Disclaimer: All datasets in this repository are simulated or pseudodata generated solely for methodological demonstration purposes. No proprietary, confidential, patient-derived, or employer-affiliated data is included. This work represents independent research and educational development conducted outside of any employment context and does not reflect the proprietary methods, data, or intellectual property of any employer or collaborator.
This repository is released under the MIT License. © 2026 Bo Ma (tjmb03). Reuse with attribution.
Multi-analyte ADC PK/PD and exposure-driven dose selection — the toxicology paradigm. An antibody-drug conjugate's therapeutic index is a two-analyte problem: efficacy is driven by conjugate exposure at the tumor, dose-limiting toxicity by free payload systemically. This repo implements the multi-analyte PK (lumped and DAR-resolved), a conjugate-driven tumor PK/PD, and a dose-selection engine that places the first-in-human window and OBD from exposure–response — the way you do it for a cytotoxic ADC, which is toxicology-anchored, not MABEL.
Companion to bispecific-fih-dosability: the same
use-mechanism-to-select-molecules theme, and the explicit contrast in FIH logic — MABEL / occupancy
for an immune agonist, toxicology / NOAEL for a cytotoxic ADC.
ADC PK is never one curve. At least three species move together, and you need all three because efficacy and safety live on different ones:
- conjugate — antibody still carrying ≥1 payload; delivers payload to the tumor → efficacy;
- total antibody — conjugated + fully deconjugated; the long-lived backbone;
- released payload — free cytotoxin on its own fast clearance → dose-limiting toxicity.
Deconjugation sheds payload, so the average DAR falls and the conjugate curve pulls away from total antibody; the released payload is a low, formation-rate-limited peak that trails the conjugate:
You can model the conjugate as one pool with an average DAR (lumped) or resolve the full
DAR0…DAR_max cascade (DAR-resolved). Which should you fit? The model settles it
(pk.py, pinned in the tests):
Under DAR-independent clearance, the total conjugated-payload PK of the DAR-resolved model is identical to the lumped model's
conjugate × DAR— both decay asexp(-(ke + k_deconj)·t)and the average DAR falls asDAR0·exp(-k_deconj·t). The two diverge only when clearance varies with DAR.
So the DAR-resolved model is only identifiable with DAR-distribution data (or demonstrably DAR-dependent clearance). Absent that, the extra compartments are unsupported and the lumped model is the honest choice — a modeling-vs-bioanalytics decision, not a matter of taste. (In the repo: agreement to ~1e-8 under DAR-independent CL; ~9% divergence once clearance rises with DAR.)
Conjugate exposure drives a tumor-growth-inhibition model (pd.py). The teaching point
is the low-dose regrowth between doses: schedule, not just cumulative dose, matters — so you model the
time course, not a single AUC.
For a cytotoxic ADC the DLT is off-target free payload, so the FIH window is built from exposure–response,
not occupancy (doseselect.py):
- start dose — toxicology-anchored: an animal HNSTD (or STD10) → human-equivalent dose ÷ a safety factor (default 6; ICH S9 HNSTD approach);
- floor (MED) — the conjugate exposure–response (minimum efficacious dose);
- ceiling (MTD) — the free-payload exposure–response (maximum tolerated dose);
- OBD / RP2D — chosen inside the window by exposure–response (Project Optimus), not at the MTD.
Conjugate target occupancy sits on the efficacy side — it sets the active dose, not the start dose. (On-target/off-tumor binding — e.g. CEACAM5 on normal GI — is the one place conjugate occupancy re-enters as a safety input; out of scope here.) This is the mirror image of the bispecific's occupancy corridor, where engagement itself is the hazard and MABEL governs.
