varCADD: large sets of standing genetic variation enable genome-wide pathogenicity prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Machine learning (ML) and artificial intelligence (AI) are crucial for identifying disease-causing genetic variations.
• Current ML models are limited by small, biased training datasets, often focusing only on protein-coding genes.
• Unbiased, comprehensive datasets are needed for accurate variant effect prediction.

Purpose of the Study:

• To develop improved training datasets for ML-based variant prioritization.
• To leverage human standing variation for creating larger and less biased training sets.
• To enhance the accuracy of predicting genetic variant deleteriousness.

Main Methods:

• Utilized whole-genome sequences from 71,156 individuals (gnomAD v3.0) to create training sets.
• Defined benign variants using frequent standing variation and deleterious variants using rare/singleton variation.
• Trained new models using the Combined Annotation Dependent Depletion (CADD) framework (v1.6).

Main Results:

• Alternative models achieved state-of-the-art accuracy, comparable to CADD v1.6/v1.7.
• Models trained on standing variation outperformed existing methods in specific genomic regions.
• Larger datasets captured broader genomic regions and rare annotations, including regulatory elements.

Conclusions:

• Human standing variation provides a powerful resource for training accurate genome-wide variant prioritization models.
• The proposed datasets are larger, less biased, and cover more genomic diversity than conventional sets.
• These datasets and models are publicly available to advance genetic research and applications.