OnSIDES database: Extracting adverse drug events from drug labels using natural language processing models

May combine two claims.

SIDER is one of the most widely utilized ADE databases derived from drug label data, constructed using dictionary-based named entity recognition to extract adverse event terms from FDA Structured Product Labeling drug labels.

“One of the most widely utilized ADE databases derived from drug label data is SIDER, constructed using a dictionary-based named entity recognition (NER) method to extract adverse event terms from FDA Structured Product Labeling (SPLs) drug labels13.”

• cites [13] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

A number of publicly available ADE databases exist, constructed from sources including drug labels, electronic health records, research literature, social media, and pharmacovigilance programs.

“A number of publicly available ADE databases exist, which have been constructed by extracting, stratifying or detecting signals from a variety of sources, including drug labels, electronic health records, research literature, social media, and adverse event reporting pharmacovigilance programs7–11.”

• cites [7] sourcePartly supported

  The Materials and Methods ‘Data source’ section describes SIDER as mined from drug labels and identifies FDA and Canadian adverse-event reporting databases. The paper also makes its own ADE databases publicly available, but it does not describe databases constructed from research literature or social media; it uses electronic medical records to validate findings, not as a source for those databases. Page numbers are unavailable in our copy. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites [8] sourcePartly supported

  The Background and Summary identifies electronic health records, spontaneous reporting systems, social media, and literature mining as sources of drug-safety evidence. The Usage Notes describe a harmonized evidence base combining product labels, published literature, observational data, and spontaneous reports; the Methods and Data Records describe a publicly available FAERS-derived dataset. The paper does not establish that publicly available ADE databases have been constructed from each listed source, particularly social media. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [9] sourceNot checked — no full text of the source
• cites [10] sourceSupported
• cites [11]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Efforts have fine-tuned BERT models to classify drug label text blocks into predefined sections and outperformed baseline models.

“For example, a number of efforts have fine-tuned BERT models to classify drug label text blocks into predefined sections of the label and outperformed baseline models for this task22–23.”

• cites [22] sourceCould not verify

  Section 2.3.5 (p. 9) describes classifying PubMed abstracts by cancer hallmark, and Section 2.4 (p. 12) describes BERT-based document classification. Neither the supplied passages nor searches for drug-label text blocks and label-section classification found a passage addressing that task or its performance against baselines. (LLM, conf=0.75)

  Experimental — accuracy not yet measured
• cites [23]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

These research efforts present promising results in extracting true ADE mentions, indicating suitability of NLP methods for this task.

“These research efforts present promising results in extracting true ADE mentions from drug label text, indicating the suitability of using NLP methods for this task.”

• cites [20] sourceNot checked — no full text of the source
• cites [21] sourceSupported
• cites [22] sourcePartly supported

  The Named Entity Recognition section (p. 7) describes NLP models for detecting drug and disease entities, and the results table (p. 17) reports performance on biomedical extraction tasks. These support the broader suitability of NLP for such tasks, but the cited paper does not report extracting true adverse drug event mentions from drug label text; its described datasets use PubMed abstracts. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [23]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The TAC 2017 task was conducted to encourage development of computational methods to recognize ADE mentions from drug labels and their relationships.

“Of note, the “Adverse Drug Reaction Extraction from Drug Labels” task of the Text Analysis Conference (TAC) 2017 was conducted to encourage the development of computational methods to recognize mentions of ADEs from drug labels and their relationships.”

• cites [18]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [19] sourcePartly supported

  The Background & Summary identifies the TAC 2017 task as an evaluation of adverse-reaction extraction from drug labels, and Technical Validation describes its mention-level annotations and training and test sets. Our copy does not say that the task aimed to recognize relationships between ADE mentions, so that part of the claim is unsupported. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Numerous biomedical models such as BioBERT, MedBERT, and BiomedBERT have been developed by training on relevant text corpora.

“In biomedical informatics, numerous models utilizing this approach, such as BioBERT, MedBERT, and BiomedBERT, have been developed by training models on various relevant text corpora, such as EHR data and biomedical academic literature15–17.”

• cites [15]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [16]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [17] sourcePartly supported

  The Introduction identifies Med-BERT as a pretrained model trained on structured EHR diagnosis data. It explicitly says those data are ICD codes, unlike the free text used to train most BERT variants, so describing Med-BERT as trained on a text corpus is inaccurate. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

A number of NLP methods have been proposed to extract ADE terms from drug label text utilizing the TAC 2017 dataset.

