@pikaa-ai/pikaa 0.3.1 → 0.3.4

package/skills/pkpd-modeling/references/source-ledger.md

@@ -1,89 +0,0 @@
-# Source ledger
-
-Provenance for the version- and date-specific claims in this skill. Everything below was checked on
-**2026-07-27**. Pharmacological and statistical methods that are stable textbook material are not
-listed; this covers the claims that go stale.
-
-## Verified by direct execution
-
-| Claim | How verified |
-| --- | --- |
-| Pharmpy 2.1.1 is the current release | `pip install pharmpy-core` into a clean environment; `pharmpy.__version__` returned `2.1.1` |
-| Pharmpy exposes 19 `run_*` tools, including `run_pdsearch`, `run_modelrank`, `run_qa`, `run_vpc` | `dir(pharmpy.tools)` on the installed package |
-| `modeling.add_placebo_model` no longer exists; `set_placebo_model` does | `hasattr` check on the installed `pharmpy.modeling` |
-| `set_tmdd` accepts `full`, `ib`, `cr`, `crib`, `qss`, `wagner`, `mmapp` | `inspect.signature(pharmpy.modeling.set_tmdd)` |
-| `add_indirect_effect` accepts `linear`, `emax`, `sigmoid` and a `prod` flag | `inspect.signature` on the installed package |
-| Bioequivalence sample sizes (2×2, GMR 0.95, 80% power): 12/20/28/40/52/66 at CV 15/20/25/30/35/40% | `bioequivalence.py --power` reproduces the published PowerTOST table exactly |
-| ABEL limits cap at 69.84–143.19% at CVwR = 50% | Computed from `exp(±0.760 · swR)` at the cap |
-| NCA recovers λz, t½, CL/F and Vz/F from an analytical one-compartment oral profile | `nca.py` against a simulated noiseless profile with known parameters |
-| The model library reproduces closed-form AUC = D/CL, Vss = ΣV, MRT = Vss/CL for 1-, 2- and 3-compartment models | Analytical checks in `tests/pkpd-modeling/test_scripts.py` |
-
-## Package versions from PyPI
-
-Retrieved from the PyPI JSON API on 2026-07-27.
-
-| Package | Version | Note |
-| --- | --- | --- |
-| `pharmpy-core` | 2.1.1 (2026-05-19) | 2.1.0 on 2026-05-08; 2.0.0 on 2026-02-12; 1.12.0 on 2025-12-05 |
-| `numpy` | 2.5.1 | |
-| `scipy` | 1.18.0 | `requires_python >= 3.12` |
-| `chi-drm` | 1.0.3 | |
-| `pints` | 0.6.1 | |
-| `lmfit` | 1.3.4 | |
-| `pkpy` | **not on PyPI** | Queried and returned no such project; PKPy is GitHub-only |
-
-## Regulatory status
-
-| Claim | Source |
-| --- | --- |
-| ICH M12 Step 4 2024; FDA adopted 2 August 2024 with a Q&A; EU effective 30 November 2024; China 29 October 2024 | FDA M12 final-guidance announcement and materials; EMA M12 scientific guideline page |
-| ICH M13A Step 4 July 2024, effective 25 January 2025 | ICH M13A Step 4 materials; EMA M13A scientific guideline page |
-| ICH M13B endorsed 13 March 2025, Step 2b, consultation 9 April – 9 July 2025 | EMA M13B scientific guideline page |
-| ICH M13C covers highly variable drugs, NTI drugs and complex designs, and follows M13B | ICH M13 workplan as described in the M13A/M13B materials |
-| ICH E11A Step 4 21 August 2024, effective 25 January 2025 | ICH E11A Step 4 guideline; EMA E11A Step 5 document |
-| FDA Population Pharmacokinetics guidance final February 2022; states model selection based on shrinkage is not necessary; lists VPC, pcVPC, NPC, NPDE | FDA guidance document and Federal Register notice of availability, 4 February 2022 |
-| FDA oncology dose-optimisation guidance (Project Optimus) finalised August 2024; PK sampling and analysis plan in each protocol sufficient for popPK and exposure-response | FDA final guidance and contemporaneous coverage |
