The proposed approaches can improve automatic detection of self-consistency and several of the most prevalent types of inconsistencies. This is the first study to explore the advantages of utilizing background knowledge and to propose a practical approach to incorporate knowledge in automatic GOA inconsistency detection. We establish a new benchmark for performance on this task. Our methods may be applicable to various tasks that involve incorporating biological background knowledge. AVAILABILITY AND IMPLEMENTATION: https://github.com/jiyuc/de-inconsistency.

Contrastive learning, which learns data representations by contrasting similar and dissimilar instances, has achieved great success in various domains including natural language processing (NLP). Recently, it has been demonstrated that incorporating class labels into contrastive learning, i.e., supervised contrastive learning (SCL), can further enhance the quality of the learned data representations. Although several works have shown empirically that incorporating SCL into classification models leads to better performance, the mechanism of how SCL works for classification is less studied. In this paper, we first investigate how SCL facilitates the classifier learning, where we show that the contrastive region, i.e., the data instances involved in each contrasting operation, has a crucial link to the mechanism of SCL. We reveal that the vanilla SCL is suboptimal since its behavior can be altered by variances in class distributions. Based on this finding, we propose aFocusedContrastiveLoss (FoCL) for classification. Compared with SCL, FoCL defines a finer contrastive region, focusing on the data instances surrounding decision boundaries. We conduct extensive experiments on three NLP tasks: text classification, named entity recognition, and relation extraction. Experimental results show consistent and significant improvements of FoCL over strong baselines on various benchmark datasets, especially in few-shot scenarios.

OBJECTIVE: With the increasing amount and growing variety of healthcare data, multimodal machine learning supporting integrated modeling of structured and unstructured data is an increasingly important tool for clinical machine learning tasks. However, it is non-trivial to manage the differences in dimensionality, volume, and temporal characteristics of data modalities in the context of a shared target task. Furthermore, patients can have substantial variations in the availability of data, while existing multimodal modeling methods typically assume data completeness and lack a mechanism to handle missing modalities. METHODS: We propose a Transformer-based fusion model with modality-specific tokens that summarize the corresponding modalities to achieve effective cross-modal interaction accommodating missing modalities in the clinical context. The model is further refined by inter-modal, inter-sample contrastive learning to improve the representations for better predictive performance. We denote the model as Attention-based cRoss-MOdal fUsion with contRast (ARMOUR). We evaluate ARMOUR using two input modalities (structured measurements and unstructured text), six clinical prediction tasks, and two evaluation regimes, either including or excluding samples with missing modalities. RESULTS: Our model shows improved performances over unimodal or multimodal baselines in both evaluation regimes, including or excluding patients with missing modalities in the input. The contrastive learning improves the representation power and is shown to be essential for better results. The simple setup of modality-specific tokens enables ARMOUR to handle patients with missing modalities and allows comparison with existing unimodal benchmark results. CONCLUSION: We propose a multimodal model for robust clinical prediction to achieve improved performance while accommodating patients with missing modalities. This work could inspire future research to study the effective incorporation of multiple, more complex modalities of clinical data into a single model.

OBJECTIVE: We explore the framing of literature-based discovery (LBD) as link prediction and graph embedding learning, with Alzheimer's Disease (AD) as our focus disease context. The key link prediction setting of prediction window length is specifically examined in the context of a time-sliced evaluation methodology. METHODS: We propose a four-stage approach to explore literature-based discovery for Alzheimer's Disease, creating and analyzing a knowledge graph tailored to the AD context, and predicting and evaluating new knowledge based on time-sliced link prediction. The first stage is to collect an AD-specific corpus. The second stage involves constructing an AD knowledge graph with identified AD-specific concepts and relations from the corpus. In the third stage, 20 pairs of training and testing datasets are constructed with the time-slicing methodology. Finally, we infer new knowledge with graph embedding-based link prediction methods. We compare different link prediction methods in this context. The impact of limiting prediction evaluation of LBD models in the context of short-term and longer-term knowledge evolution for Alzheimer's Disease is assessed. RESULTS: We constructed an AD corpus of over 16 k papers published in 1977-2021, and automatically annotated it with concepts and relations covering 11 AD-specific semantic entity types. The knowledge graph of Alzheimer's Disease derived from this resource consisted of ∼11 k nodes and ∼394 k edges, among which 34% were genotype-phenotype relationships, 57% were genotype-genotype relationships, and 9% were phenotype-phenotype relationships. A Structural Deep Network Embedding (SDNE) model consistently showed the best performance in terms of returning the most confident set of link predictions as time progresses over 20 years. A huge improvement in model performance was observed when changing the link prediction evaluation setting to consider a more distant future, reflecting the time required for knowledge accumulation. CONCLUSION: Neural network graph-embedding link prediction methods show promise for the literature-based discovery context, although the prediction setting is extremely challenging, with graph densities of less than 1%. Varying prediction window length on the time-sliced evaluation methodology leads to hugely different results and interpretations of LBD studies. Our approach can be generalized to enable knowledge discovery for other diseases. AVAILABILITY: Code, AD ontology, and data are available at https://github.com/READ-BioMed/readbiomed-lbd.

Invasive fungal infections (IFIs) are particularly dangerous to high-risk patients with haematological malignancies and are responsible for excessive mortality and delays in cancer therapy. Surveillance of IFI in clinical settings offers an opportunity to identify potential risk factors and evaluate new therapeutic strategies. However, manual surveillance is both time- and resource-intensive. As part of a broader project aimed to develop a system for automated IFI surveillance by leveraging electronic medical records, we present our approach to detecting evidence of IFI in the key diagnostic domain of histopathology. Using natural language processing (NLP), we analysed cytology and histopathology reports to identify IFI-positive reports. We compared a conventional bag-of-words classification model to a method that relies on concept-level annotations. Although the investment to prepare data supporting concept annotations is substantial, extracting targeted information specific to IFI as a pre-processing step increased the performance of the classifier from the PR AUC of 0.84 to 0.92 and enabled model interpretability. We have made publicly available the annotated dataset of 283 reports, the Cytology and Histopathology IFI Reports corpus (CHIFIR), to allow the clinical NLP research community to further build on our results.

Abstract Despite the subjective nature of semantic textual similarity (STS) and pervasive disagreements in STS annotation, existing benchmarks have used averaged human ratings as gold standard. Averaging masks the true distribution of human opinions on examples of low agreement, and prevents models from capturing the semantic vagueness that the individual ratings represent. In this work, we introduce USTS, the first Uncertainty-aware STS dataset with ∼15,000 Chinese sentence pairs and 150,000 labels, to study collective human opinions in STS. Analysis reveals that neither a scalar nor a single Gaussian fits a set of observed judgments adequately. We further show that current STS models cannot capture the variance caused by human disagreement on individual instances, but rather reflect the predictive confidence over the aggregate dataset.

