Statistical mechanics has been the workhorse that condensed matter physicists have used to make the connection between microscopic properties and macroscopic, collective phenomena. Establishing this connection requires reducing masses of microscopic information (dimensional reduction) to a few relevant microscopic variables and their distributions. Data science methods are designed for dimensional reduction, so they are a natural tool to turn to when statistical mechanics fails. But it requires physics to identify the relevant microscopic quantities as well as the most appropriate data science methods to use to access them. I will discuss two problems where we have made progress with this approach: we have applied machine learning to glassy dynamics and persistent homology to the phenomenon of allostery.