Part 5: How to Run a Dose-Finding Trial — Protocols for Eight Medications
The previous article argued that a failed trial should mean “this dose failed,” not “this drug failed” — and that the cheapest next experiment is usually a neighboring dose of the same drug. That argument is worthless without an execution layer: which doses, held for how long, measured against what, stopping when.
This article provides that layer for eight medications. It is the protocol companion to the pharmacopoeia: Part 4 tabulates what each dose band means; this article specifies how to run the trial that generates the data. The two share the same dose bands and must be read together.
1 How to read the protocols
Every protocol field carries an evidence tier, consistent with the primary paper’s convention:
| Tier | Meaning |
|---|---|
| E — established | Standard pharmacology or regulatory labeling; would not surprise any pharmacist |
| C — candidate | Inferred from ME/CFS community practice and the low-dose-window framework; plausible, not validated |
| S — speculative | Model prediction (e.g. cluster correlations) awaiting trials; treat as a bet, not a fact |
Three universal rules before any protocol:
- Define the primary outcome before the first dose. One measure. If the trial “for sleep” ends with an impression about mood, nothing was tested.
- Write the stop conditions down and date the review. A trial without a stop date is not an experiment; it is a mood (see the dose-drift pathology in the previous article).
- Do not run trials through a crash. A PEM episode during a dose trial measures the PEM, not the dose.
2 Low-Dose Naltrexone (LDN)
Targets (dose-banded): TLR4/Nrf2 hormetic priming (~0.5–1.5 mg) · opioid compensatory upregulation (~1.5–3 mg) · TRPM3 restoration (~3–4.5 mg) — see Part 1 for the full cascade. Each target is individually established in vitro or in animal models: naltrexone as a partial TLR4 antagonist on microglia, shown in a mouse cell line and rodent models (Kučić et al. 2021; Kwilasz et al. 2021), opioid-receptor blockade with compensatory endorphin rebound (Younger, Parkitny, and McLain 2014), TRPM3 calcium-flux restoration in ME/CFS NK cells (Cabanas et al. 2018, 2021), and the inflammation-to-orexin-suppression link (Grossberg et al. 2011). What is not established is the mapping of each target onto a clinical dose band — that part is model inference (Loth 2026).
| Field | Protocol | Tier |
|---|---|---|
| Dose ladder | 0.5 mg × 2 wk → 1 mg × 4 wk → 1.5 mg × 4 wk → (if partial) 2.25 mg × 4 wk → (if partial) 3 mg × 4 wk → (if no signal yet) 4.5 mg × 4 wk | E for start-low; C for steps |
| Step duration | Minimum 4 weeks per rung below 2 mg — the TLR4/Nrf2 mechanism is transcriptional; verdicts inside 2 weeks measure adjustment, not effect | C |
| Primary outcome | Fatigue or brain fog, daily 0–10 self-rating | C |
| Stop conditions | Benefit at a rung → hold ≥8 wk before any change · no benefit across 0.5–4.5 mg full sweep → mechanism read: LDN’s targets are not rate-limiting · initial flare → halve, do not abandon | C |
| Compounding | Below 2 mg: compounding pharmacy or liquid | E |
| Retrial rule | A null at 3.5–4 mg does NOT close ~1 mg — different mechanism at that band (why) | C |
Common failure: verdict declared at 3 weeks on one rung. The four-mechanism structure makes LDN the most dose-information-rich drug in the list and therefore the most wasteful to test shallowly.
3 Low-Dose Aripiprazole (LDA)
Target: dopamine D2 partial agonism — microglial and catecholamine modulation (Crosby, Kalantar, and DeRisi 2021). Single drug; the framework predicts a steep, peaky window (illustratively: sub-1 mg doses may lose protection while 1 mg holds; escalation above the working dose may add nothing) — S, model prediction (Loth 2026).
| Field | Protocol | Tier |
|---|---|---|
| Dose ladder | 0.5 mg × 3–5 d (tolerability) → 1 mg × 6 wk (the assessment window used clinically) | C (Crosby 2021 cohort mean: 1.1 mg/day (Crosby, Kalantar, and DeRisi 2021)) |
| Step duration | Days for tolerability, weeks for efficacy — do not judge effect before the prescriber’s window elapses | C |
| Primary outcome | Choose per dominant symptom: PEM frequency, cognitive clarity, or sleep — one only | C |
| Stop conditions | Akathisia, agitation, or mood switch concerns → stop and report · no benefit at 6 wk → prescriber-directed ceiling per their practice | C |
| Compounding | Below 2 mg required: compounding pharmacy (standard tablets ≥2 mg) | E |
| Monitoring | Sleep architecture and mood self-checks (documented LDA effects are qualitatively distinct from stimulant-type activation — but the distinction should be checked, not assumed) | C |
| Retrial rule | A dose that preserved sleep but lost PEM protection (the 0.75 mg pattern) means the window is symptom-specific — map per symptom, not per drug | S |
Why short can work: a 5-day trial at a single lower dose, with one prescriber-agreed outcome and defined stop conditions, can return a clean negative in a week — and still strengthen the standing of the working dose, because the window was probed and held. That is the protocol working as designed.
