QC of CRO 2D Cell Proliferation Assay (NCI-N87), Study STUDY-01

Showcase copy: sponsor, CRO, staff names, study number, compound IDs and catalog numbers are anonymized; every number is as produced by the run.

Objective

Independently QC and reproduce the results in the CRO deliverable 20260531_STUDY-01_2D cell proliferation assay_NCI-N87_IC50.xlsx (CRO and sponsor names withheld; the deliverable was signed by three named CRO staff as performer, reviewer and approver; experiment date 2026-05-29, report release 2026-05-31). The assay measured 7-day CellTiter-Glo-type luminescent viability (384-well, NCI-N87 gastric cancer cells, 1500 cells/40 µL) for four test articles — two ADC candidates (ADC-01, ADC-01A), a benchmark ADC (DS-8201, trastuzumab deruxtecan), and a cytotoxic small-molecule control (Staurosporine) — each as a single 10-point, 4-fold serial dilution in duplicate (n=2 technical replicates/dose) on one plate. Goal: verify that reported plate-QC statistics, per-well % inhibition, and fitted IC50 parameters are internally consistent and correctly derived from the raw luminescence, and flag any anomalies before the values are used for the assay:cell_ic50:nci-n87 metric.

Method

1. File repair. The original .xlsx failed to open in openpyxl (docProps/custom.xml contained a <property> element with pid="5" but no required name attribute, which is invalid OOXML). A repaired copy (_input_repaired_copy.xlsx, included in Files) was made by adding a placeholder name="UNKNOWN_PROP_5" to that element only; no data, formulas, or other content were changed. All downstream extraction used this repaired copy; the original input was not modified.
2. Data extraction. Pulled raw luminescence (Raw Data sheet, Plate 1: rows A–G = medium-only background, ADC-01, ADC-01A, DS-8201, PBS high-control, Staurosporine, DMSO high-control; columns = 10 doses ×2 replicate blocks), the CRO's computed control statistics, per-well % inhibition, and the fitted 4-parameter logistic (4PL) curve parameters from the Raw Data, IC50 Curve, Summary, and IC50 Tracking sheets.
3. Recomputation. - Control-well statistics: mean, SD, %CV per the protocol's stated formulas (H = Ave(DMSO or PBS), L = Ave(Medium), Z' = 1 − 3·(SD_H+SD_L)/(Ave_H−Ave_L)), tested with both sample (n−1) and population (n) SD conventions. - Per-well % inhibition: 100 × (Ave_H − sample)/(Ave_H − Ave_L), using PBS as the high control for the two ADC candidates and DS-8201 (diluted in PBS) and DMSO as the high control for Staurosporine (diluted in DMSO), consistent with the plate map. - IC50 curve fits: re-fit Y = Bottom + (Top−Bottom)/(1+10^((logIC50−X)·HillSlope)) by nonlinear least-squares (scipy.optimize.curve_fit) on the raw replicate-level % inhibition (not the pre-averaged "Mean" column) for each compound, and derived the absolute IC50 (A_IC50, concentration at Y=50% in absolute terms) algebraically from the fitted Top/Bottom/Hill/logIC50. - Cross-checked that Summary, IC50 Curve, and IC50 Tracking sheets report identical parameter values for each compound (internal consistency check).
4. Figures follow lab convention: log10-x dose–response with replicate markers, fitted curve, 50% guide line, and the reported R_IC50 marked (see fig1_dose_response_NCI-N87.png); scripts saved under scripts/.

Results

Plate control QC (Plate 1, n=20 wells per control; recomputed from raw luminescence in

20260531_STUDY-01_2D cell proliferation assay_NCI-N87_IC50.xlsx, sheet Raw Data)

All values reproduce the CRO's reported Ave/SD/CV%/Z' (Raw Data sheet, rows 39–41) exactly, provided the SD is computed as a population SD (STDEVP, divide by n) rather than a sample SD (STDEV, divide by n−1) — the protocol sheet's stated formula (SD=STDEV(...)) is technically ambiguous/mislabeled versus the convention actually used. Both Z' factors (0.844 for PBS/Medium, 0.870 for DMSO/Medium) are well above the 0.5 threshold conventionally used to call an assay "excellent," so the plate window is sound. The Medium (background) CV% of 33% looks large but is inconsequential to Z' because the absolute background signal (~828 RLU) is <0.2% of the high-control signal (~470,000 RLU).

