k-dense-ai/scientific-agent-skills/analytical-method-validation

Analytical Method Validation

When to use

Any time the question is whether an analytical procedure is fit for its intended purpose: designing a validation study, evaluating validation data, verifying a compendial procedure, transferring a procedure to another laboratory or instrument, or defending any of these in a report.

The two rules

1. Establish which framework governs before designing anything. The same assay validates differently under ICH Q2(R2), USP <1225>, ICH M10, CLSI EP, and ISO/IEC 17025. They differ in which characteristics are required, how the studies are laid out, and whether numeric acceptance criteria are supplied at all. Blending them produces a protocol that satisfies none of them.

2. State acceptance criteria before collecting data. Criteria chosen after seeing results are not acceptance criteria, and deciding them post hoc is a standing audit finding. ICH Q2(R2) deliberately supplies almost no numeric criteria — they have to come from the specification, the analytical target profile (ICH Q14 section 3), or development data. ICH M10 is the exception: it supplies explicit numbers, and they differ between chromatographic assays and ligand binding assays.

Scope

This skill plans studies, computes the statistics correctly, and structures the documentation. It does not decide that a procedure is validated, release a batch, accept or reject a run, close an investigation, or substitute for the analyst, the technical reviewer, the quality unit, or the regulator. Every script reports computations and supported findings; only the responsible reviewers can make the fitness-for-purpose decision. JSON uses null for unavailable statistics.

Copyright boundary

ICH guidelines are published openly and licensed for reuse with acknowledgement, so their requirements are encoded directly in this skill. USP general chapters, CLSI EP documents, and ISO standards are copyrighted and paywalled. For those, this skill supplies the designation, scope, and where to obtain an authorised copy — never the text, never invented thresholds. Do not ask an agent to retrieve, transcribe, or reconstruct their content. If a number matters and it lives in a paywalled document, read it from the authorised copy.

Frameworks

cd skills/analytical-method-validation/scripts
python3 plan_validation.py --list-frameworks

Q2(R2) replaced Q2(R1) in November 2023 and restructured the characteristics. Range is now the parent characteristic (section 3.2), containing response (linearity) and validation of lower range limits (DL/QL). Accuracy and precision are section 3.3 and may be evaluated in combination against a single criterion. Robustness is treated as a development activity and cross-refers to ICH Q14. Multivariate procedures are addressed explicitly (2.5 and 3.2.2.3), and Annex 2 adds worked examples for techniques Q2(R1) never covered — quantitative ¹H-NMR, NIR, quantitative LC/MS, qPCR, biological assays, and particle size. A Q2(R1)-shaped protocol — a flat list of linearity, range, accuracy, precision, specificity, LOD, LOQ, robustness — is out of date. Note also the error correction dated 30 November 2023 to Table 5 and Tables 6–11.

Scripts

cd skills/analytical-method-validation/scripts

Commands below use placeholder data filenames; supply your own controlled data. The printed numerical examples were re-executed with synthetic repository fixtures under tests/analytical-method-validation/fixtures/ on Python 3.13.3. They are computational smoke tests, not evidence that a physical assay is validated. All scripts take --format table|tsv|json. Provenance, guideline citations, and caveats go to stderr; data goes to stdout, so > out.tsv keeps them separate. Exit code is 0 for no findings, 1 when findings were raised, 2 for bad input — so any of them can gate a workflow.

Version 2.0 tightens the numerical input contract: QL confirmation needs explicit bias/CV limits; intercept-based limits need independent curves; M10 runs need complete labelled calibrator/QC records; grouped accuracy uses group means; unavailable JSON statistics are null. Re-run saved analyses rather than comparing their exit codes with version 1.x unchanged.

Workflow

1. Fix the framework and the required characteristics

python3 plan_validation.py --framework ich-q2r2 --attribute assay --technique hplc --range-use assay

Q2(R2) Table 1 decides what is required from the measured attribute, not from the technique. For an assay: specificity, response, accuracy, repeatability, intermediate precision. For a limit test: specificity and DL only. For an identity test: specificity alone. Attributes accepted include assay, impurity (quantitative), impurity-limit, and identity.

Reportable range comes from the specification. Q2(R2) Table 2 gives worked examples — 80–120% of declared content for an assay, 70–130% for content uniformity, reporting threshold to 120% of the specification for an impurity.