4 Low-Dose Lithium (element lithium)
Target: GSK-3β inhibition → Nrf2 derepression at micro-doses; IMPase effects enter at higher micro-doses (biphasic two-target dynamic). GSK-3β inhibition at low lithium doses is documented preclinically (Fenech et al. 2023); cognitive benefit at micro-doses is documented in Alzheimer’s and neuropsychiatric cohorts (Nunes, Viel, and Buck 2013; Wilson et al. 2017; Strawbridge et al. 2023). The dose-dependent GSK-3β-vs-IMPase split is model inference (Loth 2026).
| Field | Protocol | Tier |
|---|---|---|
| Dose ladder | 0.3 mg → 1 mg → 3 mg → 5 mg element lithium, ~2 wk steps | C |
| Step duration | 2 weeks sufficient for subjective outcomes; the Nrf2 transcriptional argument suggests longer at effective rungs | C |
| Primary outcome | Same as the patient’s responsive Nrf2-cluster drug (if LDN worked at 0.5–1.5 mg, lithium’s question is “does the window cluster hold?”) | S |
| Stop conditions | Tremor, thirst, confusion → stop (standard lithium toxicity awareness, though micro-doses rarely reach it — E for the warning, C for the threshold) · no cluster-consistent benefit across 0.3–5 mg → window-cluster hypothesis weakened for this patient | C/S |
| Monitoring | If extending above 5 mg or long-term: standard lithium monitoring per prescriber | E |
5 Pyridostigmine (Mestinon)
Target: acetylcholinesterase inhibition — peripheral cholinergic tone; relevant to orthostatic/vascular symptom layers (Joseph et al. 2022; Raj et al. 2005). Characteristic dose-dependent M3-mediated GI motility effects.
| Field | Protocol | Tier |
|---|---|---|
| Dose ladder | Start at the lowest historically tolerated dose (e.g. 1 pill/day if higher doses were GI-intolerable) — not upward from the intolerable doses | C |
| Step duration | 1–2 weeks per rung; GI effects appear fast, vascular benefit (if any) needs longer | C |
| Primary outcome | Orthostatic/vascular symptom rating — not general fatigue | C |
| Stop conditions | Excess gut motility → stop; this drug’s intolerance is dose-dependent and specific, and a lower rung can be GI-quiet where higher rungs were not | C |
| Key asymmetry | Prior intolerance at higher doses does NOT close the lower-dose question — a patient who stopped at 3×1/day for GI reasons may still find 1/day tolerable; that rung is a separate experiment | C |
| Confounders to log | Meals, motion, caffeine — GI confounders are dense for this drug | C |
6 Methylphenidate (Ritalin MR) and low-dose stimulants
Target: catecholamine reuptake inhibition. In ME/CFS the documented pattern is inverted-U and PEM induction at high stimulus doses. The double-blind crossover RCT found methylphenidate reduced fatigue in CFS (Blockmans et al. 2006), and long-term follow-up found durable benefit in about a third of continuers with a high dropout rate — consistent with the stimulant-PEM trade-off the protocol is built to measure (Blockmans and Persoons 2016).
| Field | Protocol | Tier |
|---|---|---|
| Dose ladder | Start well below doses that previously caused problems — e.g. an MR 10 mg probe if MR 30 mg induced PEM; immediate-release 5 mg equivalents for finer steps | C |
| Step duration | 3–7 days per rung — stimulant effects on both benefit and PEM induction show quickly | C |
| Primary outcome | Function-per-exertion-cost, not raw alertness. A stimulant that buys alertness with PEM is a net loss — the outcome measure must be able to see the loss | C |
| Stop conditions | PEM induction at a rung → that rung is above the window; step down, do not “push through” | C |
| Interaction note | Documented in the primary record: prior high-dose stimulant combinations in untreated ME/CFS produced excitation/mania-like states — an absolute ceiling consideration, prescriber territory | E for the risk class |
7 The MCAS stack (famotidine · rupatadine · montelukast)
Targets: H2 · H1/PAF · leukotriene receptors. Long-standing, precede most other trials, and usually already at their windows — which makes them the stack to instrument before touching. Antihistamine/PAF-antagonist combinations (H1 + H2) show clinical effect in mast-cell-driven syndromes (Salvucci et al. 2023), and up-dosed H1/PAF antagonism (rupatadine) is documented in mast-cell and urticaria contexts (Magerl et al. 2015; Izquierdo et al. 2024).
| Field | Protocol | Tier |
|---|---|---|
| Trial type | Not dose-finding but de-challenge/re-challenge: single-component withdrawal × 2 wk with symptom scoring, then restore. Establishes each component’s contribution | C |
| Dose note | Antihistamines show inverted-U behavior in the pharmacopoeia (H1 sedation/anticholinergic burden at high doses); if escalating, keep the pharmacopoeia bands open | C |
| Primary outcome | The symptom that justified the component originally (e.g. reactive symptoms, food responses) | C |
| Stop conditions | Symptom return on de-challenge → component confirmed load-bearing; no change → candidate for deprescribing discussion | C |
Why include a stack that works: a dose-finding article that only covers failures teaches the wrong lesson. Confirming that a working component is load-bearing — cheaply, one component at a time — is dose-information too, and it protects against accidental discontinuations during other trials.
