Knowledge Map
Bilingual mental-model alignment — the 5-layer framework the platform organizes around
Layer 0Therapeutic format universe
What end-product are we making? Antibody-based therapeutics fan out into several formats, but every format depends on the same foundational artifact — a binder sequence that recognizes the target with the right affinity, specificity, and developability profile.
- IASO Bio's marketed product 福可苏 (equecabtagene autoleucel, CT103a) is a scFv-based lentivirally-delivered CAR-T.
- The platform's strategic direction extends toward VHH-based in-vivo CAR delivered as mRNA-LNP, plus keeping traditional mAb / TCE optionality open.
- CAR uses an antibody fragment for the recognition end; TCE uses two antibody fragments simultaneously (one for tumor, one for CD3); a traditional mAb deploys ADCC through its Fc.
| Format | Size budget | Affinity sweet spot | Hot-button |
|---|---|---|---|
| Traditional mAb | ~150 kDa | nM–pM | aggregation, glycosylation |
| scFv-CAR (lenti) | ~25 kDa (linker matters) | nM (not too tight — see Carvykti) | tonic clustering, surface expression |
| VHH-CAR (mRNA-LNP) | ~13 kDa (mRNA fits LNP) | nM | thermal stability, no light chain |
| TCE | dual scFv + linker | balanced both arms | half-life, on-target/off-tumor |
Layer 1Common substrate: binder sequences
Antibody sequences are the foundation of every product surface. The platform's atomic unit is a sequence (FASTA) + a structural hypothesis (PDB).
- Traditional antibody: VH (~110 aa) + CH1-CH2-CH3 + VL (~110 aa) + CL. Recognition = VH + VL forming a paired Fv.
- scFv: VH—linker (e.g. (GGGGS)₃)—VL or VL—linker—VH; the linker artificially tethers VH and VL into a single chain.
- VHH / nanobody: single domain (~120 aa, VHH only). No light chain — comes from camelid heavy-chain-only antibodies. Advantages: small (fits in mRNA-LNP), monomeric (no chain pairing fail), thermostable.
- CDRs vs framework: CDR1/2/3 are the core recognition segments; the framework primarily provides scaffolding. HCDR3 is the single most important loop for specificity, antigen contact, and designability.
- Numbering schemes: IMGT, Kabat, Chothia, Martin. The platform uses ANARCI for IMGT-style numbering.
| Dimension | What it captures | Where on platform |
|---|---|---|
| Binding affinity | Kd; proxied in silico by ipTM / iPAE / p_bind | /interface, score columns |
| Specificity | Doesn't bind off-targets / homologs | --counterscreen flag, /compare |
| Developability | Manufacturable: stable, soluble, low aggregation | TAP, ThermoMPNN, AggreScan3D columns |
| Immunogenicity | Won't trigger ADA in patients | BioPhi humanness |
| Format fit | Right size / chain composition for Layer-0 format | format-specific filters |
| CDR geometry | CDR lengths in canonical ranges | cdr_geometry_ok (PR #14) |
| Buried surface area | Adequate interface contact | bsa_total/antibody/antigen (PR #13) |
| Mutation robustness | Still binds antigen variants | not one-click yet — gap (U5) |
Layer 2AI design workflow (6 canonical steps)
The operational spine of de novo design and the basis for the platform-gap-analysis W/U/C/V matrix.
W1
Input antigen structure — PDB (e.g. 4ZFO for BCMA) or AlphaFold prediction; defined epitope strongly preferred.
W2
Define binding hotspot — Residue list (e.g. A:42,A:45,A:50) or contiguous patch. Most consequential single choice in the workflow.
W3
Backbone generation — RFdiffusion / RFantibody — diffusion-based sampling of antibody backbone conformations conditioned on antigen + hotspot. Output is geometry only.
W4
Sequence design — ProteinMPNN inverse-folding — given the backbone, what amino acid sequence reliably folds into it. AntiFold is the antibody-specific variant.
W5
In-silico filtering — AF-Multimer / RF2 for interface quality, BSA, CDR geometry, TAP developability, ThermoMPNN stability, AggreScan aggregation, BioPhi humanness, counter-screen specificity.
