CAR-T Intelligence Agent¶
Cross-functional intelligence across the CAR-T cell therapy development lifecycle. Part of the HCLS AI Factory.
Overview¶
The CAR-T Intelligence Agent breaks down data silos across the 5 stages of CAR-T development. It searches across all data sources simultaneously and synthesizes cross-functional insights powered by Claude.
| Collection | Records | Source |
|---|---|---|
| Literature | 5,047 | PubMed abstracts via NCBI E-utilities |
| Clinical Trials | 973 | ClinicalTrials.gov API v2 |
| CAR Constructs | 6 | 6 FDA-approved CAR-T products |
| Assay Results | 45 | Curated from landmark papers (ELIANA, ZUMA-1, KarMMa, CARTITUDE-1, etc.) |
| Manufacturing | 30 | Curated CMC/process data (transduction, expansion, release, cryo, logistics) |
| Safety | — | Pharmacovigilance, CRS/ICANS profiles |
| Biomarkers | — | CRS prediction, exhaustion monitoring |
| Regulatory | — | Approval timelines, post-marketing requirements |
| Sequences | — | Molecular binding, scFv sequences |
| Real-World Evidence | — | Registry outcomes, real-world data |
| Genomic Evidence | (read-only) | Shared from Stage 2 RAG pipeline (Milvus) |
| Total | 6,266+ vectors | 11 collections (10 owned + 1 read-only) |
Example Queries¶
"Why do CD19 CAR-T therapies fail in relapsed B-ALL?"
"Compare 4-1BB vs CD28 costimulatory domains for DLBCL"
"What manufacturing parameters predict clinical response?"
"BCMA CAR-T resistance mechanisms in multiple myeloma"
"How does T-cell exhaustion affect CAR-T persistence?"
All queries return grounded, cross-collection answers with clickable Literature:PMID and Trial:NCT... citations.
Comparative Analysis Mode¶
Comparative queries are auto-detected and produce structured side-by-side analysis with markdown tables, advantages/limitations, and clinical context.
"Compare CD19 vs BCMA" → Target vs target
"Compare 4-1BB vs CD28 costimulatory domains" → Costimulatory domain comparison
"Kymriah versus Carvykti" → Product vs product (resolves to CD19 vs BCMA)
"Compare CRS and ICANS toxicity" → Toxicity profile comparison
How it works: The engine detects "vs/versus/compare" keywords, parses two entities, resolves each against the knowledge graph (25 antigens, 6 products, 8 toxicities, 10 manufacturing processes), runs dual retrievals with per-entity filtering, and builds a comparative prompt that instructs Claude to produce structured tables. The evidence panel groups results by entity with color-coded headers.
| Feature | Detail |
|---|---|
| Entity types | Targets, FDA products, costimulatory domains, toxicities, manufacturing |
| Entity resolution | 25 antigens + 39+ product/domain/biomarker/regulatory aliases |
| Dual retrieval | ~365ms for 46 results (24 + 22 per entity) |
| Structured output | Comparison table, advantages, limitations, clinical context |
| Fallback | Unrecognized entities gracefully fall back to normal query path |
Architecture¶
User Query
|
v
[Comparative Detection] ──── "X vs Y" detected? ──── YES ──┐
| |
NO [Parse Two Entities]
| (resolve via knowledge graph)
v |
[BGE-small-en-v1.5 Embedding] [Dual Retrieval]
(384-dim, asymmetric query prefix) (Entity A + Entity B)
| |
v v
[Parallel Search: 11 Milvus Collections] [Comparative Prompt Builder]
(IVF_FLAT / COSINE) (tables + pros/cons format)
| | | |
v v v |
Literature Trials Constructs |
5,047 973 6 |
| | | |
Assays Manufacturing | |
45 30 | |
| | | |
+-------+-------+----------+ |
| |
v v
[Query Expansion] (12 maps, 169 keywords -> 1,496 terms)
|
v
[Knowledge Graph Augmentation]
(25 antigens, 8 toxicities, 10 mfg processes)
|
v
[Claude LLM] -> Grounded response with citations
Built on the HCLS AI Factory platform:
- Vector DB: Milvus 2.4 with IVF_FLAT/COSINE indexes (nlist=1024, nprobe=16)
- Embeddings: BGE-small-en-v1.5 (384-dim)
- LLM: Claude Sonnet 4.6 (Anthropic API)
- UI: Streamlit (port 8521)
- Hardware target: NVIDIA DGX Spark ($3,999)
Setup¶
Prerequisites¶
- Python 3.10+
- Milvus 2.4 running on
localhost:19530 ANTHROPIC_API_KEYenvironment variable (or inrag-chat-pipeline/.env)
Install¶
1. Create Collections and Seed FDA Constructs¶
This creates 11 Milvus collections (10 owned + 1 read-only) with IVF_FLAT indexes and inserts 6 FDA-approved CAR-T products (Kymriah, Yescarta, Tecartus, Breyanzi, Abecma, Carvykti).
2. Ingest PubMed Literature (~15 min)¶
Fetches CAR-T abstracts via NCBI E-utilities (esearch + efetch), classifies by development stage, extracts target antigens, embeds with BGE-small, and stores in cart_literature.
3. Ingest Clinical Trials (~3 min)¶
Fetches CAR-T trials via ClinicalTrials.gov API v2, extracts phase/status/sponsor/antigen/generation, embeds, and stores in cart_trials.
