OrphanForge-AI.

We were inspired by the orphan disease crisis, where thousands of rare diseases affect millions of people but remain difficult to study because patients, biological data, researchers, and clinical trials are scattered across the world. OrphanForge-AI was built as a research prototype that combines a rare-disease knowledge graph, AI-based therapeutic candidate prioritization, federated learning, patient-cohort discovery, and in-silico clinical-trial optimization into a single architecture. We learned how generative AI and geometric deep learning can support computational drug discovery, while federated learning can enable collaboration without centralizing sensitive genomic information. The prototype workflow takes disease and biological evidence, maps gene–variant–protein–disease relationships, prioritizes potential therapeutic strategies, simulates distributed patient cohorts, and identifies more feasible trial configurations. The biggest challenges were limited rare-disease datasets, heterogeneous biological information, privacy constraints, uncertainty in AI-generated predictions, and the gap between computational results and real-world clinical validation. These challenges shaped our design toward an evidence-driven, privacy-preserving decision-support system rather than an autonomous medical system, with human and experimental validation remaining essential.