SkyWhal3/Peroxisome_PubMed_Central_Multimodal

Peroxisome PubMed Central Multimodal Dataset

Figure–caption pairs in LLaVA format for training and evaluating vision–language models on peroxisome biology — biogenesis, matrix-protein import, the RING-ligase retrotranslocation machinery, the peroxisomal disorders, and peroxisome connections to cancer, inflammation, infection, and neurodegeneration.

Every figure and caption is drawn from the PubMed Central Open Access subset and redistributed under CC-BY-4.0, with attribution carried in each image's PMC ID.

What's in it

File	Records	Contents
figure_caption.json	10,273	figure → published caption
figure_detailed.json	10,273	figure → context-grounded detailed description
figure_qa.json	10,273	visual Q&A grounded in the figure
article_multiimage.json	2,419	multi-figure, article-level summaries
combined_training.json	33,238	shuffled union of the above (the training file)

10,273 unique figure images, 0 caption/image mismatches, 0 duplicate ids.

How it was built

This set is deliberately peroxisome-wide and precision-filtered, not a naive keyword scrape. The pipeline:

1. Tiered harvest. A PubMed Central search anchored on peroxisome/peroxisomal in the title/abstract, expanded across five tiers: core peroxisome biology, the full PEX/peroxin family (PEX1–PEX26), the peroxisomal disorders (Zellweger spectrum, RCDP, X-ALD, Refsum), disease connections (cancer / inflammation / infection / ferroptosis / neurodegeneration, each requiring peroxisome co-occurrence), and model-organism mechanism (yeast-forward). → ~15,300 candidate articles.

2. PPAR-aware relevance filter. The single biggest noise source in any peroxisome corpus is PPAR — "peroxisome proliferator-activated receptor," a nuclear receptor that merely contains the word "peroxisome." A plain keyword or score threshold keeps these because they say "peroxisome" repeatedly. This build separates them by phrase, removing ~10,400 PPAR-only and body-only-mention articles while protecting single-mention disease/gene papers. → 3,219 genuinely peroxisome-focused articles.

3. Open-access image retrieval + extraction. Figure images pulled from the PMC OA service for the ~1,807 figure-bearing articles with retrievable packages, then paired to captions with a uniqueness guard (no figure-number collisions) and a 40-character minimum caption length.

Record format (LLaVA)

{
  "id": "PMC1234567_f1",
  "image": "images/PMC1234567-f1.png",
  "conversations": [
    {"from": "human", "value": "<image>\nDescribe this scientific figure in detail."},
    {"from": "gpt", "value": "Figure 1. ..."}
  ]
}

Multi-image records carry a list of image paths and a <image> tag per figure.

Source, license, and attribution

• Source: PubMed Central Open Access subset.
• License: CC-BY-4.0. Each source article is itself CC-BY (or more permissive); attribution is preserved through the PMC ID embedded in every image filename and record id.
• Reuse is permitted with attribution to the original articles.

Known limitations

• ~925 author-manuscript figures unmatched. nihms-* author-manuscript images are not always present in the OA package pull; those figures are omitted rather than paired to a wrong image. Recoverable in a future pass.
• Short captions dropped (< 40 characters) as likely figure titles rather than legends.
• Captions are the published captions; they are not re-annotated.

Intended use

Training and evaluation of vision–language models for scientific figure understanding in peroxisome cell biology and adjacent rare-disease research. Built to support the PeroxiOS project — a queryable corpus for researchers connecting their work to peroxisome pathways.