SAFE: A semi-agent system integrating feature selection, PubMed literature mining, and pathway enrichment

Abstract

We present SAFE, a computational framework that combines feature selection (FS) with PubMed-based literature mining and optional pathway enrichment. SAFE automatically detects the learning task, selects candidate features using statistical and model-based methods, and assigns literature-based evidence scores via PubMed. To enhance biological interpretability, the tool optionally performs pathway enrichment analysis through curated databases such as KEGG, Reactome, and Gene Ontology. SAFE is implemented as an interactive application that supports reproducible pipelines, visualization, and reporting. Experimental results on synthetic and real-world biomedical datasets demonstrate that SAFE not only achieves competitive predictive performance but also produces feature subsets with strong biomedical relevance. By bridging purely data-driven selection and biologically informed prioritization, SAFE offers researchers a practical solution for interpretable and reliable FS in high-dimensional biomedical data.