-| ICH E14/S7B 2022 Q&As introduced the double-negative nonclinical assessment | E14/S7B Q&A document; FDA review-experience analysis published 2025 |
-| FDA 2005 maximum-safe-starting-dose guidance supplies the Km body-surface-area conversion table | FDA guidance, Table 1 |
-
-## Software versions from vendor and project sources
-
-| Claim | Source |
-| --- | --- |
-| NONMEM 7.6 exists; adds ADVAN16 (RADAR5 stiff DDE), ADVAN17 (stiff delay DAE), NUTS Bayesian, SAEM sample storage | ICON NONMEM 7.6 workshop description and NONMEM 7.6.0 user guides dated November 2025 |
-| nlmixr2 requires rxode2 ≥ 5.0.0; CRAN package updated 30 November 2025, manual dated 9 May 2026 | CRAN nlmixr2 package page and reference manual |
-| `babelmixr2` and `monolix2rx` interchange models between nlmixr2, NONMEM and Monolix | nlmixr2 project documentation and a February 2026 methods paper |
-| Open Systems Pharmacology Suite v12 (with Update 2) is current; `ospsuite` R package needs R 4.x and .NET 8 | OSP Suite GitHub releases and v12 documentation |
-| PKPy is a 2025 Python popPK framework | PeerJ 2025 paper and its GitHub repository |
-| PKNCA and aNCA are the R NCA packages; aNCA is a Roche/Appsilon/Human Predictions pharmaverse project | Package documentation sites |
-
-## Method sources
-
-Standard methods used in the scripts, cited so the implementation can be checked:
-
-- **Lambda_z by best adjusted r-squared**, extend-backwards with a 0.0001 improvement threshold —
-  the convention implemented in Phoenix WinNonlin and PKNCA.
-- **Linear-up/log-down trapezoid** and the corresponding AUMC formula — standard NCA texts; the
-  AUMC log-down form is re-derived in the `nca.py` source comment.
-- **Hyslop's linearised upper bound** for the FDA reference-scaled criterion, with
-  `theta = ln(1.25)/0.25` — the method in FDA's progesterone product-specific guidance.
-- **ABEL** widening `exp(±0.760·swR)` capped at CVwR 50% — EMA bioequivalence guideline.
-- **Beal's M1–M7** methods for BLQ data — Beal, *J Pharmacokinet Pharmacodyn* 2001.
-- **Dayneka & Jusko** indirect response models I–IV.
-- **Sheiner** effect-compartment link model.
-- **Mager & Jusko** TMDD; **Gibiansky** QSS approximation.
-- **Anderson & Holford** sigmoidal maturation on post-menstrual age (TM50 54.2 weeks, Hill 3.92 as
-  generic clearance values).
-- **Mahmood & Balian** rule of exponents for interspecies scaling.
-- **Savic** transit-compartment absorption model.
-- **2020 ASHP/IDSA/PIDS/SIDP consensus** for vancomycin AUC₂₄/MIC 400–600.
-
-## Known gaps and deliberate omissions
-
-- USP and ISO documents are copyrighted and paywalled; where relevant they are cited by designation
-  only, never transcribed.
-- Product-specific guidances change frequently and are not enumerated here.
-- The vancomycin population parameters in `tdm_bayes.py` are a conventional illustrative
-  parameterisation, labelled as such in the code and output, not a validated published model.
-- PK/PD index targets for antimicrobials vary by organism, endpoint and study; the values given are
-  commonly cited ranges, not regulation.

package/skills/pkpd-modeling/references/special-populations.md

@@ -1,126 +0,0 @@
-# Special populations
-
-## Paediatrics
-
-### Size and maturation are separate
-
-Body size alone explains paediatric clearance well from roughly 2 years upward. Below that, enzyme
-and renal maturation dominate, and size-only scaling **overpredicts clearance — in a neonate by
-several fold**.
