Deep generative models of LDLR protein structure to predict variant pathogenicity.

Area of Science:

• Genomics and Bioinformatics
• Cardiovascular Disease Genetics
• Protein Variant Effect Prediction

Background:

• Classifying low-density lipoprotein receptor (LDLR) protein-coding missense variants is challenging due to the receptor's complex structure and function.
• Deep generative models, including Evolutionary Scale Modeling (ESM), Evolutionary model of Variant Effect (EVE), and AlphaFold 2 (AF2), show promise in predicting protein structure and function.
• ESM and EVE estimate variant likelihood, while AF2 predicts structural changes, posing different interpretability and application challenges.

Purpose of the Study:

• To evaluate the effectiveness of deep generative models (ESM, EVE, AF2) in predicting the pathogenicity of LDLR variants.
• To compare the performance of these models against established variant prediction tools and experimental measures.
• To assess the association of these models with clinical phenotypes, specifically serum LDL-C levels and atherosclerotic cardiovascular disease.

Main Methods:

• Tested ESM, EVE, and AF2 for predicting variant pathogenicity and compared their performance to established methods like Polyphen-2, SIFT, REVEL, and Primate AI.
• Assessed model correlation with experimental measures of LDL uptake.
• Utilized UK Biobank data to compare model associations with clinical phenotypes, including serum LDL-C and atherosclerotic cardiovascular disease.

Main Results:

• AF2 predicted LDLR structures poorly modeled variant pathogenicity.
• ESM and EVE showed comparable performance to established methods for binary classification in ClinVar but stronger correlations with experimental LDL uptake.
• ESM and EVE demonstrated stronger associations with serum LDL-C than Polyphen-2; ESM identified variants with more extreme LDL-C levels and showed a stronger association with atherosclerotic cardiovascular disease.

Conclusions:

• AF2-predicted LDLR structures are not reliable for modeling variant pathogenicity.
• ESM and EVE are competitive with existing methods for ClinVar classifications and superior in correlating with experimental assays and clinical phenotypes.
• ESM and EVE represent advanced tools for predicting the functional impact of LDLR variants and their association with cardiovascular risk.