Although multi-document summarization (MDS) of the biomedical literature is a highly valuable task that has recently attracted substantial interest, evaluation of the quality of biomedical summaries lacks consistency and transparency.In this paper, using systematic reviews as an example of biomedical MDS, we examine the summaries generated by two current models in order to understand the deficiencies of existing evaluation approaches in the context of the challenges that arise in the MDS task.Based on this analysis, we propose a new approach to human evaluation and identify several challenges that must be overcome to develop effective biomedical MDS systems.

Training with noisy labelled data is known to be detrimental to model performance, especially for high-capacity neural network models in low-resource domains. Our experiments suggest that standard regularisation strategies, such as weight decay and dropout, are ineffective in the face of noisy labels. We propose a simple noisy label detection method that prevents error propagation from the input layer. The approach is based on the observation that the projection of noisy labels is learned through memorisation at advanced stages of learning, and that the Pearson correlation is sensitive to outliers. Extensive experiments over real-world human-disagreement annotations as well as randomly-corrupted and data-augmented labels, across various tasks and domains, demonstrate that our method is effective, regularising noisy labels and improving generalisation performance.

Automatic generation of ophthalmic reports using datadriven neural networks has great potential in clinical practice. When writing a report, ophthalmologists make inferences with prior clinical knowledge. This knowledge has been neglected in prior medical report generation methods. To endow models with the capability of incorporating expert knowledge, we propose a Cross-modal clinical Graph Transformer (CGT) for ophthalmic report generation (ORG), in which clinical relation triples are injected into the visual features as prior knowledge to drive the decoding procedure. However, two major common Knowledge Noise (KN) issues may affect models' effectiveness. 1) Existing general biomedical knowledge bases such as the UMLS may not align meaningfully to the specific context and language of the report, limiting their utility for knowledge injection. 2) Incorporating too much knowledge may divert the visual features from their correct meaning. To overcome these limitations, we design an automatic information extraction scheme based on natural language processing to obtain clinical entities and relations directly from in-domain training reports. Given a set of ophthalmic images, our CGT first restores a sub-graph from the clinical graph and injects the restored triples into visual features. Then visible matrix is employed during the encoding procedure to limit the impact of knowledge. Finally, reports are predicted by the encoded cross-modal features via a Transformer decoder. Extensive experiments on the large-scale FFA-IR benchmark demonstrate that the proposed CGT is able to outperform previous benchmark methods and achieve state-of-the-art performances.

Thinh Truong, Timothy Baldwin, Trevor Cohn, Karin Verspoor. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Nucleotide and protein sequences stored in public databases are the cornerstone of many bioinformatics analyses. The records containing these sequences are prone to a wide range of errors, including incorrect functional annotation, sequence contamination and taxonomic misclassification. One source of information that can help to detect errors are the strong interdependency between records. Novel sequences in one database draw their annotations from existing records, may generate new records in multiple other locations and will have varying degrees of similarity with existing records across a range of attributes. A network perspective of these relationships between sequence records, within and across databases, offers new opportunities to detect-or even correct-erroneous entries and more broadly to make inferences about record quality. Here, we describe this novel perspective of sequence database records as a rich network, which we call the sequence database network, and illustrate the opportunities this perspective offers for quantification of database quality and detection of spurious entries. We provide an overview of the relevant databases and describe how the interdependencies between sequence records across these databases can be exploited by network analyses. We review the process of sequence annotation and provide a classification of sources of error, highlighting propagation as a major source. We illustrate the value of a network perspective through three case studies that use network analysis to detect errors, and explore the quality and quantity of critical relationships that would inform such network analyses. This systematic description of a network perspective of sequence database records provides a novel direction to combat the proliferation of errors within these critical bioinformatics resources.

MOTIVATION: Literature-based gene ontology annotations (GOA) are biological database records that use controlled vocabulary to uniformly represent gene function information that is described in the primary literature. Assurance of the quality of GOA is crucial for supporting biological research. However, a range of different kinds of inconsistencies in between literature as evidence and annotated GO terms can be identified; these have not been systematically studied at record level. The existing manual-curation approach to GOA consistency assurance is inefficient and is unable to keep pace with the rate of updates to gene function knowledge. Automatic tools are therefore needed to assist with GOA consistency assurance. This article presents an exploration of different GOA inconsistencies and an early feasibility study of automatic inconsistency detection. RESULTS: We have created a reliable synthetic dataset to simulate four realistic types of GOA inconsistency in biological databases. Three automatic approaches are proposed. They provide reasonable performance on the task of distinguishing the four types of inconsistency and are directly applicable to detect inconsistencies in real-world GOA database records. Major challenges resulting from such inconsistencies in the context of several specific application settings are reported. This is the first study to introduce automatic approaches that are designed to address the challenges in current GOA quality assurance workflows. The data underlying this article are available in Github at https://github.com/jiyuc/AutoGOAConsistency.

MOTIVATION: Survival risk prediction using gene expression data is important in making treatment decisions in cancer. Standard neural network (NN) survival analysis models are black boxes with a lack of interpretability. More interpretable visible neural network architectures are designed using biological pathway knowledge. But they do not model how pathway structures can change for particular cancer types. RESULTS: We propose a novel Mutated Pathway Visible Neural Network (MPVNN) architecture, designed using prior signaling pathway knowledge and random replacement of known pathway edges using gene mutation data simulating signal flow disruption. As a case study, we use the PI3K-Akt pathway and demonstrate overall improved cancer-specific survival risk prediction of MPVNN over other similar-sized NN and standard survival analysis methods. We show that trained MPVNN architecture interpretation, which points to smaller sets of genes connected by signal flow within the PI3K-Akt pathway that is important in risk prediction for particular cancer types, is reliable. AVAILABILITY AND IMPLEMENTATION: The data and code are available at https://github.com/gourabghoshroy/MPVNN. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Protein-protein interactions (PPIs) are critical to normal cellular function and are related to many disease pathways. A range of protein functions are mediated and regulated by protein interactions through post-translational modifications (PTM). However, only 4% of PPIs are annotated with PTMs in biological knowledge databases such as IntAct, mainly performed through manual curation, which is neither time- nor cost-effective. Here we aim to facilitate annotation by extracting PPIs along with their pairwise PTM from the literature by using distantly supervised training data using deep learning to aid human curation. METHOD: We use the IntAct PPI database to create a distant supervised dataset annotated with interacting protein pairs, their corresponding PTM type, and associated abstracts from the PubMed database. We train an ensemble of BioBERT models-dubbed PPI-BioBERT-x10-to improve confidence calibration. We extend the use of ensemble average confidence approach with confidence variation to counteract the effects of class imbalance to extract high confidence predictions. RESULTS AND CONCLUSION: The PPI-BioBERT-x10 model evaluated on the test set resulted in a modest F1-micro 41.3 (P =5 8.1, R = 32.1). However, by combining high confidence and low variation to identify high quality predictions, tuning the predictions for precision, we retained 19% of the test predictions with 100% precision. We evaluated PPI-BioBERT-x10 on 18 million PubMed abstracts and extracted 1.6 million (546507 unique PTM-PPI triplets) PTM-PPI predictions, and filter [Formula: see text] (4584 unique) high confidence predictions. Of the 5700, human evaluation on a small randomly sampled subset shows that the precision drops to 33.7% despite confidence calibration and highlights the challenges of generalisability beyond the test set even with confidence calibration. We circumvent the problem by only including predictions associated with multiple papers, improving the precision to 58.8%. In this work, we highlight the benefits and challenges of deep learning-based text mining in practice, and the need for increased emphasis on confidence calibration to facilitate human curation efforts.

