Protein language models lose their advantage on the genes where IVF fails

Protein language models learn from evolutionary conservation across species. A clinically important class of human genes has almost none. Maternal-effect genes such as TUBB8, PATL2, PADI6 and NLRP7 are expressed only in the oocyte and early embryo, several are primate-specific, and their phenotypes are largely invisible outside a fertility clinic: oocytes that never mature, oocytes that never fertilise, embryos that arrest. We test whether models trained on cross-species signal degrade where that signal is absent. We assembled 18 oocyte and IVF-failure genes and 10 deeply conserved disease control genes. Positives are ClinVar pathogenic missense variants; negatives are gnomAD r4 variants at allele frequency at least 1e-4, necessary because TUBB8 has 42 pathogenic missense variants in ClinVar and no benign ones. After validation against UniProt canonical sequences, 3,277 variants remain. We score each with ESM-2 650M using masked marginals, alongside a BLOSUM62 baseline, and report AUROC with stratified bootstrap intervals. ESM-2 reaches mean AUROC 0.903 (95% CI 0.851-0.946) on control genes and 0.619 (0.486-0.741) on oocyte genes, a gap of 0.284 (0.153-0.420). BLOSUM62 degrades by only 0.109 (-0.034-0.250), an interval crossing zero. The advantage of the language model over the substitution matrix is +0.196 on control genes and +0.021 on oocyte genes. Because the baseline does not degrade comparably, the effect is specific to the language model rather than to data quality. Leave-one-out across the four oocyte genes gives gaps of 0.267-0.302.