Major depressive disorder (MDD) is a significant global health burden; timely and precise diagnosis is essential to reduce relapse and mortality. However, existing approaches typically frame depression diagnosis as a monolithic classification problem, neglecting the multi-dimensional and hierarchical reasoning that underpins clinical interviews. We introduce DALE, a modular framework that mirrors psychiatrists'reasoning by explicitly disentangling four dimensions of patient narratives-psychological symptoms, somatic symptoms, protective factors and stressors. Leveraging GPT-4, we augment the dataset with dialogue-level annotations that map patient interview onto a list of attributes. On this auxiliary data, we train four domain-specialised LoRA adapters atop a frozen LLM; each adapter conducts a brief diagnostic dialogue and produces a concise report of its domain. A lightweight classifier then integrates these reports to generate a final summary and predict depression and suicide risk. Experiments on the D4 psychiatric dialogue benchmark show that DALE showing strong performance, while requiring far fewer trainable parameters, and yields interpretable, attribute-level evidence for its predictions.

Automatic detection of Adverse Drug Events (ADEs) from social media has become increasingly important for post-market drug safety surveillance and pharmacovigilance. Although existing social media-based ADE detection methods have effectively addressed some challenges such as data sparsity and class imbalance, they still suffer from spurious correlations, where models tend to learn co-occurrence patterns between drugs and symptoms, leading them to incorrectly identify drug inefficacy or therapeutic intent as ADEs. To address this issue, we propose a structured medical knowledge-guided counterfactual generation method that leverages authoritative medical databases and large language models to construct clinically plausible counterfactual samples, thereby mitigating spurious correlations at the data level. Furthermore, we propose a knowledge-enhanced instruction-tuning strategy that injects drug indications as causal cues into model inputs to enhance its causal reasoning capabilities. Experimental results on two benchmark datasets demonstrate that our method consistently outperforms state-of-the-art models, effectively alleviating spurious correlations and exhibiting superior capabilities in drug-symptom relationship identification.

With the rapid development of online medical platforms, consumer health questions (CHQs) are inefficient in diagnosis due to redundant information and frequent non-professional terms. The medical question summary (MQS) task aims to transform CHQs into streamlined doctors' frequently asked questions (FAQs), but existing methods still face challenges such as poor identification of question focus and low summary faithfulness. This paper explores the potential of large language models (LLMs) in the MQS task and finds that direct fine-tuning is prone to focus identification bias and generates summaries with low faithfulness. To this end, we propose an optimization framework based on core focus guidance. First, a prompt template is designed to drive the LLMs to extract the core focus from the CHQs that is faithful to the original text. Then, a fine-tuning dataset is constructed in combination with the original CHQ-FAQ pairs to improve the ability to identify the focus of the question. Finally, a multi-dimensional quality evaluation and selection mechanism is proposed to comprehensively improve the quality of the summary from multiple dimensions. We conduct comprehensive experiments on two widely-adopted MQS datasets using three established evaluation metrics. The proposed framework achieves state-of-the-art performance across all measures, demonstrating a significant boost in the model's ability to identify the critical focus of questions and a notable improvement in the faithfulness of the summary. The source codes are freely available at https://github.com/DUT-LiuChao/FocusMed.

Accurate recognition of biomedical named entities is critical for medical information extraction and knowledge discovery. However, existing methods often struggle with nested entities, entity boundary ambiguity, and cross-lingual generalization. In this paper, we propose a unified Biomedical Named Entity Recognition (BioNER) framework based on Large Language Models (LLMs). We first reformulate BioNER as a text generation task and design a symbolic tagging strategy to jointly handle both flat and nested entities with explicit boundary annotation. To enhance multilingual and multi-task generalization, we perform bilingual joint fine-tuning across multiple Chinese and English datasets. Additionally, we introduce a contrastive learning-based entity selector that filters incorrect or spurious predictions by leveraging boundary-sensitive positive and negative samples. Experimental results on four benchmark datasets and two unseen corpora show that our method achieves state-of-the-art performance and robust zero-shot generalization across languages. The source codes are freely available at https://github.com/dreamer-tx/LLMNER.

Large Language Models (LLMs) have shown strong abilities in language understanding and reasoning, drawing increasing attention in medical question answering (QA). While retrieval-augmented generation (RAG) methods improve factual accuracy, existing approaches still struggle to retrieve and organize relevant knowledge effectively. To overcome this, we propose NaviPath, a knowledge graph-based RAG framework for medical QA. It enhances LLM responses through a structured prompt built in three steps: (1) extended entity retrieval, (2) multiperspective reasoning path construction, and (3) natural language transformation for better comprehension. Experiments on two medical QA benchmarks show that NaviPath achieves state-of-the-art performance in diagnostic accuracy and factual consistency. The implementation is available at https://github.com/zyr319/Navipath.git.

Abstract Large language models (LLMs) demonstrate remarkable text comprehension and generation capabilities but often lack the ability to utilize up-to-date or domain-specific knowledge not included in their training data. To address this gap, we introduce KEDiT, an efficient method for fine-tuning LLMs for knowledge-grounded dialogue generation. KEDiT operates in two main phases. First, it employs an information bottleneck to compress retrieved knowledge into learnable parameters, retaining essential information while minimizing computational overhead. Second, a lightweight knowledge-aware adapter integrates these compressed knowledge vectors into the LLM during fine-tuning, updating less than 2% of the model parameters. The experimental results on the Wizard of Wikipedia and a newly constructed PubMed-Dialog dataset demonstrate that KEDiT excels in generating contextually relevant and informative responses, outperforming competitive baselines in automatic, LLM-based, and human evaluations. This approach effectively combines the strengths of pretrained LLMs with the adaptability needed for incorporating dynamic knowledge, presenting a scalable solution for fields such as medicine.1

Empathetic response generation, aiming to understand the user’s situation and feelings and respond empathically, is crucial in building human-like dialogue systems. Traditional approaches typically employ maximum likelihood estimation as the optimization objective during training, yet fail to align the empathy levels between generated and target responses. To this end, we propose an empathetic response generation framework using reinforcement learning (EmpRL). The framework develops an effective empathy reward function and generates empathetic responses by maximizing the expected reward through reinforcement learning. EmpRL utilizes the pre-trained T5 model as the generator and further fine-tunes it to initialize the policy. To align the empathy levels between generated and target responses within a given context, an empathy reward function containing three empathy communication mechanisms—emotional reaction, interpretation, and exploration—is constructed using pre-designed and pre-trained empathy identifiers. During reinforcement learning training, the proximal policy optimization algorithm is used to fine-tune the policy, enabling the generation of empathetic responses. Both automatic and human evaluations demonstrate that the proposed EmpRL framework significantly improves the quality of generated responses, enhances the similarity in empathy levels between generated and target responses, and produces empathetic responses covering both affective and cognitive aspects.

In the biomedical field, information is widely distributed across numerous pieces of literature. Extracting events between entities from biomedical texts has garnered significant attention in recent years. However, previous research primarily focus on extracting flat biomedical events, with less attention given to nested biomedical events. Moreover, existing methods for extracting nested events often overlook the long-distance dependencies and global information between trigger words and arguments within events, and they lack sufficient interaction with event type information. To address these issues, we propose a semantic segmentation-based method for extracting nested biomedical events. We introduce U-Net to capture global information and interdependencies between event entities. Additionally, we map event types to natural language text and combine them with sentences for encoding to enhance interaction. We also employ two auxiliary tasks to improve the identification of trigger words and arguments. Finally, events are extracted by identifying the four vertices of the segmented region. Experimental results on two benchmark datasets show that our method excels in recognizing nested biomedical events and outperforms current state-of-the-art methods.

The task of document-level biomedical relation extraction involves identifying relational facts between entities across sentences, given specific entities. However, most current methods overlook the associations between entity pairs and generate fixed entity representations merely through mentions, leading to irrelevant mentions interfering with the determination of relational facts. Additionally, these methods fail to consider the global information and dependencies between relational entities. To address these issues, we propose a document-level relation extraction model based on relation-guided entity-level graphs. Our model aggregates all mentions of the same entity through a relation-guided attention mechanism to obtain flexible entity representations. Furthermore, by using U-Net to generate entity-level feature graphs, it facilitates global interactions and dependency capture between entity pairs. Experimental results on two benchmark datasets demonstrate the advantages of our approach in document-level biomedical relation extraction.

Large language models (LLMs) perform well in many NLP tasks but frequently generate inaccurate information in the biomedical domain, due to hallucination issues. Retrieval-Augmented Generation (RAG) has been introduced to address this issue by integrating external knowledge, enhancing the factual accuracy of outputs. However, naive RAG encounters challenges in effectively utilizing retrieved content, particularly in specialized domains like biomedicine. LLMs often struggle to integrate retrieved content as irrelevant information can interfere with the model’s judgment. Even if relevant documents are retrieved, the model may be unable to accurately comprehend and utilize the domain-specific features due to its inherent knowledge limitations. To overcome these limitations, we propose Document Embeddings Enhanced Biomedical RAG (DEEB-RAG), a framework that incorporates document embeddings along with the original retrieved text. DEEB-RAG uses MedCPT to generate document embeddings and these embeddings are then aligned with the LLM’s semantic space using a two-stage training process on a simple projector. Experimental results on biomedical QA datasets show that DEEB-RAG improves accuracy, with an average performance increase of 2.3% over naive RAG. This demonstrates DEEB-RAG’s ability to mitigate the challenges of utilizing complex biomedical information, thereby enhancing the reliability and effectiveness of LLMs in biomedical domain.

Biomedical document-level relation extraction is a crucial technology for mining the biomedical relationships necessary for clinical diagnosis, treatment, and medical discovery. Although existing intrasentential relation extraction methods have achieved significant results, the complexity and scattered nature of information in biomedical literature require relation extraction techniques to effectively handle cross-sentence information. For example, existing methods have not been able to explicitly model the phenomena of coreference and anaphor in documents, thus affecting the model’s understanding of complex semantics within the document. To address this issue, we propose a new document-level relation extraction model with coreference and anaphor graphs. By abstracting the document into an undirected graph that includes coreference and anaphor information, the framework effectively models the interactions between entities and leverages graph convolutional network in conjunction with pretrained language model to dynamically understand graph structures. Additionally, the shift from fine-grained entity-pair level to coarse-grained document-level training and inference significantly enhances the model’s efficiency while maintaining high extraction performance. Extensive experiments demonstrate that our model achieves a 5.3% increase in F1-score over baseline models on the BioRED dataset with higher efficiency, confirming its effectiveness in handling relation extraction tasks in complex biomedical literature.