-
-```
-CL_child = CL_adult * (WT/70)^0.75 * MF
-MF = PMA^Hill / (TM50^Hill + PMA^Hill)              Anderson & Holford
-```
-
-Generic clearance values: `TM50 ≈ 54.2 weeks` post-menstrual age, `Hill ≈ 3.92`. Drug-specific
-ontogeny is much better where it exists, because individual enzymes mature on very different
-schedules.
-
-**Use post-menstrual age (gestational + postnatal), not postnatal age.** A 4-week-old born at 28
-weeks and a 4-week-old born at term have very different eliminating capacity.
-
-Enzyme ontogeny, in outline:
-
-| Enzyme | Maturation |
-| --- | --- |
-| CYP3A7 | High at birth, declines over the first year |
-| CYP3A4 | Low at birth, adult levels by ~1 year |
-| CYP2D6 | Reaches adult activity within weeks; genotype dominates thereafter |
-| CYP1A2 | Slow; adult levels around 4-5 months, and caffeine clearance in neonates is very low |
-| UGT2B7, UGT1A1 | Slow; morphine and bilirubin conjugation are limited in neonates |
-| Renal (GFR) | ~30% of adult (per surface area) at term birth; adult by 6-12 months |
-
-### ICH E11A pediatric extrapolation
-
-Step 4 adopted **21 August 2024**, effective **25 January 2025**. It formalises a framework for
-using adult (or other-population) data to support paediatric conclusions:
-
-- Build a **pediatric extrapolation concept** from the similarity of disease, response to
-  treatment, and exposure-response between the source and target populations.
-- Quantify the assumptions and the residual uncertainty; the amount of new paediatric data required
-  scales inversely with confidence in the extrapolation.
-- Where exposure matching is the basis, the standard applies the 90% CI to 80-125% bounds for AUC
-  and Cmax — but a model-informed approach using dose-response or exposure-response parameters
-  (Emax, EC50, slope) within acceptable limits is an accepted alternative.
-- Modelling and simulation, including popPK and PBPK, are central rather than supportive.
-
-The practical consequence: paediatric dose selection is expected to be model-informed, with a
-prospective plan, not a mg/kg extrapolation from the adult label.
-
-### Other paediatric points
-
-- Volume of distribution per kg is **higher** in neonates (greater total body water), so a loading
-  dose per kg is often larger while maintenance is smaller.
-- Protein binding is lower in neonates (less albumin, less alpha-1-acid glycoprotein, and
-  competition from bilirubin), raising the unbound fraction.
-- Oral absorption differs: higher gastric pH, slower gastric emptying, immature biliary function.
-
-## Renal impairment
-
-Classified by eGFR (mL/min/1.73 m²): normal ≥ 90, mild 60-89, moderate 30-59, severe 15-29, kidney
-failure < 15.
-
-- The relevant question is not only whether the **parent** drug is renally cleared, but whether an
-  **active or toxic metabolite** is. Morphine-6-glucuronide accumulating in renal failure is the
-  standard example.
-- Renal impairment also reduces some **non-renal** clearance pathways — uraemic toxins inhibit
-  CYP and transporter activity — so a low `fe` does not guarantee no effect.
-- Protein binding falls in uraemia for acidic drugs, raising unbound fraction; total concentrations
-  then understate the change in unbound exposure.
-- **Dialysis is a separate question** with its own study: whether the drug is removed depends on
-  molecular size, protein binding and volume of distribution, and the dosing implication is about
-  timing relative to the session as much as about dose.
-- Cockcroft-Gault (creatinine clearance) versus CKD-EPI (eGFR, normalised to 1.73 m²) matters. For
-  dosing, de-normalise eGFR to the individual's body surface area; using a normalised eGFR as if it
-  were an individual clearance misdoses people at the extremes of size.
-
-## Hepatic impairment
-
-Child-Pugh A/B/C is the conventional classification, though it is a crude proxy for drug-metabolic
-capacity and correlates poorly with any specific enzyme.
-
-- Effects include reduced enzyme content, reduced hepatic blood flow, portosystemic shunting
-  (raising oral bioavailability of high-extraction drugs sharply), reduced albumin, and altered
-  transporter expression.
-- For a **high-extraction** drug given orally, the dominant effect is loss of first-pass
-  extraction, and exposure can rise many-fold — much more than clearance alone would suggest.
