---
title: Cell & Gene Therapy on the platform
description: >-
  Computational tools for CAR-T design, construct assembly, NK-CAR design, KIR
  matching, and related scoring workflows.
tags:
  - cell-therapy
---

> **Some tools on this page are retired**

> This page documents working tools alongside tools that have been retired. The retired ones can no longer be run:
>
> - `cart_manufacturing_score` — use cart_construct_build for the construct composition; the manufacturing score was withdrawn 2026-10-04 and nothing replaces it
> - `cart_safety_predict` — use cart_construct_build / cart_scfv_design for construct composition; the CRS and neurotoxicity predictions were withdrawn 2026-10-04 and nothing replaces them
> - `nk_cytotoxicity_predict` — use nk_car_design / nk_kir_match; the NK cytotoxicity prediction was withdrawn 2026-10-04 and nothing replaces it
>
> Everything else on this page still works.

Design and evaluate CAR-T and NK cell therapies end to end — from antigen binder design through construct assembly, safety and manufacturing assessment, and NK-specific receptor and donor-matching analysis. Reach for these tools when your program is a cell therapy rather than a small molecule, and you need computational feasibility scores before committing to wet-lab work. They sit in the translational, experimental-biology arc of discovery, feeding downstream into safety and clinical-trial matching. See the [Cell & Tissue overview](/docs/cell-tissue) for where this fits.

## Cell & Gene Therapy overview

The Cell & Gene Therapy toolkit provides seven complementary tools spanning the CAR-T and NK cell therapy design pipeline. Tools cover antigen binder design, construct engineering, safety assessment, manufacturing evaluation, and NK cell biology. All tools run on hosted GPU compute and return results in 30-180 seconds.

> **Three tools currently always fail**

> `cart_safety_predict`, `cart_manufacturing_score`, and `nk_cytotoxicity_predict` return an error on every call. Each previously scored its output from a hand-tuned sequence-descriptor heuristic with no trained model or clinical/manufacturing/cytotoxicity dataset behind it, and the backend now refuses to return that as a real prediction rather than ship a fabricated number. Use them only to see the refusal message; the four other tools (`cart_scfv_design`, `cart_construct_build`, `nk_car_design`, `nk_kir_match`) run and return real results.

## Architecture

```
User Query (target antigen, construct config, HLA types, tumor profile)
  --> Intent Detection
       |- CAR-T design --> ScFv Design or Construct Build
       |- Safety assessment --> Safety Predict or Manufacturing Score
       |- NK cell therapy --> NK-CAR Design, KIR Match, or Cytotoxicity Predict
  --> Hosted GPU Compute (30-180 seconds)
  --> Structured Results with Feasibility Scores
```

## Tool inventory

### CAR-T design tools

| Tool | Name | Description | Key Output |
|------|------|-------------|------------|
| CAR-T ScFv Design | `cart_scfv_design` | Generates ScFv candidates with optimized CDR regions for target binding | Ranked ScFv sequences, binding affinity predictions, humanization scores |
| CAR Construct Build | `cart_construct_build` | Assembles full CAR constructs from ScFv, hinge, transmembrane, and signaling domains | Complete CAR sequence, domain annotations, codon-optimized DNA |

### CAR-T evaluation tools

| Tool | Name | Description | Key Output |
|------|------|-------------|------------|
| CAR-T Safety Predict ⚠︎ always fails | `cart_safety_predict` | Refuses to run — no trained model backs a CRS/ICANS prediction | Error only |
| CAR-T Manufacturing Score ⚠︎ always fails | `cart_manufacturing_score` | Refuses to run — no trained model backs a manufacturing-feasibility prediction | Error only |

### NK cell therapy tools

| Tool | Name | Description | Key Output |
|------|------|-------------|------------|
| NK-CAR Design | `nk_car_design` | Designs NK-optimized CAR receptors with NK-specific signaling | NK-CAR sequences, signaling configurations, activation predictions |
| NK KIR Match | `nk_kir_match` | Scores one HLA allele set against a curated KIR-HLA reference table | Per-KIR-gene match type and alloreactivity, overall score |
| NK Cytotoxicity Predict ⚠︎ always fails | `nk_cytotoxicity_predict` | Refuses to run — no trained model or assay backs a lysis/EC50 prediction | Error only |

