Learning Protein Structural Fingerprints under the Label-Free Supervision of Domain Knowledge

Finding homologous proteins is the indispensable first step in many protein biology studies. Thus, building highly efficient “search engines” for protein databases is a highly desired function in protein bioinformatics. As of August 2018, there are more than 140,000 protein structures in PDB, and this number is still increasing rapidly. Such a big number introduces a big challenge for scanning the whole structure database with high speeds and high sensitivities at the same time. Unfortunately, classic sequence alignment tools and pairwise structure alignment tools are either not sensitive enough to remote homologous proteins (with low sequence identities) or not fast enough for the task. Therefore, specifically designed computational methods are required for quickly scanning structure databases for homologous proteins.Here, we propose a novel ContactLib-DNN method to quickly scan structure databases for homologous proteins. The core idea is to build structure fingerprints for proteins, and to perform alignment-free comparisons with the fingerprints. Specifically, the fingerprints are low-dimensional vectors representing the contact groups within the proteins. Notably, the Cartesian distance between two fingerprint vectors well matches the RMSD between the two corresponding contact groups. This is done by using RMSD as the domain knowledge to supervise the deep neural network learning. When comparing to existing methods, ContactLib-DNN achieves the highest average AUROC of 0.959. Moreover, the best candidate found by ContactLib-DNN has a probability of 70.0% to be a true positive. This is a significant improvement over 56.2%, the best result produced by existing methods.GitHub: https://github.com/Chenyao2333/contactlib/