Challenge #6

Introduction：Our plan combines structure-based methods by computational biology and empirical knowledge with artificial intelligence (AI) techniques. By leveraging the powerful predictive capabilities of AI algorithms and the computational speed of GPUs, we aim to screen large molecular libraries efficiently. High-precision computational chemistry methods will enhance the hit rate of active molecules. Currently, only one compound with nanomolar activity has been reported for the triple Tudor domain of SETDB1. The discovery of additional lead compounds is urgently needed to accelerate clinical translation.Therefore, we will identify key components based on previous studies of SETDB1 to filter out the final potential active compounds. Below is a brief description of our methodology:

Methodology：1.Active Learning and Molecular Docking: We will employ active learning combined with molecular docking to screen large molecular libraries, such as Enamine.2.ADMET Property Screening: The ADMET properties of the final compounds will be checked, and compounds with low solubility or other undesirable properties will be removed.3.Redocking and Key Interaction Filtering: The results from the active learning protocol will be redocked, and compounds will be filtered based on key interactions derived from crystal structures and literature, such as hydrogen bonds and π-π stacking.4.AI-Based Force Field Analysis: Using the AI-based force field (AIMNet2), we will estimate the strain energy of docking poses, removing high-energy poses (including global strain and torsion strain) and checking the number of unsatisfied heteroatoms for H-bonding.5.Substructure Filtering: Key substructures summarized from previous research will be used to filter compounds, while maintaining the diversity of the final results.6.Molecular Dynamics Simulation: Stability and binding free energy will be assessed through MD simulations, utilizing the ASGBIE method for accurate calculations.7.For lead optimation, we will apply our development of tools (MolOpt, FragRep, and FragGrow) to optimize structures.

Detailed Workflow

1. Binding Site Analysis and Optimization:Initially, we will utilize protein structure alignment tools and long-term Molecular Dynamics (MD) simulations to identify and optimize binding sites within the SETDB1 triple Tudor domain (TTD) . We will analyze key interactions at these binding sites to refine subsequent screening results. Additionally, we will summarize key substructures from active molecules for secondary screening of hit molecules.

2. Initial Screening with Molecular Docking and Active Learning:Our strategy incorporates molecular docking and active learning to screen the molecular library. The process begins with extracting a subset of molecules from the library, which are docked into the glycine binding pocket of the LBD region to obtain docking scores and binding poses. Based on these docking scores and molecular structures, we train a graph neural network to quickly predict docking scores and associated uncertainties, forming an initial model.Using the initial model's uncertainty predictions, we select molecules with high predicted uncertainty from the molecular library for re-docking to expand the training dataset. We then retrain the model to enhance the accuracy of docking score predictions. This iterative process continues until the model's predictive performance is optimized. Using the optimized model, we select the highest-scoring molecules from the library, re-dock them, and obtain the binding poses. Compounds with low solubility and other undesirable ADMET properties are removed.

3. Secondary Screening and Energy Filtering:Finally, we conduct secondary screening and complex structure analysis based on key interactions and substructures identified from crystal structures. High-strain energy docking poses are filtered out using an AI-based force field. This process aims to select the best molecules and docking poses for further dynamic simulations.

4. Dynamic Simulations and Free Energy Calculations:Following dynamic simulations, we analyze the stability and key conformational changes of the complexes to assess the structural stability and conformational changes of key residues during dynamics. Subsequently, we calculate the binding free energy using the original alanine scanning binding interaction entropy (ASGBIE) method. Initially, ASGBIE is applied to reported complex structures to evaluate its prediction performance against experimental values and to identify hotspot amino acids. This step also involves optimizing method parameters and assessing structural reliability. For the complexes obtained from prior screenings, we calculate the binding energy and identify candidate molecules with binding energies superior to reported compounds for experimental validation.Free energy perturbation will also be used for further confirmation and screening of similar candidate compounds.