Electronic medical records are increasingly used to store patient information in hospitals and other clinical settings. There has been a corresponding proliferation of clinical natural language processing (cNLP) systems aimed at using text data in these records to improve clinical decision-making, in comparison to manual clinician search and clinical judgment alone. However, these systems have delivered marginal practical utility and are rarely deployed into healthcare settings, leading to proposals for technical and structural improvements. In this paper, we argue that this reflects a violation of Friedman's "Fundamental Theorem of Biomedical Informatics," and that a deeper epistemological change must occur in the cNLP field, as a parallel step alongside any technical or structural improvements. We propose that researchers shift away from designing cNLP systems independent of clinical needs, in which cNLP tasks are ends in themselves-"tasks as decisions"-and toward systems that are directly guided by the needs of clinicians in realistic decision-making contexts-"tasks as needs." A case study example illustrates the potential benefits of developing cNLP systems that are designed to more directly support clinical needs.

OBJECTIVE: Accurate identification of self-harm presentations to Emergency Departments (ED) can lead to more timely mental health support, aid in understanding the burden of suicidal intent in a population, and support impact evaluation of public health initiatives related to suicide prevention. Given lack of manual self-harm reporting in ED, we aim to develop an automated system for the detection of self-harm presentations directly from ED triage notes. MATERIALS AND METHODS: We frame this as supervised classification using natural language processing (NLP), utilizing a large data set of 477 627 free-text triage notes from ED presentations in 2012-2018 to The Royal Melbourne Hospital, Australia. The data were highly imbalanced, with only 1.4% of triage notes relating to self-harm. We explored various preprocessing techniques, including spelling correction, negation detection, bigram replacement, and clinical concept recognition, and several machine learning methods. RESULTS: Our results show that machine learning methods dramatically outperform keyword-based methods. We achieved the best results with a calibrated Gradient Boosting model, showing 90% Precision and 90% Recall (PR-AUC 0.87) on blind test data. Prospective validation of the model achieves similar results (88% Precision; 89% Recall). DISCUSSION: ED notes are noisy texts, and simple token-based models work best. Negation detection and concept recognition did not change the results while bigram replacement significantly impaired model performance. CONCLUSION: This first NLP-based classifier for self-harm in ED notes has practical value for identifying patients who would benefit from mental health follow-up in ED, and for supporting surveillance of self-harm and suicide prevention efforts in the population.

One unintended consequence of the Electronic Health Records (EHR) implementation is the overuse of content-importing technology, such as copy-and-paste, that creates "bloated" notes containing large amounts of textual redundancy. Despite the rising interest in applying machine learning models to learn from real-patient data, it is unclear how the phenomenon of note bloat might affect the Natural Language Processing (NLP) models derived from these notes. Therefore, in this work we examine the impact of redundancy on deep learning-based NLP models, considering four clinical prediction tasks using a publicly available EHR database. We applied two deduplication methods to the hospital notes, identifying large quantities of redundancy, and found that removing the redundancy usually has little negative impact on downstream performances, and can in certain circumstances assist models to achieve significantly better results. We also showed it is possible to attack model predictions by simply adding note duplicates, causing changes of correct predictions made by trained models into wrong predictions. In conclusion, we demonstrated that EHR text redundancy substantively affects NLP models for clinical prediction tasks, showing that the awareness of clinical contexts and robust modeling methods are important to create effective and reliable NLP systems in healthcare contexts.

Abstract State-of-the-art classification and regression models are often not well calibrated, and cannot reliably provide uncertainty estimates, limiting their utility in safety-critical applications such as clinical decision-making. While recent work has focused on calibration of classifiers, there is almost no work in NLP on calibration in a regression setting. In this paper, we quantify the calibration of pre- trained language models for text regression, both intrinsically and extrinsically. We further apply uncertainty estimates to augment training data in low-resource domains. Our experiments on three regression tasks in both self-training and active-learning settings show that uncertainty estimation can be used to increase overall performance and enhance model generalization.

Public datasets are often used to evaluate the efficacy and generalizability of state-of-the-art methods for many tasks in natural language processing (NLP).However, the presence of overlap between the train and test datasets can lead to inflated results, inadvertently evaluating the model's ability to memorize and interpreting it as the ability to generalize.In addition, such data sets may not provide an effective indicator of the performance of these methods in real world scenarios.We identify leakage of training data into test data on several publicly available datasets used to evaluate NLP tasks, including named entity recognition and relation extraction, and study them to assess the impact of that leakage on the model's ability to memorize versus generalize.The human SHBG proximal promoter was analyzed by DNase I footprinting, and the functional significance of 6 footprinted regions (FP1-FP6) within the proximal promoter was studied in human HepG2 hepatoblastoma cells.100.00

Biaoyan Fang, Christian Druckenbrodt, Saber A Akhondi, Jiayuan He, Timothy Baldwin, Karin Verspoor. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

Haplotype phasing is a critical step for many genetic applications but incorrect estimates of phase can negatively impact downstream analyses. One proposed strategy to improve phasing accuracy is to combine multiple independent phasing estimates to overcome the limitations of any individual estimate. However, such a strategy is yet to be thoroughly explored. This study provides a comprehensive evaluation of consensus strategies for haplotype phasing. We explore the performance of different consensus paradigms, and the effect of specific constituent tools, across several datasets with different characteristics and their impact on the downstream task of genotype imputation. Based on the outputs of existing phasing tools, we explore two different strategies to construct haplotype consensus estimators: voting across outputs from multiple phasing tools and multiple outputs of a single non-deterministic tool. We find that the consensus approach from multiple tools reduces SE by an average of 10% compared to any constituent tool when applied to European populations and has the highest accuracy regardless of population ethnicity, sample size, variant density or variant frequency. Furthermore, the consensus estimator improves the accuracy of the downstream task of genotype imputation carried out by the widely used Minimac3, pbwt and BEAGLE5 tools. Our results provide guidance on how to produce the most accurate phasing estimates and the trade-offs that a consensus approach may have. Our implementation of consensus haplotype phasing, consHap, is available freely at https://github.com/ziadbkh/consHap. Supplementary information: Supplementary data are available at Briefings in Bioinformatics online.

