Advancing Drug Discovery and Development through Clinical Pharmacological Approaches on Organs-on-Chips

Roadmap

KBCS 2026 · Medical Unmet Needs for BioChips

• Unmet need & regulatory shift — why preclinical systems fail, and why FDA is moving toward NAMs
• Clinical pharmacology — PBPK & QSP as the bridge
• Our engine — an LLM-augmented, open QSP library (258 diseases and growing)
• Case studies — 12 vignettes, 6 deep dives, incl. a live app (Merigolix)
• Our organ-chip work — Gut–Liver–Kidney MPS + PBPK, condensed
• Where this goes — a browser-based platform, and a challenge to this community

The unmet need: why preclinical systems fail translation

• 2D cell cultures — no tissue architecture, no organ-level function
• Animal models — persistent interspecies differences in drug disposition
• Phase 1 trials characterize healthy-volunteer exposure well, but not the patients who need it most: IBD AKI Hepatic impairment Pregnancy Pediatric Rare disease

DiMasi et al., J Health Econ 2016; Kola & Landis, Nat Rev Drug Discov 2004.

The regulatory shift: from an animal mandate to NAMs

• Since 1938, U.S. law required animal testing before human use
• FDA Modernization Act 2.0 (2022) — replaced “animal tests” with “nonclinical tests,” explicitly opening the door to in vitro/in silico methods
• FDA’s Roadmap to Reducing Animal Testing in Preclinical Safety Studies (2025) — phased rollout starting with monoclonal antibodies; goal: animal testing becomes the exception, not the default
• Motivation: many drugs that pass animal testing still fail in humans

FDA Modernization Act of 2022, Pub. L. 117–328; FDA, Roadmap to Reducing Animal Testing in Preclinical Safety Studies (2025); Van Norman GA, JACC Basic Transl Sci 4(7):845–854 (2019).

Why this matters for organs-on-chips

• FDA Modernization Act 2.0 opens the door for NAMs — including MPS — to replace some animal studies
• MIDD already expects PBPK/QSP-grade quantitative packages; folding MPS data into that workflow is the credible path to regulatory acceptance
• Organ-on-chip inside a clinical-pharmacology workflow = a translational bridge: less animal testing, faster candidate selection, more reliable human prediction

The NAM translation loop — where clinical pharmacology sits

PBPK and QSP are the mechanistic models this loop depends on — the qualification & translation step that AI/ML pattern-matching alone cannot supply.

Cao Y. New Approach Methodologies: What Clinical Pharmacologists Should Prepare For. Clin Pharmacol Ther 118(6):1269–1272 (2025).

From a chip curve to a dosing decision

• What a chip gives you: a concentration–time curve · single- or multi-organ readouts · a mechanistically rich human-cell signal
• What a dosing decision needs: clearance, tissue distribution, target engagement — for a whole human body, not a chip well
• A chip curve is a descriptive observation until it passes through a mechanistic model

Clinical pharmacology — NCA, PBPK, QSP — is that bridge, and the subject of this talk.

What is QSP & PBPK?

PBPK and QSP describe the causal chain from drug to patient, mathematically:

$$\textbf{drug} \;\rightarrow\; \textbf{target} \;\rightarrow\; \textbf{pathway} \;\rightarrow\; \textbf{disease} \;\rightarrow\; \textbf{patient}$$

• PBPK — where does the drug go, and at what concentration, in a whole human body?
• QSP — given that concentration, what happens to the disease?
• Together: dose selection, DDI risk, response in special populations, IVIVE of chip-derived parameters

Anatomy of a PBPK/QSP model

A coupled nonlinear dynamical system — an initial-value problem:

$$\frac{d\mathbf{x}}{dt} = \mathbf{f}\!\big(\mathbf{x}(t),\, \boldsymbol{\theta},\, \mathbf{u}(t)\big), \qquad \mathbf{x}(0)=\mathbf{x}_0$$

• States $\mathbf{x}$ — concentrations, receptor occupancy, signaling species, cell populations, biomarkers, endpoints
• Parameters $\boldsymbol{\theta}$ — rate constants, $EC_{50}$/$E_{max}$, Hill $n$, clearances, binding affinities
• Inputs $\mathbf{u}(t)$ — dosing regimen: the one controllable handle

Our library’s models: 15–35+ coupled ODEs, 70–100+ parameters per disease.

The recurring nonlinear structures

The same motifs recur across every disease — and every chip-to-human bridge:

Common grammar across a 258-model library and a chip–PBPK bridge.

