Computational biologists, and biologists more generally, spend a lot of time trying to more fully characterize proteins. In this talk, I will describe several of our recent efforts to use machine learning methods to gain a better understanding of proteins. First, we tackle one of the oldest problems in computational biology, the recognition of distant evolutionary relationships among protein sequences. We show that by exploiting a global protein similarity network, coupled with a latent space embedding, we can detect remote protein homologs more accurately than state-of-the-art methods such as PSI-BLAST and HHPred. Second, we use machine learning methods to improve our ability to identify proteins in complex biological samples on the basis of shotgun proteomics data. I will describe two quite different approaches to this problem, one generative and one discriminative.