A conservation score is an inference drawn from an absence, and that gap is where the models can go wrong. Nearly every tool in the audit asks one question of a position in the genome: has this stayed the same over millions of years? If it has, the software concludes the site was kept that way for a reason and that changing it must be bad [5]. Two unrelated situations produce the same stillness. Selection may have removed every variant that appeared there, or almost nothing may ever have appeared, because the site mutates rarely to begin with [9]. A model that cannot separate those carries an estimate of mutability inside its estimate of harm [6].
The heterogeneity involved is large and already partly mapped. Weghorn's earlier work put the mutation rate at the starting points of genes 35% above other regions [8], and other groups have turned up hotspots elsewhere [9].
The design decision worth underlining is the control. Ben Lehner's group, also at the CRG, made its mutations physically, hundreds of thousands of them across 500 human protein fragments, and measured the functional damage each one caused [12]. That readout comes entirely from direct measurement, independent of evolutionary conservation, which is what makes it usable as a check. It confirmed a genuine effect pointing the same way as the models' bias [13], and first author Hossameldin Ali reads that as genomes having evolved robustness against their own most frequent errors [20]. The Weghorn group then corrected for the real effect, and the skew was still there [15], which is what separates a claim about biology from a claim about software.
Scale is what lets the paper speak about gene families rather than curiosities: 13.5 million mutations across 6,659 genes [1] averages roughly 2,030 variants per gene [1]. On that base, the study reports DNA repair, cilia and sperm-function genes as likely overcalled for danger, and genes tied to intellectual disability and to conditions inherited from a single parent as likely undercalled [10]. Weghorn's summary of the two directions is "Both would be bad" [11].
How far this moves any real report is the open question, and the paper stops short of settling it: it can name the direction of the bias but not its size in any single result, and the score is one input among several that clinical experts use [17], alongside its other role in guiding treatment options and laboratory work on mechanism [19]. Weghorn's own expectation is that some variants could flip from harmless to harmful once mutation-rate biology is included, while most would shift more gradually [16]. My read is that this is a calibration fault rather than a collapse of validity, and for a specific reason: the error has a known direction and a control that measured it. What the audit does not hand you is a per-tool ranking or a per-case error rate, so nobody looking at a mid-range score today can say how far off it sits.
