Dinucleotide composition representation -based deep learning to predict scoliosis-associated Fibrillin-1 genotypes.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Scoliosis, a spinal deformation, is often idiopathic, but polygenic factors are suspected.
• Identifying genetic predispositions for Adolescent Idiopathic Scoliosis (AIS) before onset is crucial.

Purpose of the Study:

• To develop a computational framework for predicting AIS-related genetic variants.
• To leverage deep learning for identifying high-risk genotypes associated with scoliosis.

Main Methods:

• Parsed and decomposed ~58,000 ClinVar records for Dinucleotide Compositional Representation (DCR) and other genetic traits.
• Applied statistical analysis to screen for high-risk genes (e.g., FBN1, LAMA2, SPG11).
• Utilized deep learning, specifically a Convolutional Neural Network (CNN), trained on DCR features to predict scoliosis risk variants.

Main Results:

• The DCR-based CNN accurately predicted high-risk variants for genes like FBN1.
• Unsupervised machine learning showed distinct clustering of DCR for different variant types.
• Identified 179 high-risk scoliosis variants, with interpretable predictions based on 3D structural analysis.

Conclusions:

• Deep learning models utilizing DCR are effective for predicting scoliosis risk.
• The DCR-based approach shows promise for genotype-to-phenotype predictions in various diseases.
• This framework enhances the identification of genetic factors contributing to scoliosis.