Inferring direction of associations between histone modifications using a neural processes-based framework

Current technologies do not allow predicting interactions between histone post-translational modifications (HPTMs) at a system-level. We describe a computational framework, imputation-followed-by-inference, to predict directed association between two HPTMs using EpiTOF, a mass cytometry-based platform that allows profiling multiple HPTMs at a single-cell resolution. Using EpiTOF profiles of >55,000,000 peripheral mononuclear blood cells from 158 healthy human subjects, we show that neural processes (NP) have significantly higher accuracy than linear regression and k-nearest neighbors models to impute the abundance of an HPTM. Next, we infer the direction of association to show we recapitulate known HPTM associations and identify several previously unidentified ones in healthy individuals. Using this framework in an influenza vaccine cohort, we identify changes in associations between 6 pairs of HPTMs 30 days following vaccination, of which several have been shown to be involved in innate memory. These results demonstrate the utility of our framework in identifying directed interactions between HPTMs.