tiejundong
commented
10 months ago Thanks for your attention to our work.
- You can download parameters here.
- The energy minimization improves the geometric correctness of local structures, e.g., aromatics/symmetric parts. If the local structure is not correct, one may consider the following: 1. enhancing energy minimization (set
min_loop=100or setmin_type='LBFGS'). 2. decreasing constraint energy minimizion constant (setmin_constraint=0). 3. Failed energy minimization in rare cases (i.e. failed to extract force field parameters in preprocessing). Although minimizing energy can improve local structural correctness, it cannot be expected to be utilized to improve accuracy as mentioned in our article and a recent study (Zhang, Xujun, et al. Nature Computational Science 3.9 (2023): 789-804.), especially for the highly flexible polymer-like ligand you provided, which is more challenging (Wang, Zhe, et al. Physical Chemistry Chemical Physics 18.18 (2016): 12964-12975.).
xqh19970407
Thanks for the work!
I tried to download the weights for local execution, but both at command line and direct download the speed is slow (< 160 Kb/s) and fails at around 80%.
When trying the web server version, I noticed that aromatics and double-bonded in non-protein compounds tend often to violate (more or less) co-planar constraints (could be a small setting for minimization in the web server?), which is not happening in other systems such as DiffDock. Moreover, big problems occurs when docking symmetric compounds that have multi-aromatics (see the attached example). Atoms did not stay in place, as it seems that during inference the system cannot distinguish the aromatic ring to which each atom belongs to, and final things result messed up. Thanks for any support!
Marco
Input_and_server_output.zip