“A number of NLP methods have been proposed to extract ADE terms from drug label text utilizing this dataset20–21.”

• cites [20] sourceNot checked — no full text of the source
• cites [21] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

The TAC 2017 organizers published a manually annotated gold-standard reference dataset for extracting ADEs from 200 FDA SPLs.

“As part of this task, the organizers published a manually annotated, gold-standard reference dataset for extracting adverse drug events from 200 FDA SPLs18–19.”

• cites [18]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [19] sourceSupported

Classified as probably not a substantial claim. May combine two claims.

EU, UK, and Japan represent some of the largest pharmaceutical markets outside the US and have different drug regulatory processes independent of the US.

“These nations/regions represent some of the largest pharmaceutical markets in the world outside of the US, and have different drug regulatory processes independent to that of the US25–26.”

• cites [25]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [26] sourcePartly supported

  The Summary (page unavailable) describes the US and EU as major locations for drug investigation and approval and says their regulatory systems differ. The European Regulation of Drugs section describes EU approval routes distinct from the FDA’s. The paper does not establish the claimed market standing of the UK or Japan or describe Japan’s regulatory process; it discusses the UK within the EU framework rather than as a separate regulatory system. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Significant advances in chemoinformatics and AI to drug discovery and repurposing have increased research into in-silico driven development of new medications.

“Significant advances in chemoinformatics and the application of artificial intelligence to drug compound discovery and the repurposing of existing drugs have greatly increased research into in-silico driven development of new, effective medications for a variety of indications30.”

• cites [30] sourcePartly supported

  The Introduction (page unavailable) describes growing drug-discovery data resources, predictive models that may aid drug repurposing, and a need for computer-aided drug discovery. The model-development sections (page unavailable) also discuss cheminformatics and in-silico approaches, but the paper does not establish that advances in AI and repurposing have greatly increased research into new, effective medications. (LLM, conf=0.79)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DrugBank mainly contains drug indications extracted from drug labels, while MEDI uses public internet sources for drug indications.

“DrugBank primarily contains drug indications extracted from drug labels, while MEDI utilizes public internet sources to drug indications, and is a well utilized resource of drug indications.”

• cites [48] sourcePartly supported

  In “New and enhanced interface features” (page unavailable), the paper says DrugBank’s new linked indications were manually curated from drug labels, but specifies that these were available only for currently marketed products. It does not address MEDI, its use of public internet sources, or how widely MEDI is used. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [49] sourcePartly supported

  The Methods > Data sources section (page unavailable) says MEDI draws indications from four public resources, including the MedlinePlus website and Wikipedia. The paper does not identify DrugBank as a source or say that DrugBank's indications primarily come from drug labels; its discussion of label-derived indications concerns SIDER instead. It also does not establish that MEDI is well utilized. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

ChemBERTa performed better than Chemprop on several other MoleculeNet benchmark tasks.

“However, ChemBERTa was found to perform better than Chemprop on a number of other MoleculeNet benchmark tasks51”

• cites [51] sourceCould not verify

  The Results and discussion section and Table 3 in the Conclusions compare other models on MoleculeNet tasks, but neither reports a ChemBERTa–Chemprop comparison. Searches for both model names and their comparative performance found no passage addressing the claimed result. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

The SPL format includes a broad range of information such as indications, clinical pharmacology, contraindications, warnings, and ADEs.

“The SPL format incorporates a broad range of information including the drug’s indications, clinical pharmacology, contraindications, warnings, and ADEs36.”

• cites [36] sourceCould not verify

  The MEDICUS EXTENSION section mentions FDA drug labels linked to DailyMed, but does not describe the SPL format or what information it contains. Searches for SPL and the listed label contents found no passage addressing the claim. (LLM, conf=0.77)

  Experimental — accuracy not yet measured

May combine two claims.

OnSIDES contains 3,610,120 ADEs and 114,322 Boxed Warning ADEs from 47,211 drug labels, covering 3,233 unique drug ingredient combinations and 4,510 unique ADEs.

“OnSIDES contains over 3.6 million drug-ADE pairs for 3,233 unique drug ingredient combinations extracted from 47,211 labels.”

