ESTRO 2025 - Abstract Book

S3939

Radiobiology - Normal tissue radiobiology

ESTRO 2025

imputation server. A quality control test was performed, resulting in 5,897,310 SNPs for further analysis. Of the 387 participants with available ED status, 221 (57.1%) experienced ED. The cohort was stratified only once, at the outset of the modeling process, into training (70%), validation (10%), and test (20%) sets while preserving the ED rate in each set. Univariate logistic regression analysis was conducted to assess the association between ED and each SNP, adjusting for the first five principal components in the training data. Linkage disequilibrium (LD) clumping was then performed employing PLINK with a significance threshold of p<0.001, resulting in 810 lead SNPs. The ANNOVAR tool was employed to annotate genes corresponding to the identified lead SNPs. Subsequently, these genes were mapped to 146 KEGG pathways, excluding disease-specific pathways. A biologically interpretable neural network model, termed GenNet [3], was then trained using the annotated biological connections in the training set (Figure 1). Results: Age and androgen deprivation therapy (ADT) were found to be significantly different between ED and non-ED groups, yielding both p<0.0001; the mean age was 66 years for the ED group and 60 years for the non-ED group. The ADT rate was 60.1% for the ED group and 32.8% for the non-ED group. The ANNOVAR and KEGG database analyses produced annotated biological connections, consisting of 225 SNPs, 202 genes, and 103 pathways. The GenNet model was then trained with the biologically plausible connections using the set of SNPs along with age and ADT. The process was repeated 30 times, resulting in an average AUC of 0.7. The top five SNPs with the highest weights were relevant to four genes ( CHRM3 , GALR1 , LPAR1 , GRM8 ) and these genes were found to involved in the neuroactive ligand receptor interaction pathway (Figure 2). Conclusion: We demonstrated the potential of using a germline genetic neural network model to predict radiation-induced ED and identify plausible biomarkers.