Datasets:

SkyWhal3
/

STXBP1_PubMed_Central_Multimodal_Dataset

license: cc-by-4.0
task_categories:
  - image-to-text
  - visual-question-answering
  - text-generation
  - question-answering
language:
  - en
tags:
  - multimodal
  - vision-language
  - medical
  - biomedical
  - neuroscience
  - epilepsy
  - rare-disease
  - STXBP1
  - STXBP2
  - Munc18
  - Munc18-1
  - syntaxin-binding-protein
  - CRISPR
  - CRISPR-Cas9
  - gene-therapy
  - ene-editing
  - base-editing
  - prime-editing
  - AAV
  - PubMed
  - PMC
  - full-text
  - research
  - encephalopathy
  - developmental-delay
  - LLaVA
  - scientific-figures
pretty_name: STXBP1 PubMed Central Multimodal Dataset
size_categories:
  - 10K<n<100K

Dataset Description

A comprehensive multimodal collection of ~31,500 scientific articles from PubMed Central (PMC) with 175,000+ figures and images. This dataset pairs full-text articles with their associated scientific figures, making it one of the most complete publicly available multimodal resources for STXBP1 and gene therapy research.

Why Multimodal Matters

Scientific understanding often depends on visual data—protein structures, experimental results, pathway diagrams, microscopy images. This dataset preserves the connection between article text and figures, enabling:

Vision-language model training on scientific content
Figure-to-caption learning for scientific image understanding
Multimodal RAG systems that can reason over both text and images
Automated scientific figure analysis

About STXBP1

STXBP1 (also known as Munc18-1) encodes a protein essential for neurotransmitter release. Mutations cause STXBP1 Encephalopathy, a rare neurological disorder (~1 in 30,000 births) characterized by early-onset epilepsy, developmental delays, and movement disorders.

Dataset Statistics

Metric	Count
Total Articles	31,585
Articles with Images	30,139 (95.4%)
Total Images	175,404
Average Images per Article	5.5
Image References in Text	3,253,705
Image-Text Match Rate	100.00%
Total Dataset Size	~52 GB

Dataset Structure

stxbp1-pubmed-multimodal/
├── multimodal_data/           # Article JSONs with image references
│   ├── PMC10000387_multimodal.json
│   ├── PMC10002385_multimodal.json
│   └── ... (31,585 files)
├── images/                    # All figure images
│   ├── PMC10000387-Fig1.png
│   ├── PMC10000387-Fig2.png
│   └── ... (175,404 files)
├── training_llava.json        # LLaVA format training data
├── training_conversational.json   # Chat format training data
├── training_simple.json       # Simple text format
└── README.md

Data Formats

Multimodal JSON Structure

Each article JSON contains:

{
  "pmc_id": "PMC24456",
  "title": "Article title...",
  "abstract": "Abstract text...",
  "main_text": "Full text with <image>PMC24456-F1.png</image> inline references...",
  "images": [
    "PMC24456-F1.png",
    "PMC24456-F2.png"
  ],
  "image_mapping": {
    "Figure 1": "PMC24456-F1.png",
    "Fig 1": "PMC24456-F1.png",
    "F1": "PMC24456-F1.png"
  },
  "metadata": {
    "authors": ["Author One", "Author Two"],
    "journal": "Journal Name",
    "doi": "10.xxxx/xxxxx",
    "pmid": "12345678",
    "publication_date": "2024 Jan 15",
    "stxbp1_count": 5,
    "munc18_count": 2
  }
}

Pre-formatted Training Files

Ready-to-use training files are included:

1. LLaVA Format (`training_llava.json`)

For vision-language models (LLaVA, BLIP-2, etc.)

{
  "id": "PMC10000387",
  "image": ["PMC10000387-Fig1.png", "PMC10000387-Fig2.png"],
  "conversations": [
    {"from": "human", "value": "<image>\nWhat does this scientific article discuss?"},
    {"from": "gpt", "value": "Article title and full text..."}
  ]
}

2. Conversational Format (`training_conversational.json`)

For chat models (Llama, Mistral, etc.)