• Also stated in: § Generating the OnSIDES Database · plain

  “As of the February 2024 release, the OnSIDES database contains 3,610,120 ADEs and 114,322 Boxed Warning ADEs from 47,211 drug labels, listing 3,233 unique drug ingredient combinations (1,873 single-ingredient drugs, 1,360 multi-ingredient drugs) and 4,510 unique ADEs.”

May combine two claims.

The model achieved F1 0.90, AUROC 0.92, and AUPR 0.95 at extracting ADEs from the Adverse Reactions section.

“We achieve an F1 score of 0.90, AUROC of 0.92, and AUPR of 0.95 at extracting ADEs from the labels’ “Adverse Reactions” section.”

May combine two claims.

The method was used to extract pediatric-specific ADEs, serious ADEs from Boxed Warnings, and ADEs from UK, EU, and Japan labels to build OnSIDES-INTL.

“We further utilize this method to extract pediatric-specific ADEs, serious ADEs from labels’ “Boxed Warnings” section, and ADEs from drug labels of other major nations—the UK, EU, and Japan—to build a complementary OnSIDES-INTL database.”

May combine two claims.

SIDER contains 1,430 unique drugs and 4,251 unique side effects.

“SIDER contains 1,430 unique drugs and 4,251 unique side effects.”

May combine two claims.

An OnSIDES-trained linear regression model performs moderately in predicting whether a given drug has a specific drug target based on its side-effect profile (AUROC = 0.667) and outperforms a SIDER-trained model (AUROC = 0.590).

“Additionally, we find that an OnSIDES-trained linear regression model performs moderately in predicting whether a given drug has a specific drug target based on its side-effect profile (AUROC = 0.667) and outperforms a SIDER-trained model (AUROC = 0.590) (Figure 4B).”

• AssumptionRests on an assumption: Drug target and indication data from DrugBank, OpenTargets, and MEDI are accurate.

May combine two claims.

There was notable overlap in sensory organ drugs between US and UK, and antineoplastic and immunomodulating agent drugs among all data sources.

“However, there was some notable overlap in sensory organ drugs between the US and UK, and also antineoplastic and immunomodulating agent drugs among all data sources.”

May combine two claims.

OnSIDES-INTL contains 108,958, 705,917, and 237,827 unique drug-ADE pairs from 1,048, 9,003, and 6,541 drug labels, with 929, 1,889, and 1,441 unique drug ingredient combinations and 3,570, 4,848, and 3,296 unique ADEs from European, British, and Japanese labels respectively.

“By doing so, we construct additional databases with 108,958, 705,917, and 237,827 unique drug-ADE pairs from 1,048, 9,003, and 6,541 drug labels, with 929, 1,889, and 1,441 unique drug ingredient combinations and 3,570, 4,848, and 3,296 unique ADEs from European, British, and Japanese drug labels respectively.”

May combine two claims.

OnSIDES contains 1,485 unique single-ingredient drugs and 4,345 unique side effects.

“OnSIDES contains 1,485 unique single-ingredient drugs and 4,345 unique side effects”

May combine two claims.

OnSIDES performs best in predicting existing therapeutic indications identified in both DrugBank and MEDI (AUROC = 0.77) and better than SIDER (AUROC = 0.715).

“We further find that OnSIDES performs best in terms of predicting the existing therapeutic indications that are identified in both DrugBank and MEDI (AUROC = 0.77) and better than SIDER in these predictions also (AUROC = 0.715), suggesting these models perform well in predicting indications of high confidence (Figure 4D).”

• AssumptionRests on an assumption: Drug target and indication data from DrugBank, OpenTargets, and MEDI are accurate.

May combine two claims.

SOCs identified across the fewest drug labels were product issues, pregnancy and related conditions, and congenital, familial, and genetic disorders while general disorders and administration site conditions, gastrointestinal, skin and subcutaneous tissue, nervous system disorders, and vascular disorders were identified on the largest number of drug labels.

“Overall, the SOCs identified across the fewest number of drug labels were “product issues”, “pregnancy and related conditions”, and “congenital, familial, and genetic disorders” while “general disorders and administration site conditions”, “gastrointestinal”, “skin and subcutaneous tissue”, “nervous system disorders”, and “vascular disorders” were identified on the largest number of drug labels.”

May combine two claims.

The annotated dataset has many ADEs in vascular disorders, investigations, and skin and subcutaneous tissue disorders, and few in endocrine disorders, social circumstances, and congenital, familial, and genetic disorders.