MOTIVATION: Inferring gene regulatory networks (GRNs) from expression data is a significant systems biology problem. A useful inference algorithm should not only unveil the global structure of the regulatory mechanisms but also the details of regulatory interactions such as edge direction (from regulator to target) and sign (activation/inhibition). Many popular GRN inference algorithms cannot infer edge signs, and those that can infer signed GRNs cannot simultaneously infer edge directions or network cycles. RESULTS: To address these limitations of existing algorithms, we propose Polynomial Lasso Bagging (PoLoBag) for signed GRN inference with both edge directions and network cycles. PoLoBag is an ensemble regression algorithm in a bagging framework where Lasso weights estimated on bootstrap samples are averaged. These bootstrap samples incorporate polynomial features to capture higher-order interactions. Results demonstrate that PoLoBag is consistently more accurate for signed inference than state-of-the-art algorithms on simulated and real-world expression datasets. AVAILABILITY AND IMPLEMENTATION: Algorithm and data are freely available at https://github.com/gourabghoshroy/PoLoBag. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Literature-based gene ontology (GO) annotation is a process where expert curators use uniform expressions to describe gene functions reported in research papers, creating computable representations of information about biological systems. Manual assurance of consistency between GO annotations and the associated evidence texts identified by expert curators is reliable but time-consuming, and is infeasible in the context of rapidly growing biological literature. A key challenge is maintaining consistency of existing GO annotations as new studies are published and the GO vocabulary is updated. RESULTS: In this work, we introduce a formalisation of biological database annotation inconsistencies, identifying four distinct types of inconsistency. We propose a novel and efficient method using state-of-the-art text mining models to automatically distinguish between consistent GO annotation and the different types of inconsistent GO annotation. We evaluate this method using a synthetic dataset generated by directed manipulation of instances in an existing corpus, BC4GO. We provide detailed error analysis for demonstrating that the method achieves high precision on more confident predictions. CONCLUSIONS: Two models built using our method for distinct annotation consistency identification tasks achieved high precision and were robust to updates in the GO vocabulary. Our approach demonstrates clear value for human-in-the-loop curation scenarios.

We present our work on aligning the Unified Medical Language System (UMLS) to Wikipedia, to facilitate manual alignment of the two resources.We propose a cross-lingual neural reranking model to match a UMLS concept with a Wikipedia page, which achieves a recall@1 of 72%, a substantial improvement of 20% over word-and char-level BM25, enabling manual alignment with minimal effort.We release our resources, including ranked Wikipedia pages for 700k UMLS concepts, and WikiUMLS, a dataset for training and evaluation of alignment models between UMLS and Wikipedia collected from Wikidata.This will provide easier access to Wikipedia for health professionals, patients, and NLP systems, including in multilingual settings.1

BACKGROUND: Knowledge of phase, the specific allele sequence on each copy of homologous chromosomes, is increasingly recognized as critical for detecting certain classes of disease-associated mutations. One approach for detecting such mutations is through phased haplotype association analysis. While the accuracy of methods for phasing genotype data has been widely explored, there has been little attention given to phasing accuracy at haplotype block scale. Understanding the combined impact of the accuracy of phasing tool and the method used to determine haplotype blocks on the error rate within the determined blocks is essential to conduct accurate haplotype analyses. RESULTS: We present a systematic study exploring the relationship between seven widely used phasing methods and two common methods for determining haplotype blocks. The evaluation focuses on the number of haplotype blocks that are incorrectly phased. Insights from these results are used to develop a haplotype estimator based on a consensus of three tools. The consensus estimator achieved the most accurate phasing in all applied tests. Individually, EAGLE2, BEAGLE and SHAPEIT2 alternate in being the best performing tool in different scenarios. Determining haplotype blocks based on linkage disequilibrium leads to more correctly phased blocks compared to a sliding window approach. We find that there is little difference between phasing sections of a genome (e.g. a gene) compared to phasing entire chromosomes. Finally, we show that the location of phasing error varies when the tools are applied to the same data several times, a finding that could be important for downstream analyses. CONCLUSIONS: The choice of phasing and block determination algorithms and their interaction impacts the accuracy of phased haplotype blocks. This work provides guidance and evidence for the different design choices needed for analyses using haplotype blocks. The study highlights a number of issues that may have limited the replicability of previous haplotype analysis.

BACKGROUND: The large biological databases such as GenBank contain vast numbers of records, the content of which is substantively based on external resources, including published literature. Manual curation is used to establish whether the literature and the records are indeed consistent. We explore in this paper an automated method for assessing the consistency of biological assertions, to assist biocurators, which we call BARC, Biocuration tool for Assessment of Relation Consistency. In this method a biological assertion is represented as a relation between two objects (for example, a gene and a disease); we then use our novel set-based relevance algorithm SaBRA to retrieve pertinent literature, and apply a classifier to estimate the likelihood that this relation (assertion) is correct. RESULTS: Our experiments on assessing gene-disease relations and protein-protein interactions using the PubMed Central collection show that BARC can be effective at assisting curators to perform data cleansing. Specifically, the results obtained showed that BARC substantially outperforms the best baselines, with an improvement of F-measure of 3.5% and 13%, respectively, on gene-disease relations and protein-protein interactions. We have additionally carried out a feature analysis that showed that all feature types are informative, as are all fields of the documents. CONCLUSIONS: BARC provides a clear benefit for the biocuration community, as there are no prior automated tools for identifying inconsistent assertions in large-scale biological databases.

BACKGROUND: Given the importance of relation or event extraction from biomedical research publications to support knowledge capture and synthesis, and the strong dependency of approaches to this information extraction task on syntactic information, it is valuable to understand which approaches to syntactic processing of biomedical text have the highest performance. RESULTS: We perform an empirical study comparing state-of-the-art traditional feature-based and neural network-based models for two core natural language processing tasks of part-of-speech (POS) tagging and dependency parsing on two benchmark biomedical corpora, GENIA and CRAFT. To the best of our knowledge, there is no recent work making such comparisons in the biomedical context; specifically no detailed analysis of neural models on this data is available. Experimental results show that in general, the neural models outperform the feature-based models on two benchmark biomedical corpora GENIA and CRAFT. We also perform a task-oriented evaluation to investigate the influences of these models in a downstream application on biomedical event extraction, and show that better intrinsic parsing performance does not always imply better extrinsic event extraction performance. CONCLUSION: We have presented a detailed empirical study comparing traditional feature-based and neural network-based models for POS tagging and dependency parsing in the biomedical context, and also investigated the influence of parser selection for a biomedical event extraction downstream task. AVAILABILITY OF DATA AND MATERIALS: We make the retrained models available at https://github.com/datquocnguyen/BioPosDep .

