Upload folder using huggingface_hub

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+chroma.sqlite3 filter=lfs diff=lfs merge=lfs -text

0acd4040-15c6-47bb-bc79-aa7a3f4ba12d/data_level0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa916e111b5cccc428423c72666ce8cf4574852774f630766642033bdf4146da
+size 831450280

0acd4040-15c6-47bb-bc79-aa7a3f4ba12d/header.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:254c83f06634efab9cc878dc45a1ad4f7e6c41b47a69b79c809bd3f91a143072
+size 100

0acd4040-15c6-47bb-bc79-aa7a3f4ba12d/index_metadata.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c6e3ffb730fb3f5c37c0c27ab72beac5b633c983c334708190a7ad2aa56c502
+size 6629324

0acd4040-15c6-47bb-bc79-aa7a3f4ba12d/length.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f017891c97b958922a1933e93aacdf753eab40c2e5c5b7441200c5684bd5a1e
+size 399160

0acd4040-15c6-47bb-bc79-aa7a3f4ba12d/link_lists.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95c8b398048187194224a33f3ecdb69df947aea9229772125063369b434849b6
+size 850884

4793bb23-f50e-4ef9-81e1-f0d8ddcc8161/data_level0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:171c18dc9977d442c4b8664649559e180a28b621af49a7b75009f680e6c4ca51
+size 16855636

4793bb23-f50e-4ef9-81e1-f0d8ddcc8161/header.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf109cc0093a66777611721c6b78bc127a55b8b0fba96c9682f0df4148ae8395
+size 100

4793bb23-f50e-4ef9-81e1-f0d8ddcc8161/index_metadata.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59358a23e29f7c51162d9fa3b19a2b24c427e5fcf75bd766b6838a14af0b6191
+size 115948

4793bb23-f50e-4ef9-81e1-f0d8ddcc8161/length.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7dd74678cad49beed8953212de7d0561d5b1b5e1e8b036dcbcc874bf05413e6
+size 8092

4793bb23-f50e-4ef9-81e1-f0d8ddcc8161/link_lists.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35e8af60bef772c25a157a7cb1c884fca532d807c1b7e0f3f100f71eeceb987d
+size 18020

