Science Machine · PCOS, audited

Gut bacteria and androgen genes: a registered null

Do the gut bacteria of women with PCOS carry fewer genes for making androgens from cortisol? We wrote the test down, registered it, then ran it once.

The data

Public stool sequencing (PRJNA530971) from a 2019 study of 50 women with PCOS and 43 controls at one hospital in China. palace/threads/gut-androgen.md:76-77 A first, unplanned look used 61 of the 93 samples, so it counts as exploration only. The registered test used the other 32: 20 PCOS, 12 controls. palace/threads/gut-androgen.md:20,37 experiments/gut_androgen_pilot/SET_C_RESULT.txt:1

The result

Reads per million read pairs, at 50% identity or more. AUC below 0.5 means lower in PCOS.

Gene groupPCOS medianControl medianAUC (95% CI)p
Androgen-making (des pathway). The registered test19.225.50.33 (0.14 to 0.54)0.1073
Androgen-inactivating (5-beta-reductase)7.911.10.42 (0.20 to 0.63)0.4769
Control expected higher in PCOS288.9209.80.79 (0.59 to 0.96)0.0052
Control expected flat1.21.00.50 (0.30 to 0.70)1.0000
Groups G5, G1, G8 and G6 in the source. experiments/gut_androgen_pilot/SET_C_RESULT.txt:6-25

Registered hypothesis: not consistent. The direction held, but the interval includes 0.5. experiments/gut_androgen_pilot/SET_C_RESULT.txt:23-25

In the words written into the plan before the test: this study could not detect a difference of the size it was able to see. It is not evidence that no difference exists. experiments/gut_androgen_pilot/COMPARISON_PREREG.md:235-240

The test was small. 32 samples could only find a drop of about 45% reliably. The drop observed was 25%, about half the 47% the first look suggested: an effect picked because it was noticeable tends to shrink when tested on fresh data. palace/threads/gut-androgen.md:8,43

A lead, kept for later

An exploratory breakdown by gene put most of the separation in one gene, desC: 8 of 20 PCOS samples had no desC reads, against 1 of 12 controls. That is a presence-or-absence question, which needs far fewer samples to test. It needs its own registration and a new cohort before anyone treats it as more than a lead. palace/threads/gut-androgen.md:8,14

One disclosure

A code guard was meant to keep the 32 confirmatory samples closed until the plan was registered. Its own author ran the script once after recording the plan's commit but before the OSF registration, to confirm the guard opened. Only group counts and sequencing depth were seen, no outcome value. This is disclosed in the registration, and the guard now requires the OSF link as well. palace/threads/gut-androgen.md:47

Limits