Imputing abundances and inferring direction of associations between histone modifications using neural processes
Histone post-translational modifications (HPTMs) play a vital role in the regulation of numerous cellular processes. However, systems level understanding of how HPTMs coordinate and interact with each other, and the direction of interactions remain unexplored due to lack of suitable technology. EpiTOF, a high-dimensional mass cytometrybased platform, measures HPTMs and histone variants across multiple immune cell sub-types at a single-cell resolution to enable the system-wide study of HPTMs. Large number of immune cells profiled using EpiTOF present an unprecedented opportunity to learn directional networks of HPTM associations. We developed a two-step computational framework to identify direction of association between two HPTMs. In the first step, we used linear regression (LR)-, k-nearest neighbors-(kNN), or Neural Processes (NP)-based to impute the abundance of HPTMs. In the second step, we developed an interpretation framework to infer direction of association between HPTMs at a system-level using mass cytometry data. We evaluated this framework using EpiTOF profiles of more than 55 million peripheral mononuclear blood cells from 158 healthy human subjects. NP models had consistently higher imputation accuracy than LR and kNN. The inferred networks recapitulated known HPTM associations and identified several novel ones. While almost all associations were conserved across all healthy individuals, in a cohort of healthy subjects, vaccinated with the trivalent inactivated seasonal influenza vaccine (TIV), we identified changes in associations between 6 pairs of HPTMs 30 days following vaccination, many of which have been shown to be functionally involved in innate memory. These results demonstrate utility of our framework in identifying causal interactions between HPTMs that can be further tested experimentally.