README.md

@@ -1,3 +1,311 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+task_categories:
+  - text-retrieval
+  - question-answering
+language:
+  - en
+tags:
+  - medical
+  - biology
+  - rare-disease
+  - PEX10
+  - peroxisomal-biogenesis-disorder
+  - PBD-ZSD
+  - Zellweger-spectrum
+  - peroxisome
+  - RING-E3-ligase
+  - neuroscience
+  - RAG
+  - embeddings
+  - chromadb
+pretty_name: PEX10-ARIA RAG Nemotron
+size_categories:
+  - 100K<n<1M
+---
+
+# 🧬⚡ PEX10-ARIA RAG Database v1 — NVIDIA Nemotron Embeddings
+
+**The definitive retrieval corpus for PEX10 / peroxisomal biogenesis disorder research.**
+
+A pre-built ChromaDB vector database containing:
+**98,734 indexed text chunks** from **835 curated PubMed Central (PMC) biomedical papers** + **1,495 structured curated entries** (1,016 ClinVar variants + 198 truncation consequence cards + 281 ESMFold variant structural analyses),
+embedded with NVIDIA's state-of-the-art **Llama-Nemotron-Embed-1B-v2** model featuring **2048-dimensional embeddings**.
+
+> ⚡ **This is the premium GPU-accelerated version** — Nemotron embeddings deliver maximum semantic precision for rare-disease therapeutic queries, but require a GPU with 2-4GB VRAM. For a lightweight CPU-friendly alternative, see: **[PEX10-RAG-Database (BGE)](https://huggingface.co/datasets/SkyWhal3/PEX10-RAG-Database)**
+
+---
+
+## 🌐 Part of the ARIA Rare-Disease Family
+
+This is the **third gene** in the ARIA rare-disease RAG family. Same architecture, same discipline, same embedder across all three:
+
+| Gene        | Built For                  | Papers  | Chunks  | Dataset |
+|-------------|---------------------------|---------|---------|---------|
+| **STXBP1**  | Katelyn Freygang          | ~17,000 | 571,816 | [STXBP1-RAG-Nemotron](https://huggingface.co/datasets/SkyWhal3/STXBP1-RAG-Nemotron) |
+| **SNAP25**  | SNAP25 Foundation         | ~2,000  | 76,592  | [SNAP25-RAG-Nemotron](https://huggingface.co/datasets/SkyWhal3/SNAP25-RAG-Nemotron) |
+| **PEX10**   | **PBD Project (Diego)**   | **835** | **98,734** | *(this dataset)* |
+
+**"We've now done this three times."** Each rare-disease foundation partner gets a purpose-built corpus: PMC literature + ClinVar variants + structural analyses + expert curation, all in one queryable ChromaDB.
+
+---
+
+## 🧬 About PEX10
+
+**Gene**: PEX10 (also known as RNF69, Peroxin 10, Peroxisomal Biogenesis Factor 10)
+**Protein**: 346 aa (RefSeq NP_722540.1, NM_153818.2) · UniProt O60683
+**Function**: C3HC4-type RING finger E3 ubiquitin ligase integral to the peroxisomal membrane. Forms the **Pex2-Pex10-Pex12 retrotranslocation channel** essential for PEX5 receptor recycling and peroxisome matrix protein import.
+**Disease**: PEX10-related Zellweger Spectrum Disorder (PBD6A/6B, MONDO:0100079, OMIM:614870). Approximately 3% of all peroxisome biogenesis disorder cases.
+
+### Why this dataset exists
+
+The 2022 Nature paper *"A peroxisomal ubiquitin ligase complex forms a retrotranslocation channel"* (Feng et al., PubMed 35768507) established the mechanistic story of the Pex2-Pex10-Pex12 complex as a channel that extracts PEX5 back out of the peroxisomal membrane. That gave PEX10 a concrete structural framework for interpreting where variants may matter. This corpus consolidates the evidence so it can be queried in one place.
+
+---
+
+## 📊 Dataset Statistics
+
+| Metric | Value |
+|--------|-------|
+| **Paper Chunks** | 98,734 |
+| **Source Papers** | 835 PMC articles (relevance ≥ 15) |
+| **ClinVar Variants** | 1,016 |
+| **Truncation Consequence Cards** | 198 |
+| **ESMFold Structural Reports** | 281 |
+| **Total Indexed Items** | **100,229** |
+| **Database Size** | ~400-600 MB |
+| **Embedding Model** | `nvidia/llama-nemotron-embed-1b-v2` |
+| **Embedding Dimensions** | 2048 |
+| **Model Parameters** | 1B |
+| **Chunk Size** | ~1,800 chars with 200-char overlap |
+| **Index Type** | ChromaDB with HNSW (cosine) |
+| **Build Time** | 31 minutes on consumer GPU |
+| **Build Date** | April 23, 2026 |
+
+---
+
+## 🏆 Why Nemotron?
+
+NVIDIA's Llama-Nemotron-Embed-1B-v2 ranks **#2 on MTEB retrieval benchmarks** — distilled from the 8B flagship into an efficient 1B parameter model. For rare-disease work, the 2048-dim semantic space makes fine-grained distinctions that matter:
+
+- "RING-HC zinc-coordinating cysteine" vs "transmembrane helix disruption" → **clear separation**
+- "PEX10 nonsense variant" vs "PEX10 missense variant" → **distinct clusters**
+- "Zellweger severe" vs "cerebellar ataxia milder form" → **phenotype-aware retrieval**