Automatically identifying human phenotype ontology (HPO) concepts from text is important for disease analysis. Existing ontology-driven methods for phenotype concept recognition mainly rely on concept names and synonym information from the ontology, without fully exploiting the rich ontology information. In this paper, we present an improved phenotype concept recognition method by incorporating rich HPO information. We first design prompts with HPO information and use a cutting-edge large language model GPT-4 to generate synonym augmentation for expanding distant supervised training data. We then propose an ontology vector-enhanced phenotype concept classification model to efficiently integrate the taxonomic hierarchical structure of HPO. Additionally, we employ noisy data augmentation to improve the model’s recognition ability in noisy texts and implement a negation detection function. Experimental results on three standard corpora and two typo corpora show our method compares favorably to previous methods and achieves a significant improvement in noisy texts. The source code and data are freely available at https://github.com/DUTIR-BioNLP/PhenoTagger-Updates.

Polypharmacy is the combined use of multiple drugs for the treatment of diseases, which also often comes with a higher risk of side effects. In the medical industry, acquiring rich and comprehensive information about the side effects of multiple drug therapy becomes a crucial task. However, data collection for many side effects is often sparse, so the features of these data cannot be adequately learned, resulting in poor performance in side effects prediction. In this paper, we propose a framework based on knowledge graph embedding (KGE) models which improves KGE by using LTE operations and subsampling methods (called LTESampleKGE). LTESampleKGE consists of two main modules i.e., Entity embedding enhancement module and KGE subsampling module. The former applies linear transformation to entity representation instead of GCN structure to enhance entity embedding, while the latter utilizes subsampling methods for KGE negative sampling (NS) loss to pay more attention to sparse data. Thus, LTESampleKGE can effectively alleviate the problem of data sparsity in the polypharmacy side effects prediction task. Experimental evaluations indicate that our method demonstrates superior performance compared with baseline models. For example, LTESampleKGE outperforms MSTE by 1.20% in PR-AUC score on TWOSIDES dataset and by 0.46% in AP@n score on Drugbank dataset.

MOTIVATION: Significant progress has been achieved in biomedical text mining using deep learning methods, which rely heavily on large amounts of high-quality data annotated by human experts. However, the reality is that obtaining high-quality annotated data is extremely challenging due to data scarcity (e.g. rare or new diseases), data privacy and security concerns, and the high cost of data annotation. Additionally, nearly all researches focus on predicting labels without providing corresponding explanations. Therefore, in this paper, we investigate a more realistic scenario, biomedical few-shot learning, and explore the impact of interpretability on biomedical few-shot learning. RESULTS: We present LetEx-Learning to explain-a novel multi-task generative approach that leverages reasoning explanations from large language models (LLMs) to enhance the inductive reasoning ability of few-shot learning. Our approach includes (1) collecting high-quality explanations by devising a suite of complete workflow based on LLMs through CoT prompting and self-training strategies, (2) converting various biomedical NLP tasks into a text-to-text generation task in a unified manner, where collected explanations serve as additional supervision between text-label pairs by multi-task training. Experiments are conducted on three few-shot settings across six biomedical benchmark datasets. The results show that learning to explain improves the performances of diverse biomedical NLP tasks in low-resource scenario, outperforming strong baseline models significantly by up to 6.41%. Notably, the proposed method makes the 220M LetEx perform superior reasoning explanation ability against LLMs. AVAILABILITY AND IMPLEMENTATION: Our source code and data are available at https://github.com/cpmss521/LetEx.

OBJECTIVE: Most existing fine-tuned biomedical large language models (LLMs) focus on enhancing performance in monolingual biomedical question answering and conversation tasks. To investigate the effectiveness of the fine-tuned LLMs on diverse biomedical natural language processing (NLP) tasks in different languages, we present Taiyi, a bilingual fine-tuned LLM for diverse biomedical NLP tasks. MATERIALS AND METHODS: We first curated a comprehensive collection of 140 existing biomedical text mining datasets (102 English and 38 Chinese datasets) across over 10 task types. Subsequently, these corpora were converted to the instruction data used to fine-tune the general LLM. During the supervised fine-tuning phase, a 2-stage strategy is proposed to optimize the model performance across various tasks. RESULTS: Experimental results on 13 test sets, which include named entity recognition, relation extraction, text classification, and question answering tasks, demonstrate that Taiyi achieves superior performance compared to general LLMs. The case study involving additional biomedical NLP tasks further shows Taiyi's considerable potential for bilingual biomedical multitasking. CONCLUSION: Leveraging rich high-quality biomedical corpora and developing effective fine-tuning strategies can significantly improve the performance of LLMs within the biomedical domain. Taiyi shows the bilingual multitasking capability through supervised fine-tuning. However, those tasks such as information extraction that are not generation tasks in nature remain challenging for LLM-based generative approaches, and they still underperform the conventional discriminative approaches using smaller language models.

In bioinformatics, protein function prediction stands as a fundamental area of research and plays a crucial role in addressing various biological challenges, such as the identification of potential targets for drug discovery and the elucidation of disease mechanisms. However, known functional annotation databases usually provide positive experimental annotations that proteins carry out a given function, and rarely record negative experimental annotations that proteins do not carry out a given function. Therefore, existing computational methods based on deep learning models focus on these positive annotations for prediction and ignore these scarce but informative negative annotations, leading to an underestimation of precision. To address this issue, we introduce a deep learning method that utilizes a heterogeneous graph attention technique. The method first constructs a heterogeneous graph that covers the protein-protein interaction network, ontology structure, and positive and negative annotation information. Then, it learns embedding representations of proteins and ontology terms by using the heterogeneous graph attention technique. Finally, it leverages these learned representations to reconstruct the positive protein-term associations and score unobserved functional annotations. It can enhance the predictive performance by incorporating these known limited negative annotations into the constructed heterogeneous graph. Experimental results on three species (i.e., Human, Mouse, and Arabidopsis) demonstrate that our method can achieve better performance in predicting new protein annotations than state-of-the-art methods.

Emotion recognition in conversations (ERC), the task of recognizing the emotion of each utterance in a conversation, is crucial for building empathetic machines. Existing studies focus mainly on capturing context- and speaker-sensitive dependencies on the textual modality but ignore the significance of multimodal information. Different from emotion recognition in textual conversations, capturing intra- and inter-modal interactions between utterances, learning weights between different modalities, and enhancing modal representations play important roles in multimodal ERC. In this paper, we propose a transformer-based model with self-distillation (SDT)11The code is available athttps://github.com/butterfliesss/SDT.for the task. The transformer-based model captures intra- and inter-modal interactions by utilizing intra- and inter-modal transformers, and learns weights between modalities dynamically by designing a hierarchical gated fusion strategy. Furthermore, to learn more expressive modal representations, we treat soft labels of the proposed model as extra training supervision. Specifically, we introduce self-distillation to transfer knowledge of hard and soft labels from the proposed model to each modality. Experiments on IEMOCAP and MELD datasets demonstrate that SDT outperforms previous state-of-the-art baselines.

Automatic extraction of biomedical entities and their relations plays a significant role in biomedical curation tasks. Currently, the table-filling methods have received lots of attention in the general domain. However, the presence of complex lengthy sentences and overlapping relations in biomedical texts makes automatic extraction a challenging task. To address this challenge, we propose a joint extraction table-filling method based on the vertices of the triple region. We extract triples by using multi-label classification to mark the boundaries of the triples, fully utilizing the boundary information of the entities. To incorporate the information of the distance between entity pairs, distance embedding is introduced and dilated convolutions are utilized to capture multi-scale contextual information. We evaluated our model on the CHEMPROT and DDIExtraction2013 datasets. The experimental results demonstrate that our model achieves the state-of-the-art performance on both datasets.

Automatic extraction of entities and their relations from unstructured literature to form structured triples is essential for biomedical knowledge construction. Although most existing joint methods have effectively addressed some challenging problems in the biomedical corpora, i.e., the prevalent overlapping issue, they still suffer from a lack of consideration for the intrinsic correlations between entities and relations, as well as low computational efficiency. In this paper, we present a joint entity and relation extraction model with unified interaction maps. Specifically, we concatenate all relations in the natural language form with the input text to integrate the semantic information of relations through a deep Transformer-based encoder. In addition, we apply unified interaction maps to capture the correlations, which can naturally handle the overlapping issue. Extensive experiments on the CHEMPROT and DDIExtraction2013 datasets demonstrate the effectiveness of our model, achieving the state-of-the-art performance with higher efficiency.

Image-grounded dialogue systems benefit greatly from integrating visual information, resulting in high-quality response generation. However, current models struggle to effectively utilize such information in zero-resource scenarios, mainly due to the disparity between image and text modalities. To overcome this challenge, we propose an innovative multimodal framework, called ZRIGF, which assimilates image-grounded information for dialogue generation in zero-resource situations. ZRIGF implements a two-stage learning strategy, comprising contrastive pre-training and generative pre-training. Contrastive pre-training includes a text-image matching module that maps images and texts into a unified encoded vector space, along with a text-assisted masked image modeling module that preserves pre-training visual features and fosters further multimodal feature alignment. Generative pre-training employs a multimodal fusion module and an information transfer module to produce insightful responses based on harmonized multimodal representations. Comprehensive experiments conducted on both text-based and image-grounded dialogue datasets demonstrate ZRIGF's efficacy in generating contextually pertinent and informative responses. Furthermore, we adopt a fully zero-resource scenario in the image-grounded dialogue dataset to demonstrate our framework's robust generalization capabilities in novel domains.

