Challenge for Deep\nLearning: Protein Structure Prediction\nof Ligand-Induced Conformational Changes at Allosteric and Orthosteric\nSites

In the realm of biomedical research,\nunderstanding the\nintricate\nstructure of proteins is crucial, as these structures determine how\nproteins function within our bodies and interact with potential drugs.\nTraditionally, methods like X-ray crystallography and cryo-electron\nmicroscopy have been used to unravel these structures, but they are\noften challenging, time-consuming and costly. Recently, a breakthrough\nin computational biology has emerged with the development of deep\nlearning algorithms capable of predicting protein structures based\non their amino acid sequences (Jumper, J., et al. <i>Nature</i> <b>2021</b>, <i>596</i>, 583. Lane, T. J. <i>Nature Methods</i> <b>2023</b>, <i>20</i>,\n170. Kryshtafovych, A., et al. <i>Proteins: Structure, Function\nand Bioinformatics</i> <b>2021</b>, <i>89</i>, 1607). This study focuses on predicting the dynamic changes that\nproteins undergo upon ligand binding, specifically when they bind\nto allosteric sites, i.e. a pocket different from the active site.\nAllosteric modulators are particularly important for drug discovery,\nas they open new avenues for designing drugs that can target proteins\nmore effectively and with fewer side effects (Nussinov, R.; Tsai,\nC. J. <i>Cell</i> <b>2013</b>, <i>153</i>, 293). To study this, we curated a data set of 578 X-ray structures\ncomprised of proteins displaying orthosteric and allosteric binding\nas well as a general framework to evaluate deep learning-based structure\nprediction methods. Our findings demonstrate the potential and current\nlimitations of deep learning methods, such as AlphaFold2 (Jumper,\nJ., et al. <i>Nature</i> <b>2021</b>, <i>596</i>, 583), NeuralPLexer (Qiao, Z., et al. <i>Nat Mach Intell</i> <b>2024</b>, <i>6</i>, 195), and RoseTTAFold All-Atom\n(Krishna, R., et al. <i>Science</i> <b>2024</b>, <i>384</i>, eadl2528) to predict not just static protein structures\nbut also the dynamic conformational changes. Herein we show that predicting\nthe allosteric induce-fit conformation still poses a challenge to\ndeep learning methods as they more accurately predict the orthosteric\nbound conformation compared to the allosteric induce fit conformation.\nFor AlphaFold2, we observed that conformational diversity, and sampling\nbetween the apo and holo state could be increased by modifying the\nMSA depth, but this did not enhance the ability to generate conformations\nclose to the allosteric induced-fit conformation. To further support\nadvancements in protein structure prediction field, the curated data\nset and evaluation framework are made publicly available.