+- "Pex2-Pex10-Pex12 retrotranslocon" vs "PEX5 receptor recycling" → **mechanism-level routing**
+
+| Feature | BGE (sibling dataset) | **Nemotron (this dataset)** |
+|---------|----------------------|-----------------------------|
+| **Embedding Dims** | 768 | **2048** ⬆️ 2.7× |
+| **Model Params** | 110M | **1B** ⬆️ 9× |
+| **Semantic Precision** | Good | **Excellent** |
+| **Hardware** | CPU OK | **GPU recommended** |
+
+---
+
+## 📚 Corpus Composition
+
+### 1. PMC Literature (98,734 chunks)
+
+Harvested via [`SkyWhal3/PMC-Corpus-Harvester`](https://huggingface.co/datasets/SkyWhal3/PMC-Corpus-Harvester) with gene-specific scoring and exclude-terms for every non-PEX10 peroxin (PEX1, 2, 3, 5, 6, 7, 11, 12, 13, 14, 16, 19, 26 — the "secret sauce" that keeps PEX10-centric papers near the top without cross-family contamination).
+
+**Keyword families**:
+- Primary: `pex10`, `rnf69`, `peroxin-10`, `peroxisome biogenesis`, `zellweger`, `pbd-zsd`
+- Mechanism: `Pex2-Pex10-Pex12`, `retrotranslocation`, `RING finger`, `PEX5 recycling`, `PTS1 import`
+- Disease phenotype: `neonatal hypotonia`, `sensorineural hearing loss`, `cerebellar ataxia`, `VLCFA`, `plasmalogen`
+- Therapeutics: `gene therapy`, `CRISPR`, `base editing`, `ASO`, `readthrough`
+
+### 2. ClinVar Variant Records (1,016 curated entries)
+
+Every ClinVar record for the PEX10 gene as of April 22, 2026. Each entry includes:
+- `variation_id` + `accession` (VCV)
+- HGVS `c.` and `p.` notation
+- Germline classification (187 Pathogenic + 53 Likely Pathogenic + 345 VUS + 326 LB + 22 Benign + 33 Conflicting + 36 P/LP + others)
+- Review status + evaluation date
+- Associated conditions (from ClinVar TraitSet)
+- Variation type (749 SNVs + 88 deletions + 106 copy-number-loss + others)
+
+### 3. Truncation Consequence Cards (198 entries)
+
+Nonsense + frameshift + canonical splice variants, with a structured JSON card per variant describing:
+- Truncation position (first residue lost)
+- Domain it falls in (using 346-aa RefSeq frame)
+- Domains lost downstream (full vs partial)
+- Topology disruption call in plain English
+- `% protein lost` quick-read metric
+
+Truncated proteins are **deliberately not modeled** — ESMFold and OpenFold3 predictions on truncated sequences are known to produce misleadingly confident fold predictions. Consequence cards are more honest.
+
+### 4. ESMFold Variant Structural Reports (281 entries)
+
+Every foldable missense variant from ClinVar, run through ESMFold via NVIDIA NIM. Each report includes:
+- Per-residue pLDDT
+- Local Cα-Cα contact map around the mutation site (WT and mutant)
+- Delta contacts (gained, lost, net change)
+- Mean pLDDT + site pLDDT
+- Original classification + condition
+
+**Isoform handling**: PEX10 has a 20-aa segment at positions ~200-220 present in NCBI isoform 1 (NP_722540.1, 346 aa, used by ClinVar) but absent from UniProt canonical (NP_002608.1, 326 aa). 104 of the 281 variants were originally reported on the 326-aa frame; they were remapped to the 346-aa frame via +20 shift with WT-residue verification. Every remapped variant carries an `isoform2_mapping.json` audit trail. **96.4% structural coverage** of the foldable set.
+
+---
+
+## 🚀 Quick Start
+
+### Load in Python
+
+```python
+from huggingface_hub import snapshot_download
+import chromadb
+from chromadb.config import Settings
+
+# Download database (one-time, ~500 MB)
+db_path = snapshot_download(
+    repo_id="SkyWhal3/PEX10-RAG-Nemotron",
+    repo_type="dataset"
+)
+
+client = chromadb.PersistentClient(
+    path=db_path,
+    settings=Settings(anonymized_telemetry=False)
+)
+
+collection = client.get_collection("pex10_papers")
+print(f"Loaded {collection.count():,} chunks")
+```
+
+### Query with Nemotron Embeddings
+
+```python
+from sentence_transformers import SentenceTransformer
+
+# Load embedder (GPU recommended, CPU works for a handful of queries)
+embedder = SentenceTransformer(
+    "nvidia/llama-nemotron-embed-1b-v2",
+    trust_remote_code=True
+)
+
+# Ask a question
+query = "What is the role of the RING-HC domain in PEX10-mediated PEX5 recycling?"
+query_embedding = embedder.encode(query).tolist()
+
+results = collection.query(
+    query_embeddings=[query_embedding],
+    n_results=10,
+    include=["documents", "metadatas", "distances"]
+)
+
+for doc, meta, dist in zip(
+    results['documents'][0],
+    results['metadatas'][0],
+    results['distances'][0]
+):
+    src = meta.get('source', 'unknown')
+    label = meta.get('pmcid') or meta.get('variation_id') or 'curated'
