From Hypothesis to Protocol: AI-Assisted Rare Disease Trial Design and Site Feasibility

Designing a Phase 2/3 trial for a rare disease is not a linear problem. The statistical question — how many patients do I need? — cannot be answered without a credible enrollment timeline, and a credible enrollment timeline cannot be constructed without knowing which sites can actually deliver patients within the planning horizon. BioMate resolves this interdependency by chaining three analyses into a single auditable workflow: power calculation, site recruitment feasibility, and protocol summary generation.

Why Rare Disease Trial Design Is Structurally Different

FDA guidance on rare disease drug development (2019) acknowledges explicitly that "traditional approaches to demonstrating effectiveness may be modified in the rare disease setting."¹ The problem is not just regulatory — it is statistical. With a prevalent population of 14–20 per 100,000 for myasthenia gravis (MG),⁶ a sponsor designing a Phase 3 confirmatory trial cannot draw on the comfortable sample size calculations that apply to common conditions. Every assumption — effect size, standard deviation, dropout rate, screen failure rate — carries disproportionate weight when n is measured in the low hundreds rather than thousands.

The ICH E9(R1) estimand addendum² adds a further layer of complexity: the treatment policy estimand versus the hypothetical estimand behave differently under intercurrent events that are far more frequent in a rare, heterogeneous disease population with no validated natural history database. Sponsors who specify the wrong estimand early may find themselves redesigning the study after the first DSMB review.

What is typically missing from the sponsor's planning toolkit is an integrated, quantitative workflow that moves from a published effect size estimate directly to a protocol-ready statistical section — with every assumption justified by a citable source and every enrollment projection grounded in real site-level data.

Step 1: Statistical Power Calculation From Published Effect Sizes

The worked example here uses myasthenia gravis with the MG Activities of Daily Living (MG-ADL) scale as the primary endpoint — a validated 8-item patient-reported instrument scored 0–24, where higher scores indicate greater impairment.³ The ADAPT trial of rozanolixizumab (NEJM 2023) reported a least-squares mean difference of approximately 3.4 MG-ADL points versus placebo in the primary endpoint population, with a pooled standard deviation near 3.8 points.⁴ The earlier ADHERE study of efgartigimod reported comparable effect magnitudes, providing convergent evidence for the expected delta range in FcRn inhibitor trials.⁵

BioMate's power calculation module ingests these published values and generates power curves across a range of assumed effect sizes, applying a two-sided t-test with continuity correction and a 15% dropout allowance consistent with the observed retention rates in ADAPT. The analysis considers three scenarios: a conservative delta of 1.0 MG-ADL point (minimally clinically important difference), a central estimate of 1.5 points, and a liberal estimate of 2.0 points matching the ADAPT population-level effect in a more restricted patient subgroup.

At the central estimate of 1.5 MG-ADL points, a two-arm parallel trial with alpha 0.05 (two-sided), SD = 3.8, and 15% dropout allowance requires approximately 120 enrolled patients per arm (n ≈ 240 total) to achieve 80% power. A more pragmatic design of 90 per arm is shown here as a worked example, reflecting a deliberate power trade-off (approximately 75% power) against enrollment feasibility — a common decision in rare disease development where the eligible population is limited. BioMate's output includes the full parameter set, the software implementation of the power calculation (validated against published R pwr package outputs), and a structured justification paragraph referencing each input assumption to its source trial.

"Every effect size assumption in the power calculation is traceable to a published Phase 3 dataset. The system does not interpolate from preclinical data or use internal estimates that cannot be cited in a regulatory submission."

Step 2: Site Recruitment Feasibility From Real-World Enrollment Data

The power calculation produces n = 180 as the target enrollment figure. BioMate then passes this number directly to the site feasibility module, which models realistic enrollment timelines using three inputs: site-level MG patient prevalence derived from Orphanet and published epidemiology,⁶ historical screen failure rates from published MG trial reports (approximately 25–35% based on ADAPT screening data), and the competing trial landscape from ClinicalTrials.gov at the time of planning.

For a geographically distributed neuromuscular trial, the model estimates available patients per site per year as a function of the site's catchment population, the MG prevalence rate, the proportion of patients who are treatment-eligible (MGFA class II–IVa, on stable acetylcholinesterase inhibitor therapy), and the proportion willing to be randomized given competing trial availability. The resulting enrollment rate per site per month is typically 0.8–1.5 patients for a large academic neuromuscular center and 0.2–0.5 for a community site.

Across a network of 14 sites — 9 in the United States and 5 in Europe — the model projects a combined enrollment rate of 8–10 patients per month at steady state, reached approximately 6 months after site activation. At this rate, the target enrollment of 180 patients is achievable within 22–26 months of first patient in, with a 12-month follow-up period yielding a primary database lock at 34–38 months post first patient in.

Step 3: Audit-Ready Protocol Summary With Parameter Justifications

The outputs of Steps 1 and 2 — the power calculation parameters, the chosen sample size, the enrollment timeline, and the site network configuration — are passed to BioMate's protocol summary generator. This produces a structured document in the format required by FDA and EMA for the Statistical Analysis Plan (SAP) and protocol synopsis sections: estimand specification per ICH E9(R1),² sample size justification with referenced sources, enrollment timeline with per-region projections, and a summary of the competing trial landscape that contextualizes the feasibility assumptions.

Chaining as the Core Capability

What distinguishes this from running three separate tools is the structured provenance that connects each step. The n = 180 figure passed from Step 1 to Step 2 is not copied and pasted — it is a structured output with a full parameter record: alpha level, beta level, estimated standard deviation, the specific published trial it was drawn from, and the dropout assumption applied. When the feasibility module generates its enrollment timeline, it knows exactly what enrollment target it is solving for and can flag immediately if the projected network cannot achieve it within the planning horizon. When the protocol summary is generated, it has access to the full parameter chain and can produce a single self-consistent justification narrative rather than stitching together separate documents.

For regulatory reviewers, this means the statistical section of the protocol, the feasibility appendix, and the enrollment projections in the IND covering letter all reference the same underlying assumptions. Discrepancies between these sections — which are a common source of FDA Information Requests — are structurally prevented rather than caught by manual review.

What This Changes for Clinical Development Teams

A conventional protocol development process for a rare disease trial involves a biostatistician, a clinical operations feasibility analyst, and a medical writer working in sequence over 8–12 weeks to produce the statistical analysis plan, the feasibility report, and the protocol synopsis. Each handoff introduces the possibility of inconsistent assumptions. The statistical section may be based on a different effect size estimate than the one the operations team used to project enrollment. The protocol synopsis may reference a sample size that was updated after the feasibility report was written.

BioMate's chained workflow compresses the exploratory phase of this process — the iteration over effect size scenarios, the sensitivity analysis on enrollment timelines, the drafting and redrafting of justification language — from weeks to hours, while preserving the full audit trail that regulatory teams require. The biostatistician, clinical operations lead, and medical writer still own the final review and approval. What they are reviewing is a complete, internally consistent draft rather than a set of disconnected working documents.

---