BACKGROUND: Drug repositioning is the process of identifying new uses for existing drugs. Computational drug repositioning methods can reduce the time, costs and risks of drug development by automating the analysis of the relationships in pharmacology networks. Pharmacology networks are large and heterogeneous. Clustering drugs into small groups can simplify large pharmacology networks, these subgroups can also be used as a starting point for repositioning drugs. In this paper, we propose a two-tiered drug-centric unsupervised clustering approach for drug repositioning, integrating heterogeneous drug data profiles: drug-chemical, drug-disease, drug-gene, drug-protein and drug-side effect relationships. RESULTS: The proposed drug repositioning approach is threefold; (i) clustering drugs based on their homogeneous profiles using the Growing Self Organizing Map (GSOM); (ii) clustering drugs based on drug-drug relation matrices based on the previous step, considering three state-of-the-art graph clustering methods; and (iii) inferring drug repositioning candidates and assigning a confidence value for each identified candidate. In this paper, we compare our two-tiered clustering approach against two existing heterogeneous data integration approaches with reference to the Anatomical Therapeutic Chemical (ATC) classification, using GSOM. Our approach yields Normalized Mutual Information (NMI) and Standardized Mutual Information (SMI) of 0.66 and 36.11, respectively, while the two existing methods yield NMI of 0.60 and 0.64 and SMI of 22.26 and 33.59. Moreover, the two existing approaches failed to produce useful cluster separations when using graph clustering algorithms while our approach is able to identify useful clusters for drug repositioning. Furthermore, we provide clinical evidence for four predicted results (Chlorthalidone, Indomethacin, Metformin and Thioridazine) to support that our proposed approach can be reliably used to infer ATC code and drug repositioning. CONCLUSION: The proposed two-tiered unsupervised clustering approach is suitable for drug clustering and enables heterogeneous data integration. It also enables identifying reliable repositioning drug candidates with reference to ATC therapeutic classification. The repositioning drug candidates identified consistently by multiple clustering algorithms and with high confidence have a higher possibility of being effective repositioning candidates.

We explore in this paper automatic biological sequence type classification for records in biological sequence databases. The sequence type attribute provides important information about the nature of a sequence represented in a record, and is often used in search to filter out irrelevant sequences. However, the sequence type attribute is generally a non-mandatory free-text field, and thus it is subject to many errors including typos, mis-assignment, and non-assignment. In GenBank, this problem concerns roughly 18% of records, an alarming number that should worry the biocuration community. To address this problem of automatic sequence type classification, we propose the use of literature associated to sequence records as an external source of knowledge that can be leveraged for the classification task. We define a set of literature-based features and train a machine learning algorithm to classify a record into one of six primary sequence types. The main intuition behind using the literature for this task is that sequences appear to be discussed differently in scientific articles, depending on their type. The experiments we have conducted on the PubMed Central collection show that the literature is indeed an effective way to address this problem of sequence type classification. Our classification method reached an accuracy of 92.7%, and substantially outperformed two baseline approaches used for comparison.

BACKGROUND: Coreference resolution is the task of finding strings in text that have the same referent as other strings. Failures of coreference resolution are a common cause of false negatives in information extraction from the scientific literature. In order to better understand the nature of the phenomenon of coreference in biomedical publications and to increase performance on the task, we annotated the Colorado Richly Annotated Full Text (CRAFT) corpus with coreference relations. RESULTS: The corpus was manually annotated with coreference relations, including identity and appositives for all coreferring base noun phrases. The OntoNotes annotation guidelines, with minor adaptations, were used. Interannotator agreement ranges from 0.480 (entity-based CEAF) to 0.858 (Class-B3), depending on the metric that is used to assess it. The resulting corpus adds nearly 30,000 annotations to the previous release of the CRAFT corpus. Differences from related projects include a much broader definition of markables, connection to extensive annotation of several domain-relevant semantic classes, and connection to complete syntactic annotation. Tool performance was benchmarked on the data. A publicly available out-of-the-box, general-domain coreference resolution system achieved an F-measure of 0.14 (B3), while a simple domain-adapted rule-based system achieved an F-measure of 0.42. An ensemble of the two reached F of 0.46. Following the IDENTITY chains in the data would add 106,263 additional named entities in the full 97-paper corpus, for an increase of 76% percent in the semantic classes of the eight ontologies that have been annotated in earlier versions of the CRAFT corpus. CONCLUSIONS: The project produced a large data set for further investigation of coreference and coreference resolution in the scientific literature. The work raised issues in the phenomenon of reference in this domain and genre, and the paper proposes that many mentions that would be considered generic in the general domain are not generic in the biomedical domain due to their referents to specific classes in domain-specific ontologies. The comparison of the performance of a publicly available and well-understood coreference resolution system with a domain-adapted system produced results that are consistent with the notion that the requirements for successful coreference resolution in this genre are quite different from those of the general domain, and also suggest that the baseline performance difference is quite large.

BACKGROUND: Investigating and understanding drug-drug interactions (DDIs) is important in improving the effectiveness of clinical care. DDIs can occur when two or more drugs are administered together. Experimentally based DDI detection methods require a large cost and time. Hence, there is a great interest in developing efficient and useful computational methods for inferring potential DDIs. Standard binary classifiers require both positives and negatives for training. In a DDI context, drug pairs that are known to interact can serve as positives for predictive methods. But, the negatives or drug pairs that have been confirmed to have no interaction are scarce. To address this lack of negatives, we introduce a Positive-Unlabeled Learning method for inferring potential DDIs. RESULTS: The proposed method consists of three steps: i) application of Growing Self Organizing Maps to infer negatives from the unlabeled dataset; ii) using a pairwise similarity function to quantify the overlap between individual features of drugs and iii) using support vector machine classifier for inferring DDIs. We obtained 6036 DDIs from DrugBank database. Using the proposed approach, we inferred 589 drug pairs that are likely to not interact with each other; these drug pairs are used as representative data for the negative class in binary classification for DDI prediction. Moreover, we classify the predicted DDIs as Cytochrome P450 (CYP) enzyme-Dependent and CYP-Independent interactions invoking their locations on the Growing Self Organizing Map, due to the particular importance of these enzymes in clinically significant interaction effects. Further, we provide a case study on three predicted CYP-Dependent DDIs to evaluate the clinical relevance of this study. CONCLUSION: Our proposed approach showed an absolute improvement in F1-score of 14 and 38% in comparison to the method that randomly selects unlabeled data points as likely negatives, depending on the choice of similarity function. We inferred 5300 possible CYP-Dependent DDIs and 592 CYP-Independent DDIs with the highest posterior probabilities. Our discoveries can be used to improve clinical care as well as the research outcomes of drug development.

We investigate and analyse the data quality of nucleotide sequence databases with the objective of automatic detection of data anomalies and suspicious records. Specifically, we demonstrate that the published literature associated with each data record can be used to automatically evaluate its quality, by cross-checking the consistency of the key content of the database record with the referenced publications. Focusing on GenBank, we describe a set of quality indicators based on the relevance paradigm of information retrieval (IR). Then, we use these quality indicators to train an anomaly detection algorithm to classify records as "confident" or "suspicious". Our experiments on the PubMed Central collection show assessing the coherence between the literature and database records, through our algorithms, is an effective mechanism for assisting curators to perform data cleansing. Although fewer than 0.25% of the records in our data set are known to be faulty, we would expect that there are many more in GenBank that have not yet been identified. By automated comparison with literature they can be identified with a precision of up to 10% and a recall of up to 30%, while strongly outperforming several baselines. While these results leave substantial room for improvement, they reflect both the very imbalanced nature of the data, and the limited explicitly labelled data that is available. Overall, the obtained results show promise for the development of a new kind of approach to detecting low-quality and suspicious sequence records based on literature analysis and consistency. From a practical point of view, this will greatly help curators in identifying inconsistent records in large-scale sequence databases by highlighting records that are likely to be inconsistent with the literature.