“we find a significant proportion of annotated ADEs that fall into the “vascular disorders”, “investigations”, and “skin and subcutaneous tissue disorders” categories, and relatively few “endocrine disorders”, “social circumstances”, and “congenital, familial, and genetic disorders” (Supplementary Figure 1B).”

May combine two claims.

The antineoplastics and immunomodulators drug class had the largest number of ADEs while antiparasitics and insecticides consistently had the fewest across all data sources.

“We also find that the “antineoplastics and immunomodulators” drug class had the largest number of ADEs while the “antiparasitics and insecticides” class drugs consistently have the fewest identified across all data sources.”

The quote only partly states this claim.

The method was extended to extract ADE data from EU, UK, and Japanese drug labels to create additional computable resources.

“we leverage the method to expand the database to incorporate ADE data found in drug labels of other nations/regions.”

May combine two claims.

FDA SPL labels mainly present ADEs as free text, while labels from other regulatory agencies present the ADE section as structured lists, tables, or a mix.

“While the FDA SPL drug labels primarily list the ADEs in a free text format, drug labels from other regulatory agencies format the ADE section in structured lists, tables, or a combination of the three formats.”

May combine two claims.

The extraction method makes errors: some side effects are missed and some predicted adverse events are wrong.

“computational extraction methods are prone to errors, with some side effects not extracted accurately, and some predicted adverse events being incorrect.”

May combine two claims.

General ADE descriptions were standardized across drug labels and label types, but pediatric ADE descriptions were often nested in general ADE sections or absent due to lack of information.

“While general ADE descriptions were found to be standardized across drug labels / label types, pediatric ADE descriptions were often found nested in the general ADE sections, or did not exist anywhere in the label text due to the lack of information.”

May combine two claims.

Optimal input string length is 125 words, with 87.5% of the words coming after the candidate term.

“We first determine the optimal input text string format, concluding that optimal input string length is 125 words, with 87.5% of the words coming after the candidate term (Figure 1A).”

May combine two claims.

Drugs tend to have the highest degree in US data, followed by UK, while Japanese data tends to assign the lowest degree.

“Drugs tend to have the highest degree (number of ADEs) in the US data, followed by the UK data, while the Japanese data tends to assign the lowest degree.”

May combine two claims.

With the optimal parameter combination, PubMedBERT achieved an F1 score of 0.874 and an AUROC of 0.899.

“we achieve F1 Scores of 0.861 and 0.874, and AUROC Scores of 0.896 and 0.899, respectively for the ClinicalBERT and PubMedBERT models.”

May combine two claims.

In comparison between OnSIDES and OnSIDES-PED, there was enrichment of a large number of Skin and Subcutaneous Tissue and Gastrointestinal ADEs, and very few Social Circumstances and Pregnancy, Puerperium and Perinatal Conditions ADEs enriched.

“In the comparison between OnSIDES and OnSIDES-PED, we observed an enrichment of a large number of “Skin and Subcutaneous Tissue” and “Gastrointestinal” ADEs, and very few “Social Circumstances” and “Pregnancy, Puerperium and Perinatal Conditions” ADEs enriched.”

May combine two claims.

OnSIDES-PED contains 359,213 drug label-ADE pairs from 1,161 unique drug ingredient combinations, 1,526 unique ADEs, and 20,014 drug labels derived from special populations section.

“We use this adapted method to construct a large pediatric-specific ADE database of 359,213 drug label-ADE pairs from 1,161 unique drug ingredient combinations, 1,526 unique ADEs, and 20,014 drug labels derived from the drug labels’ special populations section.”

May combine two claims.

The fine-tuned PubMedBERT model outperforms all TAC 2017 models and the SIDER database.

“We find that our model outperforms all TAC 2017 models and the SIDER database as presented in (Supplementary Figure 3B).”

May combine two claims.

Between OnSIDES-PED and KidSIDES, there was enrichment of ADEs in Nervous System and General and Administration Site Conditions MedDRA SOC classes, and a relative lack of ADEs reported in Ear and Labyrinth and Social Circumstances classes in KidSIDES.

“Between OnSIDES-PED and KidSIDES, we observed an enrichment of ADEs in the “Nervous System” and “General and Administration Site Conditions” MedDRA SOC classes, and a relative lack of ADEs reported in the “Ear and Labyrinth” and “Social Circumstances” classes in KidSIDES.”

May combine two claims.

