Sen Zhang ¹ Li-Na Dai ² Qi Yin ^1,3 Xiao-Ping Kang ¹ Dan-Dan Zeng ⁴ Tao Jiang ^1* Guang-Yu Zhao ^1,3* Xiao-He Li ^2* Jing Li ^1,2*

• ¹ State Key Laboratory of Pathogen and Biosecurity, Academy of Military Medical Science, Beijing, Beijing Municipality, China
• ² College of Basic Medical Sciences, Inner Mongolia Medical University, Hohhot, Inner Mongolia Autonomous Region, China
• ³ Laboratory of Advanced Biotechnology, Academy of Military Medical Sciences, Beijing, China
• ⁴ College of Veterinary Medicine, Shanxi Agricultural University, Jinzhong, Shanxi Province, China

Scoliosis is a pathological spine structure deformation, predominantly classified as "idiopathic" due to its unknown etiology. However, it has been suggested that scoliosis may be linked to polygenic backgrounds. It's crucial to identify potential Adolescent Idiopathic Scoliosis (AIS)-related genetic backgrounds before scoliosis onset. The present study was designed to intelligently parse, decompose and predict AIS-related variants in ClinVar database. Possible AIS-related variant records downloaded from ClinVar were parsed for various labels, decomposed for Dinucleotide Compositional Representation (DCR) and other traits, screened for high-risk genes with statistical analysis, and then learned intelligently with deep learning to predict high-risk AIS genotypes. Results demonstrated that the present framework is composed of all technical sections of data parsing, scoliosis genotyping, genome encoding, machine learning (ML) / deep learning (DL) and scoliosis genotype predicting. 58, 000 scoliosis-related records were automatically parsed and statistically analyzed for high-risk genes and genotypes, such as FBN1, LAMA2 and SPG11. All variant genes were decomposed for DCR and other traits. Unsupervised ML indicated marked inter-group separation and intra-group clustering of the DCR of FBN1, LAMA2 or SPG11 for the five types of variants (Pathogenic, Pathogeniclikely, Benign, Benignlikely and Uncertain). A FBN1 DCR-based Convolutional Neural Network (CNN) was trained for Pathogenic and Benign/Benignlikely variants performed accurately on validation data and predicted 179 high-risk scoliosis variants. The trained predictor was interpretable for the similar distribution of variant types and variant locations within 2D structure units in the predicted 3D structure of FBN1. In summary, scoliosis risk is predictable by deep learning based on genomic decomposed features of DCR. DCR-based classifier has predicted more scoliosis risk FBN1 variants in ClinVar database. DCR-based models would be promising for genotype-to-phenotype prediction for more disease types.