CD-HIT is one of the most popular tools for reducing sequence redundancy, and is considered to be the state-of-art method. It tries to minimise redundancy by reducing an input database into several representative sequences, under a user-defined threshold of sequence identity. We present a comprehensive assessment of the redundancy in the outputs of CD-HIT, exploring the impact of different identity thresholds and new evaluation data on the redundancy. We demonstrate that the relationship between threshold and redundancies is surprising weak. Applications of CD-HIT that set low identity threshold values also may suffer from substantial degradation in both efficiency and accuracy.

Estimating the strength of dependency between two variables is fundamental for exploratory analysis and many other applications in data mining. For example: non-linear dependencies between two continuous variables can be explored with the Maximal Information Coefficient (MIC); and categorical variables that are dependent to the target class are selected using Gini gain in random forests. Nonetheless, because dependency measures are estimated on finite samples, the interpretability of their quantification and the accuracy when ranking dependencies become challenging. Dependency estimates are not equal to 0 when variables are independent, cannot be compared if computed on different sample size, and they are inflated by chance on variables with more categories. In this paper, we propose a framework to adjust dependency measure estimates on finite samples. Our adjustments, which are simple and applicable to any dependency measure, are helpful in improving interpretability when quantifying dependency and in improving accuracy on the task of ranking dependencies. In particular, we demonstrate that our approach enhances the interpretability of MIC when used as a proxy for the amount of noise between variables, and to gain accuracy when ranking variables during the splitting procedure in random forests.

Adjusted for chance measures are widely used to compare partitions/clusterings of the same data set. In particular, the Adjusted Rand Index (ARI) based on pair-counting, and the Adjusted Mutual Information (AMI) based on Shannon information theory are very popular in the clustering community. Nonetheless it is an open problem as to what are the best application scenarios for each measure and guidelines in the literature for their usage are sparse, with the result that users often resort to using both. Generalized Information Theoretic (IT) measures based on the Tsallis entropy have been shown to link pair- counting and Shannon IT measures. In this paper, we aim to bridge the gap between adjustment of measures based on pair- counting and measures based on information theory. We solve the key technical challenge of analytically computing the expected value and variance of generalized IT measures. This allows us to propose adjustments of generalized IT measures, which reduce to well known adjusted clustering comparison measures as special cases. Using the theory of generalized IT measures, we are able to propose the following guidelines for using ARI and AMI as external validation indices: ARI should be used when the reference clustering has large equal sized clusters; AMI should be used when the reference clustering is unbalanced and there exist small clusters.

Held each year in conjunction with one of the largest data management conferences, CIKM, the Ninth ACM International Workshop on Data and Text Mining in Biomedical Informatics (DTMBIO'15) is organized to bring together researchers interested in development and application of cutting-edge data management and analysis methods with a specific focus on applications in biology and medicine. The purpose of DTMBIO is to foster discussions regarding the state-of-the-art applications of data and text mining on biomedical research problems. DTMBIO'15 will help scientists understand emerging trends and opportunities in the evolving area of informatics related techniques and problems in the context of biomedical research.

UNLABELLED: The ISMB Special Interest Group on Linking Literature, Information and Knowledge for Biology (BioLINK) organized a one-day workshop at ISMB/ECCB 2013 in Berlin, Germany. The theme of the workshop was 'Roles for text mining in biomedical knowledge discovery and translational medicine'. This summary reviews the outcomes of the workshop. Meeting themes included concept annotation methods and applications, extraction of biological relationships and the use of text-mined data for biological data analysis. AVAILABILITY AND IMPLEMENTATION: All articles are available at http://biolinksig.org/proceedings-online/.

IN BIONLP-ST 2013: We participated in the BioNLP 2013 shared tasks on event extraction. Our extraction method is based on the search for an approximate subgraph isomorphism between key context dependencies of events and graphs of input sentences. Our system was able to address both the GENIA (GE) task focusing on 13 molecular biology related event types and the Cancer Genetics (CG) task targeting a challenging group of 40 cancer biology related event types with varying arguments concerning 18 kinds of biological entities. In addition to adapting our system to the two tasks, we also attempted to integrate semantics into the graph matching scheme using a distributional similarity model for more events, and evaluated the event extraction impact of using paths of all possible lengths as key context dependencies beyond using only the shortest paths in our system. We achieved a 46.38% F-score in the CG task (ranking 3rd) and a 48.93% F-score in the GE task (ranking 4th). AFTER BIONLP-ST 2013: We explored three ways to further extend our event extraction system in our previously published work: (1) We allow non-essential nodes to be skipped, and incorporated a node skipping penalty into the subgraph distance function of our approximate subgraph matching algorithm. (2) Instead of assigning a unified subgraph distance threshold to all patterns of an event type, we learned a customized threshold for each pattern. (3) We implemented the well-known Empirical Risk Minimization (ERM) principle to optimize the event pattern set by balancing prediction errors on training data against regularization. When evaluated on the official GE task test data, these extensions help to improve the extraction precision from 62% to 65%. However, the overall F-score stays equivalent to the previous performance due to a 1% drop in recall.

by stating that they required an external tag dictionary.We have corrected these inaccuracies to reflect their modest data requirements.

Mutual information is a very popular measure for comparing clusterings. Previous work has shown that it is beneficial to make an adjustment for chance to this measure, by subtracting an expected value and normalizing via an upper bound. This yields the constant baseline property that enhances intuitiveness. In this paper, we argue that a further type of statistical adjustment for the mutual information is also beneficial - an adjustment to correct selection bias. This type of adjustment is useful when carrying out many clustering comparisons, to select one or more preferred clusterings. It reduces the tendency for the mutual information to choose clustering solutions i) with more clusters, or ii) induced on fewer data points, when compared to a reference one. We term our new adjusted measure the *standardized mutual information*. It requires computation of the variance of mutual information under a hypergeometric model of randomness, which is technically challenging. We derive an analytical formula for this variance and analyze its complexity. We then experimentally assess how our new measure can address selection bias and also increase interpretability. We recommend using the standardized mutual information when making multiple clustering comparisons in situations where the number of records is small compared to the number of clusters considered.

BACKGROUND: Ontological concepts are useful for many different biomedical tasks. Concepts are difficult to recognize in text due to a disconnect between what is captured in an ontology and how the concepts are expressed in text. There are many recognizers for specific ontologies, but a general approach for concept recognition is an open problem. RESULTS: Three dictionary-based systems (MetaMap, NCBO Annotator, and ConceptMapper) are evaluated on eight biomedical ontologies in the Colorado Richly Annotated Full-Text (CRAFT) Corpus. Over 1,000 parameter combinations are examined, and best-performing parameters for each system-ontology pair are presented. CONCLUSIONS: Baselines for concept recognition by three systems on eight biomedical ontologies are established (F-measures range from 0.14-0.83). Out of the three systems we tested, ConceptMapper is generally the best-performing system; it produces the highest F-measure of seven out of eight ontologies. Default parameters are not ideal for most systems on most ontologies; by changing parameters F-measure can be increased by up to 0.4. Not only are best performing parameters presented, but suggestions for choosing the best parameters based on ontology characteristics are presented.

