Biomedical Literature and Social Media Mining

Our group has been involved in this field from its very start, having participated successfully in the first four BioCreAtIvE (Critical Assessment for Information Extraction) between 2004 and 2012. Much of the research presently conducted in the biomedical domain relies on the inference of correlations and interactions from data at multiple levels of the biological organization: from the molecular to the social. Because we ultimately want to increase our knowledge of the biochemical, functional and behavioral roles of genes and proteins in organisms, there is a clear need to integrate the associations and interactions among biological entities that have been reported and accumulated in experimental databases, the literature, electronic health records, and non-traditional data sources such as social media.

Biomedical literature mining is an important informatics methodology for large scale information extraction from repositories of textual documents, as well as for integrating information available in various domain-specific databases and ontologies, ultimately leading to knowledge discovery. It helps uncover relationships and interactions buried in the literature and other media, from experimental to phenomenological data. The goal is then to tap into the vast biomedical collective knowledge available in various data sources, which we can think of as the "bibliome" or the "digital phenotype" in health and disease.

Our contributions to the Biomedical literature mining have been the development of novel methods based on network science or bio-inspired computing. This data-driven approach has enabled the automatic discovery, classification and annotation of protein-protein and drug-drug interactions, health risks including gender and age biases, pharmacokinetic parameters in drug interaction and adverse reaction studies, population and epidemiological studies, protein sequence and structure prediction, functional annotation of transcription data, enzyme annotation publications, etc. Examples of these are shown below, together with links to additional resources and publications.

Public health monitoring using Social Media data

Social media and mobile application data enable population-level observation tools with the potential to speed translational research. Our group has been one of the first to use social media data to study collective social behavior in biomedical problems [Correia, Wood, Bollen & Rocha, 2020]---see also our related work in computational social science. For instance, our group was the first to use Instagram to build public health monitoring and surveillance tools for discovering drug interactions, adverse reactions, and behavior pathology, focusing on depression and epilepsy [Correia, Li, & Rocha, 2016]. This recent work demonstrates that the universe of social media provides a very promising source of large-scale data that can help with monitoring and understanding public health in ways that have not been hitherto possible. Indeed, given the large number of users, social media data allows us to identify under-reported, population-level pathology.

Our methodology is based on the longitudinal analysis of social media user timelines at different timescales: day, week or month. Our approach enables extraction and study of social media cohorts with great demographic, geographic, and situational precision, which is important for social media mining in biomedical studies as these require attention to cohort segmentation. Most social mining approaches focus on collecting large sets of random tweets, selected by hashtags or other textual elements, which confound geographical and demographic cohorts. Instead, we focus on large-scale, longitudinal cohort datasets comprised of the entire timelines of users selected for having a tweet with a hashtag (e.g. a medication [Correia, Li & Rocha, 2016 ]). We have shown that such longitudinal social media datasets are better suited for building precise cohorts, as they provide enough data and resolution to predict adverse drug reactions and interactions, as well characterize subpopulations with specific health concerns (e.g. epilepsy) [Correia, Li & Rocha, 2016 ] and other problems of biomedical relevance [Correia, Wood, Bollen & Rocha, 2020].

Typically, our methodology involves building knowledge (weighted) graphs from the co-occurrence of terms from various biomedical dictionaries (drugs, symptoms, natural products, side-effects, and sentiment) at various timescales. We showed that spectral methods, shortest-paths, and distance closures reveal relevant drug-drug and drug-symptom pairs, as well as clusters of terms and drugs associated with the complex pathology associated with epilepsy and depression. We validate inferences about drug interactions and adverse reactions via curated bioinformatics databases (e.g. DrugBank and SIDER), and develop demo tools to share our analysis with the community [Correia, Li, & Rocha, 2016, Min et al, 2021]. We currently analyze various web search and social media sources such as: Google Trends, Wikipedia, Twitter, Facebook, ChaCha, Reddit, and the Epilepsy Foundation public forums, and have focused on studying depression, epilepsy, and opioid abuse, as well as other health-related problems such as human-reproduction and even automated online fact-checking. Very important for this area is also the development of biomedical corpora and dictionaries to mine social media and the literature, as well as analysis of electronic health records described below. See publications below for details on all these threads.

