Curation
In vivo CAR-T delivery intelligence: candidate intake infrastructure, target-hypothesis context, technical evidence, and upgrade decisions.
Primary mission
ICL quality gate
for in vivo CAR-T
Target status
unresolved hypothesis space
BCMA / CD19 / CD20 / CD22 / dual
Intake base
submission-ready infrastructure
schema / evidence / QA
Upgrade target
Decision engine
not just scoring
Candidate Review Console
high confidenceCandidate intake
Neutral intake shell for any candidate. Partner/source names stay outside this front-end layer.
Formulation snapshot
Decision packet
Intake, ICL, verification and panel dissent in one view.
Run a candidate review to generate a decision packet.
Model Lifecycle Gate
inferenceFine-tune / retraining readiness
Blocks weight updates until labels, splits and metadata are strong enough for the requested strategy.
Activation decision
Policy minimums and current blockers.
Check a dataset before changing model weights.
Prior Local Work
high confidence| Asset | What it contains | Internalized use |
|---|---|---|
| competitor deck | 54-page in vivo CAR-T progress deck; duplicate copies de-weighted | Converted into benchmark context and source-dedup rules |
| prior analyst portal | Previous in vivo CAR-T analyst workflow, knowledge base, radar feed and internalization architecture | Absorbed into entity model, review roles and evidence workflow |
| knowledge base | Deals, programs, clinical readouts, canonical facts and post-cutoff updates | Converted into curated program/readout/deal entities |
| competitive radar | Long-form in vivo CAR-T radar reports and weekly competitive updates | Absorbed into watchlist and update cadence |
| clinical planning | Competitive landscapes, CDP logic and regulatory/trial-design reasoning | Used for clinical readout definitions and risk register |
| AI portal patterns | RAG ingestion, citation, taxonomy and eval-harness patterns | Reused for source registry, claim ledger and eval contracts |
| external narrative | Public pipeline and technology-platform narrative | Used only for sanitized external display language |
Target Hypothesis Context
high confidence| Program | Target | Decision status | What Curation must support |
|---|---|---|---|
| IASO206 | BCMA | Context, not final target call | Use MM / plasma-cell benchmarks to define CAR-mRNA assay expectations and safety bar |
| IASO207 | CD19 | Context, not final target call | Use CD19 as the densest LNP/LVV comparator space for delivery and repeat-dosing logic |
| IASO208 | CD20 | Context, not final target call | Use CD20/CD19-CD20 readouts to track depletion, CAR kinetics and antigen-escape questions |
| Dual / multi-target | CD19/CD20, CD19/CD22, BCMA/CD19 | Hypothesis space | Payload size, expression balance and efficacy/safety tradeoffs become central LNP constraints |
Target choice is intentionally unresolved here. Internal deck context and prior in vivo work have been absorbed into the evidence model; AI-positioning slides are excluded from evidence.
Competitive Benchmarks
Kelonia / Lilly
KLN-1010
BCMA · LVV
Current BCMA in vivo benchmark and safety/efficacy bar
EsoBiotec / AZ
ESO-T01
BCMA · LVV
Closest BCMA/MM comparator; watch CRS and MRD/CAR expansion details
Legend
LB2501
CD19/CD20 · LVV
China-origin dual-target clinical comparator
Capstan / AbbVie
CPTX2309
CD19 · tLNP + mRNA
Non-viral mRNA benchmark for targeted T-cell engineering
Orna / Lilly
ORN-252
CD19 · LNP + circRNA
circRNA durability comparator
Domestic LNP IITs
GT801 / WGb-0301 / others
mostly CD19 · LNP + mRNA
safety, repeat dosing and CAR kinetics reference
Upgrade Plan
inference| Layer | Build | Decision value |
|---|---|---|
| Source registry | source_id, as_of, confidence, confidentiality, citation | Stops model claims and facts from being mixed |
| Entity model | company, program, asset, target, modality, candidate, assay, readout, deal | Makes competitor and candidate evidence comparable |
| Candidate intake infrastructure | structure, batch, formulation, targeting ligand, payload, assay data, IP | Turns future submissions into reviewable decision packets |
| ICL gate | pKa, tail geometry, degradability, SA, safety, novelty, OOD | Keeps the first decision focused on lipid quality |
| Model lifecycle | calibration -> adapter -> ensemble retrain -> AGILE fine-tune | Upgrades predictions only when assay data supports it |
| Workflow harness | intake -> model outputs -> evidence -> claim ledger -> decision | Makes every review reproducible and replayable |
| Multi-agent panel | ICL chemist, formulation scientist, CAR-T analyst, clinical translator, BD/IP reviewer, verifier | Separates expert judgments and exposes dissent |
| Eval suite | target uncertainty, citation presence, OOD, ICL-only overclaim, benchmark caveats | Prevents judgment drift as the platform grows |
| Verification gates | citation, evidence class, target uncertainty, delivery overclaim, benchmark hygiene | Blocks final decisions when evidence hygiene fails |
Full planning detail is kept in the internal source registry; this page only exposes the operational upgrade surface.
