Transfer learning to integrate multiomic data to enhance nemolizumab response assessment in atopic dermatitis and prurigo nodularis

Summary: Abstract Body: The effectiveness of nemolizumab, an antagonist of IL-31 receptor, has recently been demonstrated in atopic dermatitis (AD) and prurigo nodularis (PN). However, both AD and PN are highly heterogeneous clinically and pathologically, and therefore identifying biomarkers of response to IL-31 targeting is important. To address this, we set up a large-scale prospective cohort. Over 1,200 omics experiments were conducted to profile gene expression and protein in blood, skin, and tape strips from 126 patients with AD and 113 patients with PN receiving nemolizumab treatment. Different clinical outcome variables, including investigator global assessment and weekly average peak pruritus numerical rating scale measured at baseline and week 16 post treatment were assessed. To overcome limitations of machine learning approaches secondary to the multi-layer partially overlapping data, we devised a hybrid transfer learning model, combining LASSO with additional penalty for the discrepancy with prior tissue information and the factor decomposition approach. Benchmarking results showed that the hybrid model significantly outperformed traditional LASSO. Furthermore, we illustrate that combining prior non-lesional tissue information enhances the model performance in contrast to using only lesional tissue. Notably, the drug response prediction for AD was more stable compared with PN, where we achieved AUROC of 0.819 for AD, compared to 0.581 with the baseline LASSO model; while for PN the AUROC ranged between 0.652 and 0.749. This suggests that the biological mechanisms underlying IL-31-targeting therapy in AD are more consistent compared to PN. Significantly, our analysis identified the interferon signature genes ISG15 and IFI6 as critical predictive genes for AD, highlighting their presence in up to 90% of the random cross-validation sets. Our study highlights the value of integrating inter-tissue information for enhancing prediction accuracy in clinical outcomes. Yifei Dai^{1, 2}, Qinmengge Li², Nicolas Delaleu³, Zhi He², Valérie Julia³, Johann E. Gudjonsson², Lam C. Tsoi² 1. University of Southern California, Los Angeles, CA, United States. 2. University of Michigan, Ann Arbor, MI, United States. 3. Galderma SA, Zug, ZG, Switzerland. Bioinformatics, Computational Biology, and Imaging