Fig. 1

The SEPIRA algorithm and application to smoking and lung cancer. a The first step involves construction and validation of a tissue-specific regulatory network using the SEPIRA algorithm. This network consists of TFs significantly overexpressed in the given tissue compared to other tissue types and corresponding downstream gene targets. This network is constructed from computing co-expression correlations across a large gene expression compendium encompassing many different tissue types and subsequently using partial correlations to remove likely indirect associations. The inferred high-quality regulatory network can be used to infer TF activity in any given sample by regressing the sample’s gene expression profile against the gene target profile, encoded as 1 for activating interactions, – 1 for repression, and 0 for no significant association. SEPIRA also allows TF binding activity to be estimated from genome-wide DNAm data, regressing the gene-target promoter DNAm profile (suitably normalized, i.e. centered) of the sample against the gene-target binding profile (reversing signs relative to the gene-expression case, since lower promoter DNAm usually reflects binding activity). Finally, the tissue-specific regulatory network is validated against an independent dataset (messenger RNA expression or DNAm) encompassing many different tissue-types including the tissue-type of interest. b Application of SEPIRA to the case scenario of lung cancer and smoking. SEPIRA results in a lung-specific regulatory network (called LungNet, which is then used to infer TF activity in normal-adjacent (NADJ) and LSCC, as well as in lung carcinoma in situ (LCIS) (a precursor cancer lesion). This identifies TFs which become inactivated in LSCC and LCIS. A subset of these would be expected to also exhibit inactivation in the normal cell-of-origin samples exposed to the major risk factor for LSCC (i.e. smoking). We propose that inactivation of this subset of TFs could be causal mediators between smoking and LSCC