Protein design has emerged as an important field in contemporary biology, driven in part by the accumulation of vast amounts of protein data in public databases like the Protein Data Bank (PDB). The challenge now is to use this data to decipher the principles underlying protein design, as guided by nature, and to develop novel proteins with desired properties. To this end, we investigated the design principles of orthologs and paralogs of a small binding protein - Sho1-SH3 - in the yeast osmosensing pathway. Using this natural system as a template, we employed deep learning models to design novel functional osmosensing orthologs. Our results demonstrate that these models not only accurately captured the distribution of functionality of natural proteins, but also expanded the functional space by designing novel proteins that extended beyond the functional constraints of natural proteins. This work provides valuable insights into the principles governing protein design and opens up new avenues for the development of novel proteins with desirable functions.