## Capabilities

### ScFv candidate design

The ScFv design tool generates single-chain variable fragment candidates for a target antigen (sequence or UniProt ID) using deep learning. Optionally seed it with a starting `vh_sequence`/`vl_sequence` to optimize an existing lead instead of designing de novo. Choose the linker with `linker_type` (`G4S`, `flexible`, or `rigid`) and how many candidates to generate with `num_designs` (default 10).

### CAR construct assembly

The construct builder assembles a complete chimeric antigen receptor sequence from an scFv plus modular components: hinge (`CD8`, `IgG4`, or `CD28`), transmembrane domain (`CD28` or `CD8`), a costimulatory domain (`41BB`, `CD28`, `OX40`, or `ICOS`), and the CD3-zeta signaling domain. `generation` selects `2`, `3`, or `4` — there's no first-generation (CD3-zeta-only) option.

### Safety and toxicity prediction — always fails

`cart_safety_predict` takes a full CAR `construct_sequence` and `target_antigen`, but every call returns an error rather than a CRS/neurotoxicity risk level: the previous scores came from an instability-index/GRAVY heuristic with no clinical dataset behind them, and the backend now refuses to ship that as a prediction.

### Manufacturing feasibility scoring — always fails

`cart_manufacturing_score` takes a CAR `construct_sequence` and a delivery `vector_type` (`lentiviral`, `retroviral`, or `transposon`; default lentiviral), but every call returns an error rather than a feasibility score: the previous scores came from sequence-length/instability formulas with no manufacturing dataset behind them.

### NK-CAR receptor design

NK-CAR design creates a CAR receptor for a `target_antigen` using NK-specific components: an `activating_receptor` (`NKG2D`, `DNAM1`, `NKp30`, or `NKp44`; default NKG2D), a transmembrane domain (`NKG2D`, `DAP10`, or `CD28`), and an intracellular `signaling_domain` (`DAP10`, `DAP12`, or `CD3z_2B4`; default DAP10).

### KIR-HLA scoring

`nk_kir_match` scores a single set of HLA alleles (`hla_alleles`) against a curated KIR-HLA reference table, returning per-KIR-gene match type and predicted alloreactivity plus an overall `alloreactivity_score`. It does not compare a donor genotype against a recipient genotype — the tool interface also accepts `donor_kir_genotype`, `recipient_hla` (as gene/allele flag objects), and `analysis_type`, but the executing backend reads none of them; only `hla_alleles` affects the result.

### NK cytotoxicity prediction — always fails

`nk_cytotoxicity_predict` takes the target cell's HLA expression (`target_cell_hla`) and optional stress ligands (`target_ligands`) and NK receptor profile (`nk_receptors`), but every call returns an error rather than a lysis prediction: the previous numbers came from a Hill-equation-shaped heuristic with no cytotoxicity assay behind them.

## When to use each tool

| I want to... | Use this tool |
|--------------|---------------|
| Design a new CAR-T targeting a specific antigen | `cart_scfv_design` |
| Assemble a complete CAR construct | `cart_construct_build` |
| Assess CRS and neurotoxicity risk | `cart_safety_predict` — currently always returns an error |
| Evaluate if a construct can be manufactured | `cart_manufacturing_score` — currently always returns an error |
| Design a CAR for NK cells | `nk_car_design` |
| Score KIR-HLA alloreactivity for one allele set | `nk_kir_match` |
| Predict how well NK cells will kill a tumor | `nk_cytotoxicity_predict` — currently always returns an error |
| Run the working part of the CAR-T design pipeline | `cart_scfv_design` + `cart_construct_build` |

