Expert-curated gene signatures through biological insights and an interactive AI agent

Advances in molecular biology, the rapid growth of high-dimensional data, and new computational technologies continue to transform how diseases are studied and treated. Within this landscape, identifying gene signatures from expression data remains challenging because of the high dimensionality and biological complexity of such datasets. Existing tools are often driven primarily by mathematical and statistical criteria, so their outputs may contain genes of uncertain biological or study-specific relevance, requiring biomedical experts to repeatedly consult computational specialists to modify the analysis and interpret its results. To address this, we introduce GeneCurator, a human-in-the-loop web platform for expert-guided gene signature construction and refinement. The platform applies established statistical and machine learning feature-selection methods to expression data and produces an editable ranked list of candidate genes. The expert evaluates this list using two complementary sources of support: biological knowledge, including functional annotations, pathway and ontology enrichment, protein-protein interaction networks, and ligand-receptor relationships; and an interactive AI agent that considers the study objective, current gene list, and downstream outcomes, and proposes genes for addition, removal, or reinstatement with a rationale for each. The expert retains control throughout, modifying the signature and observing the effects in real time. In an experimental evaluation on a real transcriptomic dataset, expert curation reshaped the signature and every downstream result, showing that the tool contributes biologically meaningful, study-relevant refinements beyond automated selection. By uniting computational methods, biological knowledge, and interactive AI in one environment, GeneCurator strengthens the expert's analytical autonomy while keeping human judgement central to signature curation.