DBSAV database

DBSAV database provides computational annotations and deleteriousness scoring of single amino acid variations (SAVs) in the human proteome using DeepSAV-derived variant scores and gene-level GTS metrics to assess impacts on protein function and disease associations.

Key Features:

• DeepSAV scoring: DeepSAV scores derived from a deep convolutional neural network predictor estimate the deleterious effects of SAVs.
• Sequence, structural and functional context: Evaluations incorporate sequence, structural, and functional properties for mechanistic interpretation of SAV impacts.
• Gene-level GTS scores: Aggregation of individual DeepSAV variant scores into gene-level GTS (Gene Tolerance of rare SAVs) scores quantifies gene tolerance to deleterious missense mutations.
• Variant datasets: Dataset includes pathogenic, benign, and population-level variants observed through exome sequencing.
• Multiple sequence alignments: Multiple sequence alignments are constructed from vertebrate-level orthologs identified by BLAST searches with a 50% sequence identity cutoff to exclude distant homologs.
• Model optimization: Prediction accuracy was improved by expanding variant datasets with pathogenic and benign examples and optimizing neural network parameters.

Scientific Applications:

• SAV interpretation: Interpret single amino acid variations to assess potential deleteriousness and mechanistic effects on proteins.
• Gene-disease association studies: Support analysis of gene-disease associations using gene-level GTS metrics derived from aggregated SAV scores.
• Diagnostic and genomics research: Contextualize SAVs for diagnostic analyses and genomic studies by combining DeepSAV scores with population exome data.

Methodology:

DeepSAV uses a deep convolutional neural network to evaluate SAV deleteriousness using sequence, structural, and functional features; individual variant scores are aggregated into gene-level GTS metrics; multiple sequence alignments are built from vertebrate orthologs identified by BLAST with a 50% sequence identity cutoff; datasets were expanded with pathogenic and benign variants and neural network parameters were optimized.