CASE STUDY

This case study documents a model-first, stage-locked decision system built for a dual-platform ABCD1 mRNA therapy program in X-linked adrenoleukodystrophy (ALD) using a Peripheral-First → CNS strategy. The package includes:

• Stage-0: feasibility + decision risks + what must be locked early
• Stage-1: a complete A→G architecture (delivery → expression/targeting → PD → safety → CMC → uncertainty → decision)
• Stage-2: a full mathematical formalization (ODEs + derived metrics + constraints + Go/No-Go surfaces)
• Stage-3 Core (zip repo): an executable implementation (mouse branch, Blocks A–E) with verification evidence (“proof of match”)

Note: This case study is a decision-infrastructure deliverable. It is not a claim of clinical efficacy, and it does not constitute medical advice.

Source: internal Method2Model case-study draft. :contentReference[oaicite:0]{index=0}

Method2Model: We sell decisions without risk — what decisions were de-risked here?

In this program, Method2Model converted “uncertain research activity” into explicit decisions that could be made early, defended later, and updated systematically as evidence accumulates.

The decisions this process made possible (and locked)

• Platform allocation decision (LNP vs EV vs combination): minimum evidence thresholds to justify dual-platform vs single-platform focus; platform-to-objective assignment (peripheral correction vs CNS relevance).
• Route strategy decision (IV only vs IN vs IV+FUS): under what exposure conditions FUS meaningfully increases CNS delivery probability; explicit safety policy and pivot triggers for FUS (stop / reduce / re-route).
• Targeting evidence decision (what counts as “real rescue”): locked “targeting triad” = localization + membrane insertion + functional rescue (prevents false confidence from non-decision-grade signals).
• CMC and batch-release decision (what batch quality is usable): potency drift, stability decay, and handling translate into effective dose; enables batch classification and prevents “dose illusion.”
• Phenotype targeting decision (AMN vs very-early cALD vs early cALD): phenotype treated as a scenario-defined initial condition and progression choice (prevents mixing incompatible endpoints and timelines).
• Go / No-Go / Pivot decision under uncertainty: decision surfaces based on \(p_{\mathrm{eff}}\) and \(p_{\mathrm{safe}}\), where Pivot is a governed region with defined corrective actions (platform, route, dose, endpoint strategy, phenotype focus).

What would have happened without this process (hidden risks and costs)

• False Go (worst case): advancing a platform due to non-decision-grade signals → wasted animal studies, manufacturing cycles, and time.
• False No-Go: killing a viable approach because variability (FUS, EV loading, batch potency) was not modeled → “null” appears decisive when it is not.
• Batch-driven contradictions: two “same dose” studies disagree due to potency/handling differences → reruns, interpretability collapse.
• Endpoint drift: redefining success midstream after seeing results → non-defensible narrative to reviewers/sponsors.
• Late discovery of operational constraints: learning too late that CNS strategy is infeasible under safety/throughput constraints.
• Rework spiral: repeated protocol tweaking because the system cannot deterministically explain what failed and what to change.

Method2Model turns those risks into governed rules and scenario outputs—so decisions are made earlier, cheaper, and defensibly.

What we built (Stage-Locked Deliverables)

1) Stage-1 Architecture (Blocks A–G)

A review-ready architecture capturing the mechanistic chain:

• A: Delivery & Exposure (PBPK / BBB / cell uptake)
• B: Expression & Peroxisomal Targeting (mRNA → ALDP → peroxisomal insertion + targeting score)
• C: Disease / PD (mouse PD + human-projected branch)
• D: Safety & Immunogenicity (cytokines, ALT/AST, ADA; FUS safety policies)
• E: CMC / batch variability (potency multipliers + stability/handling)
• F: Uncertainty & scenario engine (Monte Carlo across phenotype × route × regimen × batch × immune background)
• G: Decision logic (chain of evidence + Go/No-Go/Pivot surfaces)

2) Stage-2: Formal Model Specification (architecture → explicit mathematics)

The goal is not to drown teams in equations, but to ensure every term has an operational meaning and can be locked before coding.

Block A — Delivery & Exposure (coarse PBPK + BBB modulation + uptake)

Compartments (Stage-3 mouse core): \[ c \in \{\text{plasma, liver, periphery, CNS, CSF}\} \]

Platforms: \[ k \in \{\text{LNP, EV}\} \]

Core state variables:

• \(C_{k,c}(t)\): carrier concentration in compartment \(c\)
• \(U_{k,c,\text{cell}}(t)\): cumulative uptake into a cell type in compartment \(c\)
• \(P_{\text{BBB}}(t)\): effective BBB permeability (time-dependent; FUS window)

PBPK transport (generic form):

\[ \frac{dC_{k,c}(t)}{dt} = \sum_{c’} Q_{c’\to c}C_{k,c’}(t) – \sum_{c”} Q_{c\to c”}C_{k,c}(t) – CL_{c}C_{k,c}(t) – \text{Uptake}_{k,c}(t) + \text{DoseInput}_{k,c}(t) \]

BBB flux (plasma → CNS):

\[ Q_{k,\text{plasma}\to \text{CNS}}(t)=P_{\text{BBB}}(t)A_{\text{BBB}}\left(C_{k,\text{plasma}}(t)-C_{k,\text{CNS}}(t)\right) \]