MOTIVATION: Few-shot learning that can effectively perform named entity recognition in low-resource scenarios has raised growing attention, but it has not been widely studied yet in the biomedical field. In contrast to high-resource domains, biomedical named entity recognition (BioNER) often encounters limited human-labeled data in real-world scenarios, leading to poor generalization performance when training only a few labeled instances. Recent approaches either leverage cross-domain high-resource data or fine-tune the pre-trained masked language model using limited labeled samples to generate new synthetic data, which is easily stuck in domain shift problems or yields low-quality synthetic data. Therefore, in this article, we study a more realistic scenario, i.e. few-shot learning for BioNER. RESULTS: Leveraging the domain knowledge graph, we propose knowledge-guided instance generation for few-shot BioNER, which generates diverse and novel entities based on similar semantic relations of neighbor nodes. In addition, by introducing question prompt, we cast BioNER as question-answering task and propose prompt contrastive learning to improve the robustness of the model by measuring the mutual information between query-answer pairs. Extensive experiments conducted on various few-shot settings show that the proposed framework achieves superior performance. Particularly, in a low-resource scenario with only 20 samples, our approach substantially outperforms recent state-of-the-art models on four benchmark datasets, achieving an average improvement of up to 7.1% F1. AVAILABILITY AND IMPLEMENTATION: Our source code and data are available at https://github.com/cpmss521/KGPC.

Rapid and effective utilization of biomedical literature is paramount to combat diseases like COVID19. Biomedical named entity recognition (BioNER) is a fundamental task in text mining that can help physicians accelerate knowledge discovery to curb the spread of the COVID-19 epidemic. Recent approaches have shown that casting entity extraction as the machine reading comprehension task can significantly improve model performance. However, two major drawbacks impede higher success in identifying entities (1) ignoring the use of domain knowledge to capture the context beyond sentences and (2) lacking the ability to deeper understand the intent of questions. In this paper, to remedy this, we introduce and explore external domain knowledge which cannot be implicitly learned in text sequence. Previous works have focused more on text sequence and explored little of the domain knowledge. To better incorporate domain knowledge, a multi-way matching reader mechanism is devised to model representations of interaction between sequence, question and knowledge retrieved from Unified Medical Language System (UMLS). Benefiting from these, our model can better understand the intent of questions in complex contexts. Experimental results indicate that incorporating domain knowledge can help to obtain competitive results across 10 BioNER datasets, achieving absolute improvement of up to 2.02% in the f1 score.

The automatic extraction of the chemical-disease relation (CDR) from the text becomes critical because it takes a lot of time and effort to extract valuable CDR manually. Studies have shown that prior knowledge from the biomedical knowledge base is important for relation extraction. The method of combining deep learning models with prior knowledge is worthy of our study. In this paper, we propose a new model called Knowledge Guided Attention and Graph Convolutional Networks (KGAGN) for CDR extraction. First, to make full advantage of domain knowledge, we train entity embedding as a feature representation of input sequence, and relation embedding to capture weighted contextual information further through the attention mechanism. Then, to make full advantage of syntactic dependency information in cross-sentence CDR extraction, we construct document-level syntactic dependency graphs and encode them using a graph convolution network (GCN). Finally, the chemical-induced disease (CID) relation is extracted by using weighted context features and long-range dependency features both of which contain additional knowledge information We evaluated our model on the CDR dataset published by the BioCreative-V community and achieves an F1-score of 73.3%, surpassing other state-of-the-art methods. the code implemented by PyTorch 1.7.0 deep learning library can be downloaded from Github: https://github.com/sunyi123/cdr.

Ontologies are widely utilized in the biological domain for data annotation, integration, and analysis. Some representation learning methods have been proposed to learn the representation of entities to assist intelligent applications, such as knowledge discovery. However, most of them neglect the class information of entities in the ontology. In this article, we propose a unified framework, named ERCI, which jointly optimizes the knowledge graph embedding model and self-supervised learning. In this way, we can generate embeddings of bio-entities by fusing the class information. Moreover, ERCI is a pluggable framework that can be easily incorporated with any knowledge graph embedding model. We validate ERCI in two different ways. In the first way, we utilize the protein embeddings learned by the ERCI to predict protein-protein interactions on two different datasets. In the second way, we leverage the gene and disease embeddings generated by the ERCI to predict gene-disease associations. In addition, we create three datasets to simulate the long-tail scenario and evaluate ERCI on these. Experimental results show that ERCI has superior performance on all metrics compared with the state-of-the-art methods.

Proteins are the main undertakers of life activities, and accurately predicting their biological functions can help human better understand life mechanism and promote the development of themselves. With the rapid development of high-throughput technologies, an abundance of proteins are discovered. However, the gap between proteins and function annotations is still huge. To accelerate the process of protein function prediction, some computational methods taking advantage of multiple data have been proposed. Among these methods, the deep-learning-based methods are currently the most popular for their capability of learning information automatically from raw data. However, due to the diversity and scale difference between data, it is challenging for existing deep learning methods to capture related information from different data effectively. In this paper, we introduce a deep learning method that can adaptively learn information from protein sequences and biomedical literature, namely DeepAF. DeepAF first extracts the two kinds of information by using different extractors, which are built based on pre-trained language models and can capture rudimentary biological knowledge. Then, to integrate those information, it performs an adaptive fusion layer based on a Cross-attention mechanism that considers the knowledge of mutual interactions between two information. Finally, based on the mixed information, DeepAF utilizes logistic regression to obtain prediction scores. The experimental results on the datasets of two species (i.e., Human and Yeast) show that DeepAF outperforms other state-of-the-art approaches.

Automatic extraction of drug-drug interactions and their adverse effects can promote the research of pharmacovigilance and thus attracts attention from both academia and industry. Recent efforts focus on span-based approaches and show more promising results. However, span-based methods enumerate all possible candidate entity spans while ignoring boundary information of spans. Meanwhile, lacking sufficient interactions in intra-span and inter-span further hinders the performance of the nested entity and overlapping relation extraction. To this end, we propose a syntactic type-aware graph attention network for drug-drug interactions and their adverse effects extraction. Specifically, a boundary heuristic module is designed firstly to generate the boundary of linguistically legitimate entity spans. And then, different from the general syntactic graph (i.e., only considering dependency edges), we construct a syntactic type-aware graph attention network (STG) to capture interactions in intra-span and inter-span by considering syntactic edges and types simultaneously. Results1achieved on two biomedical benchmark datasets, including drug-drug interaction (DDI) and adverse drug effect (ADE), indicate that our model obtains significantly more performance than the state-of-the-art methods, achieving improvements in the relation F1 score of 1.63% on ADE and 2.07% on DDI dataset, respectively.

As an important task of natural language processing, medication recommendation aims to recommend medication combinations according to the electronic health record, which can also be regarded as a multi-label classification task. But patients often have multiple diseases simultaneously, and the model must consider drug-drug interactions (DDI) of medication combinations when recommending medications, making medication recommendation more difficult. There is little existing work to explore the changes in patient conditions. However, these changes may point to future trends in patient conditions that are critical for reducing DDI rates in recommended drug combinations. In this paper, we proposed the Patient Condition Change Network (PCCNet), which models the current core medications of patient by mining the temporal and spatial changes of patient medication order and patient condition vector, and allocates some auxiliary medications as the currently recommended medication combination. The experimental results show that the proposed model greatly reduces the recommended DDI of medications while achieving results no lower than the state-of-the-art results.11The code is available at https://github.com/master032/PCCNet

Medicine recommendation assists physicians in automatically providing medicine combinations, which is critical in health care. Existing efforts focus on making medicine recommendations based on the patient’s electronic health record(EHR). However, they ignore external medicine knowledge and are vulnerable to the missing EHR. In this paper, a knowledge-enhanced dual graph neural network (KDGN) is proposed to recommend medicine sets. KDGN combines diagnosis-level and procedure-level attention mechanisms to encode multiple types of medical codes. In order to mine medicine from medical knowledge, KDGN further designs a dual-graph neural network, which constructs a medicine co-occurrence graph and molecular connection graph, and retrieves potential therapeutic drugs. Furthermore, during the training phase, we introduce the automatic correction loss based on maximum likelihood estimation to mitigate the impact of missing EHR and enhance the robustness of KDGN. We evaluate the proposed model on the public MIMIC-III dataset, and experimental results show that KDGN outperforms the state-of-the-art model in 4 out of 5 evaluation metrics. Our dataset and code are available at: https://github.com/Benjamin-cell/KDGN.

Biomedical named entity recognition (Bio-NER) is an essential task for biomedical information extraction. In this paper, we regard word-level features and character-level features as two different modalities from a novel perspective and propose a biomedical named entity recognition model based on character feature enhancement with multimodal method (called BioNER-CFEM). BioNER-CFEM can not only capture interactions between modalities, but also learn interactions within modalities. In addition, our proposed cross-attention based sparse selection mechanism can effectively alleviate the noise in the interaction process of the two ‘modalities’. Experimental results show the effectiveness of BioNER-CFEM for the Bio-NER task: it achieves performance boost over SOTA models with competitive efficiency on all six Bio-NER datasets, i.e., $+0.89, +0.64, +0.40$, $+1.40, +5.57, +2.81$ on NCBI-Disease, BC5CDR-Disease, BC5CDR-Chem, BC2GM, JNLPBA, BC4CHEMD, respectively.

Automatic extraction of biomedical relations is important for many tasks, such as drug discovery, protein prediction. However, since the biomedical corpus contains a large number of complex long sentences and overlapping triples, most general domain joint modeling methods are unapplicable to the biomedical domain. Aiming at the problem of overlapping triples in the biomedical field, we design a token pair tagger to achieve single-stage joint biomedical entity and relation extraction. In addition, in order to make the tagger better understand the semantic information and structural information, we design both global-aware module and distance-aware module to introduce location information. Compared with previous methods, our model makes better use of location information, enhances the understanding of complex long sentences, and improves the ability to extract overlapping triples. We evaluated our model on the CHEMPROT and DDIExtraction2013 datasets. Experimental results show that our model significantly outperforms a range of baseline models, achieving the state-of-the-art performance.