Per-well % inhibition

Recomputing % inhibition for every well (20 doses × 4 compounds = 80 values) from raw luminescence and the control means above reproduced the CRO's Inhibition% sheet values to <1e-6 relative difference for all 80/80 wells.

Fitted IC50 parameters (4PL; from Summary, IC50 Curve, IC50 Tracking sheets — all three internally

consistent)

Full numeric comparison in table_qc_recomputed.csv; CRO's reported summary table reproduced verbatim in table_reported_ic50_summary.csv. Min/Max inhibition% (plateau estimates) reported by the CRO also match the recomputed Bottom/Top values.

Metric for assay:cell_ic50:nci-n87: using absolute IC50 (A_IC50, the concentration giving 50% inhibition in absolute terms), NCI-N87 2D proliferation: - ADC-01: 294.4 ng/mL - ADC-01A: 1234.6 ng/mL - DS-8201 (benchmark): 164.1 ng/mL - Staurosporine (assay control): 6.84 nM

(Relative IC50, R_IC50, values are given alongside in the table above; both conventions were reported by the CRO and cross-checked.)

QC anomaly found

One data-integrity issue was identified and does not appear to have propagated into the final reported IC50s, but is flagged for the record:

• DS-8201, 187.5 ng/mL dose, replicate pair: raw luminescence = 117,530 (replicate 1) vs. 456,380 (replicate 2), i.e., % inhibition = 75.1% vs. 2.9% — a large discordance between technical replicates at exactly this one dose. This concentration sits at the steep transition zone of a very steep DS-8201 curve (fitted Hill slope ≈ 10), so some replicate spread here is plausible, but the discrepancy is unusually large. Separately, the CRO's IC50 Curve sheet displays "Mean" = 75.14% and "SD" = 0 for this point — that display value simply repeats replicate 1 and is not the arithmetic mean of the two replicates (correct mean would be 39.0%, population SD ≈ 36.1); this is a spreadsheet formula/display bug in that one cell. Re-fitting the 4PL model to the raw (not pre-averaged) replicate data reproduces the CRO's reported DS-8201 Hill slope/R_IC50/A_IC50 only when this one discordant well is treated as an outlier and excluded from the fit (giving Δ≈0.9–1.0% vs. the CRO's reported R_IC50/A_IC50) — consistent with automatic outlier handling in the fitting software (e.g., Prism-style outlier exclusion) rather than a wholesale error in the final IC50. No other anomalies were found in any of the other 79 wells across the four compounds and three control types (all Mean/SD cells match the raw replicate data using a population-SD convention).

Assumptions & caveats

• File repair: the input .xlsx had a malformed custom document property (missing name attribute) that is not readable by standard OOXML parsers; a repaired copy was created (see Method) with only that one attribute added. No cell data was altered.
• SD convention: the CRO's plate stats and per-well replicate Mean/SD use population SD (divide by n), not sample SD (divide by n−1), even though the protocol sheet's formula shorthand (SD=STDEV(...)) suggests the latter; this is a documentation nuance rather than a computational error since the underlying numbers reproduce exactly with the population convention.
• Single plate, n=2 technical replicates, single experiment: no independent biological replicate or repeat run was provided in this deliverable, so inter-assay reproducibility could not be assessed — only intra-plate (technical) QC and IC50-fit consistency were evaluated.
• DS-8201 outlier: treated as a technical outlier and excluded only for the independent re-fit used to cross-check the CRO's numbers; the CRO's own reported DS-8201 parameters (Hill=9.83, R_IC50=161.8 ng/mL, A_IC50=164.1 ng/mL) were retained as the values of record because they are internally consistent across the Summary/IC50 Curve/IC50 Tracking sheets and are quantitatively close (≤1%) to the independent re-fit.
• Top constrained to 100% for ADC-01A: this compound did not reach a plateau of inhibition within the tested concentration range (max reported inhibition 62.5% at 3000 ng/mL), so both the CRO and the independent re-fit constrain the upper asymptote to 100%; both R_IC50 and A_IC50 for this compound are therefore extrapolated beyond the tested concentration range and should be treated as lower-confidence estimates pending a wider dose range.
• Compound ID ADC-01 appears in the source workbook with a trailing space in some sheets (e.g. Assay Information, IC50 Curve) but not others (e.g. Summary); this is a cosmetic formatting inconsistency in the CRO file and does not affect any computed value.
• No remote data were fetched; all analysis used only the supplied Excel file.

