A detectability floor for treatment-response labels in rheumatology transcriptomics

Objectives

Before-and-after transcriptomics in rheumatic disease often yields no reproducible treatment signal, though disease-state stratification is robust. We asked whether exposure and response-prediction labels fall below the detection floor of the designs defining them, while matched disease-state labels do not.

Methods

Eight public datasets were analysed (paired juvenile dermatomyositis n = 10; RA before/after methotrexate, n = 28 pairs; lupus nephritis n = 10; controlled exposures; a single-cell dose series). Response labels in current use (one benchmark trajectory, five published signatures) were audited in every analysable source cohort, using effect sizes (dz) against each design’s exact minimum detectable effect.

Results

Disease-state labels were detectable: the interferon signature declined in 10/10 juvenile dermatomyositis patients (p = 0.002; 59.5% of score variance). No exposure label cleared its design floor in any clinical cohort; a drug-matched methotrexate prior was nominally significant in juvenile dermatomyositis yet did not clear its floor (dz =-0.80; CI spanning it). Exposure became detectable only where prior–context and intensity were jointly favourable (multi-day target tissue dz = 1.24, q = 0.018; well-level pseudobulk dz = +3.70, floor 1.80). Audited response labels failed in 14 of 15 cohorts (93%; descriptive binomial 95% CI 68–100): eleven below their floors, three sign-reversed; the largest training cohort pooled to zero (dz = +0.07; I² = 69%). Effect size ordered reported AUCs monotonically.

Conclusion

Under conventional treatment intensities, exposure and response-prediction labels are either undetectable below the design floor or detected with a sign that does not replicate; disease-state labels are detectable. Part of the negative literature reflects a design floor.

Key messages

• Disease-state labels are detectable in paired rheumatologic designs; exposure labels are not.

• Audited response-prediction labels failed in 14 of 15 cohorts, mostly below design floors.

• A clean design, not a larger cohort, moves weak labels above the detection floor.