Bayesian and systems-level reanalysis of public tuberculosis progression transcriptomes reveals latent host-response programs

Background Public tuberculosis progression transcriptomic datasets contain more biological information than can be captured by ranked-gene lists alone. We performed a systems-level reanalysis to assess latent structure, uncertainty-aware gene ranking, pathway convergence, and bias-sensitive interpretation in the harmonizable public cohorts. Methods Two cohorts with shared gene symbols and binary progressor labels (GSE107994 and GSE193777) were reanalyzed. We applied joint principal component analysis before and after cohort centering, factor analysis on the most variable genes, Bayesian hierarchical synthesis of within-cohort differential expression effects, pathway-level posterior modeling, marker-based NNLS deconvolution, WGCNA-style coexpression analysis, signature correlation analysis, and a directed acyclic graph to clarify potential bias pathways. Results The advanced analysis included 301 samples, comprising 87 progressors and 214 non-progressors. Raw PC1 remained strongly cohort structured, but cohort-centered PC1 separated non-progressors and progressors more clearly (mean PC1 2.9 vs -7.0 before centering; -13.2 vs 32.4 after centering). Bayesian synthesis prioritized MILR1, VSIG4, FZD5, CD36, CCR2, ASGR2, with MILR1 showing the strongest pooled effect (posterior mean 1.229, 95% credible interval 1.108 to 1.351). The leading pathway signals were angiogenesis, blood vessel development, blood vessel morphogenesis. All three latent factors remained associated with progressor status, with the strongest evidence for Factor1 (p = 5.78e-11). Marker-based deconvolution suggested higher monocyte and lower lymphoid-associated scores in progressors. Exploratory remapping of GSE79362 yielded 10,419 overlapping genes but shifted the strongest pooled signal toward FCGR3B. Conclusions The harmonizable public datasets support a coordinated tuberculosis progression signal that combines myeloid regulation with vascular-remodeling biology. The findings are stronger as uncertainty-aware biological evidence than as a clinical prediction claim, because the shared-gene advanced layer currently rests on two directly comparable cohorts and should be expanded before clinical translation is considered. The deconvolution and coexpression analyses are supportive interpretation layers, not direct measures of leukocyte fractions or causal network effects.