Gene Ontology (GO) containing a set of standard concepts (or terms) is launched to unify the functional descriptions of proteins. Developing computational models based on GO to automatically annotate protein functions has been a longstanding active research area. In this paper, we propose a novel method to adaptively fuse functional and topological information between GO Terms. Our method is composed of a pre-trained language model for encoding protein sequences and an adaptive multi-view graph convolutional network (Multi-view GCN) for representing GO terms. Particularly, the Multi-view GCN considers multiple views from functional information, topological structures, and their combinations, and extracts multiple corresponding representations of GO terms. Then, an attention mechanism is applied to adaptively learn the importance weights of these representations. Finally, the predicted scores are calculated by using a dot product between protein sequence features and GO term representations. Experimental results on the datasets of two species (i.e., Human and Yeast) show that our method outperforms other state-of-the-art methods. The code of our proposed method is available at: https://github.com/Candyperfect/Master.

Intelligent medical services have attracted great research interests for providing automated medical consultation. However, the lack of corpora becomes a main obstacle to related research, particularly data from real scenarios. In this paper, we construct RealMedDial, a Chinese medical dialogue dataset based on real medical consultation. RealMedDial contains 2,637 medical dialogues and 24,255 utterances obtained from Chinese short-video clips of real medical consultations. We collected and annotated a wide range of meta-data with respect to medical dialogue including doctor profiles, hospital departments, diseases and symptoms for fine-grained analysis on language usage pattern and clinical diagnosis. We evaluate the performance of medical response generation, department routing and doctor recommendation on RealMedDial. Results show that RealMedDial are applicable to a wide range of NLP tasks with respect to medical dialogue.

Chinese Named Entity Recognition (NER) has continued to attract research attention. However, most existing studies only explore the internal features of the Chinese language but neglect other lingual modal features. Actually, as another modal knowledge of the Chinese language, English contains rich prompts about entities that can potentially be applied to improve the performance of Chinese NER. Therefore, in this study, we explore the bilingual enhancement for Chinese NER and propose a unified bilingual interaction module called the Adapted Cross-Transformers with Global Sparse Attention (ACT-S) to capture the interaction of bilingual information. We utilize a model built upon several different ACT-Ss to integrate the rich English information into the Chinese representation. Moreover, our model can learn the interaction of information between bilinguals (inter-features) and the dependency information within Chinese (intra-features). Compared with existing Chinese NER methods, our proposed model can better handle entities with complex structures. The English text that enhances the model is automatically generated by machine translation, avoiding high labour costs. Experimental results on four well-known benchmark datasets demonstrate the effectiveness and robustness of our proposed model.

BACKGROUND: Electronic medical records (EMR) contain detailed information about patient health. Developing an effective representation model is of great significance for the downstream applications of EMR. However, processing data directly is difficult because EMR data has such characteristics as incompleteness, unstructure and redundancy. Therefore, preprocess of the original data is the key step of EMR data mining. The classic distributed word representations ignore the geometric feature of the word vectors for the representation of EMR data, which often underestimate the similarities between similar words and overestimate the similarities between distant words. This results in word similarity obtained from embedding models being inconsistent with human judgment and much valuable medical information being lost. RESULTS: In this study, we propose a biomedical word embedding framework based on manifold subspace. Our proposed model first obtains the word vector representations of the EMR data, and then re-embeds the word vector in the manifold subspace. We develop an efficient optimization algorithm with neighborhood preserving embedding based on manifold optimization. To verify the algorithm presented in this study, we perform experiments on intrinsic evaluation and external classification tasks, and the experimental results demonstrate its advantages over other baseline methods. CONCLUSIONS: Manifold learning subspace embedding can enhance the representation of distributed word representations in electronic medical record texts. Reduce the difficulty for researchers to process unstructured electronic medical record text data, which has certain biomedical research value.

External document knowledge is helpful for dialogue systems to generate high-quality responses. Although several knowledge-grounded dialogue models have been designed, external knowledge cannot be comprehensively exploited due to the complex relationships among dialogue context, knowledge, and responses. To this end, we propose a novel transformer-based model, named TransIKG, which incorporates external document knowledge for dialogue generation. TransIKG comprises a two-step integration mechanism, including correlation integration and overall integration. Correlation integration is designed to fully exploit the pairwise mutual information among dialogue context, knowledge, and responses, while overall integration adopts an integration gate to capture global information. Furthermore, we utilize the positional information of dialogue turns to better represent the dialogue context and enhance the generalization ability of our model on out-of-domain documents. Finally, we propose a novel knowledge-aware pointer network to generate knowledge-enhanced response tokens. Experimental results on two benchmark datasets demonstrate that our model outperforms state-of-the-art models on both open-domain and domain-specific dialogues.

Biomedical Named Entity Recognition (BioNER) is often modeled as a sequence labeling task, which assigns the predefined label to each token in given input sequence. Although these sequential labeling models achieve significant achievements, they often fail to give precise boundaries of the named entity. In addition, a vast amount of work focuses more on textual sequence representation but ignores label information. To tackle these problems, in this paper, we directly model span-level named entity recognition, specifically, we treat the BioNER as a joint task of boundary detection and span classification under a multitask framework. In order to enhance boundary supervision, we introduce an entity type label as an additional guide and propose a co-attentive interactive mechanism to improve the span representation. Extensive experiments1on four benchmark datasets demonstrate that our proposed method obtains competitive results, achieving 90.26%, 78.04%, 90.21%, and 86.58% on BC5CDR, JNLPBA, NCBI, and BC2GM datasets, respectively, in terms of F1 score.

The emergence and spread of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) have created an enormous socioeconomic impact. Although there are several promising drug candidates in clinical trials, none of them are approved yet. Thus, the drug repositioning approach may help to overcome the current pandemic. However, the sparse dataset of COVID-19 limits the accuracy of existing drug repositioning. To overcome this problem, we propose a novel drug repositioning framework (named Drug2Cov). Drug2Cov can learn an effective representation via integrating self-supervised learning with sparse data. Meanwhile, Drug2Cov uses a heterogeneous graph neural network to capture the complex interaction between viruses, targets, and drugs that enhance the accuracy of drug repositioning. The experimental results demonstrate the effectiveness and feasibility of our proposed Drug2Cov framework. Source code and dataset are freely available at https://github.com/lhf3291109/Drug2Cov.

Automatically extracting entity-pair interactions from biomedical literature plays an important role in promoting the development of the biomedical field. For instance, the interactions between drugs can guide patients to take drugs correctly and avoid clinical adverse drug reactions; The interactions between proteins can help researchers design therapeutic drugs and discover disease mechanisms. However, it is found that relation instances of different classes generated from the same sentence are easily classified into the same class since their context information is almost the same. To address this issue, Triplet Loss based BERT (TL-BERT) approach is proposed in this paper, where the Triplet Loss training strategy is first introduced into the biomedical relation extraction field. Triplet Loss training strategy will increase the distances between these instances generated from the same sentence but belonging to different classes, and decrease the distances between these instances generated from different sentences but belonging to the same class. As a result, our approach can classify these instances generated from the same sentence more correctly. TL-BERT was evaluated on AIMed, BioInfer, and DDI Extraction-2013 corpus, the experimental results demonstrate that the Triplet Loss training strategy can improve the performance on both Protein-Protein interactions extraction tasks and Drug-Drug interactions detection tasks.

Intent detection and slot filling are two main tasks for building a spoken language understanding (SLU) system. Since the two tasks are closely related, the joint models for the two tasks always outperform the pipeline models in SLU. However, most joint models directly incorporate multiple intent information for each token, which introduces intent noise into the sentence semantics, causing a decrease in the performance of the joint model. In this paper, we propose a Dynamic Graph Model (DGM) for joint multiple intent detection and slot filling, in which we adopt a sentence-level intent-slot interactive graph to model the correlation between the intents and slot. Besides, we design a novel method of constructing the graph, which can dynamically update the interactive graph and further alleviate the error propagation. Experimental results on several multi-intent and single-intent datasets show that our model not only achieves the state-of-the-art (SOTA) performance but also boosts the speed by three to six times over the SOTA model.

BACKGROUND: The recognition of pharmacological substances, compounds and proteins is essential for biomedical relation extraction, knowledge graph construction, drug discovery, as well as medical question answering. Although considerable efforts have been made to recognize biomedical entities in English texts, to date, only few limited attempts were made to recognize them from biomedical texts in other languages. PharmaCoNER is a named entity recognition challenge to recognize pharmacological entities from Spanish texts. Because there are currently abundant resources in the field of natural language processing, how to leverage these resources to the PharmaCoNER challenge is a meaningful study. METHODS: Inspired by the success of deep learning with language models, we compare and explore various representative BERT models to promote the development of the PharmaCoNER task. RESULTS: The experimental results show that deep learning with language models can effectively improve model performance on the PharmaCoNER dataset. Our method achieves state-of-the-art performance on the PharmaCoNER dataset, with a max F1-score of 92.01%. CONCLUSION: For the BERT models on the PharmaCoNER dataset, biomedical domain knowledge has a greater impact on model performance than the native language (i.e., Spanish). The BERT models can obtain competitive performance by using WordPiece to alleviate the out of vocabulary limitation. The performance on the BERT model can be further improved by constructing a specific vocabulary based on domain knowledge. Moreover, the character case also has a certain impact on model performance.

The World Health Organization underlines the significance of adverse drug reaction (ADR) reports for patients' safety. Actually, many potential ADRs tend to be under-reported in post-market ADR surveillance. Recognizing ADRs from social media is indispensably important and could complement post-market ADR surveillance for more effective pharmacovigilance studies. However, previous approaches pose two challenges: 1) ADRs show high expression variability in social media, and thus, many potential ADRs are out-of-lexicon ones, which are difficult to be recognized, and 2) most phrasal ADRs are non-standard mentions and their boundaries are difficult to identify accurately. To tackle these challenges, we design three interaction graphs and propose a neural network approach, i.e., Interaction Graph Network (IGN). Specifically, to recognize more out-of-lexicon ADRs, besides the mentions in ADR lexicon, noun phrases in the input sentence are regarded as candidate phrases and their features are taken into considerations. Moreover, in an attempt to accurately identify ADR boundaries, three word-phrase interaction graphs are designed to represent lexicon knowledge and are encoded using graph attention networks (GATs) to directly integrate various boundary and contextual information of candidate phrases into ADR recognition. Experimental results on two benchmark datasets show that IGN can recognize ADR accurately and consistently outperforms other state-of-the-art approaches.

