(197g) Representation Learning Approach for Mapping Protein Surface Patches

When designing proteins or molecules that target proteins, a common issue is that
proteins may dimerize and aggregate, blocking active sites and limiting intended interactions.
Hydrophobicity plays an important role in protein aggregation and dimerization, but current
methods for high resolution mapping of hydrophobicity on protein surfaces are still lacking. To
predict hydrophobicity, we need to understand the surface configuration of proteins. In this
presentation, we develop a patch based characterization of protein surfaces through
representation learning. First, we break down proteins from the PDB database into surface
patches. Then we construct and train a variational autoencoder to represent patches in a vector
form. Understanding similarities and differences in surface patches begins to suggest
relationships between molecular features and protein function. This is a crucial step toward
protein hydrophobicity mapping because it reduces the latent space of all possible proteins
down to common surface patches. Mapping the constructed surface patch space allows us to
more easily predict protein function as well as similarities between proteins or different regions
of the same protein.