Combining heterogeneous sources of data is essential for accurate prediction of protein function. The task is complicated by the fact that while sequence-based features can be readily compared across species, most other data are species-specific. In this paper, we present a multi-view extension to GOstruct, a structured-output framework for function annotation of proteins. The extended framework can learn from disparate data sources, with each data source provided to the framework in the form of a kernel. Our empirical results demonstrate that the multi-view framework is able to utilize all available information, yielding better performance than sequence-based models trained across species and models trained from collections of data within a given species. This version of GOstruct participated in the recent Critical Assessment of Functional Annotations (CAFA) challenge; since then we have significantly improved the natural language processing component of the method, which now provides performance that is on par with that provided by sequence information. The GOstruct framework is available for download at http://strut.sourceforge.net.

Information about verb subcategorization frames (SCFs) is important to many tasks in natural language processing (NLP) and, in turn, text mining. Biomedicine has a need for high-quality SCF lexicons to support the extraction of information from the biomedical literature, which helps biologists to take advantage of the latest biomedical knowledge despite the overwhelming growth of that literature. Unfortunately, techniques for creating such resources for biomedical text are relatively undeveloped compared to general language. This paper serves as an introduction to subcategorization and existing approaches to acquisition, and provides motivation for developing techniques that address issues particularly important to biomedical NLP. First, we give the traditional linguistic definition of subcategorization, along with several related concepts. Second, we describe approaches to learning SCF lexicons from large data sets for general and biomedical domains. Third, we consider the crucial issue of linguistic variation between biomedical fields (subdomain variation). We demonstrate significant variation among subdomains, and find the variation does not simply follow patterns of general lexical variation. Finally, we note several requirements for future research in biomedical SCF lexicon acquisition: a high-quality gold standard, investigation of different definitions of subcategorization, and minimally-supervised methods that can learn subdomain-specific lexical usage without the need for extensive manual work.

BACKGROUND: Manually annotated corpora are critical for the training and evaluation of automated methods to identify concepts in biomedical text. RESULTS: This paper presents the concept annotations of the Colorado Richly Annotated Full-Text (CRAFT) Corpus, a collection of 97 full-length, open-access biomedical journal articles that have been annotated both semantically and syntactically to serve as a research resource for the biomedical natural-language-processing (NLP) community. CRAFT identifies all mentions of nearly all concepts from nine prominent biomedical ontologies and terminologies: the Cell Type Ontology, the Chemical Entities of Biological Interest ontology, the NCBI Taxonomy, the Protein Ontology, the Sequence Ontology, the entries of the Entrez Gene database, and the three subontologies of the Gene Ontology. The first public release includes the annotations for 67 of the 97 articles, reserving two sets of 15 articles for future text-mining competitions (after which these too will be released). Concept annotations were created based on a single set of guidelines, which has enabled us to achieve consistently high interannotator agreement. CONCLUSIONS: As the initial 67-article release contains more than 560,000 tokens (and the full set more than 790,000 tokens), our corpus is among the largest gold-standard annotated biomedical corpora. Unlike most others, the journal articles that comprise the corpus are drawn from diverse biomedical disciplines and are marked up in their entirety. Additionally, with a concept-annotation count of nearly 100,000 in the 67-article subset (and more than 140,000 in the full collection), the scale of conceptual markup is also among the largest of comparable corpora. The concept annotations of the CRAFT Corpus have the potential to significantly advance biomedical text mining by providing a high-quality gold standard for NLP systems. The corpus, annotation guidelines, and other associated resources are freely available at http://bionlp-corpora.sourceforge.net/CRAFT/index.shtml.

BACKGROUND: We introduce the linguistic annotation of a corpus of 97 full-text biomedical publications, known as the Colorado Richly Annotated Full Text (CRAFT) corpus. We further assess the performance of existing tools for performing sentence splitting, tokenization, syntactic parsing, and named entity recognition on this corpus. RESULTS: Many biomedical natural language processing systems demonstrated large differences between their previously published results and their performance on the CRAFT corpus when tested with the publicly available models or rule sets. Trainable systems differed widely with respect to their ability to build high-performing models based on this data. CONCLUSIONS: The finding that some systems were able to train high-performing models based on this corpus is additional evidence, beyond high inter-annotator agreement, that the quality of the CRAFT corpus is high. The overall poor performance of various systems indicates that considerable work needs to be done to enable natural language processing systems to work well when the input is full-text journal articles. The CRAFT corpus provides a valuable resource to the biomedical natural language processing community for evaluation and training of new models for biomedical full text publications.

BACKGROUND: Bio-molecular event extraction from literature is recognized as an important task of bio text mining and, as such, many relevant systems have been developed and made available during the last decade. While such systems provide useful services individually, there is a need for a meta-service to enable comparison and ensemble of such services, offering optimal solutions for various purposes. RESULTS: We have integrated nine event extraction systems in the U-Compare framework, making them intercompatible and interoperable with other U-Compare components. The U-Compare event meta-service provides various meta-level features for comparison and ensemble of multiple event extraction systems. Experimental results show that the performance improvements achieved by the ensemble are significant. CONCLUSIONS: While individual event extraction systems themselves provide useful features for bio text mining, the U-Compare meta-service is expected to improve the accessibility to the individual systems, and to enable meta-level uses over multiple event extraction systems such as comparison and ensemble.

BACKGROUND: We report the Gene Normalization (GN) challenge in BioCreative III where participating teams were asked to return a ranked list of identifiers of the genes detected in full-text articles. For training, 32 fully and 500 partially annotated articles were prepared. A total of 507 articles were selected as the test set. Due to the high annotation cost, it was not feasible to obtain gold-standard human annotations for all test articles. Instead, we developed an Expectation Maximization (EM) algorithm approach for choosing a small number of test articles for manual annotation that were most capable of differentiating team performance. Moreover, the same algorithm was subsequently used for inferring ground truth based solely on team submissions. We report team performance on both gold standard and inferred ground truth using a newly proposed metric called Threshold Average Precision (TAP-k). RESULTS: We received a total of 37 runs from 14 different teams for the task. When evaluated using the gold-standard annotations of the 50 articles, the highest TAP-k scores were 0.3297 (k=5), 0.3538 (k=10), and 0.3535 (k=20), respectively. Higher TAP-k scores of 0.4916 (k=5, 10, 20) were observed when evaluated using the inferred ground truth over the full test set. When combining team results using machine learning, the best composite system achieved TAP-k scores of 0.3707 (k=5), 0.4311 (k=10), and 0.4477 (k=20) on the gold standard, representing improvements of 12.4%, 21.8%, and 26.6% over the best team results, respectively. CONCLUSIONS: By using full text and being species non-specific, the GN task in BioCreative III has moved closer to a real literature curation task than similar tasks in the past and presents additional challenges for the text mining community, as revealed in the overall team results. By evaluating teams using the gold standard, we show that the EM algorithm allows team submissions to be differentiated while keeping the manual annotation effort feasible. Using the inferred ground truth we show measures of comparative performance between teams. Finally, by comparing team rankings on gold standard vs. inferred ground truth, we further demonstrate that the inferred ground truth is as effective as the gold standard for detecting good team performance.

