RegPathTopic: deterministic regulatory-path topic modeling with traceable regulatory vocabularies for single-cell annotation

Single-cell RNA sequencing annotation can achieve high label accuracy when suitable references are available, but assigned labels do not necessarily expose the regulatory features associated with cell identity or state. Existing regulator-activity methods address related biological questions, whereas graph-informed topic models face a separate reproducibility problem when regulatory vocabulary terms are generated stochastically. We developed RegPathTopic to construct deterministic graph-derived vocabulary terms and evaluate them under matched controls. RegPathTopic converts TF-target graphs into regulatory paths and TF-centered modules, scores and filters these terms, calibrates them against degree-matched controls, and learns topic representations from gene-only, regulatory-only, or combined matrices. Across seven benchmark settings, RegPathTopic achieved a macro-F1 of 0.975 in the filtered CellTypist internal benchmark, matching the best gene-only topic baseline. In CellTypist and scIB transfer, it achieved macro-F1 values of 0.970 and 0.830, respectively, with positive margins over the displayed predecessor-style and degree-matched graph-derived controls. These margins did not extend to all comparator families: gene-only transfer, several established annotation tools, and decoupler TRRUST activity baselines were stronger in relevant settings. Vocabulary audits showed that retained paths and modules remained explicit and traceable through topic-level outputs; these analyses establish feature traceability rather than validated TF activity or causal regulatory interpretation. The present evidence supports RegPathTopic as a reproducible framework for constructing and auditing graph-derived topic vocabularies and testing them against matched graph controls. It does not establish RegPathTopic as a superior annotation method or as a replacement for dedicated regulator-activity approaches. Perturbation-supported or orthogonal biological validation is required before the traced vocabulary terms can be interpreted as validated regulatory programs.