Why this matters — Model-Informed Drug Development

FDA/EMA actively endorse MIDD — an established pathway, not a novelty.

The bottleneck: building one QSP model, traditionally

• Deep literature synthesis across pathophysiology + pharmacology
• Careful ODE formulation, parameterization, implementation, calibration
• Cost: expert-months to years per disease

Consequences:

• only a handful of diseases ever get a model
• models are siloed, rarely open, rarely reproducible
• the long tail — most chip programs’ target diseases — never gets one

One disease ≈ one expert-year+. Scarce · slow · closed.

Thesis — LLM-augmented model generation

An autonomous LLM coding agent builds complete, quality-gated, referenced QSP models — one disease per day — every artifact under version control.

• 🗺️ Graphviz map · ⚙️ mrgsolve ODE model · 📊 Shiny dashboard · 📚 curated references
• enforced quality gates, literature grounding for every parameter
• standardized schema → comparable, auditable, reproducible

The Claude Code Routine — an autonomous daily loop

Each run completes one disease end-to-end and pushes it:

1. Select an uncovered disease
2. Research — ~50 PubMed sources
3. Map — Graphviz network (≥100 nodes)
4. Model — mrgsolve ODEs (≥15 compartments)
5. Dashboard — Shiny (≥6 tabs)
6. Ground — sectioned references (≥30)
7. Commit & push — under a git stop-hook that won’t allow a half-done session

Guardrails — why trust code an LLM wrote

• Quality gates: map ≥100 nodes/≥8 clusters · ODE ≥15 compartments/≥5 scenarios · dashboard ≥6 tabs · ≥30 references
• Structural templates: reusable ODE motifs (Hill, turnover, TMDD) shrink the failure space
• Literature grounding: every parameter set cites a calibration trial — the main anti-hallucination mechanism
• Reproducibility: diffable, auditable, re-runnable code in git; human review before use

Why an LLM/AI is the right engine

• Tireless & continuous — a full disease model nearly every day
• Vast literature synthesis — ~12,800 citations (~50/model)
• Breadth — 258 diseases, hundreds of drugs/targets, one framework
• Consistency — identical schema and quality bar
• Economics — expert-months → hours

The scarce resource is modeling expertise. The LLM scales it.

What’s inside each model

• 🗺️ Mechanistic map — clustered directed graph, targets to disease
• ⚙️ mrgsolve ODE model — PK coupled to PD; 15–35+ states, ≥5 scenarios
• 📊 Shiny dashboard — patient profile, PK, endpoints, scenarios (6–8 tabs)
• 📚 ~50 curated references per model

The exact structure a chip-to-human bridge needs to plug into.

Scale: a 258-model open QSP library

Traditional QSP: a handful of diseases, expert-months each. This library: 258 diseases, hours each.

Breadth of coverage

Oncology Autoimmune/rheumatic Vasculitis Cardiovascular Respiratory Renal/urologic GI/hepatobiliary Endocrine/metabolic Neurologic Psychiatric Dermatologic Infectious Ophthalmic Rare/genetic

• Cancers — breast, NSCLC, SCLC, glioblastoma, CML, multiple myeloma, melanoma…
• Rare & genetic — Fabry, Gaucher, DMD, SMA, Huntington, ATTR amyloidosis…
• Common chronic — diabetes, heart failure, COPD, CKD…
• Immune / hematologic — lupus, RA, sickle cell, ITP, myelofibrosis…

Hundreds of drugs and molecular targets — very likely including yours.

Case studies — one modeling insight each (1/4)

Case studies — one modeling insight each (2/4)

Case studies — one modeling insight each (3/4)

Case studies — one modeling insight each (4/4)

Deep dive — IgA nephropathy: the “four-hit” cascade

Deep dive — sickle cell disease: polymerization to vaso-occlusion

Deep dive — multiple myeloma: an oncology exemplar

Deep dive — rheumatoid arthritis: IL-6/TNF/JAK-driven synovitis

Deep dive — Duchenne muscular dystrophy: a genetic exemplar

Deep dive — Alzheimer’s disease: amyloid to cognition

Flagship case — Merigolix, a live drug–disease model

Merigolix — the estrogen “threshold” trade-off

QSP turns this trade-off into a tunable optimization problem — a chip program’s question too.

The gallery — a fraction of the library

A sample of mechanistic maps — each one autonomous build, spanning oncology, autoimmune, cardiovascular, renal, metabolic, neurological, and rare disease.

