Introduction and Objectives Differentiating tuberculosis and sarcoidosis can be difficult, particularly in the context of mediastinal lymphadenopathy, because both diseases are characterised by overlapping clinical phenotypes and histologically similar granulomatous inflammation. Currently, diagnosis relies heavily on microbiological confirmation of tuberculosis which is only available in <50% of cases. Therefore, novel diagnostic strategies are needed to prevent morbidity associated with delayed or inappropriate treatment. We tested the hypothesis that genome wide transcriptional profiling of mediastinal lymph node samples obtained by minimally invasive endobronchial ultrasound guidance could identify gene signatures that differentiate tuberculosis and sarcoidosis.

Methods In vivo immune responses were compared in mediastinal lymph node biopsies obtained via endobronchial ultrasound guidance from patients with tuberculosis, sarcoidosis or non-granulomatous disease using genome-wide transcriptional profiling. Machine learning algorithms were used to test the discriminatory power of identified gene signatures which distinguished granulomatous from non-granulomatous disease or tuberculosis from sarcoidosis.

Results Comparison of lymph node genome‑wide transcriptional profiles by principal component analysis revealed clear differences between granulomatous and non-granulomatous disease. Granulomatous profiles showed significant enrichment for genes involved in antigen presentation, inflammatory responses, innate immune responses and T cell activation, in keeping with the processes involved in granuloma generation. As expected, sarcoidosis and tuberculosis sample profiles were very similar, however, significant gene expression differences were still evident between these two groups. In particular, several genes related to development of granuloma architecture were more highly expressed in sarcoidosis samples. Next we used machine learning tools in order to test the discriminatory power of differentially expressed gene signatures and found that the support vector machines algorithm correctly classified up to 97% of granulomatous and non-granulomatous disease cases. Importantly, this technique successfully distinguished sarcoidosis from tuberculosis in up to 100% cases.

Conclusions Transcriptomic analysis of lymph node samples from the site of disease identifies gene signatures that can reliably distinguish tuberculosis from sarcoidosis using computational classification tools. Our data highlight the superior discriminatory power of multiple gene expression differences over a single marker in complex disease and generate a pathway for biomarker discovery in the management of tuberculosis and sarcoidosis.