Scientific Methodology

This paper describes the published standards and models each MolWard tool implements, and how outputs are verified — so a reviewer can trace every result to a citable method. It intentionally omits implementation detail (code, tuning, and optimisation), which is proprietary. This is a methodology overview, not a validation certificate.

1. Purpose

MolWard's premise is that computational tools used in regulated pharmaceutical work should be transparent and traceable, not black boxes. This document sets out, for each tool, the recognised guideline or published model it implements, and the way we confirm its output is correct.

2. Design principles

Validated chemistry, not opaque prediction. Each engine implements an established ICH/USP/FDA method or a peer-reviewed model, cited below.
Deterministic science. The scientific calculations are deterministic and reproducible; only the optional narrative is AI-drafted, and it is numerically guarded.
Human in the loop. Outputs are decision support for a qualified reviewer, never an automated regulatory decision.
Zero-retention. Confidential inputs are computed in memory and not stored (see our Security Statement).

3. Method basis, tool by tool

Tool	Method / model implemented	Primary reference
Degradation prediction	Forced-degradation pathway reasoning under ICH stress conditions; mass-balance framing	ICH Q1A(R2); ICH Q3A/B
ICH M7 toxicology	Structure-based hazard assessment for mutagenic impurities; cohort-of-concern; CPCA nitrosamine limits	ICH M7(R2); FDA nitrosamine guidance
ICH Q1E stability	Least-squares regression with poolability testing and guideline extrapolation rules for shelf life	ICH Q1E; ICH Q1A(R2)
ICH Q3 impurity limits	Reporting / identification / qualification thresholds; residual-solvent and elemental limits scaled to dose	ICH Q3A/B/C/D
Dissolution equivalence	Model-independent f2 similarity factor with bootstrap confidence; SUPAC change context	FDA/EMA f2 guidance; FDA SUPAC
Column selectivity	Hydrophobic-Subtraction Model; Factor-of-Similarity (Fs) column comparison over published column parameters	Snyder, Dolan & Carr HSM; USP/PQRI database
QbD DoE & design space	Classical designs, response-surface modelling with ANOVA, Monte-Carlo edge-of-failure	ICH Q8(R2); ICH Q9
ICH Q14 risk assessment	ATP-driven FMEA (RPN = S×O×D); physics-informed sensitivity for risk scoring; MODR robustness	ICH Q14; ICH Q9; ICH Q2(R2)
QC calculators	System suitability, LOD/LOQ, content-uniformity acceptance value, and outlier tests	USP <621>, <905>; ICH Q2(R2); Grubbs / Dixon
HED / MRSD dose scaling	Allometric body-surface-area scaling to human-equivalent dose and maximum recommended starting dose	FDA 2005 FIH guidance
PDE / ADE / OEL	Permitted-daily-exposure derivation with the adjustment-factor framework	ICH Q3C/Q3D; EMA shared-facility guideline
RegIntel	Retrieval-augmented question answering restricted to a curated corpus; every answer returns its verbatim source passage	ICH, 21 CFR, FDA guidances & annual reports

4. Verification approach

Each engine is checked by reproducing worked examples and reference values from the source standards and literature, and — where a legacy method exists — by cross-comparison against it. New releases are exercised against these reference cases before shipping. Verification confirms that a computed output matches the published method; it is distinct from, and does not replace, the customer's own computer-system validation.

5. Guard-railed AI methodology

Where a tool offers an AI narrative, the model operates under strict constraints: it may use only the numeric values from the computed result (a numeric guard flags any number that is not traceable to the result), it may cite only an allow-listed set of guidelines (a citation guard flags any others), it runs at low temperature under a role-locked prompt, and every output carries a mandatory "draft for expert review" disclaimer. The RegIntel assistant goes further: it answers only from retrieved source passages and returns the exact quoted text and reference for each answer, and declines when the corpus does not contain the answer.

6. Intended use & limitations

MolWard tools are decision-support aids for qualified scientists. They do not replace formal method validation (ICH Q2(R2)), a multivariate design-space study, wet-lab confirmation, or regulatory judgement. Representative models and reference data are used where noted; results give correct relative and methodological output for review, not a certified conclusion. The control strategy, validation, and regulatory responsibility remain with the sponsor.

7. Standards & references

ICH Q1A(R2), Q1E, Q2(R2), Q3A/B/C/D, Q8(R2), Q9, Q14, M7(R2); USP General Chapters <621>, <905>, <1220>, <1225>; FDA guidances on nitrosamine impurities, dissolution, SUPAC, and estimating the maximum safe starting dose (2005); EMA shared-facility (PDE) guideline; U.S. Code of Federal Regulations Title 21; the Hydrophobic-Subtraction Model (Snyder, Dolan & Carr) and the USP/PQRI column database; and the Grubbs and Dixon outlier tests. Full citations available on request.