onsidesdb.org

What It Is

OnSIDES v3.1 is an adverse drug reaction database. Its core data comes from FDA Structured Product Labels, with adverse drug events/effects automatically extracted using a fine-tuned PubMedBERT language model. The site provides browsing entry points such as Drug Products, Ingredients, and Adverse Reactions, and lists basic statistics for drugs, adverse reactions, and drug–adverse reaction pairs. It is more of a research data resource than a traditional IDE, CI/CD tool, or general-purpose developer SaaS.

Core Capabilities and Data Quality

According to the main text, OnSIDES uses 200 manually annotated labels to fine-tune its model and processes downloadable DailyMed labels up to November 2023. It also provides OnSIDES-INTL, covering labels from regions such as Japan, the United Kingdom, and the European Union, as well as OnSIDES-PED, which focuses on adverse events in pediatric patient labeling. The disclosed model metrics are fairly detailed: for extraction from the ADVERSE REACTIONS section of FDA labels, it reports an F1 score of 0.90, AUROC of 0.92, and AUPR of 0.95. However, performance is lower in the BOXED WARNINGS and WARNINGS AND PRECAUTIONS sections, indicating that data reliability varies by label section. Users should evaluate it carefully based on the specific section and task.

Developer Perspective

The site includes Download and Source Code navigation, but the main text does not provide a repository URL, license, installation instructions, or data schema. It also does not describe an API, SDK, query endpoints, authentication method, or self-hosting deployment steps. As a result, developers who want to integrate it into pharmacovigilance pipelines, medical NLP benchmarks, or knowledge graph construction workflows will likely need to download the data and code first and explore them independently. In terms of documentation, the homepage introduction, statistics, model accuracy, and paper citation are clear, but engineering-oriented documentation is limited.

Pricing and Access

The main text does not mention commercial pricing, free or paid plans, payment methods, or usage restrictions. Access from mainland China cannot be determined from the crawled text, so it is rated as unknown. For research scenarios that require stable access, it is advisable to prepare mirrored downloads, institutional network access, or alternative data sources. Comparable resources include SIDER, openFDA, FAERS, DailyMed, and RxNorm.

Pros, Cons, and Ideal Users

Its strengths are clear data provenance, a paper-backed methodology, publicly reported metrics, and extensions to international and pediatric labeling. Its drawbacks are the lack of API/SDK details, licensing information, self-hosting guidance, and field-level documentation, while the model-extracted results are not a manually curated gold standard. It is best suited for drug safety researchers, medical NLP teams, clinical pharmacology data scientists, and research projects that need to build adverse reaction knowledge bases from drug labels.