Precisely Predicting Bio-aware 3D Conformation for Cancer-related Proteins with Deep Active Learning (Artificial Intelligence in Medicine Alliance)

Accurate protein 3D conformations are critical to many cancer research areas, including but not limited to discovering new druggable sites for regulating tumor cell metabolism and screening drugs for protein mutation-induced therapy resistance of cancers. However, obtaining protein conformation is notoriously difficult and expensive through wet-lab experiments. Deep learning tools, such as alphaFold2 (AF2) and RosettaFold (RF), have been recently developed for predicting protein conformations on a large scale. With accuracy comparable to X-ray crystallography, AF2 and RF are reshaping cancer research by instantly providing protein conformations as long as amino acid sequences are known. This proposal will integrate AF2 for initial structure prediction, molecular dynamics (MD) simulation for protein conformation annotation, and deep graph neural networks to reconstruct cancer proteins to their conformations in a biological solvent (Aim 1). The proposed algorithm will leverage the power of the recent advances in deep learning and in-parallel MD simulation. The results will be systematically validated regarding accuracy and efficiency in refining protein conformations at different scales and structural complexities (Aim 2). In summary, we propose to develop a deep active learning algorithm that refines AF2-generated conformations into NMR-like derived conformations by utilizing a molecular dynamics generated conformational search-space. This project will result in a novel AI algorithm and a repository of bio-aware conformations of cancer proteins.