Cell-line-derived and published transcriptomic signatures benchmarked against transcriptome-based clinical axes for neoadjuvant response prediction in breast cancer

Abstract Motivation Signatures from cell-line drug screens are deployed to patient transcriptomes to predict response. In breast cancer this may re-encode routine clinical variables rather than drug-specific biology. Results We tested whether ten transcriptomic signatures, scored without refitting, add pathological complete response information beyond four transcriptome-derived axes in five neoadjuvant cohorts. The axes are proxies for proliferation, hormone-receptor status, HER2 and subtype, and alone reached an area under the receiver operating characteristic curve of 0.67–0.79, 0.67–0.77 in the two held-out cohorts. No signature added value after multiplicity correction, and for the 60-gene doxorubicin signature the pooled added odds ratio per standard deviation was 0.98 [0.81, 1.18], P = 0.84. GSE41998 was the exception. Within the ten tests on that cohort, two immune scores and a second cell-line doxorubicin signature survived correction (false-discovery rate ≤ 0.061). At 80% power the minimum detectable residual area under the curve was 0.60–0.75, so smaller effects were not detected rather than excluded. These analyses are exploratory. A two-gene score, MKI67−ESR1, outperformed the 60-gene signature on held-out data (pooled difference in area under the curve +0.231 [0.172, 0.289], P = 1.31 × 10−14), so such signatures need validation against transcriptome-derived clinical baselines. Availability and implementation Code and analysis outputs are available at https://doi.org/10.5281/zenodo.20726402.