Review Harness Status
high confidence| Endpoint | Capability | Gate |
|---|---|---|
| /api/v1/review/intake/validate | CandidateIntake schema and missing-data checklist | Blocks missing SMILES, source owner and structure provenance |
| /api/v1/review/icl-gate | Existing reverse smoke-test wrapped as ICL quality gate | Returns pass / watch / reject with grade, composite score and blockers |
| /api/v1/review/verify | Claim ledger verification | Checks citation, evidence class, target uncertainty, delivery overclaim and benchmark hygiene |
| /api/v1/review/run | Full review harness | Runs intake, ICL gate, static multi-agent panel and decision synthesis |
| /api/v1/review/eval/run | Regression eval suite | Tests target uncertainty, overclaim, delivery context, benchmark caveats and duplicate sources |
| /api/v1/model-lifecycle/retrain/readiness | Fine-tune / retrain readiness gate | Blocks small datasets, mixed endpoints and weak split strategies before model activation |
P1 backend started in pipeline/review; source-specific candidate names are intentionally excluded from this layer.
Internal Lipid Benchmark Signals
inference| Axis | Finding | Interpretation |
|---|---|---|
| Lead signal | Internal benchmark 08 | C18:1 oleyl tails, composite 0.683, B+, predicted pKa 6.59 |
| SAR axis | Tail length / unsaturation | C8 to C18:1 drives pKa down toward the ICL optimal window |
| Tropism tension | Liver-primary prediction | Internal SORT-style model predicts liver-primary behavior across 8/8 examples |
| Model caveat | Rigid aromatic OOD | AGILE/TransMA likely need scaffold-specific wet-lab data for transfection confidence |
| Strategic asset | Clean chemistry | Safety assessor finds no structural alerts across the eight examples |
Original source links and local file paths are intentionally hidden from the front-end; only verified, internalized technical signals are displayed here.
Literature Radar
In vivo CAR-T tLNP
Hunter et al., Science, 2025
CD8-targeted LNPs delivered anti-CD19 CAR mRNA to generate functional CAR T cells in vivo; this is the product-class benchmark.
LNP-mRNA foundation
Hou et al., Nature Reviews Materials, 2021
Canonical map of ionizable lipid, helper lipid, cholesterol, PEG-lipid, routes, barriers and manufacturing.
Organ tropism / SORT
Cheng et al., Nature Nanotechnology, 2020
Tissue delivery is tunable by SORT molecules and formulation composition; route and composition must be modeled explicitly.
AGILE design loop
Xu et al., Nature Communications, 2024
Deep learning plus combinatorial chemistry shortens ionizable lipid discovery and supports cell-line-specific optimization.
3D conformation route
Su et al., Nature Biomedical Engineering, 2026
Dynamic spatial conformation of ionizable lipids is linked to assembly, protein corona, organ targeting and AI screening.
TransMA
Wu et al., Briefings in Bioinformatics, 2025
Multi-modal 3D Transformer + Mamba model for transfection prediction; strong but still constrained by dataset coverage.
Formulation immunology
Vadovics et al., Nature Nanotechnology, 2025
PEG-lipid ratio and phospholipid identity can shift uptake, biodistribution and immune response quality.
Strategic Signals
IASO in vivo CAR-T focus
The Reindeer deck frames Fucaso globalization and in vivo CAR-T innovation as dual engines; LVV is near-term, while tLNP is a non-viral extension.
Internal tLNP capability
The deck lists IVT mRNA, LNP packaging, antibody conjugation, primary T-cell transfection, targeting assessment and in vitro killing as key capabilities.
PD-L1 mRNA LNP tolerance
Curated literature signals in vivo tolerogenic APC programming through PD-L1 mRNA LNPs, pointing beyond vaccine-only mRNA use cases.
JK / Keyuan excipient collaboration
Public reports emphasize cationic lipid, polymeric excipient, lyophilization and room-temperature stability as translation themes.
Internal lipid benchmark
Eight patent-style examples are curated locally with Validator, Safety and Tropism outputs; benchmark 08 is the current internal lead signal.
Platform Upgrade Decisions
Candidate intake infrastructure
Prepare neutral fields for source owner, structure provenance, batch, IP status, measured data and confidentiality boundary.
ICL Quality Gate
Pass/watch/reject an ICL before downstream formulation: pKa, lipophilicity, tail geometry, degradability, SA, safety and novelty.
OOD disclosure
Surface scaffold novelty and model-domain mismatch beside every AGILE/TransMA score.
Primary T-cell assay schema
Standardize CAR%, MFI, CD4/CD8 split, viability, activation, exhaustion, killing and cytokine readouts.
Target-hypothesis scoring
Treat BCMA, CD19, CD20, CD22 and dual CAR payloads as unresolved context, not a fixed target conclusion.
Experiment requests
Generate the smallest wet-lab data package needed to decide whether a candidate should advance.
Knowledge Framework
Candidate ICL
quality gate
CAR payload
target hypothesis
tLNP design
PEG / Ab / N:P
T-cell delivery
CD4 / CD8 / off-target
Clinical readout
CAR% / depletion / safety