-- Reduced albumin raises the unbound fraction; for a low-extraction, highly bound drug the unbound
-  concentration may be nearly unchanged while total concentration falls. Interpreting total
-  concentrations alone gives the wrong dose adjustment.
-
-## Obesity
-
-Which size descriptor to scale by depends on the parameter and the drug:
-
-| Descriptor | Use |
-| --- | --- |
-| Total body weight | Volume of distribution for lipophilic drugs |
-| Lean body weight | Clearance, most of the time; the best general-purpose descriptor |
-| Fat-free mass + a fraction of fat mass ("normal fat mass") | Where lean weight under-predicts |
-| Body surface area | Conventional in oncology; poorly justified for most agents |
-| Ideal body weight | Older convention, largely superseded |
-
-Fixed allometric exponents derived across species do not automatically apply within a species
-across the obesity range. Fitting the descriptor and letting the data choose is legitimate here.
-
-## Pregnancy
-
-Physiological changes across gestation are large and progressive: plasma volume up ~50%, GFR up
-~50%, albumin down, CYP3A4 and CYP2D6 induced, CYP1A2 and CYP2C19 inhibited. A single "pregnancy"
-covariate is inadequate — the effect is gestational-age dependent. PBPK with a pregnancy population
-model is the usual approach, since dedicated PK studies in pregnancy are rare.
-
-## Geriatric
-
-Age effects are mostly mediated: declining renal function, reduced hepatic blood flow and mass,
-changed body composition (less water, more fat), lower albumin. **Include the mediators as
-covariates rather than age itself** where possible — a model with age standing in for renal
-function will mispredict a fit 80-year-old and a frail 60-year-old in opposite directions.
-
-## Organ impairment study design
-
-Both regulators accept a reduced ("staged") design: study severe impairment first, and if exposure
-is unchanged, the intermediate categories can often be waived. A full design covers each category
-against matched controls. Match on age, weight and sex; unmatched controls are the usual reason an
-organ-impairment study is uninterpretable.

package/skills/pkpd-modeling/references/structural-models.md

@@ -1,140 +0,0 @@
-# Structural PK models: solutions, parameterisations, and the ADVAN map
-
-## Always parameterise in clearance and volume
-
-Micro-constants (`k10`, `k12`, `k21`) and macro-constants (`A`, `alpha`, `B`, `beta`) are outputs,
-not parameters to estimate. Clearance and volume are the parameters that:
-
-- have physiological meaning and known covariate relationships (CL scales with weight^0.75,
-  V with weight^1.0, CL with renal function);
-- are comparable across studies, formulations and populations;
-- keep their meaning when a compartment is added — `k10` changes when you add a peripheral
-  compartment, `CL` does not.
-
-`_models.py` accepts only `cl, v1, q, vp`. `micro_constants()` converts for reporting.
-
-## Linear mammillary disposition
-
-Any linear model with a central compartment and *n* peripheral compartments has a unit-bolus
-impulse response that is a sum of *n+1* exponentials:
-
-```
-C(t) = (D) * sum_i  coef_i * exp(lambda_i * t)
-```
-
-`_models.disposition()` gets `lambda_i` and `coef_i` by eigendecomposition of the rate matrix
-rather than from the textbook quadratic/cubic root formulas. Same answer, no special cases, works
-for any number of compartments.
-
-Given that impulse response, every input is a convolution with a closed form:
-
-| Input | Solution |
-| --- | --- |
-| IV bolus, dose D | `D * sum coef_i exp(lambda_i t)` |
-| Infusion, rate R over T | `R * sum (coef_i / -lambda_i)(1 - exp(lambda_i min(t,T))) exp(lambda_i (t - min(t,T)))` |
-| First-order absorption, ka | `F*D*ka * sum coef_i (exp(lambda_i t) - exp(-ka t)) / (ka + lambda_i)` |
-
-**The absorption term has a removable singularity at `ka = -lambda_i`.** The limit is
-`coef_i * t * exp(lambda_i t)`. This is not an edge case: it is precisely the flip-flop boundary,
-and an optimiser walking through it returns `inf` or `nan` without the branch. `_models.py` handles
-it; a hand-written Bateman function usually does not.
