Abstracts

Rationale: In Asians, the human leukocyte antigen HLA-B*15:02 is strongly predictive of severe cutaneous reactions to carbamazepine (CBZ) including Stevens-Johnson syndrome (SJS) and toxic epidermal necrolysis (TEN). Pre-treatment screening of this risk allele is currently recommended by the US Food and Drug Administration. However, HLA-B*15:02 has a low positive predictive value (PPV) suggesting that other genetic variants may modify risk. Transcription factors (TFs) bind to the DNA and play a critical role in gene regulation. Genetic variants within TF binding regions can modify cellular response to stimuli. We hypothesised that variation with TF-binding sites modify risk of CBZ-induced SJS.

Methods: Han Chinese people with epilepsy who either developed CBZ-related SJS/TEN or tolerated CBZ were recruited to undergo whole-genome sequencing. The binding regions for each of the 1,361 human TFs, listed in the Gene Transcription Regulation Database, were used as input to a machine-learning model. The resulting performance metrics (AUC, weighted F1 score) and feature importance were used to identify and rank variants with strong predictive capacity. The top 242 variants in DNA binding regions of 19 TFs were selected as input to a final predictive model and bootstrapping with 10,000 iterations was applied.

Results: 84 individuals with CBZ-induced SJS/TEN (80.95%, 68/84 HLA-B*15:02 carriers) and 73 CBZ-tolerant (38.36%, 28/73 HLA-B*15:02 carriers) were included. The final model achieved a median AUC of 98.37%, weighted F1-score of 92.62%, sensitivity of 96.30%, and specificity of 89.66%. The median negative predictive value (NPV) and PPV were higher as compared to those based on carrier status of the HLA-B*15:02 risk allele [NPV=99.99%, PPV=2.20% vs. NPV=99.93%, PPV=0.50%], with the PPV of the TF-based model being higher in 99.96% of the bootstrap instances [9,996/10,000; sign-rank test p< 0.001]. PPV and NPV calculations were adjusted to the 0.24% prevalence of CBZ-induced SJS/TEN.