W6
Experimental validation — Outside platform. Yeast / phage display → SPR / BLI → specificity panel → functional assay (cytotoxicity for CAR-T; cell-based for TCE).
Important caveat: RFdiffusion output is not a final drug candidate; it gives structural hypotheses. Filters narrow the funnel; they don't replace experimental validation.
Layer 3Operating modes
How Step 3 + Step 4 are configured depends on what you already have on hand. The canonical BCMA workflow (regenerate antibody for BCMA using a wet-lab-validated sequence) operates in warm-start regenerate mode.
| Mode | What you have | What you do | Flag / source |
|---|---|---|---|
| Full de novo | Only target + epitope, no parent binder | All CDRs + framework diffused; entire candidate set is novel | (default) |
| CDR redesign (HCDR3-only) | A parent VHH framework you trust | Freeze framework, diffuse HCDR3 only | --mode partial_cdrs --cdr-regions hcdr3 (PR #7) |
| CDR redesign (all CDRs) | A parent framework you trust | Freeze framework, diffuse H1+H2+H3 (+L1+L2+L3 for scFv) | --mode partial_cdrs --cdr-regions all |
| Warm-start regenerate | A wet-lab-validated parent sequence | Partial diffusion + ProteinMPNN biased to parent | (BCMA workflow) |
| Sequence-only re-score | Existing candidate sequences | Skip Steps 1-3, directly score in Step 5 | /import → /interface |
Layer 4Use cases & quality filters
The original framework named four use cases + five CAR-T-specific filters. Subsequent internal review surfaced three more that aren't in the original framework but are real.
Use cases
| # | Use case | Example | Platform |
|---|---|---|---|
| U1 | Epitope-specific design | Force binding to CD20 extracellular loop, BCMA membrane-distal epitope, GPRC5D tumor-selective region | /epitope-picker curated library |
| U2 | CDR grafting / redesign | Keep antibody framework fixed, redesign HCDR3/LCDR loops | partial diffusion |
| U3 | De novo VHH or scFv | CAR binder discovery, smaller binders, non-natural epitope targeting | RFantibody full mode |
| U4 | Avoiding cross-reactivity | Design against unique surface patch, counter-screen against homologs | --counterscreen |
| U5 | Mutation-resistance prediction | Post-treatment BCMA mutation (e.g. pos 27); want CT103a still binds, competitors don't | manual workaround — gap |
| U6 | Competitor differentiation analysis | CT103a vs Cilta-cel: epitope overlap, structural rationale for clinical narrative | manual via /compare — gap |
| U7 | Sequence-input candidate ranking | Bench team has 10 candidates, which is best? | /import + /interface + /candidates |
CAR-T specific filters
| # | Filter | Note | Status |
|---|---|---|---|
| C1 | High tumor binding | Affinity within therapeutic window | ipTM/iPAE |
| C2 | Low tonic signaling (binder) | Binder shouldn't cluster CAR without antigen | heuristic v0 (PR #9) |
| C3 | Proper surface expression | Folds and traffics correctly | TAP-derived approx; dedicated model open |
| C4 | Low antigen-independent clustering | Relates to tonic signaling | folded into C2 |
| C5 | Activity under low antigen density | Functional CAR-T activity at low Ag | not modeled — Phase 2+ |
Methodology lineage
Generation 1 — Animal immunization → hybridoma → humanization
Bottleneck: 1-2 years per target; mouse repertoire limits
Generation 2 — In vitro display (phage / yeast / mammalian / ribosome)
Bottleneck: Library size limited; epitope-specific design hard
Generation 3 — ML-augmented (sequence priors + structure prediction)
Bottleneck: VAE, transformer, IgLM, AntiBERTy — still needs experimental closing loop
Generation 4 (we are here) — Generative AI de novo design (RFdiffusion + ProteinMPNN + AF-Multimer/RF2)