-
-Useful identities that hold for every linear model, and make good unit tests:
-
-```
-AUC(0-inf) after an IV bolus = D / CL          (independent of the number of compartments)
-Vss = V1 + sum(Vp)
-MRT after an IV bolus = Vss / CL
-```
-
-## Flip-flop kinetics
-
-When absorption is slower than elimination (`ka < lambda_z`), the terminal slope of an oral profile
-reflects **absorption**, not elimination. The two exponentials are mathematically interchangeable:
-the fit is identical if you swap `ka` and `k`. Consequences:
-
-- t½ from an oral profile is the absorption half-life, and Vz/F is meaningless;
-- which root the optimiser lands on is a starting-value accident;
-- **the assignment cannot be resolved from extravascular data alone.** It needs IV data, or a
-  formulation with faster absorption, or an external argument.
-
-`fit_compartmental.py` reports `flip_flop_suspected` and raises a finding. Depot formulations,
-extended-release products and subcutaneous biologics are routinely flip-flop.
-
-## Absorption models
-
-| Model | Parameters | Use when |
-| --- | --- | --- |
-| First-order | `ka` | Default; adequate for most immediate-release oral data |
-| First-order with lag | `ka`, `tlag` | A genuine delay before any drug appears; a discontinuous derivative that some estimation methods dislike |
-| Zero-order into central | `duration` | Absorption that looks constant-rate; often fits an IR tablet better than expected |
-| Sequential zero- then first-order | `duration`, `ka` | Delayed then first-order |
-| Transit compartments (Savic) | `MTT`, `n` | Smooth delay; `n` is estimated as a continuous parameter, `ktr = (n+1)/MTT` |
-| Weibull | scale, shape | Empirical, flexible, no mechanistic reading |
-
-The transit model must be evaluated with `lgamma`, not a literal factorial: fitted `n` routinely
-exceeds 20 and the direct form overflows. `_models.conc_transit()` does this.
-
-A lag time and a transit chain describe the same phenomenon differently. The transit model is
-continuous and usually estimates better; the lag is easier to explain. Do not fit both.
-
-## Nonlinear elimination
-
-Michaelis-Menten: `dA/dt = -Vmax * C / (Km + C)`, with `C = A/V1`.
-
-- Below `Km`, clearance is approximately `Vmax/Km` and the drug looks linear;
-- above `Km`, clearance falls and exposure rises faster than the dose;
-- **superposition is invalid**, so multiple-dose behaviour cannot be derived from a single dose,
-  and the accumulation ratio is dose-dependent.
-
-Phenytoin is the classic example: within the therapeutic range a 10% dose increase can produce a
-much larger exposure increase. If a dose-proportionality analysis shows AUC rising faster than
-dose, MM elimination is one explanation; saturable first-pass metabolism and solubility-limited
-absorption (which produce *less* than proportional increases) are others.
-
-## Combining PK with a delayed effect
-
-| Structure | Distinguishing feature |
-| --- | --- |
-| Direct effect | Effect tracks concentration with no hysteresis |
-| Effect compartment | Counter-clockwise hysteresis; one parameter `ke0`; the effect site is a modelling construct with no mass |
-| Indirect response | Delay arises from turnover of the response; return to baseline is governed by `kout` |
-| Transit/signal transduction | A chain of compartments producing a smooth, longer delay |
-
-An effect compartment and an indirect response model can both fit a hysteresis loop. They differ in
-what happens when the drug is stopped, and in how the delay scales with dose — an indirect response
-model's onset is dose-dependent while `ke0` is not. See `pd-and-exposure-response.md`.