2. Generate the protocol and fill in the criteria

python3 plan_validation.py --framework ich-q2r2 --attribute impurity --protocol > protocol.md

Every bracketed field is a decision to make and record before data collection. The protocol skeleton deliberately refuses to pre-fill acceptance criteria for Q2(R2) work, because there is no defensible default.

3. Evaluate the response

python3 check_response.py -i calibration.csv --max-back-calc-error 2

Input is level,response, one row per injection; repeated rows at the same level are replicates, and supplying them is what makes the linearity test possible.

Real output from a curve that a coefficient of determination would wave through:

statistic                           value
distinct levels                     5
slope                               166.6000
intercept                           2495.0000
intercept CI includes 0             no
coefficient of determination (r2)   0.9830
lack-of-fit F                       469.5294
lack-of-fit p                       1.5139e-06
runs test p                         0.0492
 
level     n  mean_response  mean_back_calculated  relative_error_pct
50.0000   2  10075.0000     45.4982               -9.0036
75.0000   2  15150.0000     75.9604               1.2805
100.0000  2  20050.0000     105.3721              5.3721
125.0000  2  24050.0000     129.3818              3.5054
150.0000  2  26450.0000     143.7875              -4.1417

r² = 0.983 and the model is unusable: −9.0% back-calculated error at the bottom of the range, lack-of-fit p = 1.5 × 10⁻⁶, non-random residual signs. r² is not evidence of linearity — it rises with range and is nearly insensitive to curvature. The lack-of-fit F test against pure error and the residual pattern provide diagnostics, which is why Q2(R2) 3.2.2.1 asks for an analysis of the deviation of points from the line rather than a correlation coefficient alone.

Use --weight 1/x2 only when supported by the variance model for a wide-range curve. The script flags heteroscedasticity when the residual variance in the top third of the range exceeds the bottom third by more than 10×, as a heuristic warning. Unequal variance does not itself bias OLS coefficients; choose weights from a justified error model, then assess low-end performance. Weighted lack-of-fit assumes the weights represent inverse variances. A significant test is not a practical acceptance criterion.

4. Evaluate accuracy and precision

python3 check_accuracy_precision.py -i ap.csv --accuracy-limit 2 --rsd-limit 1.0 --design-check assay

Input is level,measured,group, where group is the intermediate-precision factor — day, analyst, or instrument. Record independent sample-preparation IDs separately from repeat injections: reinjecting one preparation estimates injection repeatability, not the whole procedure. The bundled one-way model estimates one between-group component. If day, analyst, and instrument change together, it cannot identify their separate contributions; use a planned crossed or nested study and a matching model when those components matter. Grouped accuracy intervals use independent group means with equal group weighting; individual injections are not counted as independent evidence across days. To check the six-at-100% design alternative, state --test-concentration 100 when nominal values are percentages (or the actual concentration).

level  component                       sd      rsd_pct  df      ci90_low_sd  ci90_high_sd
100    repeatability (within group)    0.0707  0.0705   3       0.0438       0.2065
100    between-group                   1.6515  1.6458   2       n/a          n/a
100    intermediate precision (total)  1.6530  1.6473   2.0037  0.9554       7.2821

Repeatability of 0.07% RSD looks superb; intermediate precision is 1.65%, twenty-three times larger, because the variability lives entirely between days. Reporting the within-day figure as the procedure's precision would understate routine performance by more than an order of magnitude. This is why the script fits a one-way random-effects model rather than pooling.

Two traps the script handles for you:

• Precision is estimated within each level, never pooled across levels. Pooling 80/100/120%

results into one standard deviation turns the range itself into apparent imprecision. The script reports per level; even pooling percent recoveries can confound level-specific bias with imprecision.

• `--require-ci-within-limit` enforces that the whole confidence interval sits inside the

limit, not just the mean. Q2(R2) 3.3.1.4 asks for the interval to be compatible with the criterion; a mean that scrapes inside on six replicates has not demonstrated much.

5. Establish DL and QL, and confirm them

python3 check_detection_limits.py --calibration lowcal.csv --blanks blanks.csv \
    --confirm-ql 0.05 --confirm-data ql_check.csv --reporting-threshold 0.05 \
    --confirm-accuracy-limit 10 --confirm-rsd-limit 10

approach                                          sigma   slope      DL      QL
sd-and-slope (sigma = residual SD of regression)  7.2816  5033.3490  0.0048  0.0145
sd-and-slope (sigma = SD of 8 blanks)             3.7702  5033.3490  0.0025  0.0075