KeloidBench: Interpretable machine learning on fibroblast program features for keloid detection

We present KeloidBench, a demonstration pipeline that predicts whether a tissue transcriptomic profile is keloid versus unaffected/normal skin under true cross-study generalization. KeloidBench collapses high-dimensional gene expression into compact fibroblast activation program features motivated by a replication-first observation: single marker genes are unstable across public cohorts, while aggregated fibroblast-state programs transfer more reliably. Moreover, KeloidBench acts as an auditable GenAI copilot for cross-cohort keloid transcriptomic analysis. The model initializes a deterministic program-score to produce predictions, confidence estimates, abstention decisions, and feature contributions, and follows up with a constrained language framework that converts these structured outputs into traceable biological summaries without generating or modifying the underlying predictions. We evaluate linear classifiers with leakage-safe nested leave-one-study-out (LOSO) selection, accession×donor-balanced weighting, and Platt-calibrated probabilities. On a transferable subset of public studies, program features achieve approximately 0.92 LOSO macro F1, while the same held-out LOSO predictions pooled sample-level accuracy achieves approximately 0.84. These results support KeloidBench as a research-facing public-cohort triage tool and motivate prospective matched-cohort or biobank validation before any clinical use.