FUS modulation (windowed multiplier):

\[ P_{\text{BBB}}(t)=P_{\text{baseline}}\times \begin{cases} M_{\text{FUS}}, & t\in[t_{\text{FUS}},t_{\text{FUS}}+\Delta t_{\text{FUS}}] \\ 1, & \text{otherwise} \end{cases} \]

Cellular uptake (saturable driver):

\[ \frac{dU_{k,c,\text{cell}}(t)}{dt}=f\!\left(C_{k,c}(t),\theta_{k,r,c,\text{cell}}\right) \]

Block B — Expression & Peroxisomal Targeting (mRNA → ALDP → peroxisome)

• \(m(t)\): intracellular ABCD1 mRNA
• \(P_{\text{tot}}(t)\): total ALDP protein
• \(P_{\text{peri}}(t)\): peroxisomal membrane ALDP
• \(P_{\text{mis}}(t)\): mislocalized ALDP

mRNA kinetics:

\[ \frac{dm(t)}{dt}=k_{\text{in}}(t)-k_{\text{deg,m}}m(t) \]

Translation:

\[ \frac{dP_{\text{tot}}(t)}{dt}=k_{\text{tr}}m(t)-k_{\text{deg,p}}P_{\text{tot}}(t) \]

Peroxisomal insertion (capacity-limited):

\[ \frac{dP_{\text{peri}}(t)}{dt}=k_{\text{ins}}(P_{\text{tot}})P_{\text{tot}}(t)-k_{\text{turnover,peri}}P_{\text{peri}}(t) \]

Mislocalization:

\[ P_{\text{mis}}(t)=P_{\text{tot}}(t)-P_{\text{peri}}(t) \]

Targeting score (audit-friendly triad):

\[ S_{\text{target}} = f_{\text{localization}}\cdot f_{\text{insertion}}\cdot f_{\text{function}} \]

Each factor maps to concrete assays and is normalized to \([0,1]\).

Block C — Disease / PD (mouse implemented; human-projected defined)

VLCFA dynamics (generic):

\[ \frac{dC_{26}(t)}{dt}=P_{\text{prod}}(p,s,t)-V_{\beta}(t)\,h(C_{26}(t)) \]

Latent injury:

\[ \frac{d\,\text{AxonInjury}(t)}{dt}=F\!\left(C_{26,\text{CNS}}(t),\text{Inflamm}_{\text{CNS}}(t)\right) \]

Block D — Safety & Immunogenicity (rule-based + state dynamics)

Example structure:

\[ \frac{dALT(t)}{dt}=f_{\text{hepatotox}}\!\left(C_{\text{LNP,liver}}(t),\text{repeat-dose history}\right) \]

With explicit safety constraints and pivot triggers.

Block E — CMC / Batch Variability (dose meaning is governed)

\[ D_{\text{effective}} = D_{\text{nominal}}\cdot \text{Pot}_{k,\text{batch}} \]

\[ \text{Pot}(t)=\text{Pot}_0\cdot e^{-k_{\text{decay}}t}\cdot f_{\text{FT}}(\#\text{freeze-thaw}) \]

Block F — Uncertainty & Scenario Engine

Distributions (priors) are defined for uncertain parameters; scenarios span phenotype × route × regimen × batch × immune background. Outputs include first-break maps and robustness curves.

Block G — Decision Logic (probabilistic Go/No-Go surfaces)

\[ p_{\text{eff}}(d,s)=\mathbb{P}\big(\text{efficacy criteria hold}\mid d,s\big) \quad\quad p_{\text{safe}}(d,s)=\mathbb{P}\big(\text{safety criteria hold}\mid d,s\big) \]

Stakeholder-defined regions: Go, No-Go, Pivot in \((p_{\text{eff}},p_{\text{safe}})\) space.

3) Stage-3 Core (Executable Implementation) — What is actually in the repo

The Stage-3 repository implements the mouse branch of Blocks A–E as a reusable Python package with templates and verification evidence.

Implemented state vector (high-level)

• Block A: \(C_{k,c}(t)\), \(U_{k,c,\text{cell}}(t)\)
• Block B: \(m(t)\), \(P_{\text{tot}}(t)\), \(P_{\text{peri}}(t)\)
• Block C: \(C26(t)\), \(\text{Inflamm}(t)\), \(\text{AxonInjury}(t)\)
• Block D: cytokines, ALT/AST, ADA (policy + mapping)
• Block E: batch potency scaling wired into dosing

Inputs

• example_params_mouse.yaml
• example_initial_state.yaml
• example_doses_mouse.yaml

Outputs

• state trajectories (.npy)
• metrics summary (.json)
• time series (.csv)

Verification (“Proof of Match”)

• traceability matrix (Architecture → Formula → Code)
• acceptance tests + scenario equivalence checks
• version mapping between Stage IDs and code

What a medical researcher should take away

This case is not “a model.” It is a decision engine:

• It defines what success means before experiments.
• It integrates CMC into the biological chain.
• It outputs defensible Go/No-Go/Pivot decisions under uncertainty.
• It provides executable code and verification assets, so results can be reproduced and reviewed.

Links

• GitHub: https://github.com/RamyarAzar/ALD_mRNA_Therapy
• Zenodo (Method2Model ALD mRNA community): https://zenodo.org/communities/method2model-ald-mrna/records?q=&l=list&p=1&s=10&sort=newest