4. Seed Assay Data (~30 sec)¶
Inserts 45 curated assay records from landmark CAR-T publications (ELIANA, ZUMA-1, KarMMa, CARTITUDE-1, etc.) covering cytotoxicity, cytokine, persistence, exhaustion, and resistance data.
5. Seed Manufacturing Data (~30 sec)¶
Inserts 30 curated manufacturing/CMC records covering transduction, expansion, harvest, cryopreservation, release testing, logistics, cost analysis, and emerging platforms (POC, allogeneic, non-viral).
6. Validate¶
Runs 5 tests: collection stats, single-collection search, multi-collection search_all(), filtered search (target_antigen == "CD19"), and all demo queries.
7. Run Integration Test (requires API key)¶
Tests the full RAG pipeline: embed -> search_all -> knowledge graph -> Claude LLM response generation. Validates both synchronous and streaming modes.
8. Launch UI¶
Project Structure¶
cart_intelligence_agent/
├── Docs/
│ └── CART_Intelligence_Agent_Design.md # Architecture design document
├── src/
│ ├── models.py # Pydantic data models (16 models + enums)
│ ├── collections.py # 11 Milvus collection schemas + manager
│ ├── knowledge.py # Knowledge graph (25 targets, 8 toxicities, 10 mfg)
│ ├── query_expansion.py # 12 expansion maps (169 keywords -> 1,496 terms)
│ ├── rag_engine.py # Multi-collection RAG engine + comparative analysis + Claude
│ ├── agent.py # CAR-T Intelligence Agent (plan -> search -> synthesize)
│ ├── ingest/
│ │ ├── base.py # Base ingest pipeline (fetch -> parse -> embed -> store)
│ │ ├── literature_parser.py # PubMed NCBI E-utilities ingest
│ │ ├── clinical_trials_parser.py # ClinicalTrials.gov API v2 ingest
│ │ ├── construct_parser.py # CAR construct data parser
│ │ ├── assay_parser.py # Assay data parser
│ │ └── manufacturing_parser.py # Manufacturing/CMC data parser
│ └── utils/
│ └── pubmed_client.py # NCBI E-utilities HTTP client
├── app/
│ └── cart_ui.py # Streamlit chat interface (NVIDIA theme, comparative mode)
├── config/
│ └── settings.py # Pydantic BaseSettings configuration
├── data/
│ └── reference/
│ ├── assay_seed_data.json # 45 curated assay records from landmark papers
│ └── manufacturing_seed_data.json # 30 curated manufacturing/CMC records
├── scripts/
│ ├── setup_collections.py # Create collections + seed FDA constructs
│ ├── ingest_pubmed.py # CLI: ingest PubMed CAR-T literature
│ ├── ingest_clinical_trials.py # CLI: ingest ClinicalTrials.gov trials
│ ├── seed_assays.py # Seed assay data from published papers
│ ├── seed_manufacturing.py # Seed manufacturing/CMC data
│ ├── validate_e2e.py # End-to-end data layer validation
│ ├── test_rag_pipeline.py # Full RAG + LLM integration test
│ └── seed_knowledge.py # Export knowledge graph to JSON
├── requirements.txt
└── LICENSE # Apache 2.0
55 Python files | ~16,748 lines | Apache 2.0
Knowledge Graph¶
| Component | Count | Examples |
|---|---|---|
| Target Antigens | 25 | CD19, BCMA, CD22, CD20, CD30, HER2, GPC3, EGFR, Mesothelin, GPRC5D, ... |
| FDA-Approved Products | 6 | Kymriah, Yescarta, Tecartus, Breyanzi, Abecma, Carvykti |
| Entity Aliases | 39+ | Product names, generic names, costimulatory domains (for comparative resolution) |
| Toxicity Profiles | 8 | CRS, ICANS, B-cell aplasia, HLH/MAS, cytopenias, TLS, GvHD, on-target/off-tumor |
| Manufacturing Processes | 10 | Transduction, expansion, leukapheresis, cryopreservation, release testing, ... |
| Biomarkers | 15 | CRS prediction, T-cell exhaustion, persistence, cytokine panels, ... |
| Regulatory Histories | 6 | Approval timelines, post-marketing requirements for all FDA products |
| Immunogenicity Topics | 6 | ADA, immunogenicity assays, risk factors, ... |
| Query Expansion Maps | 12 | Target Antigen, Disease, Toxicity, Manufacturing, Mechanism, Construct, Safety, Biomarker, Regulatory, Sequence, RealWorld, Immunogenicity |
| Expansion Keywords | 169 | Mapping to 1,496 related terms |
Performance¶
Measured on NVIDIA DGX Spark (GB10 GPU, 128GB unified memory):
| Metric | Value |
|---|---|
| PubMed ingest (5,047 abstracts) | ~15 min |
| ClinicalTrials.gov ingest (973 trials) | ~3 min |
| Assay seed ingest (45 records) | ~30 sec |
| Manufacturing seed ingest (30 records) | ~30 sec |
| Vector search (11 collections, top-5 each) | 12-16 ms (cached) |
| Comparative dual retrieval (2x11 collections) | ~365 ms |
| Full RAG query (search + Claude) | ~24 sec |
| Comparative RAG query (dual search + Claude) | ~30 sec |
| Cosine similarity scores | 0.74 - 0.90 |
Credits¶
- Adam Jones
- Apache 2.0 License