Any Promoter Will Do: Auditing a Sequence-to-Expression Model on Patient-Derived Data

Deep models that predict mRNA abundance from promoter sequence work well on reference genomes, but it is not obvious that they keep working when the input is a patient's own exome. We audit a publicly released convolutional model of this kind, trained on patient-derived whole-exome data for 15,370 human genes. The released checkpoint reports $R^2 = 0.329$ and we reproduce it (0.330). We then add the baselines and ablations the original release did not include. Three findings follow. First, evaluating a single released checkpoint understates the model: retrained networks reach $R^2 = 0.352 \pm 0.018$, and two of three significantly beat gradient boosting on eight scalar mRNA-stability features ($\Delta R^2 = +0.05$, $p = 0.01$), whereas the released checkpoint does not ($p = 0.39$). Second, and more surprisingly, the advantage survives destroying the promoter. Swapping promoters between genes, shuffling each promoter while holding its dinucleotide content fixed, and scrambling it to bare base composition all change $R^2$ by less than 0.010 in either direction, with no consistent sign, against the 0.029 that removing the promoter altogether costs. Shifting the input window by thousands of base pairs is almost as free, in seven of eight model-window combinations. What the branch supplies is scale, not information: zeroing it largely preserves the ranking of predictions while collapsing their spread, the signature of a calibration term. We also note that the released files carry one row per gene with no per-sample column, so the artifact is a gene-level dataset derived from patient sequencing rather than a per-patient one. Third, the architecture fails to train from a cold start in 3 of 6 attempts, collapsing to constant mean-prediction through a two-unit ReLU bottleneck, a failure the original release also reported but did not diagnose. We argue these results point at the data-construction pipeline rather than the architecture as the limiting factor.