Improving Inference via Perturbations

Cellular networks in biological systems are complex and as such, identifying the molecular interactions that give rise to the complex behavior observed can require an immense amount of data. Often, statistical and machine learning techniques are used to analyze this data and extract a global picture of network dynamics. One of the challenges of network analysis in systems biology is finding the connections between genes, proteins, or both, and predicting additional ones that have not yet been detected experimentally. This problem is easily mappable to the inverse problem of statistical physics: inferring the microscopic particle-particle interactions given macroscopic observations of a system. In particular, the focus of this work is to investigate whether perturbations can be introduced into the system so as to improve the output data quality. Specifically, we explore how perturbations in the form of magnetic field can be used to improve the inference of interactions for a three-spin Ising system. Utilizing a maximum likelihood approach, we empirically show that there exists an optimal field where learning is most ecient. Such a field seems to counteract the individual interactions between spins, allowing for optimal inference.