Estimating the similarity of biomedical sentence pair is an important component in such natural language processing (NLP) tasks as text retrieval and text summarization with great amount of biomedical information growing. Deep learning-based approaches have been successfully applied to the task, but they often rely on traditional pre-trained context-independent word embedding. Bidirectional Encoder Representations from Transformers (BERT) is recently employed to pre-train contextualized word/sentence representation models via bidirectional Transformers, outperforming the state-of-the-art for many NLP tasks. The mutual semantic influence between sentences is important for estimating semantic textual similarity, which is neglected in existing methods including BERT. On the other hand, biomedical corpora mainly consist of syntactic complex and long sentences. Owing to the above-mentioned issues, we proposed a hybrid architecture, integrating the pre-trained BERT and downstream bidirectional recurrent neural network (bi-RNN). The proposed model enhanced the sentence semantic representation via employing the self-attention instead of global attention to perform cross attention between sentences. Meanwhile, bi-RNN reduced redundant information in the output of BERT. Experimental results show that the best fine-tuned models consistently outperform previous methods and advance the state-of-the-art for clinical semantic textual similarity in OHNLP 2018 task 2, with up to 0.6% increase in Pearson correlation coefficient.

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is the strain of coronavirus that causes coronavirus disease 2019 (COVID-19), which leads to over 800,000 deaths and is still no specific medicines. Drug repositioning aiming to infer potential drugs for diseases and achieve much attention during the SARS-CoV-2 epidemic. However, find a specific drug of SARS-CoV-2 is still a large challenge that cannot be addressed well with current methods. To overcome this problem, we present a novel drug repositioning framework of heterogeneous graph convolutional networks for SARS-CoV2. The deep2CoV model can effectively search the potential drugs for SARS-CoV-2, which reduce the number of clinical trials and drug development cycles. The experimental results demonstrate the effectiveness and feasibility of our proposed deep2CoV framework.

Extracting protein-protein interaction affected by genetic mutation from biomedical literature automatically is an essential step toward the ultimate goal of precision medicine. However, the existing methods fail to be accurate enough to meet the needs in practice. In this paper, considering the significant progress made by the pre-training model in a wide variety of NLP tasks, we use BioBERT to obtain the representation of the text and adopt a multi-task learning strategy to improve the performance. Evaluated on the BioCreative VI PPIm data set, our proposed model achieves a new state-of-the-art performance that surpassed the previous one by 4.86% in F1-score. The source code is available at https://github.com/dlutwy/ppim.

Automatic extraction of biomedical relations is important for many tasks, such as drug discovery, protein prediction and knowledge graph construction. However, due to the complex and noisy expressions in biomedical texts, existing traditional neural networks, such as recurrent neural networks and convolutional neural networks, fail to capture syntactic information effectively. In this paper, we introduce a multi-head attention mechanism into graph convolutional networks to extract biomedical relations. In our method, the graph convolutional network is exploited to encode the dependency structure of an input sentence and the multi-head attention mechanism is utilized to alleviate the influence of noisy words. We evaluated our method on the ChemProt corpus and the protein-protein interaction corpus which includes five separate sub-datasets and it achieves F-scores of 67.37% and 84.8% on ChemProt corpus and PPI corpora, respectively. The experimental results suggest that Our model can not only alleviate the influence of noisy words, but also obtain more semantic and syntactic information from dependency graph than previous proposed models.

Relations between mutations and diseases hiding in biomedical literature are valuable for the analysis and interpretation of many complex diseases, which can help explore more effective treatment options for corresponding diseases. Most current document-level mutation-disease relation extraction methods are based on classification approaches and suffer the lack of ability to extract inter-sentential relations. To solve this problem, we regard extracting document-level mutation-disease relations as a sequence tagging task and propose a neural network-based method called Star-BiLSTM-LAN. By combining the star transformer and the Bi-directional Long Short-Term Memory network, this method achieves a strong ability to capture semantic and syntactic information at the document level from different aspects, and it can discover internal representations that prove useful for the task of interest. Star-BiLSTM-LAN is evaluated on EMU BCa and PCa datasets, and achieves the state-of-the-art F-scores of 89.20% and 90.43%, which are 4.70% and 2.43% higher than the baseline, respectively. Also, the proposed method achieves an F-score of 94.40% on BRONCO dataset.

In this paper, we propose a gated iterative capsule network model for the ADR detection task, named GICN. To alleviate the impact caused by abbreviations and misspelled words, we add character embedding as part of the input. Most ADRs consist of multiple words, e.g., short-term memory dysfunction. Hence, we apply a convolutional neural network (CNN) to obtain the complete phrase information. To effectively extract deep semantic information, we introduce a capsule network with a gated iteration unit that clusters features from underlying to high capsules. The gated iteration mechanism can remember contextual information, which will be introduced when clustering features. Experimental results show that our approach can achieve significant performance improvement for ADR detection from social media text compared with other state-of-the-art works.

MOTIVATION: The biomedical literature contains a wealth of chemical-protein interactions (CPIs). Automatically extracting CPIs described in biomedical literature is essential for drug discovery, precision medicine, as well as basic biomedical research. Most existing methods focus only on the sentence sequence to identify these CPIs. However, the local structure of sentences and external biomedical knowledge also contain valuable information. Effective use of such information may improve the performance of CPI extraction. RESULTS: In this article, we propose a novel neural network-based approach to improve CPI extraction. Specifically, the approach first employs BERT to generate high-quality contextual representations of the title sequence, instance sequence and knowledge sequence. Then, the Gaussian probability distribution is introduced to capture the local structure of the instance. Meanwhile, the attention mechanism is applied to fuse the title information and biomedical knowledge, respectively. Finally, the related representations are concatenated and fed into the softmax function to extract CPIs. We evaluate our proposed model on the CHEMPROT corpus. Our proposed model is superior in performance as compared with other state-of-the-art models. The experimental results show that the Gaussian probability distribution and external knowledge are complementary to each other. Integrating them can effectively improve the CPI extraction performance. Furthermore, the Gaussian probability distribution can effectively improve the extraction performance of sentences with overlapping relations in biomedical relation extraction tasks. AVAILABILITY AND IMPLEMENTATION: Data and code are available at https://github.com/CongSun-dlut/CPI_extraction. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Biomedical document triage is the foundation of biomedical information extraction, which is important to precision medicine. Recently, some neural networks-based methods have been proposed to classify biomedical documents automatically. In the biomedical domain, documents are often very long and often contain very complicated sentences. However, the current methods still find it difficult to capture important features across sentences. RESULTS: In this paper, we propose a hierarchical attention-based capsule model for biomedical document triage. The proposed model effectively employs hierarchical attention mechanism and capsule networks to capture valuable features across sentences and construct a final latent feature representation for a document. We evaluated our model on three public corpora. CONCLUSIONS: Experimental results showed that both hierarchical attention mechanism and capsule networks are helpful in biomedical document triage task. Our method proved itself highly competitive or superior compared with other state-of-the-art methods.

BACKGROUND: Most biomedical information extraction focuses on binary relations within single sentences. However, extracting n-ary relations that span multiple sentences is in huge demand. At present, in the cross-sentence n-ary relation extraction task, the mainstream method not only relies heavily on syntactic parsing but also ignores prior knowledge. RESULTS: In this paper, we propose a novel cross-sentence n-ary relation extraction method that utilizes the multihead attention and knowledge representation that is learned from the knowledge graph. Our model is built on self-attention, which can directly capture the relations between two words regardless of their syntactic relation. In addition, our method makes use of entity and relation information from the knowledge base to impose assistance while predicting the relation. Experiments on n-ary relation extraction show that combining context and knowledge representations can significantly improve the n-ary relation extraction performance. Meanwhile, we achieve comparable results with state-of-the-art methods. CONCLUSIONS: We explored a novel method for cross-sentence n-ary relation extraction. Unlike previous approaches, our methods operate directly on the sequence and learn how to model the internal structures of sentences. In addition, we introduce the knowledge representations learned from the knowledge graph into the cross-sentence n-ary relation extraction. Experiments based on knowledge representation learning show that entities and relations can be extracted in the knowledge graph, and coding this knowledge can provide consistent benefits.

In order to save time and cost, many disease gene prediction methods have been proposed in recent years. However, the traditional network model uses a binary relationship to represent the relationship between different proteins or gene molecules and phenotypes, which leads to the loss of information. Recently, hypergraph shows that it can overcome this loss of information to some extent and preserve the multivariate relationship, so we transformed the disease gene prediction problem into the problem of ranking the multivariate-relationship object. In this paper, we propose a method of Heterogeneous Probabilistic Hypergraph Ranking (HPHR) to predict disease genes. Firstly, fix a graph centroid for each hyperedge and according to different associations, and add other nodes related to the graph centroid to hyperedges with a certain probability. Then transform the problem of predicting disease genes into the problem of ranking heterogeneous objects, and the candidate genes are sorted by hypergraph ranking. The method is then applied to the integrated disease gene network. Compared with other prediction methods achieved better results, which was verified by this experiment.

BACKGROUND: The number of biomedical research articles have increased exponentially with the advancement of biomedicine in recent years. These articles have thus brought a great difficulty in obtaining the needed information of researchers. Information retrieval technologies seek to tackle the problem. However, information needs cannot be completely satisfied by directly introducing the existing information retrieval techniques. Therefore, biomedical information retrieval not only focuses on the relevance of search results, but also aims to promote the completeness of the results, which is referred as the diversity-oriented retrieval. RESULTS: We address the diversity-oriented biomedical retrieval task using a supervised term ranking model. The model is learned through a supervised query expansion process for term refinement. Based on the model, the most relevant and diversified terms are selected to enrich the original query. The expanded query is then fed into a second retrieval to improve the relevance and diversity of search results. To this end, we propose three diversity-oriented optimization strategies in our model, including the diversified term labeling strategy, the biomedical resource-based term features and a diversity-oriented group sampling learning method. Experimental results on TREC Genomics collections demonstrate the effectiveness of the proposed model in improving the relevance and the diversity of search results. CONCLUSIONS: The proposed three strategies jointly contribute to the improvement of biomedical retrieval performance. Our model yields more relevant and diversified results than the state-of-the-art baseline models. Moreover, our method provides a general framework for improving biomedical retrieval performance, and can be used as the basis for future work.