+    print(f"[{src} · {label}] (distance: {dist:.3f})")
+    print(f"{doc[:200]}...\n")
+```
+
+### Combined Literature + Variant Query
+
+Because ClinVar records, consequence cards, ESMFold reports, and paper chunks all share the same embedding space, one query retrieves across all four sources:
+
+```python
+query = "PEX10 RING domain cysteine mutations that disrupt zinc coordination"
+# Returns: relevant PMC paper chunks + ClinVar C306Y/C306F entries + ESMFold reports for those specific variants
+```
+
+That cross-source retrieval is what makes this corpus more queryable than any single database alone.
+
+---
+
+## 🧠 ARIA Family Navigator Integration
+
+This dataset is the canonical RAG backend for the **PEX10 / PBD Research** knowledge base chip in [ARIA Family Navigator](https://aria-app.webflow.io). When a family has a PEX10 profile, the chip auto-enables in Research Mode and `rag_search` queries route to this corpus.
+
+Config (from `src/shared/rag-config.ts`):
+```typescript
+{
+  id: 'pex10',
+  displayName: 'PEX10 / PBD Research',
+  hfDatasetId: 'SkyWhal3/PEX10-RAG-Nemotron',
+  collectionName: 'pex10_papers',
+  defaultForGenes: ['PEX10', 'PEX-10', 'RNF69'],
+  embeddingModel: 'nvidia/llama-nemotron-embed-1b-v2',
+  embeddingDim: 2048,
+}
+```
+
+---
+
+## 🔗 Related Resources
+
+| Resource | Link |
+|----------|------|
+| **CPU Version (BGE)** | [PEX10-RAG-Database](https://huggingface.co/datasets/SkyWhal3/PEX10-RAG-Database) |
+| **Harvester Tool** | [PMC-Corpus-Harvester](https://huggingface.co/datasets/SkyWhal3/PMC-Corpus-Harvester) |
+| **STXBP1 RAG (sibling)** | [STXBP1-RAG-Nemotron](https://huggingface.co/datasets/SkyWhal3/STXBP1-RAG-Nemotron) |
+| **SNAP25 RAG (sibling)** | [SNAP25-RAG-Nemotron](https://huggingface.co/datasets/SkyWhal3/SNAP25-RAG-Nemotron) |
+| **ARIA Family Navigator** | [aria-app.webflow.io](https://aria-app.webflow.io) |
+
+---
+
+## 📁 Files
+
+| File | Description |
+|------|-------------|
+| `chroma.sqlite3` | ChromaDB metadata + document store |
+| `{uuid}/data_level0.bin` | HNSW vector index |
+| `{uuid}/header.bin` | Index metadata |
+| `{uuid}/length.bin` | Vector dimensions |
+| `curated_entries.json` | ClinVar + truncation + ESMFold entries (for transparency; already embedded) |
+| `metadata.json` | Build provenance + chunking parameters + source paths |
+| `README.md` | This file |
+
+---
+
+## 🧬 The Story Behind This Dataset
+
+Built by **Adam Freygang** — AI/ML engineer, STXBP1 parent, maker of the ARIA Family Navigator. This is the third rare-disease corpus I've published under the same architecture. The first was for my daughter Katelyn (STXBP1). The second was at the request of the SNAP25 Foundation. This one is for the **PBD Project** — Andrew Longenecker's son Diego has PEX10-related peroxisomal biogenesis disorder, and I built this corpus in a single night so our meeting the next morning wouldn't start from zero.
+
+The pitch is not "we found secret PEX10 facts nobody can access." The pitch is: *"We assembled a consolidated, queryable PEX10 knowledge-and-structure atlas that most research groups do not currently have in one interoperable package — combining literature, variant evidence, structural modeling, membrane topology interpretation, and variant-level exports that can extend into a live research tool."*
+
+If your family or foundation has a rare disease and this platform would help you, reach out. Each new gene takes ~1 day of compute time once the template is in place.
+
+---
+
+## 📜 License
+
+Apache 2.0. Use freely. Attribution requested.
+
+---
+
+## 🙏 Acknowledgments
+
+- **Andrew Longenecker** (PBD Project) — for trusting a stranger with your son's cause
+- **Lior Dior** — for the introduction
+- **NVIDIA** — for Nemotron embeddings and NIM API access
+- **The PEX10 research community** — especially the Feng et al. 2022 team whose mechanistic Nature paper gave this corpus its structural backbone
+- **Katelyn, Diego, and every rare-disease child whose families are doing this in the margins of their lives**
+
+---
+
+## 📄 Citation
+
+```bibtex
+@dataset{pex10_rag_nemotron_2026,
+  author = {Freygang, Adam},
+  title = {PEX10-ARIA RAG Database v1: Nemotron-Embedded Biomedical Literature + ClinVar + Structural Analysis for PBD Research},
+  year = {2026},
+  publisher = {HuggingFace},
+  url = {https://huggingface.co/datasets/SkyWhal3/PEX10-RAG-Nemotron}
+}
+```
+
+---
+
+<div align="center">
+
+*Built with ❤️ for the PBD community. For Diego.*
+
+**Last Updated: April 23, 2026**
+
+</div>

