Specifically, we will first conduct a performance study comparing representative protein structure alignment methods, including both traditional non deep learning (non-DL) algorithms and recent deep learning (DL) models. We will not only measure the alignment result accuracy but also the performance of downstream tasks including protein classification, evolution history construction, and protein function inference. These downstream tasks exercise alignment intensively, and their results are of scientific or practical importance. With insights from the proposed performance study, we will develop a new GPU-accelerated non-DL protein structure alignment algorithm with high alignment accuracy. Finally, we will design a new DL-based evolution-aware protein structure alignment algorithm. DL models perform poorly on pair-wise alignment accuracy, but they may embed evolutionary information and can potentially assign higher scores to amino acids that are evolutionarily more important.
In summary, our proposed research will advance the state of art by (1) revealing the performance characteristics of a wide selection of existing structure alignment tools, (2) accelerating a non-DL method with high alignment accuracy on the GPU, and (3) developing a new DL method that is effective for downstream tasks. Our code packages will be put in the public domain for researchers and practitioners to follow up and apply in real-world protein structure alignment and downstream tasks.