MEDAKA: Leaflet-Derived Biomed Graph
- MEDAKA is a biomedical knowledge graph derived from standardized drug leaflets, capturing patient-focused data like side effects, dosage guidance, and contraindications.
- It employs an end-to-end pipeline with web scraping, LLM-based text extraction, and majority-vote filtering to convert unstructured content into a directed, labeled graph.
- MEDAKA complements traditional biomedical graphs by providing real-world, patient-facing medication details often missing from molecular or mechanism-centric databases.
MEDAKA is a biomedical knowledge graph built from publicly available drug leaflets using a largely automated, LLM-based pipeline. It was introduced together with a curated dataset and a “hackable,” end-to-end construction workflow designed to transform unstructured online content into a directed labeled graph. The central premise is that drug leaflets contain a clinically useful, patient-facing layer of information that is often absent from biomedical knowledge graphs centered on molecular interactions, drug–target relations, or adverse-event reporting. MEDAKA therefore emphasizes side effects, warnings, contraindications, ingredients, dosage guidance, storage instructions, and physical characteristics such as color and shape (Sengupta et al., 30 Sep 2025).
1. Motivation and conceptual position
The motivation for MEDAKA is a coverage gap in existing biomedical knowledge graphs and databases. Resources such as DrugBank, SIDER, FAERS, PrimeKG, and DRKG are described as valuable, but as tending to focus on narrower slices of biomedical knowledge, including molecular interactions, drug–target relations, drug–disease links, side effects, or curated mechanisms. They often do not preserve the contextual, real-world drug-use information found in leaflets and package inserts (Sengupta et al., 30 Sep 2025).
Drug leaflets are treated as a particularly suitable source because they are standardized, regulated documents created as part of the approval process and include information on safety, efficacy, dosage, storage, and product appearance. MEDAKA is therefore positioned not as a replacement for molecularly oriented biomedical graphs, but as a complementary resource that captures practical medication-use knowledge. This suggests a shift from predominantly mechanism-centric graph construction toward a more patient-centered slice of drug knowledge.
The paper’s contribution is dual. First, it presents a modular, reproducible, and adaptable pipeline for constructing knowledge graphs from unstructured web content. Second, it instantiates that pipeline as a leaflet-derived biomedical graph whose novelty lies both in its source material and in its focus on clinically relevant attributes that matter in prescribing and medication use (Sengupta et al., 30 Sep 2025).
2. End-to-end construction pipeline
The MEDAKA pipeline has three main stages: web scraping and parsing, LLM-based extraction, and knowledge-graph construction. For data collection, the authors scrape drug leaflets from the Health Products Regulatory Authority in Ireland using a lightweight Python scraper built with BeautifulSoup. The scraper parses HPRA HTML pages, extracts direct PDF links, and stores the PDFs locally. The reported corpus is about 13,000 drug leaflets for human medicine (Sengupta et al., 30 Sep 2025).
The PDFs are converted to plain text using PyMuPDF. That text is then processed by a prompt-based extraction system that emits subject–relation–object triples. A key design choice is single-pass processing of each leaflet rather than chunking. The paper explains that leaflets are typically 4,000–8,000 words; chunking caused redundancy or omissions, especially for smaller models. In model comparison on 200 leaflets, Mistral 7B required chunking and introduced redundancy, Qwen-32B was too verbose, and LLaMA 3.3 70B Instruct was the most consistent choice for complete, concise extraction. The prompt was iteratively refined, and in-context examples were added to improve stability and reduce redundancy. Extraction is constrained to clinically relevant triples with the drug name as subject and one of the defined relation types (Sengupta et al., 30 Sep 2025).
To reduce hallucinations and other LLM errors, the same leaflet is processed five times. Each triple is tracked across the five generations, and its confidence score is defined by the number of generations containing the triple divided by 5. Triples are retained only if their confidence score is at least 0.5, so a retained triple must appear in at least 3 of 5 generations. This majority-vote filter serves as a postprocessing safeguard against unstable or spurious outputs.
After filtering, subjects, relations, and objects are normalized by converting them to lowercase. The final graph is then assembled as a directed labeled graph from the normalized triples. The paper also reports an operational limitation: the same simple scraping approach did not work well on a second pharmacy website because of heavy JavaScript rendering and restrictive crawling rules, and more advanced tools such as ScraperAPI, Crawl4AI, and Selenium had to be explored. A plausible implication is that the pipeline is portable in principle but remains sensitive to source-site web architecture (Sengupta et al., 30 Sep 2025).
3. Schema and graph structure
The MEDAKA schema is intentionally simple and drug-centric. Its entities are organized around drug nodes connected to safety, composition, administration, storage, and appearance attributes.
| Node types | Relations |
|---|---|
| Drug | hassideeffect |
| ActiveIngredient | haswarning |
| InactiveIngredient | hascontraindication |
| SideEffect | hasactiveingredient |
| Warning | hasinactiveingredient |
| Contraindication | hasdosageinfo |
| DosageInfo | hasstorageinfo |
| StorageInfo | hascolor |
| Color | hasshape |
| Shape |
This schema captures clinically relevant product information at a granularity that the paper characterizes as often absent from conventional biomedical graphs. In particular, it includes inactive ingredients, dosage information, storage instructions, and physical attributes such as color and shape, in addition to more familiar safety-related categories such as side effects, warnings, and contraindications (Sengupta et al., 30 Sep 2025).
The graph’s semantic organization is therefore centered on practical medication-use attributes rather than on molecular or biological mechanisms alone. This suggests that MEDAKA is structured for tasks in which interpretability and patient-facing drug attributes are central, rather than for exclusively mechanistic reasoning.