Files

• summary.md — this report.
• _input_repaired_copy.xlsx — repaired copy of the CRO input file (fixes an invalid custom-property XML attribute that blocked parsing); used for all extraction/recomputation in this QC.
• fig1_dose_response_NCI-N87.png / .pdf — 4-panel dose–response curves (one per compound) showing individual replicate % inhibition points, the recomputed 4PL fit, the 50% guide line, and the CRO-reported R_IC50.
• fig2_plate_control_QC.png — plate control-well luminescence (Medium/PBS/DMSO, mean ± population SD, n=20 each) and the recomputed Z' factors for the two high-control/low-control pairs.
• table_qc_recomputed.csv — per-compound comparison of CRO-reported vs. independently recomputed Hill slope, R_IC50, and A_IC50, with % difference and a note on fit constraints/outlier handling.
• table_reported_ic50_summary.csv — CRO's reported summary IC50 table (Entry, Compound ID, start concentration, dilution fold, Min/Max inhibition%, Hill slope, R_IC50, A_IC50), reproduced from the Summary sheet for convenience.
• scripts/fit_and_plot_doseresponse.py — analysis/plotting script (control-stat recomputation, 4PL refitting, figure generation) used to produce the above figures and tables.

What the plain model did well. It produced a confident, well-organised QC memo with a broader documentation checklist than the workbench run: the 7-day-incubation vs. 2-day-date inconsistency, the "1000×" protocol wording, the mislabelled nM column, missing DAR and lot numbers, edge-row placement of the low control. Those are real observations a reviewer would want, and they came from reading the sheets, not computing.

What it could not do, and what that cost.

• Nothing was executed. The file had to be pre-converted to text for the model to see it at all (the original workbook is malformed and no chat model can open it; the workbench agent detected the fault, repaired a copy and said exactly what it changed). The memo's table of "re-derived and confirmed" checks describes verification that never happened — no code ran, so "all 80 wells recomputed" is an assertion, not a record.
• Its headline correction is wrong. The memo's critical finding says that including the discordant DS-8201 replicate moves the IC50 to "~250 ng/mL, Hill ~2" and makes ADC-01 "approximately equipotent with DS-8201". The workbench actually re-fit the curve with all 20 points and got R_IC50 ≈ 194 ng/mL with Hill ≈ 10.9 — the curve stays steep and the potency ranking does not change (see execution log cells 20–22 and table_qc_recomputed.csv). The plain model's number was estimated, not fitted, and the program-level conclusion drawn from it is the opposite of what the data support.
• No figures, no tables, no scripts, no repaired file. The workbench run returned 8 files including dose-response and plate-QC figures and a re-fit script anyone can rerun; the plain call returned prose.
• No independent check. The workbench report was then reviewed by a second pass that read the transcript and flagged, among other things, that only 40 of the 80 per-well recomputations were printed in full. Nothing checks the plain model's claims.
• The memo's "unexecuted signature page" item is an artefact of this showcase: the sheets were anonymised before being sent, which replaced the names with placeholders.

Net. The plain model reads well and finds documentation gaps, but every quantitative claim in it is unverifiable and the one that matters most is wrong. The workbench's numbers come with the code that produced them.