BACKGROUND: The adverse reactions that are caused by drugs are potentially life-threatening problems. Comprehensive knowledge of adverse drug reactions (ADRs) can reduce their detrimental impacts on patients. Detecting ADRs through clinical trials takes a large number of experiments and a long period of time. With the growing amount of unstructured textual data, such as biomedical literature and electronic records, detecting ADRs in the available unstructured data has important implications for ADR research. Most of the neural network-based methods typically focus on the simple semantic information of sentence sequences; however, the relationship of the two entities depends on more complex semantic information. METHODS: In this paper, we propose multihop self-attention mechanism (MSAM) model that aims to learn the multi-aspect semantic information for the ADR detection task. first, the contextual information of the sentence is captured by using the bidirectional long short-term memory (Bi-LSTM) model. Then, via applying the multiple steps of an attention mechanism, multiple semantic representations of a sentence are generated. Each attention step obtains a different attention distribution focusing on the different segments of the sentence. Meanwhile, our model locates and enhances various keywords from the multiple representations of a sentence. RESULTS: Our model was evaluated by using two ADR corpora. It is shown that the method has a stable generalization ability. Via extensive experiments, our model achieved F-measure of 0.853, 0.799 and 0.851 for ADR detection for TwiMed-PubMed, TwiMed-Twitter, and ADE, respectively. The experimental results showed that our model significantly outperforms other compared models for ADR detection. CONCLUSIONS: In this paper, we propose a modification of multihop self-attention mechanism (MSAM) model for an ADR detection task. The proposed method significantly improved the learning of the complex semantic information of sentences.

The explosive growth of biomedical literature has created a rich source of knowledge, such as that on protein-protein interactions (PPIs) and drug-drug interactions (DDIs), locked in unstructured free text. Biomedical relation classification aims to automatically detect and classify biomedical relations, which has great benefits for various biomedical research and applications. In the past decade, significant progress has been made in biomedical relation classification. With the advance of neural network methodology, neural network-based approaches have been applied in biomedical relation classification and achieved state-of-the-art performance for some public datasets and shared tasks. In this review, we describe the recent advancement of neural network-based approaches for classifying biomedical relations. We summarize the available corpora and introduce evaluation metrics. We present the general framework for neural network-based approaches in biomedical relation extraction and pretrained word embedding resources. We discuss neural network-based approaches, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs). We conclude by describing the remaining challenges and outlining future directions.

Yufeng Diao, Hongfei Lin, Di Wu, Liang Yang, Kan Xu, Zhihao Yang, Jian Wang, Shaowu Zhang, Bo Xu, Dongyu Zhang. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.

Motivation: In biomedical research, chemical is an important class of entities, and chemical named entity recognition (NER) is an important task in the field of biomedical information extraction. However, most popular chemical NER methods are based on traditional machine learning and their performances are heavily dependent on the feature engineering. Moreover, these methods are sentence-level ones which have the tagging inconsistency problem. Results: In this paper, we propose a neural network approach, i.e. attention-based bidirectional Long Short-Term Memory with a conditional random field layer (Att-BiLSTM-CRF), to document-level chemical NER. The approach leverages document-level global information obtained by attention mechanism to enforce tagging consistency across multiple instances of the same token in a document. It achieves better performances with little feature engineering than other state-of-the-art methods on the BioCreative IV chemical compound and drug name recognition (CHEMDNER) corpus and the BioCreative V chemical-disease relation (CDR) task corpus (the F-scores of 91.14 and 92.57%, respectively). Availability and implementation: Data and code are available at https://github.com/lingluodlut/Att-ChemdNER. Contact: [email protected] or [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

Motivation: Adverse events resulting from drug-drug interactions (DDI) pose a serious health issue. The ability to automatically extract DDIs described in the biomedical literature could further efforts for ongoing pharmacovigilance. Most of neural networks-based methods typically focus on sentence sequence to identify these DDIs, however the shortest dependency path (SDP) between the two entities contains valuable syntactic and semantic information. Effectively exploiting such information may improve DDI extraction. Results: In this article, we present a hierarchical recurrent neural networks (RNNs)-based method to integrate the SDP and sentence sequence for DDI extraction task. Firstly, the sentence sequence is divided into three subsequences. Then, the bottom RNNs model is employed to learn the feature representation of the subsequences and SDP, and the top RNNs model is employed to learn the feature representation of both sentence sequence and SDP. Furthermore, we introduce the embedding attention mechanism to identify and enhance keywords for the DDI extraction task. We evaluate our approach using the DDI extraction 2013 corpus. Our method is competitive or superior in performance as compared with other state-of-the-art methods. Experimental results show that the sentence sequence and SDP are complementary to each other. Integrating the sentence sequence with SDP can effectively improve the DDI extraction performance. Availability and implementation: The experimental data is available at https://github.com/zhangyijia1979/hierarchical-RNNs-model-for-DDI-extraction. Contact: [email protected] or [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

BACKGROUND: Biomedical event extraction is a crucial task in biomedical text mining. As the primary forum for international evaluation of different biomedical event extraction technologies, BioNLP Shared Task represents a trend in biomedical text mining toward fine-grained information extraction (IE). The fourth series of BioNLP Shared Task in 2016 (BioNLP-ST'16) proposed three tasks, in which the Bacteria Biotope event extraction (BB) task has been put forward in the earlier BioNLP-ST. Deep learning methods provide an effective way to automatically extract more complex features and achieve notable results in various natural language processing tasks. RESULTS: The experimental results show that the presented approach can achieve an F-score of 57.42% in the test set, which outperforms previous state-of-the-art official submissions to BioNLP-ST 2016. CONCLUSIONS: In this paper, we propose a novel Gated Recurrent Unit Networks framework integrating attention mechanism for extracting biomedical events between biotope and bacteria from biomedical literature, utilizing the corpus from the BioNLP'16 Shared Task on Bacteria Biotope task. The experimental results demonstrate the potential and effectiveness of the proposed framework.

BACKGROUND: Protein complexes are one of the keys to deciphering the behavior of a cell system. During the past decade, most computational approaches used to identify protein complexes have been based on discovering densely connected subgraphs in protein-protein interaction (PPI) networks. However, many true complexes are not dense subgraphs and these approaches show limited performances for detecting protein complexes from PPI networks. RESULTS: To solve these problems, in this paper we propose a supervised learning method based on network node embeddings which utilizes the informative properties of known complexes to guide the search process for new protein complexes. First, node embeddings are obtained from human protein interaction network. Then the protein interactions are weighted through the similarities between node embeddings. After that, the supervised learning method is used to detect protein complexes. Then the random forest model is used to filter the candidate complexes in order to obtain the final predicted complexes. Experimental results on real human and yeast protein interaction networks show that our method effectively improves the performance for protein complex detection. CONCLUSIONS: We provided a new method for identifying protein complexes from human and yeast protein interaction networks, which has great potential to benefit the field of protein complex detection.

BACKGROUND: Drug discovery is the process through which potential new medicines are identified. High-throughput screening and computer-aided drug discovery/design are the two main drug discovery methods for now, which have successfully discovered a series of drugs. However, development of new drugs is still an extremely time-consuming and expensive process. Biomedical literature contains important clues for the identification of potential treatments. It could support experts in biomedicine on their way towards new discoveries. METHODS: Here, we propose a biomedical knowledge graph-based drug discovery method called SemaTyP, which discovers candidate drugs for diseases by mining published biomedical literature. We first construct a biomedical knowledge graph with the relations extracted from biomedical abstracts, then a logistic regression model is trained by learning the semantic types of paths of known drug therapies' existing in the biomedical knowledge graph, finally the learned model is used to discover drug therapies for new diseases. RESULTS: The experimental results show that our method could not only effectively discover new drug therapies for new diseases, but also could provide the potential mechanism of action of the candidate drugs. CONCLUSIONS: In this paper we propose a novel knowledge graph based literature mining method for drug discovery. It could be a supplementary method for current drug discovery methods.

BACKGROUND: In biomedical information extraction, event extraction plays a crucial role. Biological events are used to describe the dynamic effects or relationships between biological entities such as proteins and genes. Event extraction is generally divided into trigger detection and argument recognition. The performance of trigger detection directly affects the results of the event extraction. In general, the traditional method is used to address the trigger detection as a classification task, as well as the use of machine learning or rules method, which construct many features to improve the classification results. Moreover, the classification model only recognizes triggers composed of single words, whereas for multiple words, the result is unsatisfactory. RESULTS: The corpus of our model is MLEE. If we were to only use the biomedical LSTM and CRF model without other features, the F-score would reach about 78.08%. Comparing entity to part of speech (POS), we find the entity features more conducive to the improvement of performance of detection, with the F-score potentially reaching about 80%. Furthermore, we also experiment on the other three corpora (BioNLP 2009, BioNLP 2011, and BioNLP 2013) to verify the generalization of our model. Hence, F-scores can reach more than 60%, which are better than the comparative experiments. CONCLUSIONS: The trigger recognition method based on the sequence annotation model does not require initial complex feature engineering, and only requires a simple labeling mechanism to complete the training. Therefore, generalization of our model is better compared to other traditional models. Secondly, this method can identify multi-word triggers, thereby improving the F-scores of trigger recognition. Thirdly, details on the entity have a crucial impact on trigger detection. Finally, the combination of character-level word embedding and word-level word embedding provides increasingly effective information for the model; therefore, it is a key to the success of the experiment.

