Kidney Week

Abstract: SA-PO0643

Computational Pathogenicity Prediction of Autosomal Recessive Aquaporin 2 (AQP2) Mutations (ArM)

Session Information

• Monogenic Kidney Diseases: Tubular and Other
  November 08, 2025 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Genetic Diseases of the Kidneys

• 1201 Genetic Diseases of the Kidneys: Monogenic Kidney Diseases

Authors

• Rodriguez, Esteban, University of South Florida Morsani College of Medicine, Tampa, Florida, United States

Esteban Rodriguez, DO

• Durr, Jacques A., University of South Florida Morsani College of Medicine, Tampa, Florida, United States

Jacques A. Durr, MD

• Amerson, John K., University of South Florida Morsani College of Medicine, Tampa, Florida, United States

John K. Amerson, DO

• Audi, Akram, University of South Florida Morsani College of Medicine, Tampa, Florida, United States

Akram Audi, MD

• Reddy, Purushottam, Tampa General Hospital Brooksville, Brooksville, Florida, United States

Purushottam Reddy, MD

• P Narayanankutty, Naveen, University of South Florida Morsani College of Medicine, Tampa, Florida, United States

Naveen P Narayanankutty, MD, FASN

Introduction

ArM pathogenicity typically requires homozygous/compound heterozygous NDI or functional assays. What is the role of emerging AI/machine learning in predicting ArM pathogenicity for VUS, variants of unknown significance?

Case Description

A cognitively delayed adult female patient presented with severe hypernatremia (177 mmol/L) and 15L/d dDAVP-unresponsive polyuria. She had normal pituitary MRI, high copeptin, and a history of recurrent infant dehydrations (aunt's report), but negative family history. She improved on chlorthalidone/low Na diet.

She was homozygous for AQP2 ArM c.499T>C (p.Ser167Pro), a VUS by ClinVar. Wildtype (wt) AQP2 is a homotetramer of a 271 aa-long 6 TransMemberane helix chain with S167 in the middle of TM5 (UniProt P41181). AlphaMissense gives this ArM a high pathogenicity score, 0.98/1! We modeled its 3D structure using AlphaFold3. S167P abolishes S167’s H-bonds to Q140 (TM4) and its intraTM H-bond (red arrow). This, with P’s steric distortion likely breaks the α helix or misfolds the ArM, hence the transition in per-residue confidence color from dark- (Very High, wt) to light blue (Confident only, ArM below P167). AI predicts novel C144(TM5)-T54 (chain D) H-bond (green arrow). Missense3D predicts 3 buried H-bond losses and a switch from buried S167 to solvent accessible P167. gnomAD lists only 5 heterozygous ArM alleles/1609744. Strong in-silico pathogenicity (REVEL, PolyPhen-2, DynaMut, Varsome, and others with meta-scoring) contrasts VUS database classification due to absent NDI phenotype reports.

Discussion

The scarcity of in vivo NDI phenotypes and limited access to in vitro ArM functional assays require in silico VUS pathogenicity predictions by emerging AI/neural nets using evolutionary/ensemble/protein-LLMs. Yet, while “AI solved the fold,” modeling folding accuracy lags, as exemplified by a 3D-improbable ArM with a water-accessible nonpolar Pro167 in lieu of a buried highly conserved polar aa in a water channel, a striking irony.

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025mm284g53