GPT 6.2 Astra swarm agents found a 10× conversion error in a breast cancer trial
Arth Singh
, 6 min read
Summary
197.3 ng/g of total SN-38 converts to 0.523 µM. The trial and a Gilead medical information table print 0.0523 µM. We checked the units, reconstructed the patient groups and published the data and code.
How much of a cancer drug reaches a breast tumor that has spread to the brain? Balinda and colleagues had a rare way to measure it directly. They gave patients sacituzumab govitecan the day before brain surgery, then measured its payload, SN-38, in the removed tissue. In the breast cancer cohort, the Results report a median of 197.3 ng per gram of tissue. Using the authors’ own conversion factors, that is 0.523 µM. The paper prints 0.0523 µM, and Gilead’s medical information for clinicians repeats it. Original trial · Gilead Medical Information, Table 8, p. 8
Twenty GPT-6.2 Astra agents investigated open questions in breast cancer using Optimism, our research coordination system. Daria flagged the conversion and a second discrepancy in the abstract. Another agent extracted the measurements independently, and a third checked the arithmetic. The trial team made those checks possible by publishing individual tissue and blood measurements. Those records also explain why one apparent additional error deserves a closer look. Supplementary Table 2, p. 2
Where the factor of ten comes from
The trial enrolled 13 patients with breast cancer brain metastases and 12 with recurrent glioblastoma. The investigators obtained paired tissue and serum measurements from 24 of the 25 participants. For brain tissue they assumed a density of 1.04 g/mL. They used a molecular weight of 392.4 g/mol for SN-38. With those two quantities, the breast tissue conversion takes three steps. Study methods and Results
| Step | Calculation |
|---|---|
| Convert drug mass per gram of tissue to drug mass per milliliter | 197.3 ng/g × 1.04 g/mL = 205.192 ng/mL |
| Express the result per liter | 205.192 ng/mL = 205.192 µg/L |
| Divide by the mass of a micromole of SN-38 | 205.192 µg/L ÷ 392.4 µg/µmol = 0.5229 µmol/L |
There is a quick check in the same paragraph of the paper. The glioblastoma tissue median was 104.5 ng/g, a little over half the breast tissue value. It converts to 0.277 µM, which the paper reports as 0.28 µM. At the same density, the higher breast tissue measurement must also have the higher molarity. The printed figures put them in the opposite order. The breast serum figure provides a second check. Serum is already measured per milliliter, so 2,462.4 ng/mL divided by 392.4 gives about 6.275 µM, close to the published 6.27 µM. Trial Results
The abstract’s median matches a mean
The abstract calls 249.8 ng/g the breast tissue median. Supplementary Table 2 lists 13 breast cases. For case 23, the surgical specimen showed radiation effects and the pathologists found no malignant cells. The paper reports that case’s tissue level separately. Setting it aside leaves these 12 published values, sorted from lowest to highest. Supplementary Table 2 · Supplementary Note 1, p. 21
| Lower six tissue concentrations (ng/g) | Upper six tissue concentrations (ng/g) |
|---|---|
| 66.9, 86.5, 102.2, 118, 122, 182.6 | 212, 236, 308.3, 372.1, 539.4, 652 |
The middle two values average to 197.3 ng/g, exactly the Results median. The mean of the 12 is 249.83 ng/g, matching the abstract’s 249.8 after rounding. Including case 23 gives a median of 182.6 ng/g. A reader taking the figure from the abstract will therefore get a different statistic from a reader taking it from the Results. Download the transcribed trial data (CSV)
Why eight patients and twelve patients give the same median
The Results give a breast tissue range of 86.5–652 ng/g. Case 13 measured 66.9 ng/g, below that minimum. At first glance that looks like a third error. But Supplementary Note 1 identifies four patients, including case 13, who had no measurable disease after surgery. It names the eight patients eligible for the radiographic response assessment. We reconstructed both groups and compared them with the published summaries. Supplementary Note 1
| Breast cancer cases | Tissue median (ng/g) | Tissue range (ng/g) | Tissue IQR (ng/g) | Serum median (ng/mL) |
|---|---|---|---|---|
| All 13 with measurements | 182.6 | 21.2–652 | 206.1 | 2,266.8 |
| 12 after setting aside case 23 | 197.3 | 66.9–652 | 230.1 | 2,462.4 |
| 8 eligible for response assessment | 197.3 | 86.5–652 | 230.1 | 2,682.4 |
| Published in the Results | 197.3 | 86.5–652 | 230.1 | 2,462.4 |
The four cases removed from the 12 include two values below the middle pair and two above it. 182.6 and 212 remain the middle pair. Even the interquartile range (IQR) is identical using Tukey’s hinges, the medians of each half of the ordered data. In both groups the hinges are 110.1 and 340.2 ng/g, a difference of 230.1.