Biomedical relation extraction can automatically extract high-quality biomedical relations from biomedical texts, which is a vital step for the mining of biomedical knowledge hidden in the literature. Recurrent neural networks (RNNs) and convolutional neural networks (CNNs) are two major neural network models for biomedical relation extraction. Neural network-based methods for biomedical relation extraction typically focus on the sentence sequence and employ RNNs or CNNs to learn the latent features from sentence sequences separately. However, RNNs and CNNs have their own advantages for biomedical relation extraction. Combining RNNs and CNNs may improve biomedical relation extraction. In this paper, we present a hybrid model for the extraction of biomedical relations that combines RNNs and CNNs. First, the shortest dependency path (SDP) is generated based on the dependency graph of the candidate sentence. To make full use of the SDP, we divide the SDP into a dependency word sequence and a relation sequence. Then, RNNs and CNNs are employed to automatically learn the features from the sentence sequence and the dependency sequences, respectively. Finally, the output features of the RNNs and CNNs are combined to detect and extract biomedical relations. We evaluate our hybrid model using five public (protein-protein interaction) PPI corpora and a (drug-drug interaction) DDI corpus. The experimental results suggest that the advantages of RNNs and CNNs in biomedical relation extraction are complementary. Combining RNNs and CNNs can effectively boost biomedical relation extraction performance.

In these years, the number of biomedical articles has increased exponentially, which becomes a problem for biologists to capture all the needed information manually. Information retrieval technologies, as the core of search engines, can deal with the problem automatically, providing users with the needed information. However, it is a great challenge to apply these technologies directly for biomedical retrieval, because of the abundance of domain specific terminologies. To enhance biomedical retrieval, we propose a novel framework based on learning to rank. Learning to rank is a series of state-of-the-art information retrieval techniques, and has been proved effective in many information retrieval tasks. In the proposed framework, we attempt to tackle the problem of the abundance of terminologies by constructing ranking models, which focus on not only retrieving the most relevant documents, but also diversifying the searching results to increase the completeness of the resulting list for a given query. In the model training, we propose two novel document labeling strategies, and combine several traditional retrieval models as learning features. Besides, we also investigate the usefulness of different learning to rank approaches in our framework. Experimental results on TREC Genomics datasets demonstrate the effectiveness of our framework for biomedical information retrieval.

Trigger detection plays a key role in the extraction of biomedical events, so it will influence the results of biomedical events extraction directly. The traditional biomedical event trigger recognition method is based on artificial design features and construct feature vectors; Not only does it consume great amounts of manpower, it also lacks system generalization ability. Most of methods of trigger detection are based on the convolutional neural network that identify each word in the text, and regard it as a multi-classification task. However for the multi-word composed of the trigger, there is no useful recognition effect. In this paper, we will use the IBO format and consider the trigger detection as a task of sequence annotation, a solution that improves the recognition accuracy of multi-word triggers by bidirectional LSTM and CRF.

BACKGROUND: Drug-drug interactions (DDIs) often bring unexpected side effects. The clinical recognition of DDIs is a crucial issue for both patient safety and healthcare cost control. However, although text-mining-based systems explore various methods to classify DDIs, the classification performance with regard to DDIs in long and complex sentences is still unsatisfactory. METHODS: In this study, we propose an effective model that classifies DDIs from the literature by combining an attention mechanism and a recurrent neural network with long short-term memory (LSTM) units. In our approach, first, a candidate-drug-oriented input attention acting on word-embedding vectors automatically learns which words are more influential for a given drug pair. Next, the inputs merging the position- and POS-embedding vectors are passed to a bidirectional LSTM layer whose outputs at the last time step represent the high-level semantic information of the whole sentence. Finally, a softmax layer performs DDI classification. RESULTS: Experimental results from the DDIExtraction 2013 corpus show that our system performs the best with respect to detection and classification (84.0% and 77.3%, respectively) compared with other state-of-the-art methods. In particular, for the Medline-2013 dataset with long and complex sentences, our F-score far exceeds those of top-ranking systems by 12.6%. CONCLUSIONS: Our approach effectively improves the performance of DDI classification tasks. Experimental analysis demonstrates that our model performs better with respect to recognizing not only close-range but also long-range patterns among words, especially for long, complex and compound sentences.

Adverse drug reactions between chemicals and diseases make chemical-disease relations (CDR) become a research focus. In this paper, we present a chemical-induced disease (CID) relation extraction system, CIDExtractor, to extract CID relations from biomedical literature. CIDExtractor first employs a sentence-level classifier to extract the CID relations located in the same sentence. To construct the classifier, a sentence-level training set is manually annotated and then Co-Training algorithm is used to exploit the unlabeled data with the feature kernel and graph kernel as two independent views. Then CIDExtractor uses a document-level classifier to extract the CID relations spanning multiple sentences. The classifier utilizes the document level information (features) of the chemical and disease pair. Finally, some post-processing rules are applied to the union set of two classifiers and generate the final outputs. Experimental results on the test set of BioCreative V CDR CID subtask show that CIDExtractor can achieve better performance (an F-score of 67.72%) than the state-of-the-art methods. The online CIDExtractor demonstration system is available at http://202.118.75.18:8888/cdr-dut-ir/cid.html.

The two main problems of biomedical event extraction are trigger identification and argument detection which can both be considered as classification problems. In this paper, we propose a distributed representation method, which combines context, consisted by dependency-based word embedding, and task-based features represented in a distributed way on deep learning models to realize biomedical event extraction. The experimental results on Multi-Level Event Extraction (MLEE) corpus show higher F-scores compared to the state-of-the-art SVM method. This demonstrates that our proposed method is effective for biomedical event extraction.

In this paper, we present a deep learning based disease named entity recognition architecture. First, the word-level embedding, character-level embedding and lexicon feature embedding are concatenated as input. Then multiple convolutional layers are stacked over the input to extract useful features automatically. Finally, multiple label strategy, which is firstly introduced, is applied to the output layer to capture the correlation information between neighboring labels. Experimental results on both NCBI and CDR corpora show that ML-CNN can achieve the state-of-the-art performance.

High-throughput experimental techniques have produced a large amount of human protein-protein interactions, making it possible to construct a large-scale human PPI network and detect human protein complexes from the network with computational approaches. However, most of current complex detection methods are based on graph theory which can't utilize the information of the known complexes. In this paper, we present a supervised learning method to detect protein complexes in a human PPI network. In this method, biological characteristics and properties of the network are taken into consideration to construct a rich feature set to train a regression model for protein complex detection. In addition, the specific disease related PPIs are extracted from biomedical literatures and then integrated into the original PPI network for detecting the disease-specific protein complexes more effectively. Experimental results show that the performance of our method is superior to other existing state-of-the-art methods. Furthermore, through the analysis of the breast cancer specific complexes detected with our method, more biological insights for breast cancer (e.g., some candidate susceptible genes of breast cancer) are provided.

MOTIVATION: Detecting drug-drug interaction (DDI) has become a vital part of public health safety. Therefore, using text mining techniques to extract DDIs from biomedical literature has received great attentions. However, this research is still at an early stage and its performance has much room to improve. RESULTS: In this article, we present a syntax convolutional neural network (SCNN) based DDI extraction method. In this method, a novel word embedding, syntax word embedding, is proposed to employ the syntactic information of a sentence. Then the position and part of speech features are introduced to extend the embedding of each word. Later, auto-encoder is introduced to encode the traditional bag-of-words feature (sparse 0-1 vector) as the dense real value vector. Finally, a combination of embedding-based convolutional features and traditional features are fed to the softmax classifier to extract DDIs from biomedical literature. Experimental results on the DDIExtraction 2013 corpus show that SCNN obtains a better performance (an F-score of 0.686) than other state-of-the-art methods. AVAILABILITY AND IMPLEMENTATION: The source code is available for academic use at http://202.118.75.18:8080/DDI/SCNN-DDI.zip CONTACT: [email protected] information: Supplementary data are available at Bioinformatics online.

BACKGROUND: Recently, high-throughput experimental techniques have generated a large amount of protein-protein interaction (PPI) data which can construct large complex PPI networks for numerous organisms. System biology attempts to understand cellular organization and function by analyzing these PPI networks. However, most studies still focus on static PPI networks which neglect the dynamic information of PPI. RESULTS: The gene expression data under different time points and conditions can reveal the dynamic information of proteins. In this study, we used an active probability-based method to distinguish the active level of proteins at different active time points. We constructed dynamic probabilistic protein networks (DPPN) to integrate dynamic information of protein into static PPI networks. Based on DPPN, we subsequently proposed a novel method to identify protein complexes, which could effectively exploit topological structure as well as dynamic information of DPPN. We used three different yeast PPI datasets and gene expression data to construct three DPPNs. When applied to three DPPNs, many well-characterized protein complexes were accurately identified by this method. CONCLUSION: The shift from static PPI networks to dynamic PPI networks is essential to accurately identify protein complex. This method not only can be applied to identify protein complex, but also establish a framework to integrate dynamic information into static networks for other applications, such as pathway analysis.

BACKGROUND: Accurate determination of protein complexes has become a key task of system biology for revealing cellular organization and function. Up to now, the protein complex prediction methods are mostly focused on static protein protein interaction (PPI) networks. However, cellular systems are highly dynamic and responsive to cues from the environment. The shift from static PPI networks to dynamic PPI networks is essential to accurately predict protein complex. RESULTS: The gene expression data contains crucial dynamic information of proteins and PPIs, along with high-throughput experimental PPI data, are valuable for protein complex prediction. Firstly, we exploit gene expression data to calculate the active time point and the active probability of each protein and PPI. The dynamic active information is integrated into high-throughput PPI data to construct dynamic PPI networks. Secondly, a novel method for predicting protein complexes from the dynamic PPI networks is proposed based on core-attachment structural feature. Our method can effectively exploit not only the dynamic active information but also the topology structure information based on the dynamic PPI networks. CONCLUSIONS: We construct four dynamic PPI networks, and accurately predict many well-characterized protein complexes. The experimental results show that (i) the dynamic active information significantly improves the performance of protein complex prediction; (ii) our method can effectively make good use of both the dynamic active information and the topology structure information of dynamic PPI networks to achieve state-of-the-art protein complex prediction capabilities.