The OnSIDES-trained model was able to predict drug indications from each database, and outperforms a SIDER-trained model.

“We find that the OnSIDES-trained model was able to predict drug indications from each database, and outperforms a SIDER-trained model (Figure 4C).”

• AssumptionRests on an assumption: Drug target and indication data from DrugBank, OpenTargets, and MEDI are accurate.

May combine two claims.

The PubMedBERT-based model gives a modest performance improvement over the ClinicalBERT-based model and performs more consistently across the two evaluations.

“The PubMedBERT-based model achieves a modest performance improvement over the ClinicalBERT-based model, and notably it performs more consistently across the two different evaluations.”

May combine two claims.

The annotated dataset overrepresents antineoplastic and immunomodulating agents and underrepresents dermatologicals, systemic hormonal preparations, and antiparasitic products.

“we find that there is a significant representation of drugs that are “antineoplastic and immunomodulating agents”, and very few “dermatologicals”, “systemic hormonal preparations”, and “antiparasitic products” (Supplementary Figure 1A).”

May combine two claims.

With the optimal parameter combination, ClinicalBERT achieved an F1 score of 0.861 and an AUROC of 0.896.

“we achieve F1 Scores of 0.861 and 0.874, and AUROC Scores of 0.896 and 0.899, respectively for the ClinicalBERT and PubMedBERT models.”

May combine two claims.

US drug labels tend to have more listed ADEs than other nations, although considerable variation exists between drugs.

“These results show that US drug labels tend to have more listed ADEs than that of other nations, although considerable variation exists between drugs.”

May combine two claims.

MEDI has 1,257 drugs with indications for OnSIDES and 1,155 for SIDER.

“There are 1,474 and 1,347 drugs with indications listed in DrugBank, and 1,257 and 1,155 drugs with indications listed in MEDI for OnSIDES and SIDER respectively.”

May combine two claims.

DrugBank has 1,474 drugs with indications for OnSIDES and 1,347 for SIDER.

“There are 1,474 and 1,347 drugs with indications listed in DrugBank, and 1,257 and 1,155 drugs with indications listed in MEDI for OnSIDES and SIDER respectively.”

Classified as probably not a substantial claim.

Pediatric mentions in the general AR section are difficult to systematically separate from non-pediatric-specific mentions.

“as the pediatric mentions in this section are difficult to systematically separate from those that are non-pediatric specific.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Significant advancements in NLP methods, including BERT models, have been made.

“In recent years, significant advancements have been made in Natural Language Processing (NLP) methods, including the development of BERT (Bidirectional Encoder Representations from Transformers) models14.”

• cites [14] sourceNot checked — no full text of the source

May combine two claims.

Few machine-readable databases of adverse drug events exist, which limits broad systematic study of drug safety.

“However, few machine-readable databases of ADEs exist, limiting our capacity to study drug safety on a broader, systematic scale.”

• cites [5]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [6] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

OnSIDES can be utilised as a comprehensive resource to study and enhance drug safety.

“OnSIDES can be utilised as a comprehensive resource to study and enhance drug safety.”

May combine two claims.

Drug label repositories are collections of raw text or PDF files and are difficult for secondary research because they are not standardized or optimized for machine readability.

“However, these repositories are a collection of individual raw text or PDF files, and are difficult to utilize for secondary research purposes as they are not standardized or optimized for machine readability.”

Probably not a substantial claim: may not pass S2 (checkable).

To comprehensively study ADEs, high-quality machine-readable databases encompassing the breadth of drugs currently used are required.

“As such, to comprehensively study ADEs, we require high-quality, machine-readable databases encompassing the breadth of drugs currently used.”

May combine two claims.

More context helps performance, but an overly long input string may dilute relevant context with unrelated information.

“This suggests that more context aids performance, but too long of an input string may dilute the relevant context with unrelated information.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Some known ADEs may not appear on drug labels and may be found in other sources like adverse event reporting systems.

“there may be known ADEs that are not yet reflected on drug labels that may appear in other sources, such as adverse event reporting systems.”

May combine two claims.

Knowledge of ADEs in pediatric patients is limited, with few RCTs conducted on these populations due to restricted population and ethical concerns.

“For example, while the risk of serious ADEs is greater for pediatric patients, our knowledge is limited for this population, with few RCTs being conducted on these populations given the restricted population and ethical concerns24.”

