An automated identification and analysis of ontological terms in gastrointestinal diseases and nutrition-related literature provides useful insights

Search strategy for GI diseases and nutrition-related literature

The abstracts used for analysis were retrieved from PubMed database using the list of keywords described in Fig. 1, using a time frame from 1991 to 2016 (searches were performed in July 2017). Composite keywords were constructed through Boolean logic, with four by three possibilities (four disease groups × three nutritional keywords, yielding twelve possible combinations). The returned abstracts were then grouped together in pairs of years, extracted and stored in external files, using the “Citation Manager” function in MEDLINE (tagged) format. The complete search for all the possible combinations from 1991 to 2016, yielded a total number of 24,559 PubMed abstracts. These were then imported into EndNote^® X7 citation management software to export them in BibTeX format (input format for our software), where every abstract was described by a number of records including the PubMed ID, i.e., a unique identifier used in PubMed and assigned to each article record when it enters the PubMed system. In total, 156 BibTeX files were generated for all the possible combinations of composite keywords and pairs of years (i.e., 12 possibilities in a 26-year timeline).

Annotation process

The BibTeX files were then processed by a custom written script called pyTag that accepts these files in a given folder as input (using --input_dir) and supports nine ontologies inherited from EXTRACT 2.0, a custom named entity recognition system (Pafilis et al., 2017) as a switch to the program, with --onto_types all specifying all of them. In general, one is at liberty to choose a subset of these ontologies in a given piece of text. These supported ontologies can recover mentions for: organisms (NCBI Taxonomy using --onto_types -2; Federhen, 2011); environments (Environment Ontology using --onto_types -27; Buttigieg et al., 2016); diseases and phenotypes (Disease Ontology also incorporating Mammalian Phenotype Ontology using --onto_types -26; Kibbe et al., 2014; Smith & Eppig, 2012); tissues and cell lines (BRENDA Tissue Ontology using --onto_types -25, Placzek et al., 2016); biological processes (--onto_types -21), cellular components (--onto_types -23) and molecular functions (--onto_types -22) using the Gene Ontology (The Gene Ontology Consortium, 2015); genes and proteins (STRING and RAIN databases using --onto_types 0; Szklarczyk et al., 2016; Junge et al., 2017); and small molecule compounds (STITCH database using --onto_types -1, Szklarczyk et al., 2015). pyTag then uses these ontologies on the abstracts recovered from the NCBI database using the associated PubMed IDs. After the annotation of all the relevant abstracts, the resulting frequency of the identified terms (a formal textual representation for a given concept including all its possible synonyms identifiable by a unique ID for a given ontology) are converted to a two-dimensional abundance table for multiple search criteria, with enough replicates per search to ensure that ecological statistics including alpha and beta diversities can be calculated as well as differential analysis can be performed. This whole workflow is summarized in Fig. 2.

For the annotation of the literature, all the ontologies listed above were employed. Out of 24,559 abstracts, 21,035 of them were annotated, i.e., at least one term was found in their content (for terms appearing more than once in an abstract only one occurrence was considered). From the identified terms, those with low or rare frequencies were removed (<5 total hits across all searches). After preprocessing, this resulted in 2,399 unique terms from which terms related to nutrition were selected and considered for statistical analysis. It should be noted here that in the absence of any specific ontology available for nutrition related terms, with this study being the first of its kind, we sought to manually annotate all the available ontological terms using clinical and expert opinion, thus resulting in a list of 445 nutrition related terms given in the Table S3. Therefore, in the remainder of this paper, whenever we use the word “terms,” it is implicitly assumed that they are relevant to nutrition only.

Statistical analysis

Statistical analysis was performed in R software. To account for the variation of the number of publications over time, the counts of each term found in a search for a pair of years for a specific disease condition, were adjusted with respect to the number of the papers published in literature for this condition and annotated from the workflow for this specific year (document-based normalization). To explore the significance of the variability of ontological terms between the disease conditions, the Vegan package (Oksanen et al., 2017) was used, particularly, the function adonis for PERMANOVA (ANOVA for distance matrices). Clustering between the disease groups, how dissimilar the terms for a given search (e.g., year or condition) are from each other and temporal changes in literature were assessed using the reduced-order representation of the datasets using the non-metric multidimensional scaling (NMDS), which reduces the multivariate dataset to two or more dimensions (similar to Principal Component Analysis (PCA)) based on dissimilarity (Bray–Curtis distance) between the terms for a given search. The Local Contributions to Beta Diversity (LCBD) was also used with a Hellinger transformation (Legendre & De Cáceres, 2013), where the overall beta diversity is divided into individual contributions from samples to identify outliers. The smaller the LCBD value is, the closer the sample is to the group average. To identify ontology-based terms that were significantly different between the conditions, Kruskal–Wallis test (Kruskal & Wallis, 1952) was used. The Benjamini–Hochberg correction was used on the returned P-values to correct for multiple testing and Dunn’s test as a post hoc procedure for pair-wise comparisons, where appropriate.

Ontological terms clustered IBD separately from non-IBD conditions with temporal changes observed in the literature of each disease group

When the composition of the ontological terms for the disease conditions was assessed using NMDS plots, findings demonstrated an evident clustering of IBD related ontological terms distinct from non-IBD (Fig. 3A). The clusters for CCD and IBS stood well apart from those of CD and UC. CD and UC showed a degree of overlap, suggesting a degree of similarity in the ontological terms between these two conditions. Temporal variability was also noticeable from the NMDS plot (Fig. 3B). The beta diversity analysis revealed that the nutrition-related literature for each disease group has shifted over time. For all groups, the between-year variability was higher in the earlier dates, but gradually decreased, as we moved forward in time. This was clearer in the case of CD, UC, and IBS. It could be seen that the proximity between CD and UC was increasing more for the later years and that the two IBD groups were further converging to a similar set of ontological terms.