chroma.sqlite3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2564b01f02419a432f17c9f062af25868f65e219a96465e3141525dbc71ea1bf
+size 843980800

metadata.json

@@ -0,0 +1,29 @@
+{
+  "gene": "PEX10",
+  "build_session": "S94",
+  "build_date": "2026-04-23",
+  "embedder": "nvidia/llama-nemotron-embed-1b-v2",
+  "embedding_dim": 2048,
+  "collection_name": "pex10_papers",
+  "paper_count": 835,
+  "chunk_count": 98734,
+  "curated_count": 1495,
+  "curated_by_source": {
+    "clinvar": 1016,
+    "consequence_card": 198,
+    "esmfold_report": 281
+  },
+  "build_elapsed_s": 1861.0,
+  "device": "cuda",
+  "chunking": {
+    "target_chars": 1800,
+    "overlap_chars": 200,
+    "min_chars": 120
+  },
+  "sources": {
+    "pmc_corpus": "D:\\PEX10_datasets\\corpus\\PEX10\\PEX10",
+    "clinvar_master_csv": "D:\\PEX10_datasets\\clinvar\\pex10_clinvar_master.csv",
+    "truncation_cards": "D:\\PEX10_datasets\\structures\\truncations",
+    "esmfold_reports": "D:\\PEX10_datasets\\structures\\variants"
+  }
+}