• cites [24]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

To the best of the authors' knowledge, OnSIDES is the largest and only current up-to-date drug label-derived ADE database.

“To the best of our knowledge, OnSIDES is the largest and only, current, up-to-date, drug label-derived ADE database.”

May combine two claims.

PubMedBERT and ClinicalBERT were selected because they are trained on biomedical language similar to drug label text and have performed well on relevant NLP tasks.

“We chose to consider and evaluate these models as they are trained on biomedical language similar to drug label text, and have been demonstrated to perform well on relevant NLP tasks37.”

• cites [37]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

The manually annotated dataset was used to evaluate the model's accuracy in the generated datasets.

“We use this manually annotated dataset to evaluate the accuracy of our model in the generated datasets.”

Classified as probably not a substantial claim.

Campillios et al. previously suggested the similarity of two drugs' side effect profiles is positively correlated with the probability that these drugs share a common target.

“The work of Campillios et al. previously suggested the similarity of two drugs’ side effect profiles is positively correlated with the probability that these drugs share a common target28.”

• cites [28] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Yang and Agarwal showed side-effect profile similarities can also predict shared indications.

“In a similar manner, Yang and Agarwal showed side-effect profile similarities can also predict shared indications29.”

• cites [29]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The limited description and subsequent extraction of pediatric-specific ADEs in drug label text is highlighted by KidSIDES having more drugs in each category.

“Again, KidSIDES had more drugs in each category, highlighting the limited description and subsequent extraction of pediatric-specific ADEs in drug label text.”

May combine two claims.

OnSIDES could be used as a potential resource for drug repurposing applications, identifying existing drugs that may be used for novel therapeutic indications, and to identify potential drug off-targets.

“This suggests that OnSIDES could be used as a potential resource for drug repurposing applications, identifying existing drugs that may be used for novel therapeutic indications, and to identify potential drug off-targets.”

Probably not a substantial claim: may not pass S2 (checkable).

Further challenges lie in applying these methods to systematically construct a comprehensive and readily accessible database to aid further pan-medication study of ADEs.

“However, further challenges lie in applying these methods to systematically construct a comprehensive and readily accessible database to aid further pan-medication study of ADEs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The accuracy of machine translation for Japanese labels is a limitation of OnSIDES-INTL.

“Another limitation is the machine translation accuracy of the Japanese labels.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Chemprop performed better than ChemBERTa for ClinTox classification.

“Ahmad et al.32 found that Chemprop performed better than ChemBERTa for ClinTox classification, a drug-toxicity classification task.”

• cites Ahmad et al.32Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The EMA, MHRA, and PMDA are the regulatory bodies overseeing pharmaceutical product approval in the EU, UK, and Japan, respectively.

“In each country/region, the EMA (European Medicines Agency), MHRA (Medicines and Healthcare products Regulatory Agency), and PMDA (Pharmaceuticals and Medical Devices Agency) are the respective regulatory bodies overseeing pharmaceutical product approval38–40.”

• cites [38]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [39]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [40]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

All entities submitting drug approval applications to the FDA must submit documentation in SPL format.

“All entities that submit drug approval applications to the FDA are required to submit documentation in this format35.”

• cites [35]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Enrichment in the number of adverse events per adverse event type for each drug class was evaluated using Fisher's exact test with a Bonferroni adjustment.

“we evaluated enrichment in the number of adverse events per adverse event type for each drug class using Fisher’s exact test, as listed in the axis labels in Figure 3C. We used a Bonferroni adjustment to p-values from these tests due to testing all drug classes in parallel.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

ChemBERTa is a transformer-based RoBERTa model that takes SMILES strings as input and can be fine-tuned to predict chemical and physical targets.

“ChemBERTa is a transformer-based model of RoBERTa architecture52 that takes SMILES strings as input and can be fine-tuned to predict various chemical and physical targets”

• cites [52]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The SPL drug label format is an XML-based HL7 clinical and administrative standard document markup format used by the FDA.

“The SPL drug label format is a XML-based HL7 clinical and administrative standard document markup format utilized by the FDA to uniformly present relevant pharmaceutical information text34.”

• cites [34]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The SHAP method is a game theory-based approach that explains model predictions by assigning importance values to individual features.

“The SHAP method is a game theory-based approach that explains model predictions by assigning individual features an importance value for a specific prediction, allowing for a better understanding of how individual features impact model decisions.”

• cites [45]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.