A reproducible computational pipeline integrating network toxicology, nested machine learning, single-cell transcriptomics and molecular docking nominates cytoskeletal and BMP genes (ACTA2, BMP4, FLNA, FLNB) as bisphenol A–responsive markers of menin

Abstract Background Bisphenol A (BPA) is a high-production-volume endocrine-disrupting chemical, and meningioma is a common hormone-sensitive intracranial tumour. A systematic, reproducible molecular map linking BPA-responsive genes to meningioma transcriptomes has been lacking. We therefore developed a fully computational pipeline that integrates multi-cohort transcriptomics, chemical–target databases, nested machine learning, single-cell localisation and molecular docking to prioritise candidate markers while treating all docking and classification outputs as computational evidence rather than clinical causality. Results Three GEO microarray cohorts (GSE43290, GSE54934, GSE84263; 165 tumours, 23 normals) were merged on 10,938 genes after ComBat correction. Differential expression (682 DEGs) and WGCNA tumour modules (153 genes) were intersected with BPA targets from SwissTargetPrediction (96), SEA (19), PharmMapper (174), CTD and ChEMBL (five-source union 2,250) and meningioma genes from Open Targets/CTD. Twenty-five BPA-responsive core candidates emerged; consensus PPI×SHAP hubs were ACTA2 , BMP4 , FLNA and FLNB (Table 2 ranks 7.0 / 12.0 / 13.0 / 12.0). Nested tumour-versus-normal classifiers approached a performance ceiling (best external AUC 1.000; 83/109 models = 1.0), interpreted cautiously under class imbalance, whereas WHO-grade prediction was weak (external AUC ≤ 0.677). In GSE183655, BMP4 / FLNB mapped to tumour mesenchymal cells and ACTA2 to pericyte/smooth-muscle niches. AutoDock Vina affinities were weaker at hubs (− 5.17 to − 5.70 kcal/mol) than at ESR1 / ESR2 / PTGS2 (− 8.21 to − 8.49), with symmetry-aware redock RMSDs supporting control-site validity. Conclusions This reproducible pipeline provides computational evidence that cytoskeletal and BMP genes mark meningioma tissue within a BPA-responsive intersection. Direct high-affinity BPA–hub binding is not supported relative to classical oestrogen receptors and PTGS2 . Wet-lab and exposure-linked studies are required before translational inference.