Nonalcoholic fatty liver disease (NAFLD) is a globally prevalent metabolic disorder for which no pharmacotherapy has been approved. We applied an integrated computational framework to a ten-herb traditional Chinese medicine (TCM) transdermal formula administered via umbilical (Shenque) acupoint therapy, combining network pharmacology, multi-dataset machine learning, immune-infiltration analysis, molecular docking, and molecular dynamics (MD). Network pharmacology of 5276 drug-likeness-filtered compounds yielded 380 shared herb–disease targets. Across three training cohorts (n=346, ComBat-corrected) and an independent validation cohort (GSE167523, n=98), LASSO and random forest identified four hub genes—FOS, HSP90AB1, HIF1A, and MAPK8 (validation AUC 0.760, 0.701, 0.802, and 0.744)—all dysregulated in NAFLD. GSEA highlighted upregulated ECM–receptor interaction and suppressed ribosome, IL-17, and NF-κB signalling, whereas ssGSEA revealed elevated follicular helper and exhausted CD4+ T cells. Three-dimensional screening (transdermal ADMET, ligand similarity, AutoDock Vina) prioritised danshenspiroketallactone as the strongest HSP90AB1 binder (ΔG=−11.7 kcal/mol). Across six complexes, 3×100 ns MD with MM-PBSA confirmed favourable binding (ΔGbind −14.2 to −18.0 kcal/mol). These findings propose a multitarget inflammation–proteostasis–fibrosis mechanism and a “thermal activation–chemical fine-tuning” model for Shenque therapy, generating hypotheses that warrant experimental validation.