Biomedical events can reveal crucial processes in biomedical research. As an important step in biomedical event extraction, biomedical event trigger detection has become a research hotspot. Traditional machine learning methods, which aim to manually design powerful features fed to the classifiers, greatly depend on the understanding of the specific task. In this paper, we propose an approach to automatically learn good features from raw input without manual intervention. The approach is based on dependency-based word embedding and first learns dependency-based word embedding from all available PubMed abstracts. The word embedding contains rich functional and semantic information. Then neural network architecture is used to learn better feature representation based on raw dependency-based word embedding. Meanwhile, we dynamically adjust the embedding while training for adapting to the trigger classification task. Finally, softmax classifier labels the examples by specific trigger class using the features learned by the model. The experimental results show that our approach achieves a micro F1 score of 78.27% and a macro F1 score of 76.94% in significant trigger classes, and performs better than baseline methods. In addition, we can achieve the semantic distributed representation of every trigger word.

Research articles in biomedicine domain have increased exponentially, which makes it more and more difficult for biologists to manually capture all the information they need. Information retrieval technologies can help to obtain the users' needed information automatically. However, it is a great challenge to apply these technologies to biomedicine domain directly because of some domain specific characteristics, such as the abundance of terminologies. To enhance the effectiveness of the biomedical information retrieval, we propose a novel framework based on the state-of-the-art information retrieval methods, called learning to rank, which has been proved effective to rank documents based on their relevance degree. In the framework, we attempt to tackle the problem of the abundance of terminologies by constructing ranking models, which focus on not only retrieving the most relevant documents but also diversifying the searching results to increase the completeness of the resulting list for a given query. In the model training, we propose two novel document labeling strategies, and combine several traditional retrieval models as learning features. Besides, we also investigate the usefulness of different learning to rank approaches in our framework. Experimental results on TREC Genomics datasets demonstrate our proposed framework is effective in improving the performance of biomedical information retrieval.

In these years, the adverse drug reactions (ADRs) have seriously impacted the people's health, and adverse drug event reporting systems become a key means to monitor the drug safety, in which healthcare professionals or drug consumers can submit the adverse drug event reports based on their experience or professional knowledge. However, with the increase of drugs, the number of the submitted reports increases rapidly, making it more and more difficult to capture all the ADRs manually. To tackle the problem, we develop a novel system to compute the similarities among the drugs and adverse reactions automatically from the reports. In the method, we represent the mentions of drugs and adverse reactions as distributed vectors using the skip-gram model, and discover the most potential adverse drug reactions based on the similarities.

This paper presents a deep neural network-based protein-protein interactions (PPIs) information extraction approach which can learn complex and abstract features automatically from unlabeled data by unsupervised representation learning methods. This approach first employs the training algorithm of auto-encoders to initialize the parameters of a deep multilayer neural network. Then the gradient descent method using back-propagation is applied to train this deep multilayer neural network model. Experimental results on five public PPI corpora show that our method can achieve better performance than can a multilayer neural network. In addition, the performance comparison with APG also verifies the effectiveness of our method.

BACKGROUND: Revealing protein complexes are important for understanding principles of cellular organization and function. High-throughput experimental techniques have produced a large amount of protein interactions, which makes it possible to predict protein complexes from protein-protein interaction (PPI) networks. However, the small amount of known physical interactions may limit protein complex detection. METHODS: The new PPI networks are constructed by integrating PPI datasets with the large and readily available PPI data from biomedical literature, and then the less reliable PPI between two proteins are filtered out based on semantic similarity and topological similarity of the two proteins. Finally, the supervised learning protein complex detection (SLPC), which can make full use of the information of available known complexes, is applied to detect protein complex on the new PPI networks. RESULTS: The experimental results of SLPC on two different categories yeast PPI networks demonstrate effectiveness of the approach: compared with the original PPI networks, the best average improvements of 4.76, 6.81 and 15.75 percentage units in the F-score, accuracy and maximum matching ratio (MMR) are achieved respectively; compared with the denoising PPI networks, the best average improvements of 3.91, 4.61 and 12.10 percentage units in the F-score, accuracy and MMR are achieved respectively; compared with ClusterONE, the start-of the-art complex detection method, on the denoising extended PPI networks, the average improvements of 26.02 and 22.40 percentage units in the F-score and MMR are achieved respectively. CONCLUSIONS: The experimental results show that the performances of SLPC have a large improvement through integration of new receivable PPI data from biomedical literature into original PPI networks and denoising PPI networks. In addition, our protein complexes detection method can achieve better performance than ClusterONE.

In this paper, we analyze six protein interaction networks widely used for protein complex detection, and compare the performance of six classic computational methods on them in order to find the relations between network characteristics and the performances of these complex detection methods. Furthermore, we explore the difference among the two complexes detected by different methods and the real complex by using a visualization approach which can easily find the difference of two methods and locate the undetected proteins.

In this paper, we present an approach to automatically construct disease related knowledge summarization from biomedical literature. In this approach, first Kullback-Leibler divergence combined with mutual information metric is used to extract disease salient information. Then deep search based on depth first search (DFS) is applied to find hidden relations between biomedical entities. Finally random walk algorithm is exploited to filter out the weak relations. The experimental results show that our approach achieves a precision of 60% and a recall of 61% on salient information extraction, and outperforms the method of Combo. In addition, the method of deep search obtains more hidden relations than the original correlation extraction methods.

The rapidly growing biomedical literature provides a significantly large and readily available source of PPI data. In this paper, we present supervised learning and data integration based complex detection approach. In this approach, a sophisticated natural language processing system, PPIExtractor, is employed to extract new PPI interactions from biomedical literature which are then integrated into original PPI networks. Then a supervised learning model, built by via of the information of available known complexes, is used in the multiple complex detection stages, e.g. the cliques filtering, growth, and candidate complex filtering. The experimental results on three yeast PPI networks demonstrate the effectiveness of our approach.

BACKGROUND: The amount of biomedical literature available is growing at an explosive speed, but a large amount of useful information remains undiscovered in it. Researchers can make informed biomedical hypotheses through mining this literature. Unfortunately, popular mining methods based on co-occurrence produce too many target concepts, leading to the declining relevance ranking of the potential target concepts. METHODS: This paper presents a new method for selecting linking concepts which exploits statistical and textual features to represent each linking concept, and then classifies them as relevant or irrelevant to the starting concepts. Relevant linking concepts are then used to discover target concepts. RESULTS: Through an evaluation it is observed textual features improve the results obtained with only statistical features. We successfully replicate Swanson's two classic discoveries and find the rankings of potentially relevant target concepts are relatively high. CONCLUSIONS: The number of target concepts is greatly reduced and potentially relevant target concepts gain higher ranking by adopting only relevant linking concepts. Thus, the proposed method has the potential to help biomedical experts find the most useful and valuable target concepts effectively.

In this paper, we present a supervised learning-based method for predicting protein complexes in protein interaction network. The method extracts rich features from protein interaction network to train a Regression model, which is then used for the cliques filtering, growth, and candidate complex filtering. The experimental results on several protein interaction networks show that our method outperforms other state-of-the-art protein complex detection methods.

Prediction of protein complexes from protein-protein interaction (PPI) networks is crucial to unraveling the principles of cellular organization. Most existing approaches only exploit high-throughput experimental PPI data to predict protein complexes. In this paper, we integrate the multiple biomedical resources for protein complex prediction by constructing attributed PPI networks, which include high-throughput data, co-expression data, genomic data, text mining data and gene ontology data. Multiple biomedical resources are complementary in attributed PPI networks. We propose a novel approach called IMBP based on attributed PPI networks. IMBP can effectively learn the degree of contributions of different biomedical resource for complex prediction. The experimental results show that IMBP can make good use of multiple biomedical data and achieve state-of-the-art performance.

Protein complexes are important for unraveling the secrets of cellular organization and function. Many computational approaches have been developed to predict protein complexes in protein-protein interaction (PPI) networks. However, most existing approaches focus mainly on the topological structure of PPI networks, and largely ignore the gene ontology (GO) annotation information. In this paper, we constructed ontology attributed PPI networks with PPI data and GO resource. After constructing ontology attributed networks, we proposed a novel approach called CSO (clustering based on network structure and ontology attribute similarity). Structural information and GO attribute information are complementary in ontology attributed networks. CSO can effectively take advantage of the correlation between frequent GO annotation sets and the dense subgraph for protein complex prediction. Our proposed CSO approach was applied to four different yeast PPI data sets and predicted many well-known protein complexes. The experimental results showed that CSO was valuable in predicting protein complexes and achieved state-of-the-art performance.

Extracting protein-protein interaction (PPI) from biomedical literature is an important task in biomedical text mining (BioTM). In this paper, we propose a hash subgraph pairwise (HSP) kernel-based approach for this task. The key to the novel kernel is to use the hierarchical hash labels to express the structural information of subgraphs in a linear time. We apply the graph kernel to compute dependency graphs representing the sentence structure for protein-protein interaction extraction task, which can efficiently make use of full graph structural information, and particularly capture the contiguous topological and label information ignored before. We evaluate the proposed approach on five publicly available PPI corpora. The experimental results show that our approach significantly outperforms all-path kernel approach on all five corpora and achieves state-of-the-art performance.

Protein complexes play a key role in many biological processes. Various computational approaches have been developed to identify complexes from protein-protein interaction (PPI) networks. However, high false-positive rate of PPIs makes the identification challenging. In this paper, we propose a protein semantic similarity measure based on the ontology structure of Gene Ontology (GO) terms and GO annotations to estimate the reliability of interactions in PPI networks. Interaction pairs with low GO semantic similarity are removed from the network as unreliable interactions. Then, a cluster-expanding algorithm is applied to identify complexes with core-attachment structure on the filtered network. We have applied our method on three different yeast PPI networks. The effectiveness of our method is examined on two benchmark complex datasets. Experimental results show that our method outperforms other state-of-the-art approaches in most evaluation metrics. Removing interactions with low similarity significantly improves the performance of complex identification.