## When to use other tools instead

| I want to... | Use this instead |
|--------------|------------------|
| Validate target antigen expression | OpenTargets associations or other [Genomics](/docs/genomics) lookups |
| Predict 3D structure of ScFv | [ESMFold](/docs/esmfold), or [AlphaFold DB](/docs/alphafold) for a known protein |
| Dock ScFv to target antigen | [DiffDock](/docs/diffdock) |
| Search for antibody sequences | [UniProt](/docs/uniprot) or the [antibody](/docs/antibodies) databases |
| Assess small molecule drug safety | [ADMETlab 3](/docs/admetlab3) or a Safety Passport |
| Search for cell therapy clinical trials | [ClinicalTrials.gov](/docs/clinical-trials) |
| Analyze protein-protein interactions | A STRING interaction lookup |

## CAR architecture reference

`cart_construct_build`'s `generation` parameter accepts `2`, `3`, or `4` (default `2`) and pairs with a single `costimulatory` domain choice (`41BB`, `CD28`, `OX40`, or `ICOS`) plus the fixed CD3-zeta signaling domain. There's no first-generation (CD3-zeta-only, no costimulatory domain) option in this tool.

## Citations

> June CH, et al. CAR T cell immunotherapy for human cancer. *Science*. 2018;359(6382):1361-1365. doi: [10.1126/science.aar6711](https://doi.org/10.1126/science.aar6711)

> Liu E, et al. Use of CAR-Transduced Natural Killer Cells in CD19-Positive Lymphoid Tumors. *N Engl J Med*. 2020;382(6):545-553. doi: [10.1056/NEJMoa1910607](https://doi.org/10.1056/NEJMoa1910607)

> Ruggeri L, et al. Effectiveness of donor natural killer cell alloreactivity in mismatched hematopoietic transplants. *Science*. 2002;295(5562):2097-2100. doi: [10.1126/science.1068440](https://doi.org/10.1126/science.1068440)

## Chat Usage

### Example Prompts

- "Design ScFv candidates for targeting BCMA on multiple myeloma cells"
- "Build a generation-2 4-1BB/CD3-zeta CAR construct using this ScFv sequence"
- "Design an NK-CAR with DAP12 signaling for targeting GPC3"
- "Find the best NK cell donor match for this patient's HLA type"

### Workflow Chaining

- **Feeds into**: Safety Passport (toxicity assessment), Compound Passport (combination therapy), Clinical Trial matching, ESMFold/AlphaFold (ScFv structure prediction)
- **Accepts from**: Genomics tools (target gene data), UniProt (protein sequences), OpenTargets (target validation), Literature search (clinical evidence), Antibody database tools (existing binder sequences)

## FAQ

### What target antigens are supported for CAR-T ScFv design?

  The ScFv design tool accepts any target antigen specified by protein name, UniProt ID, or amino acid sequence. Common targets include CD19, BCMA, CD22, HER2, EGFR, and mesothelin. The tool designs ScFv candidates by predicting variable heavy (VH) and variable light (VL) domains with optimized CDR regions.

### What signaling domains can I use in CAR constructs?

  The construct builder's costimulatory domain (`costimulatory`) is a single choice of 4-1BB (`41BB`), CD28, OX40, or ICOS, paired with the fixed CD3-zeta activation domain. `generation` selects `2`, `3`, or `4` — there's no first-generation, CD3-zeta-only option in this tool.

### How does the safety prediction tool assess CRS risk?

  It doesn't currently — `cart_safety_predict` always returns an error. The previous CRS/neurotoxicity scores came from a sequence-descriptor heuristic with no clinical dataset behind them, and the backend now refuses to return that as a real prediction.

### What is KIR-HLA matching and why does it matter for NK cells?

  KIR (Killer Immunoglobulin-like Receptor) genes on NK cells interact with HLA class I molecules on target cells. When a cognate HLA ligand is missing, the inhibitory KIR isn't engaged and NK cells become more alloreactive — the basis for KIR-HLA mismatch strategies. `nk_kir_match` scores one HLA allele set against a curated KIR-HLA reference table; it doesn't compare two people's genotypes or rank candidate donors in a single call.

## Read next