How this helps your chip program, directly

• Working on a cytokine, transporter, receptor, or disease? A model likely already exists
• Use it to generate a hypothesis before the chip experiment
• Use it as a PD/disease layer on top of your own chip→PBPK bridge

Not “run all 258” — a platform to accelerate translation of what you already work on.

Fully open — every line of code

• github.com/pipetcpt/qsp — every map, ODE model, dashboard, reference, openly versioned
• Clone it, run a model, adapt it, or just browse for ideas
• QR code, top-right corner of every slide, links to the repository

Where this goes: a browser-based PBPK/QSP platform

A web platform where any researcher can, from a browser alone —

• search by disease, target, drug, or cytokine
• browse the map and ODE structure with no install
• run a simple simulation — a dose, a tweak, an instant plot
• export parameters into your own workflow

Prerequisites for trusting a chip-derived number

Four checks before a chip number belongs inside a PBPK/QSP model:

1. Chip-specific drug behavior — adsorption to plastics/membranes
2. Scaling of cellular and fluidic dimensions — correct per-organ factors
3. Verification against held-out clinical data
4. Explicit, falsifiable correction rules — not curve-fitting

The next case study tests whether this discipline works in practice.

Our case study: Gut–Liver–Kidney MPS + PBPK

• Modular gut–liver–kidney MPS (hiEC–HepaRG–RPTEC), shared media recirculation
• Chip-fitted parameters propagate to whole-body PBPK — no human PK enters the prediction
• All 24 observations reserved for post-hoc validation

The model predicts human PK well

Predicted vs. observed AUC and Cmax across 24 observations, 6 drugs. Inner band = 2-fold; outer band = 4-fold.

How it works — chip kinetics to whole-body PBPK

• Chip → kinetics: 33-state ODE fits chip data (parent + metabolite); $K_{puu,hep}$ captures OATP-style hepatic uptake
• Whole-body PBPK: 10-state perfusion-limited; Rodgers–Rowland $K_p$; Varma clearance; ACAT absorption
• Inputs: physchem only (logP, $pK_a$, $f_{u,p}$, blood:plasma, MW, dose) — no clinical PK to fit

How it works — the falsifiable scaling rule

• Apply a relative-activity factor (RAF) only if CYP3A4 ≥90%, biliary clearance <50%, hepatic uptake not already captured
• Mechanism-based test, not curve-fitting
• Uniform correction → GMFE 9.7×; selective rule → 1.85×
• Disease-state extension: re-fit under DSS colitis / cisplatin injury; unaffected-organ parameters locked; same pipeline

Disease-state extension: gut vs. hepatic localization

• Gefitinib, DSS colitis chip: AUC 0.49× — gut-localized (Fg shifts, Fa/Fh unchanged)
• Simvastatin, cisplatin-injury chip: AUC 2.19× — hepatic-localized ($K_{puu,liver}$ halved, OATP1B1 suppression)
• Same pipeline distinguishes gut vs. hepatic shifts — hypothesis-generating for IBD/AKI patients usually excluded from Phase 1

Limitations, and what this means

• No mucus/peristalsis layer — affects chip $P_{app}$
• HepaRG 2D CYP3A4 ~10% of in vivo (RAF-corrected) · RPTEC OAT1/3 ~1% (PTC-anchor corrected)
• Sparse clinical PK in disease-state patients — roadmap: prospective IBD/AKI cohorts
• Open-source Python — full run reproduces in <5 minutes

Augments, does not replace, animal/clinical PK.

Clinical pharmacology’s unmet needs

Conclusion 1 — Unmet needs in our own field

• A real shortage of clinical pharmacologists trained to interpret NAM signal through PBPK/QSP
• No standardized workflow yet for FDA’s four NAM validation principles: context-of-use, biological relevance, technical characterization, fit-for-purpose
• Few chip/organoid models validated head-to-head against human clinical data
• The bottleneck isn’t technology — it’s people who can translate

Not a data problem, not a technology problem — a workforce problem.

A challenge to KBCS members

Conclusion 2 — A direct ask

• Design chips with PBPK/QSP-ready outputs in mind — clearance, permeability, absorption
• Let’s validate your chip data against held-out clinical data, together
• Try the matching model in our 258-model library first — tell us where it’s wrong
• Waiting to collaborate: github.com/pipetcpt/qsp · shan@catholic.ac.kr

If your chip already speaks a language PBPK/QSP can read, half the problem is solved before we meet.

Conclusion & an open invitation

Contact