Accelerating Drug Discovery: AI-Powered Indication Exploration for Ropeginterferon Alfa-2b

Traditional indication finding relies on pre-existing biological data and in vitro/vivo validation. In contrast, drug repurposing that utilizes in silico approaches including transcriptional signature mapping and knowledge graphs can connect existing therapeutics with novel indications beyond their originally intended medical purpose. These approaches can significantly accelerate drug discovery by leveraging the well-characterized and clinically validated pharmacology and toxicology attributes of approved drugs. Ropeginterferon alfa-2b (ropeg) is a mono-PEGylated recombinant human interferon alpha approved by the FDA for the treatment of polycythemia vera (PV) in 2021 and is currently being clinically evaluated in other myeloproliferative neoplasms (MPNs). Ropeg targets mutant hematopoietic stem cells that drive MPNs, decreasing the associated chronic inflammation, elevated blood cell counts and risk for thromboembolic events. To identify additional indication opportunities for ropeg, we first generated a gene signature by bulk RNA sequencing of in vitro ropeg-treated primary human peripheral blood mononuclear cells (PBMCs). Differentially expressed genes (N = 76) were included in the ropeg signature based on an adjusted p<0.05 and absolute log₂ FC > 2 compared to vehicle-treated samples. We then employed a state-of-the-art machine learning Causal Embedding Model leveraging the QIAGEN Biomedical Knowledge Base – HD to access a manually curated knowledge graph of causal gene–disease/function relationships. Using literature-curated causal gene expression relationships, we constructed unsupervised causal embeddings of genes, the ropeg gene signature and diseases by approximating a bipartite graph’s bi-adjacency matrix, which maps genes to their expression-regulated targets. We then calculated the cosine distance between embedding vectors to determine the (dis)similarity of the ropeg gene signature to diseases/biological functions. This approach revealed ten distinct clusters of diseases that are negatively associated with ropeg. These disease embeddings are dissimilar to the ropeg-induced gene signature embedding, suggesting that these disease states could potentially be reversed by ropeg. These diseases include one cluster with known/approved indications, MPNs, as well as new indications not previously well documented in the clinic. To further validate our approach, we selected another top ranked indication, cutaneous T cell lymphoma (CTCL), with subtypes characterized by malignant T cell accumulation in the skin (mycosis fungoides) or in the skin and blood (Sezary syndrome). In a pilot study, we demonstrated that ropeg potently (28.01 U/mL) and dose-dependently inhibited viability in a human CTCL cell line derived from a patient with Sezary syndrome. In conclusion, we generated a ropeg gene signature and developed a machine learning model, employing a causal knowledge graph to discover novel indications. Furthermore, we have conducted experimental validation for one such indication, CTCL, thereby demonstrating the promise and potential of our AI-powered approach to identify new indications beyond traditional approaches.