Bottleneck: Validation rate still <10% on hard targets; framework not yet calibrated against IASO-internal data
Bilingual glossary(43 entries)
| EN | 中文 / CN | Note |
|---|---|---|
| Antibody / antibody fragment | 抗体 / 抗体片段 | Generic; the platform's universe |
| Antigen | 抗原 | The target protein, e.g. BCMA |
| Epitope | 表位 | The specific surface patch the antibody binds |
| Paratope | 互补位 | The antibody's surface that touches the epitope |
| Affinity (Kd) | 亲和力 | Equilibrium dissociation constant; lower = tighter |
| ADCC | 抗体依赖性细胞毒性 | Fc-mediated cell killing via NK |
| CAR | 嵌合抗原受体 | Chimeric antigen receptor; antibody fragment + signaling domain |
| scFv | 单链抗体 | VH-linker-VL or VL-linker-VH, single chain |
| VHH / nanobody | 纳米抗体 / VHH | Single-domain antibody from camelid heavy-chain-only Ig |
| TCE / T-cell engager | T 细胞拉合剂 | Bispecific binding tumor + CD3 |
| CDR (1 / 2 / 3) | 互补决定区 | Three hyper-variable loops per chain; CDR-H3 is most variable |
| Framework region (FR) | 框架区 | Conserved scaffold around CDRs |
| ANARCI / IMGT numbering | ANARCI / IMGT 编号 | Sequence numbering assigning CDR boundaries |
| De novo design | 从头设计 | Designing a binder without a parent sequence |
| Partial diffusion | 部分扩散 | RFdiffusion mode: freeze framework, diffuse only specified CDRs |
| Warm-start regenerate | 序列锚定重新生成 | Validated parent as anchor + partial diffusion to explore variants |
| Backbone / scaffold | 骨架 / 支架 | The 3D geometry without amino acid identities |
| Inverse folding | 反向折叠 | Given backbone, predict sequence (ProteinMPNN does this) |
| ipTM | (no canonical CN) | AF-Multimer interface confidence; higher = more confident complex |
| iPAE | (no canonical CN) | AF-Multimer interface predicted aligned error; lower = better |
| pLDDT | (no canonical CN) | Per-residue confidence; >0.7 acceptable, >0.85 excellent |
| p_bind | (no canonical CN) | RF2 binding probability proxy |
| Buried surface area (BSA) | 埋藏表面积 | Surface area buried at the binding interface |
| Developability | 可开发性 | Manufacturable: stable, soluble, low aggregation |
| TAP score | TAP 评分 | Therapeutic Antibody Profiler; gates for clinical antibodies |
| ThermoMPNN ΔΔG | ThermoMPNN ΔΔG | Stability change from mutation; negative = stabilizing |
| Aggregation propensity | 聚集倾向 | Tendency to form non-functional aggregates |
| Humanness (OASis) | 人源化程度 | Sequence similarity to natural human repertoire |
| Tonic signaling | 紧张性信号 | CAR-T self-activation without antigen — bad |
| Antigen-independent clustering | 抗原非依赖聚集 | CARs clustering on T-cell surface w/o antigen — drives tonic |
| Cross-reactivity | 交叉反应 | Off-target binding — usually bad |
| Specificity | 特异性 | On-target only — good |
| Counter-screen | 反向筛选 | Screen out candidates that bind homologs / off-targets |
| Mutation resistance | 突变抗性 | Binder still binds when antigen mutates at known positions |
| Yeast / phage display | 酵母 / 噬菌体展示 | In vitro library screening platforms |
| SPR / BLI | (no canonical CN) | Surface plasmon resonance / bio-layer interferometry — Kd measurement |
| ADA | 抗药抗体 | Anti-drug antibody — immunogenicity readout |
| Lentivirus / lenti | 慢病毒 | Delivers CAR to ex-vivo T cells (CT103a uses this) |
| mRNA-LNP | mRNA 脂质纳米颗粒 | The delivery modality for in-vivo CAR |
| MHC-I / MHC-II | MHC I / MHC II | Antigen-presenting molecules; relevant for TCR-mimetic Ab + neoantigen |
| Neoantigen | 新抗原 | Mutation-derived antigen unique to the tumor |
| BCMA / TNFRSF17 | BCMA / TNFRSF17 | B-cell maturation antigen; IASO + Carvykti + Abecma target |
| CT103a / 福可苏 | CT103a / 福可苏 | IASO's marketed BCMA CAR-T (equecabtagene autoleucel) |