Studying Drug-Drug interaction from the Literature and Electronic Health Records

Drug-drug interactions (DDIs) are major causes of morbidity and mortality and a subject of intense scientific interest. Biomedical literature mining can aid DDI research by extracting evidence for large numbers of potential interactions from published literature and clinical databases. We started with the estimation of pharmacokinetics numerical data from literature to mine drug-specific (e.g. Midazolam (MDZ)) pharmokinetic (PK) clearance data (systemic and oral) from the literature. We obtained 88% precision rate and 92% recall rate are achieved, with an F-score = 90%. Out-performs support vector machine (F-score of 68.1%). Further investigation on 7 other drugs showed comparable performance [Wang et al, 2009]. Recently, via a four-year ($1.7M) R01 grant from from NIH/NLM we have studied the large-scale extraction of drug-Interaction from medical text. This is a collaboration with Prof. Lang Li from Ohio State Medical School, and Prof. Hagit Shatkay from the University of Delaware. While evidence for DDI ranges in scale from intracellular biochemistry to human populations, literature mining methods have not been used to extract specific types of experimental evidence which are reported differently for distinct experimental goals. We have developed and used the team's manually curated corpora [Wu et al, 2013; Zhang et al, 2022] of PubMed abstracts and annotated sentences with three types of experimental DDI evidence: in vitro, in vivo, and clinical. The goal is the production of a text mining pipeline using several linear classifiers and a variety of feature transformation methods. Preliminary results [Kolchinsky et al 2015] on pharmacokinetics DDI experimental evidence in PubMed has yielded excellent classification performance in distinguishing relevant and irrelevant abstracts (reaching F1 ~= 0.93, MCC ~= 0.74, iAUC ~= 0.99) and sentences (F1 ~= 0.76, MCC ~= 0.65, iAUC ~= 0.83). New results on all three DDI types are forthcoming.

Our group and collaborators have also studied the DDI phenomenon using other sources of large-scale data such social media [Correia, Wood, Bollen & Rocha, 2020; Correia, Li, & Rocha, 2016] (see above) and electronic health records(EHR) [Correia et al, 2019 ]. In addition to uncovering the most worrisome DDI prescribed to large populations, at great health and financial cost to individuals and communities, our analysis of EHR revealed very significant gender and age biases, whereby women and older people are prescribed many more DDIs than expected by random given the same rates of medication co-prescription. In forthcoming work with Prof. Alfonso Valencia's group, we extend this study to various worldwide populations and also study comorbidity in medical care.

Protein-Protein Interaction Discovery (PPI)

We have worked in the discovery and automatic annotation of relationships among biochemical entities, e.g. protein-protein and gene-disease interactions. The Biocreative challenges II, II.5, and III, which we participated in [Abi-Haidar et al,2008], [Kolchinsky et al, 2010], [Lourenco et al, 2011]), included a series of tasks on extraction of protein-protein interaction information from the literature. As the field moves to uncovering relations rather than entities, our complex network approach to biomedical literature mining [Verspoor et al,2005], which we tried on the first BioCreative competition, makes all the more sense. Additionally, since literature mining hinges on the quality of available sources of literature as well as their linkage to other electronic sources of biological knowledge, it is particularly important to study the quality of the inferences it can provide. We were among most competitive teams in the PPI tasks of BioCreative II, II.5 and III. We have also developed a bio-inspired solution for binary classification of textual documents inspired by T-cell cross-regulation in the vertebrate adaptive immune system. See our publications below for additional details

Characterizing gene regulation

Spectral methods such as Singular Value Decomposition (SVD), are very useful for tasks ranging from gene expression analysis [Wall, Rechtsteiner and Rocha, 2003] to automatic functional annotation of genes and proteins from the literature [Rechtesteiner, 2005; Maguitman, A. et al, 2006; Haidar et al, 2008;]. We have studied SVD-based methods for visualization of gene expression data, representation of the data using a smaller number of variables, and detection of patterns in noisy gene expression data. SVD ("eigen-clustering") of microarray data produces sets of co-expressed genes, which were then characterized with annotations automatically extracted from literature .

More recently, we have used SVD and information theory to cluster very large knowledge networks of gene regulation obtained from bioinformatics databases and the literature. This allows us to identify overlapping functional clusters that occur in various scales of complex networks [Correia, Navarro-Costa and Rocha, 2020], such as those characterizing gene regulation. Together with our distance backbone methodology, this has lead to the discovery of novel genes involved in human infertility [Correia et al, 2022].

Protein Structure Prediction (PSP)

Linking of information from different data sources, specifically literature, becomes increasingly important to annotate the growing number of new genome sequences. For the large percentage of genes with no known sequence homologs, new, possibly integrative, methods need to be developed. Ab-initio structure prediction and comparison is a method some of us pursued previously for functional annotation of sequences with no known homologs. We used a large set of sequences of known structure to evaluate a literature-based method against previously used ab-initio structure prediction methods. The Literature-mining prediction is comparable to best ab-initio methods in lack of sequence homology. Combining text-mining with ab-initio method leads to 35% improvement over ab-initio method alone. See [Rechtsteiner et al, 2006]