4. Dataset scale and empirical organization
MEDAKA contains 41,142 nodes and 466,359 directed, labeled edges. It is constructed from the extracted leaflet information and is semantically organized around drug nodes linked to clinically relevant attributes, including side effects, warnings, contraindications, active ingredients, inactive ingredients, dosage information, storage instructions, and physical characteristics such as color and shape (Sengupta et al., 30 Sep 2025).
The paper further notes that drugs with similar connected attributes can be clustered using Jaccard similarity over binary vectors of their associated biomedical entities. This indicates that the graph is not merely a repository of isolated leaflet facts; it also supports similarity-based organization of drugs by shared safety and usage profiles. Example subgraphs centered on Toltertan SR and Catasart Plus are used to illustrate how a single drug node can connect to multiple clinically relevant attribute types.
Because MEDAKA is assembled from normalized triples and organized around attribute-bearing drug nodes, it exposes a representation that is directly compatible with graph-based comparison at the product-information level. A plausible implication is that such a structure can complement graphs aimed at pharmacological mechanism by supplying orthogonal information about how drugs are described for real-world use.
5. Evaluation methodology and results
Because no gold-standard reference knowledge graph exists for this exact task, the authors use three complementary evaluation strategies: manual inspection, recall estimation through false-negative analysis on a small subset, and LLM-as-a-Judge evaluation using gpt-oss-120b (Sengupta et al., 30 Sep 2025).
In the human evaluation, 100 drug leaflets were randomly sampled, yielding 3,549 extracted triples. Each triple was labeled as Correct, Incorrect, or Partially Correct. For recall estimation, 10 randomly selected leaflets containing 619 relations were manually inspected for false negatives. The LLM-as-a-Judge protocol supplied the leaflet text and extracted triples to gpt-oss-120b, which then labeled each triple as correct, incorrect, or partially correct with a short justification; the prompt was refined iteratively and supplemented with in-context examples.
| Evaluation | Correct | Incorrect | Partially correct |
|---|---|---|---|
| Human labels | 3,427 (96.6%) | 34 (0.9%) | 88 (2.5%) |
| LLM judge | 3,439 (96.9%) | 24 (0.7%) | 86 (2.4%) |
For recall estimation on the manually inspected subset, the authors found 79 false negatives among 619 relations, corresponding to a recall of 87.2%. The close agreement between human labels and the LLM judge is presented as evidence that LLM-based evaluation can be a scalable proxy for manual inspection, although human evaluation remains the reference standard (Sengupta et al., 30 Sep 2025).
These results support a characterization of MEDAKA as high-precision and promising-recall under the reported protocol. At the same time, the restricted scope of recall estimation indicates that coverage remains harder to certify exhaustively than precision.
6. Relation to existing biomedical knowledge graphs and intended uses
MEDAKA is explicitly framed as complementary to existing biomedical knowledge graphs and databases rather than as a replacement. The comparison is centered less on overall graph scale than on attribute coverage. According to the paper’s comparison table, iKraph, PrimeKG, DRKG, DrugMechDB, SIDER, and FAERS all contain drug entities, but none jointly cover the full set of leaflet-level features captured by MEDAKA. SIDER and FAERS include side effects or adverse events but lack contraindications, warnings, dosage, storage, and physical attributes. PrimeKG and DRKG are described as oriented more toward molecular or precision-medicine use cases and as not modeling visual attributes or many leaflet-based usage instructions. DrugBank includes rich curated drug information, but MEDAKA is presented as unique in its direct extraction from leaflets and in its broader coverage of patient-facing attributes (Sengupta et al., 30 Sep 2025).
The main novelty claim therefore lies in the simultaneous representation of warnings, contraindications, side effects, active ingredients, inactive ingredients, dosage information, storage instructions, and physical attributes in a single leaflet-derived graph. This fills what the paper describes as a gap between molecular biomedical graphs and practical medication-use knowledge.
The intended downstream uses are patient safety monitoring, drug recommendation, and broader medication support tasks, especially where warnings, contraindications, and dosage or storage information matter. Because the graph groups drugs by connected attributes, it could help identify drugs with similar safety profiles or usage constraints. The paper also suggests that future graph-based reasoning could support drug–disease tasks if MEDAKA is enriched with disease-related information (Sengupta et al., 30 Sep 2025).
7. Limitations, future directions, and terminological disambiguation
The paper identifies several limitations. The approach depends on the reliability and completeness of online drug leaflets; some leaflets may be incomplete or inconsistently structured. Website variability affects scraping reliability. LLM extraction remains subject to hallucinations, motivating majority voting and postprocessing. Recall was estimated only on a small subset because exhaustive manual validation is expensive (Sengupta et al., 30 Sep 2025).
Future work is described along three lines: enriching MEDAKA with disease-related information such as symptoms, risk factors, and definitions from external biomedical sources; extending the pipeline to multilingual corpora; and applying the modular pipeline to domains beyond biomedicine. These directions are consistent with the presentation of the method as a reusable framework for graph construction from text-rich, unstructured content.
A recurrent source of confusion is terminological rather than methodological. In arXiv literature, “medaka” also denotes the fish Oryzias latipes, a vertebrate model organism used in embryo phenotyping, tomographic segmentation, polarization-sensitive OCT, and population genetics (Genest et al., 2019, Bhatt et al., 2023, Zhu et al., 2023, Spivakov et al., 2013). MEDAKA in (Sengupta et al., 30 Sep 2025), by contrast, denotes a leaflet-derived biomedical knowledge graph. The overlap is lexical only; the acronymic graph resource and the model organism belong to distinct research contexts.