{
  "conversations": [
    {"from": "human", "value": "Analyze this scientific article about Neurology."},
    {"from": "assistant", "value": "Article title and full text..."}
  ]
}

3. Simple Format (`training_simple.json`)

For base LLM training or embeddings

{
  "text": "Article title and full text..."
}

Usage Examples

Loading with Hugging Face Datasets

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("SkyWhal3/stxbp1-pubmed-multimodal")

Loading Images with Articles

import json
from PIL import Image
from pathlib import Path

# Load an article
with open("multimodal_data/PMC24456_multimodal.json") as f:
    article = json.load(f)

# Load associated images
for img_name in article['images']:
    img_path = Path("images") / img_name
    if img_path.exists():
        img = Image.open(img_path)
        print(f"{img_name}: {img.size}")

Filtering High-Relevance Articles

import json
from pathlib import Path

# Find articles with 5+ STXBP1 mentions
high_relevance = []
for json_file in Path("multimodal_data").glob("*.json"):
    with open(json_file) as f:
        data = json.load(f)
    if data.get('metadata', {}).get('stxbp1_count', 0) >= 5:
        high_relevance.append(data)

print(f"Found {len(high_relevance)} high-relevance articles")

Fine-tuning LLaVA

# The training_llava.json is ready to use with standard LLaVA fine-tuning
# See: https://github.com/haotian-liu/LLaVA

# Example training command:
# python llava/train/train.py \
#     --data_path training_llava.json \
#     --image_folder images/ \
#     --model_name_or_path liuhaotian/llava-v1.5-7b \
#     ...

Search Terms Used

This dataset was collected using comprehensive PMC full-text search:

STXBP Family Terms:

STXBP1, STXBP2, STXBP3, STXBP4, STXBP5, STXBP6, STXBP11
Munc18, Munc18-1, Munc-18, Munc 18
Munc18 syntaxin, syntaxin binding protein

CRISPR/Gene Therapy Terms:

CRISPR Cas9, CRISPR-Cas9, CRISPR/Cas9
CRISPR Cas12, CRISPR-Cas12, CRISPR/Cas12
CRISPR Cas13, CRISPR-Cas13
base editing, prime editing

Combination Searches:

STXBP1/Munc18/syntaxin binding protein + CRISPR/gene therapy/AAV/antisense

Data Quality

Image-Text Alignment Audit

We performed comprehensive validation of image references:

Check	Result
Total image references in text	3,253,824
Successfully matched to files	3,253,705
Unmatched references	119 (removed)
Match rate	99.996%

The 119 mismatched references (from 3 articles) were removed to ensure clean training data.

Filtering Recommendations

Filter	Description	Use Case
`stxbp1_count >= 1`	Mentions STXBP1 at least once	General STXBP1 research
`stxbp1_count >= 5`	Substantial STXBP1 discussion	Core STXBP1 papers
`munc18_count >= 1`	Uses Munc18 nomenclature	Older literature
`stxbp1_count == 0 AND munc18_count == 0`	No direct mentions	CRISPR methodology papers

Limitations

What This Dataset Contains

✅ Full article text with inline image references
✅ 175,000+ scientific figures (PNG format)
✅ Structured image-to-text mapping
✅ Pre-formatted training files (LLaVA, conversational, simple)
✅ Rich metadata (authors, DOIs, journals, dates)
✅ Relevance scoring (STXBP1/Munc18 mention counts)

What This Dataset Does NOT Contain

❌ Supplementary data files (Excel, raw data, etc.)
❌ Video content (some articles may reference videos)
❌ Interactive figures (3D viewers, etc.)
❌ Table structure (tables are linearized text)
❌ LaTeX equations (flattened to text)

Image Considerations

All images are PNG format (converted from original TIF/JPG/etc.)
Some complex multi-panel figures may be single images
Figure quality varies by source journal
Inline <image> tags in text mark where figures appear

Ethical Considerations

All articles are from the PMC Open Access Subset under various open licenses
This dataset is intended for research and educational purposes
Users should cite original articles when using specific findings
This dataset should not replace professional medical advice
Patient data in case studies has been de-identified by original authors

Citation

@dataset{stxbp1_pubmed_multimodal_2025,
  title={STXBP1 PubMed Central Multimodal Dataset},
  author={SkyWhal3},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/SkyWhal3/stxbp1-pubmed-multimodal},
  note={31,585 articles with 175,404 scientific figures}
}

Related Datasets

Dataset	Description
SkyWhal3/stxbp1-pubmed-central-fulltext	Text-only version (1.3 GB) - faster to download

Acknowledgments

This dataset was curated by a parent advocate in the STXBP1 community to accelerate research into treatments and potential cures for STXBP1-related disorders.

Special thanks to:

The researchers whose work is represented in this dataset
PubMed Central for providing open access to scientific literature
The STXBP1 Foundation and patient advocacy community
ClaudeCode Opus 4.5

Version History

Version	Date	Changes
1.0.0	2025-06	Initial release with 31,585 articles and 175,404 images

Contact

For questions, corrections, or contributions, please open an issue on this repository.

This dataset is dedicated to all children and families affected by STXBP1 Encephalopathy.

"The cure is in the data."