All three printed tissue statistics match the eight cases. The printed serum median matches the twelve. This gives the authors a precise question to clarify. Which patients contributed to each tissue and serum summary? The range could reflect the eight-person group, and the published summaries alone cannot establish which analysis the investigators ran. Trial Results · Patient-level measurements
The glioblastoma data illustrate why the endpoint matters. All 11 measured specimens give the reported median of 104.5 ng/g. Omitting the three rows marked for radiation injury would give 176.85 ng/g in eight specimens. The notes say that case 4 continued on study and was included in all analyses, while case 18 had different inclusion rules. The alternative median illustrates the effect of that choice. The authors’ pathology notes give reason to keep those cases distinct. Supplementary Table 2 · Supplementary Note 1
Where the numbers traveled
Gilead’s Table 8 pairs 197.3 ng/g with 0.0523 µM. It lists 13 tissue and 13 serum samples. Thirteen paired samples were indeed collected, but summarizing all 13 produces neither median in that table. Another column in the same row gives 0.662 µM as an observed SN-38 level. That agrees numerically with converting 249.8 ng/g using the stated density. The table does not specify how that entry was derived. Gilead Medical Information, Table 8, p. 8
The abstract’s number traveled separately. A 2025 review in Pharmaceuticals and a 2026 review in The Breast Journal both describe 249.8 ng/g as a median. There is also a record predating the final paper. At the 2020 San Antonio Breast Cancer Symposium, the original investigators reported a mean of 626 nM (0.626 µM) in the first four breast tumors assayed, with a range of 173–1,160 nM. That early mean gives context for the scale of the measurements. It comes from the same trial. 2025 review · 2026 review · Preliminary report by Brenner and colleagues
What changes for a reader of the trial
The paper compares tumor exposure with 0.00517 µM, the lowest concentration it cites for 50% inhibition in invasive breast cancer cell lines. Using the printed tissue molarity, the ratio is about 10 to 1. Using the recalculated molarity, it is about 101 to 1. The correction strengthens the paper’s numerical case for drug delivery. A researcher using the published molarity to benchmark another compound would otherwise start from a value tenfold too low. Trial Results and Discussion
The investigators measured total SN-38 in surgically removed tumor tissue. How much was available as free drug inside cancer cells is a different measurement. A cell-line comparison also cannot establish a dose, a survival benefit, or how many patients’ care changed. What the public evidence establishes is a tenfold conversion error, a mean labeled as a median, and a specific question about which patients contributed to each summary. Original trial
As of November 19, 2026, the publisher’s Crossmark record lists the article as current. We found no correction to these two numbers in that record. A correction to the molarity and the abstract’s label, together with the case counts for each summary, would give future readers one consistent set of figures to cite. Publisher’s update record
Run the reanalysis yourself
The technical note, CSV, and runnable Python are archived together under the DOI 10.5281/zenodo.24372633. Cite that record as Arth Singh (2026), Reanalysis of SN-38 concentrations in breast cancer brain metastases. Zenodo archive
Download the CSV and run this Python 3 code in the same directory. It uses the standard library. Each row contains the published breast case number, tissue and serum measurements, and two flags derived from Supplementary Note 1. The flags let you reproduce each row of our comparison table. Download the transcribed trial data (CSV)
import csv
from statistics import mean, median
with open("sn38-tissue.csv", newline="") as source:
cases = list(csv.DictReader(source))
groups = {
"All breast cases": cases,
"Excluding case 23": [
r for r in cases if r["pathology_consistent_with_recurrence"] == "yes"
],
"Response-evaluable": [
r for r in cases if r["radiographic_response_evaluable"] == "yes"
],
}
for name, rows in groups.items():
tissue = sorted(float(r["total_sn38_ng_g"]) for r in rows)
serum = [float(r["total_sn38_serum_ng_ml"]) for r in rows]
half = (len(tissue) + 1) // 2 # include the middle value in both halves for odd n
iqr = median(tissue[-half:]) - median(tissue[:half])
print(f"{name} (n={len(rows)})")
print(f" tissue mean {mean(tissue):.2f}, median {median(tissue):.1f}, "
f"range {min(tissue):g}–{max(tissue):g}, IQR {iqr:.1f} ng/g")
print(f" serum median {median(serum):.1f} ng/mL")
tissue_median = median(float(r["total_sn38_ng_g"]) for r in groups["Excluding case 23"])
molarity = tissue_median * 1.04 / 392.4
print(f"Tissue concentration: {molarity:.3f} µM")
print(f"Ratio to the cited minimum cell-line IC50: {molarity / 0.00517:.1f}")The clinical measurements are from Balinda and colleagues. We transcribed the breast cancer rows, encoded the case notes as flags, and calculated the summaries above. The source article and supplement are available under CC BY 4.0. Source publication · Supplementary Information · CC BY 4.0 license