We approached the problems of event detection, argument identification, and negation and speculation detection in the BioNLP'09 information extraction challenge through concept recognition and analysis. Our methodology involved using the OpenDMAP semantic parser with manually written rules. The original OpenDMAP system was updated for this challenge with a broad ontology defined for the events of interest, new linguistic patterns for those events, and specialized coordination handling. We achieved state-of-the-art precision for two of the three tasks, scoring the highest of 24 teams at precision of 71.81 on Task 1 and the highest of 6 teams at precision of 70.97 on Task 2. We provide a detailed analysis of the training data and show that a number of trigger words were ambiguous as to event type, even when their arguments are constrained by semantic class. The data is also shown to have a number of missing annotations. Analysis of a sampling of the comparatively small number of false positives returned by our system shows that major causes of this type of error were failing to recognize second themes in two-theme events, failing to recognize events when they were the arguments to other events, failure to recognize nontheme arguments, and sentence segmentation errors. We show that specifically handling coordination had a small but important impact on the overall performance of the system. The OpenDMAP system and the rule set are available at http://bionlp.sourceforge.net.

SUMMARY: The Unstructured Information Management Architecture (UIMA) framework and web services are emerging as useful tools for integrating biomedical text mining tools. This note describes our work, which wraps the National Center for Biomedical Ontology (NCBO) Annotator-an ontology-based annotation service-to make it available as a component in UIMA workflows. AVAILABILITY: This wrapper is freely available on the web at http://bionlp-uima.sourceforge.net/ as part of the UIMA tools distribution from the Center for Computational Pharmacology (CCP) at the University of Colorado School of Medicine. It has been implemented in Java for support on Mac OS X, Linux and MS Windows.

BACKGROUND: An increase in work on the full text of journal articles and the growth of PubMedCentral have the opportunity to create a major paradigm shift in how biomedical text mining is done. However, until now there has been no comprehensive characterization of how the bodies of full text journal articles differ from the abstracts that until now have been the subject of most biomedical text mining research. RESULTS: We examined the structural and linguistic aspects of abstracts and bodies of full text articles, the performance of text mining tools on both, and the distribution of a variety of semantic classes of named entities between them. We found marked structural differences, with longer sentences in the article bodies and much heavier use of parenthesized material in the bodies than in the abstracts. We found content differences with respect to linguistic features. Three out of four of the linguistic features that we examined were statistically significantly differently distributed between the two genres. We also found content differences with respect to the distribution of semantic features. There were significantly different densities per thousand words for three out of four semantic classes, and clear differences in the extent to which they appeared in the two genres. With respect to the performance of text mining tools, we found that a mutation finder performed equally well in both genres, but that a wide variety of gene mention systems performed much worse on article bodies than they did on abstracts. POS tagging was also more accurate in abstracts than in article bodies. CONCLUSIONS: Aspects of structure and content differ markedly between article abstracts and article bodies. A number of these differences may pose problems as the text mining field moves more into the area of processing full-text articles. However, these differences also present a number of opportunities for the extraction of data types, particularly that found in parenthesized text, that is present in article bodies but not in article abstracts.

We introduce a system developed for the BioCreative II.5 community evaluation of information extraction of proteins and protein interactions. The paper focuses primarily on the gene normalization task of recognizing protein mentions in text and mapping them to the appropriate database identifiers based on contextual clues. We outline a ""fuzzy" dictionary lookup approach to protein mention detection that matches regularized text to similarly regularized dictionary entries. We describe several different strategies for gene normalization that focus on species or organism mentions in the text, both globally throughout the document and locally in the immediate vicinity of a protein mention, and present the results of experimentation with a series of system variations that explore the effectiveness of the various normalization strategies, as well as the role of external knowledge sources. While our system was neither the best nor the worst performing system in the evaluation, the gene normalization strategies show promise and the system affords the opportunity to explore some of the variables affecting performance on the BCII.5 tasks.

MOTIVATION: It is important for the quality of biological ontologies that similar concepts be expressed consistently, or univocally. Univocality is relevant for the usability of the ontology for humans, as well as for computational tools that rely on regularity in the structure of terms. However, in practice terms are not always expressed consistently, and we must develop methods for identifying terms that are not univocal so that they can be corrected. RESULTS: We developed an automated transformation-based clustering methodology for detecting terms that use different linguistic conventions for expressing similar semantics. These term sets represent occurrences of univocality violations. Our method was able to identify 67 examples of univocality violations in the Gene Ontology. AVAILABILITY: The identified univocality violations are available upon request. We are preparing a release of an open source version of the software to be available at http://bionlp.sourceforge.net.

BACKGROUND: Recent years have seen an increased amount of natural language processing (NLP) work on full text biomedical journal publications. Much of this work is done with Open Access journal articles. Such work assumes that Open Access articles are representative of biomedical publications in general and that methods developed for analysis of Open Access full text publications will generalize to the biomedical literature as a whole. If this assumption is wrong, the cost to the community will be large, including not just wasted resources, but also flawed science. This paper examines that assumption. RESULTS: We collected two sets of documents, one consisting only of Open Access publications and the other consisting only of traditional journal publications. We examined them for differences in surface linguistic structures that have obvious consequences for the ease or difficulty of natural language processing and for differences in semantic content as reflected in lexical items. Regarding surface linguistic structures, we examined the incidence of conjunctions, negation, passives, and pronominal anaphora, and found that the two collections did not differ. We also examined the distribution of sentence lengths and found that both collections were characterized by the same mode. Regarding lexical items, we found that the Kullback-Leibler divergence between the two collections was low, and was lower than the divergence between either collection and a reference corpus. Where small differences did exist, log likelihood analysis showed that they were primarily in the area of formatting and in specific named entities. CONCLUSION: We did not find structural or semantic differences between the Open Access and traditional journal collections.

BACKGROUND: We participated in the BioCreAtIvE Task 2, which addressed the annotation of proteins into the Gene Ontology (GO) based on the text of a given document and the selection of evidence text from the document justifying that annotation. We approached the task utilizing several combinations of two distinct methods: an unsupervised algorithm for expanding words associated with GO nodes, and an annotation methodology which treats annotation as categorization of terms from a protein's document neighborhood into the GO. RESULTS: The evaluation results indicate that the method for expanding words associated with GO nodes is quite powerful; we were able to successfully select appropriate evidence text for a given annotation in 38% of Task 2.1 queries by building on this method. The term categorization methodology achieved a precision of 16% for annotation within the correct extended family in Task 2.2, though we show through subsequent analysis that this can be improved with a different parameter setting. Our architecture proved not to be very successful on the evidence text component of the task, in the configuration used to generate the submitted results. CONCLUSION: The initial results show promise for both of the methods we explored, and we are planning to integrate the methods more closely to achieve better results overall.