The convergence between the groups was also obvious when LCBD (Legendre & De Cáceres, 2013) analysis was applied. The findings, in this case, showed a decreasing trend of the LCBD values over the years for each disease group, more noticeable for the case of CD, UC and, IBS (Figs. 4A–4D). This indicated that the relative contribution of each sample (search for a pair of years) in every group was shifting towards the mean value (multivariate centroid) of the sample space when approaching more recent dates, suggesting their gradual convergence in recent years. This pattern indicated a relative consensus on a particular nutrition research theme for these disease conditions.

Most frequent topics and conserved patterns in the literature of the disease conditions

Permutational multivariate analysis of variance (distances between groups) suggested that most of the variability was explained significantly by the different disease conditions (R² = 27%, p = 0.001). To further explore this and inspect for terms that stratify the groups, we first looked at the twenty most frequent terms in the literature of each condition for the entire time frame. Findings showed that CD and UC, shared more than a half (65%) of their most common topics and terms such as growth (Freq. CD = 3.90; UC = 3.00) and fatty acids (Freq. CD = 2.08, UC = 2.78) were listed as the top two most frequent in the literature of the IBD groups (Figs. 5B and 5D). In a similar way, wheat (Freq. = 6.16) and gliadin (Freq. = 5.12) were unsurprisingly some of the most prevalent in the literature of CCD research (Fig. 5A). Likewise, the ontological terms fibre (Freq. = 4.57) and lactose (Freq. = 3.10) were found very common in IBS (Fig. 5C).

Moreover, differential analysis performed over separate time intervals (see Table 1) showed that the above findings were fairly conserved between the groups over the years (Fig. 6). This can suggest a continuous scientific interest for these topics in the research of each disease. In addition, results showed multiple terms becoming significant between the disease conditions for each time interval (Padj < 0.05; see Table 1 and Tables S1A–S1D). Specifically, in CCD, terms for gliadin, wheat, rye, barley, and oats were found to be stably frequent between 1991 and 2016 and clearly more common compared to the other groups (CCD > other diseases; Fig. 6 and Tables S1A–S1D). In a similar way, a considerable presence of fibre and lactose was observed in IBS throughout the years with findings also indicating a decrease in the frequency of both terms for the more recent dates (Fig. 6).

In the case of the IBD groups, terms such as omega-3 fatty acids and n-6 fatty acids were evidently more frequent compared to IBS and CCD where they were less common (CD and UC > CCD and IBS; Fig. 6; Tables S1B–S1D). For omega-3 fatty acids, the pattern was relatively stable over time (between 1991 and 2016) with some slight decrease for both CD and UC between 2011 and 2016 (Fig. 6). In a similar way, n-6 fatty acids were very common in CD and UC between 1999 and 2016 (Tables S1B–S1D). Growth term was also observed to be significantly different between the disease groups (Fig. 6). In CD, the same term had the highest prevalence with UC and CCD following respectively, appearing the least in IBS (CD > UC > CCD > IBS; Fig. 6). However, only during 1991–1998, this term appeared in CCD almost in similar levels to CD and UC literature.

Ontological terms showing temporal changes in the literature of the disease groups

Analysis of variance using the adonis function showed that also temporal variability (expressed as in pairs of years) explained up to 19% of the changes in the use of ontological terms (R² = 19.0%, p = 0.001). To investigate this further, differential analysis was performed on each term (see Table 1). Findings showed a number of terms differentiating over time in the literature of the disease conditions (Padj < 0.05; see Table 1 and Tables S2A–S2D).

More specifically, results revealed a considerable increase in the frequency of obesity term for all disease conditions (Padj CCD = 0.025, CD = 0.01083, IBS = 0.00691, UC = 0.01108; Fig. 7A). This was more evident after years 2008–2009, for each disease group. Obesity was higher in IBS between 2009–2010 and 2015–2016 compared to the other groups, with CCD being next and CD and UC following respectively. Similarly, wheat allergy was found becoming more common between the disease conditions over the years (Padj CCD = 0.01712, CD = 0.03986, IBS = 0.00381, UC = 0.04184; Fig. 7B). This term was noticed more frequent for CCD and IBS in the more recent dates (2011–2012 and thereafter). CCD seemed to be the group where wheat allergy was increasing the most with IBS being next. In the case of CD and UC, the same term was found to be equally prevalent between 2015 and 2016 for both groups, but clearly in a lower frequency when compared to the non-IBD types.

The frequency of several terms was also found to change temporally in relation to CD and UC (Fig. 7C). This was the case for butyrate (Padj CCD = 0.025, UC = 0.01108) and curcumin (Padj CCD = 0.01549, UC = 0.04184). Butyrate showed an increasing trend in the literature, most prominently in UC, with a peak frequency noticed in 2001–2002 and becoming considerably less common onwards (Fig. 7C). The same term was notably less common in CD compared to UC, where it became frequent between 1999–2000 and 2001–2002 and it was found in similar levels to UC in 2015–2016 (Fig. 7C). In addition, a partially transient prevalence over time was seen for the term short-chain fatty acids (SCFAs). SCFAs (Padj CCD = 0.01404, UC = 0.01763) were noticed to be more frequent for both groups between 1993–1994 and 2003–2004 and decreasing rapidly onwards, particularly in the case of UC (Fig. 7C). Moreover, the term vitamin D (Padj CCD = 0.01242, UC = 0.04184) was found more common in CD compared to UC and becoming frequent over the years for both groups showing a notable increase between 2009–2010 and 2013–2014 (Fig. 7C). However, after these dates, a slight decrease could be observed in both cases for the years 2015–2016.