8 Melatonin (low-dose)
Target: chronobiotic at low doses; hypnotic (and different receptor profile) at high doses — a textbook inverted-U that most patients take backwards. Low-dose (0.5 mg) melatonin phase-shifts the circadian clock without suppressing endogenous secretion (Burgess et al. 2010); in CFS, DLMO-guided melatonin timing improved fatigue in patients with delayed phase (Heukelom et al. 2006). The “low dose is chronobiotic, high dose is hypnotic” framing, and its application in ME/CFS, is model inference (Loth 2026).
| Field | Protocol | Tier |
|---|---|---|
| Dose ladder | 0.3 mg → 0.5 mg → 1 mg → 3 mg, descending if higher doses worked less (the common surprise) | C (0.3 mg physiologic-dose rationale: E) |
| Step duration | 1 week per rung, scored nightly | C |
| Primary outcome | Sleep-onset latency or night waking — the chronobiotic claim is about timing, not sedation depth | C |
| Timing | 2–4 h before target sleep onset at low doses (chronobiotic window), not at bedtime (hypnotic habit) — timing is half the dose for this drug | C/E mixed |
| Stop conditions | Groggy mornings at a rung → above window, step down · no effect at any rung → try timing change before abandoning | C |
9 Modafinil
Target: wakefulness promotion, catecholaminergic/orexin-adjacent (Volkow et al. 2009). The dose logic applies as for the other stimulants: inverted-U with anxiety/overstimulation on the upper arm. In chronic fatigue syndrome a double-blind crossover found standard-dose modafinil produced no fatigue benefit and worsened cognition at 400 mg (Randall et al. 2005) — the anchor for starting below the standard range.
| Field | Protocol | Tier |
|---|---|---|
| Dose ladder | Quarter-tablet (25–50 mg) → half (100 mg) — start below the standard 100–200 mg, unlike general population dosing | C |
| Step duration | Single-day probes, spaced ≥48 h, before consecutive-day trials | C |
| Primary outcome | Alertness and next-day PEM check — this drug most tempts the “borrow from tomorrow” error the Ritalin outcome measure is built to catch | C |
| Stop conditions | Anxiety, overstimulation → above window · next-day crash → rung adds PEM debt; the drug is costing more than it pays | C |
10 Reading a null across all eight protocols
When a properly run trial returns null, walk the alternative-explanations table from the previous article in order of cost: neighboring dose → longer duration → re-scored outcome → mechanism argument. The protocols above are built so that the first three are usually already answered by the trial itself — the ladder covered neighboring doses, the step durations covered the timescale, the pre-defined outcome covered the measurement. What remains is the mechanism read, which is not a failure of the trial but its product: a drug whose targets are not rate-limiting in your case, crossed off at adequate cost.
That is the goal state of the whole search: every trial — positive or null — leaves a map with one more region colored in.
11 Assumptions and limits
Every number above is tiered for a reason. The dose bands come from the pharmacopoeia framework (Part 4) and, ultimately, from the primary paper’s inverted-U analysis — established pharmacology where marked E, community-and-model inference where marked C, and explicit bets where marked S. The cluster-correlation idea (windows tracking across mechanistically related drugs) is the framework’s core falsifiable prediction, not a result. If the framework is wrong, the dose-response data you collect with these protocols remains valid — the interpretive layer on top is what would fall. The motivation is not hypothetical: the controlled LDN trials that exist are single-dose and returned null on their primary endpoints (Due Bruun et al. 2024; Ologunowa et al. 2025) — no within-range dose-response trial has been run in any condition, which is precisely the gap these protocols are designed to fill (Loth 2026).
The protocols are templates, not prescriptions. Doses interact — with each other, with illness phase, with the patient’s own history. The illustrative numbers come from pharmacological reasoning and ME/CFS community experience, not from validated trials; the interpatient extrapolation is hypothesis. Your prescriber sets the actual numbers; these protocols structure the conversation, and several of these drugs (compounded low doses, off-label use, stimulants) are inaccessible without one.
PEM risk is a trial variable, not a footnote. Every rung of every ladder is an experiment conducted on an energy budget that does not forgive overexertion. Exertion during a dose trial is exertion during an experiment on your own threshold — the protocols’ short durations exist partly because of this.
Patient-to-patient analysis grounded in the pharmacological framework of Loth (2026) — not medical advice. Every trial requires prescriber agreement. Dose bands cross-reference Part 4 of this series; strategy and failure modes are covered in the companion article, Dismiss the Dose, Not the Drug.
Dose-response series: Part 1: Why More Isn’t Better · Part 2: The Inverted-U Is Not One Thing · Part 3: Your LDN Dose Is a Diagnosis · Part 4: The Pharmacopoeia · Part 5 (this post) · Related: Dismiss the Dose, Not the Drug