-
-## NONMEM ADVAN/TRANS map
-
-For translating a model built here into a control stream:
-
-| ADVAN | Model | Common TRANS |
-| --- | --- | --- |
-| ADVAN1 | One compartment, IV | TRANS2 → `CL`, `V` |
-| ADVAN2 | One compartment with depot | TRANS2 → `CL`, `V`, `KA` |
-| ADVAN3 | Two compartment, IV | TRANS4 → `CL`, `V1`, `Q`, `V2` |
-| ADVAN4 | Two compartment with depot | TRANS4 → `CL`, `V2`, `Q`, `V3`, `KA` |
-| ADVAN11 | Three compartment, IV | TRANS4 → `CL`, `V1`, `Q2`, `V2`, `Q3`, `V3` |
-| ADVAN12 | Three compartment with depot | TRANS4 |
-| ADVAN13 | General nonlinear ODE | user-written `$DES` |
-| ADVAN6, ADVAN8, ADVAN9 | General ODE (non-stiff, stiff, equilibrium) | user-written `$DES` |
-| ADVAN15 | General with equilibrium compartments | |
-| ADVAN16, ADVAN17 | **New in NONMEM 7.6**: stiff delay differential equations (RADAR5) and stiff delay differential-algebraic equations | |
-
-Watch the compartment numbering: in ADVAN4 the central compartment is 2 and `V2` is the central
-volume, whereas in ADVAN3 the central compartment is 1 and `V1` is central. Mixing the two
-conventions when converting a model is a standard source of a silently wrong volume.
-
-Analytical ADVANs are far faster and more numerically stable than `$DES` and should be used
-whenever the model is linear. Reach for ADVAN13 only when the structure genuinely is not.
-
-## Choosing the number of compartments
-
-Use, in order: the residual pattern (runs of the same sign mean the shape is wrong), the F test for
-nested models, BIC, and the parameter precision. Do not use AIC alone — its fixed penalty of 2 per
-parameter frequently selects an extra compartment whose intercompartmental clearance has an RSE
-above 50%. `fit_compartmental.py --compare` prints all four side by side.
-
-Adding a compartment is justified when it changes the *conclusions* — Vss, the terminal half-life,
-the accumulation ratio, the predicted trough — not merely when it improves the objective function.

package/skills/pkpd-modeling/references/tmdd-and-biologics.md

@@ -1,115 +0,0 @@
-# Target-mediated drug disposition and biologics PK
-
-## The full TMDD model
-
-When a drug binds its target with high affinity and the target is present at a concentration
-comparable to the drug's, binding is not just pharmacology — it is a clearance pathway.
-
-```
-dL/dt  = In - kel*L - kon*L*R + koff*RL          free drug
-dR/dt  = ksyn - kdeg*R - kon*L*R + koff*RL       free target
-dRL/dt = kon*L*R - (koff + kint)*RL              complex
-```
-
-The characteristic profile has four phases: a rapid initial drop as the target is bound, a slower
-linear phase while the target is saturated, a steep terminal drop as drug falls below target
-capacity and target-mediated clearance resumes, and a final linear phase. **Dose-normalised
-profiles that do not superimpose, with the low dose disappearing faster, is the signature.**
-
-The model is stiff by construction — `kon` is typically 10³ to 10⁶ times `kel` — which is why
-`_models.simulate_tmdd()` uses LSODA rather than a fixed-step explicit method.
-
-## Approximations, in order of increasing assumption
-
-| Approximation | Assumes | Parameters |
-| --- | --- | --- |
-| Full | Nothing | `kon`, `koff`, `kint`, `ksyn`, `kdeg`, `kel`, `V` |
-| Rapid binding (QE) | Binding at equilibrium: `Kd = koff/kon` | replaces `kon`, `koff` with `Kd` |
-| Quasi-steady-state (QSS) | Complex at steady state: `Kss = (koff + kint)/kon` | replaces `kon`, `koff` with `Kss` |
-| Michaelis-Menten | Target dynamics fast and target constant | `Vmax`, `Km` — loses all target information |
-| Wagner / constant Rtot | Total target constant | |
-| Irreversible binding (IB) | `koff` negligible | |
-
-**QSS is the usual practical choice.** The full model is rarely identifiable from plasma drug
-concentrations alone: `kon` and `koff` appear almost exclusively as their ratio, and estimating them
-separately requires target or complex measurements. Pharmpy 2.1.1 exposes exactly this hierarchy —
-`set_tmdd(model, type=...)` accepts `'full'`, `'ib'`, `'cr'`, `'crib'`, `'qss'`, `'wagner'` and
-`'mmapp'`, with `dv_types` to map observations to drug, total drug, target, total target, or
-complex.
