Computational prediction of pharmacological chaperone rescuability in GPCR missense variants with cross-protein transfer from V2R to Rhodopsin

Pharmacological chaperones (PCs) rescue the surface expression of destabilizedprotein variants by binding and stabilizing the target protein. This mechanismaffords a promising therapeutic strategy for rare diseases caused by proteinmisfolding. Recent deep mutational scanning (DMS) experiments have mapped PCrescue across entire mutational landscapes. Yet, to our knowledge, no DMS-scalecomputational predictor benchmarked on rescue landscapes exists. Here, we developa machine learning approach to predict PC rescuability of missense variants using acompact, transferable feature set. We used 2,133 folding/trafficking-defective V2Rvariants treated with tolvaptan as a training benchmark. Our Random Forestclassifier integrates pathogenicity scores, thermodynamic stability forecasts, andbaseline expression severity. It achieves an AUROC of 0.751 (95% CI: 0.730–0.772),with no evidence of positional information leakage (leave-positions-out AUROC:0.745). Mechanistic analysis shows two classes of non-rescued variants: thosedisrupting the drug-binding site (moderate destabilization, small hydrophobicitychanges), and those with severe fold disruption exceeding the chaperone's stabilizingcapacity. A frozen V2R-trained model using four transferable features(AlphaMissense score, baseline expression, hydrophobicity change, and ESM1blanguage model score) transferred to Rhodopsin, a different GPCR treated with adifferent chaperone (YC-001), measured by a different assay. It achieved an AUROCof 0.843 (95% CI: 0.816–0.868; permutation p < 0.001). Results were robust acrossclassification thresholds (AUROC range: 0.783–0.843). Our findings show thatrescuability is partially predictable from a compact transferable feature setcombining computational and assay-derived information, and is mechanisticallydistinct from pathogenicity. This suggests computational pre-screening couldaccelerate identification of PC-responsive variants for rare disease therapeutics.