Blast percentage confounds transcriptome to drug-response associations in Beat AML: DHODH and sorafenib as a worked example

Abstract Beat AML pairs ex vivo dose response for 125 compounds with transcriptomes from the same patients and is widely mined for markers of drug sensitivity. Those analyses adjust for general drug sensitivity, differentiation state, proliferation and genotype, but not for the blast content of the specimen, and a recent four-cohort study concluded that such an adjustment was not warranted. Across 12,000 gene-by-drug pairs, 35 per cent of the associations significant under the usual adjustment set lost significance when blast percentage was added, against 3.7 per cent gained. The pairs came from 800 genome-wide genes and the 15 most tested compounds, with the sample set fixed per drug so that nothing turns on a change in n. Blast percentage correlated with both transcript abundance and residual drug response. The comparable figures for tumor purity in solid tumors are about 14 and 11 per cent. A worked example shows the cost to a candidate marker. Under the usual model DHODH expression was the strongest correlate of sorafenib response among the genes tested, sorafenib ranked first of 125 drugs (rho − 0.259, FDR 6.5 × 10⁻⁷) and the empirical p against an expression-matched genome-wide null was below 0.001. With blast percentage added, rho fell to − 0.186 (n 414), sorafenib to rank four and the empirical p to 0.012. What survived was selective: five of the six other FLT3-directed compounds showed nothing. Blast percentage belongs in the adjustment set for these analyses. It is already in the diagnostic record and costs nothing to include.