Six ADCs (data/adc_candidates.csv), each perturbing one lever, scored for
window and OBD and ranked by therapeutic index (examples/dose_selection.py):
| ADC | ED50 / f_payload / TD50 | MED | MTD | OBD | TI | verdict |
|---|---|---|---|---|---|---|
| ADC-04 | 0.3 / 0.05 / 0.5 | 0.30 | 6.55 | 1.30 | 21.8 | GO — high antigen + potent ⇒ widest index |
| ADC-01 | 1.0 / 0.05 / 0.5 | 1.00 | 6.55 | 4.33 | 6.5 | GO — baseline, OBD below MTD |
| ADC-05 | 1.0 / 0.05 / 0.2 | 1.00 | 2.62 | 2.62 | 2.6 | GO* — toxic payload, tox-limited (OBD = MTD) |
| ADC-02 | 1.0 / 0.15 / 0.5 | 1.00 | 2.18 | 2.18 | 2.2 | GO* — unstable linker drops the ceiling |
| ADC-03 | 3.0 / 0.05 / 0.5 | 3.00 | 6.55 | 6.55 | 2.2 | GO* — low potency raises the floor |
| ADC-06 | 5.0 / 0.15 / 0.3 | 5.00 | 1.31 | — | 0.26 | NO-GO — MED above MTD, no window |
Three mechanistic reads the screen makes explicit:
- ADC-01 vs ADC-02 — identical but for linker stability (
f_payload0.05 → 0.15). More free payload per dose drops the MTD 6.55 → 2.18 and the index 6.5 → 2.2. Linker stability is a therapeutic-index lever, through the payload ceiling. - ADC-01 vs ADC-03 — lower potency / antigen (ED50 1 → 3) raises the MED, shrinking the window from the floor while the ceiling is unchanged.
- ADC-05 — a more toxic payload (TD50 0.5 → 0.2) drops the MTD so the window is tox-limited: the OBD is pinned at the MTD, you cannot reach the efficacy plateau safely. ADC-06 stacks low potency and an unstable linker until the floor rises above the ceiling — no window at all.
pip install -e . # or: pip install -e ".[dev]" for the tests
python -m pytest -q # 13 tests: multi-analyte PK, the identifiability result, the decision rule
python -m adcti.figures # regenerate every figure in figures/ from the model
python examples/multianalyte_pk.py # PK summary + the lumped-vs-DAR identifiability check
python examples/dose_selection.py # the ranked panel + rebuild adc_panel.pngProgrammatic use:
from adcti import select_dose
r = select_dose("my-adc", ED50=1.0, f_payload=0.05, TD50=0.5, hnstd_hed=6.0)
print(r.dosable, round(r.therapeutic_index, 1), r.verdict)
# True 6.5 GO: wide therapeutic index- First-order compartmental PK; single average-DAR release. A full model would add DAR-resolved clearance from data, tumor disposition, and payload distribution kinetics.
- Dose-selection uses the exposure–response reduction (efficacy ∝ conjugate/dose, DLT ∝ free-payload
exposure =
dose × f_payload); the multi-analyte ODE PK is the mechanism behind that reduction, shown separately. - Toxicology thresholds are placeholders. The HNSTD, safety factor, efficacy target, and DLT rate stand in for a program-specific analysis, not regulatory values.
- Soluble-target sink not included. For a shed antigen (CEACAM5 → serum CEA), a soluble-target TMDD term buffers the conjugate and belongs in the PK — a natural extension, omitted here for clarity.
- Parameters are representative, not measured. The panel is synthetic and exists to exercise the levers (potency/antigen, linker stability, payload toxicity).
src/adcti/
pk.py # multi-analyte PK: lumped + DAR-resolved; the identifiability result
pd.py # conjugate-driven tumor growth inhibition
doseselect.py # exposure-response window (HED start, MED, MTD, OBD) + panel screen
figures.py # regenerates all figures from the model
data/adc_candidates.csv
examples/ # multianalyte_pk.py, dose_selection.py
figures/ # generated figures + the two ADC schematics
tests/ # PK behaviour, the identifiability result, the decision rule
Three repos on one theme — model-informed molecule selection, using mechanistic PK/PD to choose between molecular designs before any clinical data exists:
- bispecific-fih-dosability — the first-in-human question for an immune-agonist bispecific, where engagement itself is the hazard and MABEL/occupancy governs — the mirror image of this repo's toxicology paradigm.
- ocular-tmdd-format-selection — format selection for a posterior-segment target: naked peptide vs Fab vs Fc-fusion on integrated target coverage, via a four-compartment intravitreal TMDD model with FcRn recycling.
MIT © 2026 Bo Ma