-
-**Which quantity was measured is a first-order question.** A ligand-binding assay typically reports
-**total** drug (free + complex), while the model's natural state is free drug. Fitting a total-drug
-observation to a free-drug prediction produces a badly wrong Kd, and nothing in the fit statistics
-reveals it. `_models.simulate_tmdd()` returns free drug, free target, complex and total drug
-separately for this reason.
-
-## Monoclonal antibody pharmacokinetics
-
-Typical IgG behaviour, useful as a sanity check on any fitted mAb model:
-
-| Property | Typical value |
-| --- | --- |
-| Clearance | 0.1-0.5 L/day (linear component) |
-| Central volume | ~3 L, close to plasma volume |
-| Vss | 5-10 L; distribution is largely confined to plasma and interstitial fluid |
-| Terminal half-life | 2-4 weeks for a typical IgG1 |
-| Subcutaneous bioavailability | 50-80% |
-| Time to SC Tmax | 2-8 days |
-
-Mechanisms that matter:
-
-- **FcRn recycling** is what gives IgG its long half-life. Antibodies engineered for higher FcRn
-  affinity at endosomal pH (YTE, LS mutations) extend half-life several-fold.
-- **Catabolism** is nonspecific proteolysis, not renal or hepatic clearance. Renal impairment does
-  not meaningfully change mAb clearance; molecules below ~60 kDa are a different story.
-- **Target-mediated clearance** dominates at low doses; the drug looks nonlinear until the target
-  is saturated.
-- **Subcutaneous absorption** is via the lymphatics, slow and incomplete, and produces flip-flop
-  kinetics: the apparent terminal slope after SC dosing can reflect absorption.
-
-Allometric exponents for mAbs are often closer to 0.85-0.9 for clearance and near 1.0 for volume
-rather than the small-molecule 0.75.
-
-## Immunogenicity
-
-Anti-drug antibodies increase clearance, sometimes by an order of magnitude, and typically appear
-after weeks. Handling in a model:
-
-- Treat ADA status as a **time-varying** covariate, not a baseline one. A subject who seroconverts
-  at week 8 has two different clearances in one profile.
-- ADA-positive subjects often show a bimodal concentration distribution rather than a shifted one;
-  a covariate on clearance may fit poorly where a mixture model fits well.
-- Neutralising versus binding ADA, and titre, matter more than the binary status.
-- Assay drug tolerance limits ADA detection when drug is present, so ADA-negative at trough is not
-  the same as ADA-negative.
-
-## Antibody-drug conjugates
-
-An ADC needs at least three analytes modelled, and they answer different questions:
-
-- **ADC** (conjugated antibody) — the dosed entity
-- **Total antibody** (conjugated + unconjugated) — the deconjugation rate is the difference
-- **Unconjugated payload** — usually drives systemic toxicity
-
-Drug-to-antibody ratio changes over time as the conjugate deconjugates, so "the ADC" is a
-distribution of species, not one molecule. Payload exposure is generally the safety-relevant
-metric.
-
-## Bispecifics and cell engagers
-
-Ternary complex formation (drug + target + effector) is not captured by standard TMDD. The
-concentration-effect relationship is typically **bell-shaped**: at high concentrations the drug
-saturates both arms separately and forms fewer ternary complexes ("hook effect"). A monotone Emax
-model fitted to such data will mislead about the optimal dose in exactly the region that matters.
-
-## Practical guidance
-
-- Do not fit a full TMDD model to plasma drug data alone. Start with QSS; move up only if target or
-  complex measurements exist.
-- Check dose-normalised profiles first. If they superimpose across the clinical dose range, TMDD is
-  saturated throughout and a linear model is adequate for that range — but it will not extrapolate
-  to lower doses.
-- Baseline target concentration `R0 = ksyn/kdeg` is often measurable and should be fixed to the
-  measurement rather than estimated.
-- Report which analyte each observation is. This single piece of metadata resolves more confused
-  biologics models than any structural change.