TRIDS: AI-native molecular docking framework unified with binding site identification, conformation sampling and scoring
By Xuhan Liu & Hong Zhang, on JUN 26th 2026
Please see the LICENSE file for the license terms for the software. Basically it's free to academic users. If you do wish to sell the software or use it in a commercial product, then please contact us:
Xuhan Liu (First Author): xuhan.liu@xtalpi.com
Hong Zhang (Correspondent Author): zhangh@pku.edu.cn
Molecular docking is a cornerstone of drug discovery for unveiling the mechanism of ligand-receptor interactions. With the recent advances of deep learning (DL), AI-powered molecular docking methods have achieved higher accuracy for binding pose prediction and virtual screening compared with classical physics-based methods. However, there is still a scarcity of approaches to strike a balance among accuracy, computational efficiency, and rigorous physical validity of the output conformations. In the previous two versions of DSDP, we demonstrated the effectiveness of guiding conformation sampling with the gradient of an analytic scoring function. As the third version, TRIDS is devised as an AI-native docking framework that expand the similar strategy to unify conformation sampling and docking processes with DL-based model for improving accuracy of docking and screening. Furthermore, it is tailored for seamless cooperation of AI and physics to guarantee the physical validity of predicted binding poses. Being user-friendly, TRIDS predicts the binding site, parses multiple file formats, and supports Python programming and PyMOL graphical interaction. It improves docking accuracy and passes physical validation with high computational efficiency, i.e. a single docking task is done in a fraction of second while maintaining a highly lightweight GPU memory footprint of merely hundreds of megabytes, facilitating high-throughput virtual screening in reality. As a proof of concept, TRIDS allowed us to obtain hit compounds with novel scaffolds for tumor necrosis factor-alpha (TNFα) inhibitor through a large-scale virtual screening.
TRIDS is implemented with PyTorch C++ (LibTorch). It combined multiple optimization of streams parallel computing，CUDA graph-operator merge, automatic differentiation, CUDA kernel function to accelerate the program dramatically. General speaking, any differentiable ML-based scoring function could be compatible with this framework for conformation sampling.
The main SF in the present work is the mixture density network (MDN), which consists of three key components: a feature extraction module, a feature concatenation module, and a feed-forward output layer. The graphs obtained from the previous step are processed by multiple Graph Transformer blocks to ensure that the processed node features contain not only information about an individual atom or residue in the molecule, but also information about the surrounding nodes. The resulting graphs are then pairwise-concatenated to model the interaction and fed into a feed-forward layer to parameterize a Gaussian mixed model.
The sampling space is defined on the degrees of freedom for each molecule, including translation, rotation and torsion angles of rotatable bonds, the ranges of which are [box_min, box_max], [-π, π) and [-π, π), respectively. In the beginning, each copy of conformation was initiated with randomization of the degrees of freedom involved. During the computational loop, one variable was randomly perturbed to update the conformation at first. Subsequently, the updated conformation was optimized through the gradient descent method. Adam with learning rate 0.1 was used to update the variables, in which the above-mentioned ML-based scoring model was differentiated to provide gradients. At the end of each loop, the perturbation was accepted according to Metropolis criterion.
To run the compiled TRIDS, some dependent packages need to be installed. You could create an new Conda environment with these required packages:
For CUDA-12.6:
conda env create -f env-cu126.yml
or CUDA-11.8:
conda env create -f env-cu118.yml
After creation, activating the Conda environment trids:
conda activate trids
Usually, you had better to check if the following required packages have been installed:
-
PyTorch (version >= 2.6)
-
OpenBabel (version >= 3.1.1)
-
CLI11 (version >= 2.0)
-
spdlog (version >= 1.0)
-
pybind11 (version >= 2.11) Note: Pybind11 is required only when you want to install PyTrids for Python.
python setup.py install
or
pip install -e . --no-build-isolation
Note: If you want to uninstall this package, run:
pip uninstall trids
First of all, make sure that PyTrids and PyMOL has been installed in the same Conda environment
conda install pymol-open-source -c conda-forge
Method 1: (Recommended)
- Run PyMOL
- Menu: Plugin -> Plugin Manager -> Install New Plugin -> Choose File ...
- Choose: ~/plugins/pymol/trids_gui.py
- Restart PyMOL
- If installation is complete, the menu in plugin has "TRIDS Settings"
Method 2: (Manual) Copy trids_gui.py into the startup folder of PyMOL.
On Linux:
cp plugins/pymol/trids_gui.py ~/.pymol/startup/
on Windows:
copy plugins/pymol/trids_gui.py C:\Users\<Username>\.pymol\startup\
Method 3: (Temporary) Run the following code in the command line of PyMOL
run ~/plugins/pymol/trids_gui.py
trids -h
on GPU:
python tests/test_trids.py -g 0
on CPU:
python tests/test_trids.py -g -1
python tests/test_pymol.py
Note:
-
If usage instruction outputs after running trids -h, it has been installed successfully. Otherwise, you have to set some environmental variable manually, on Linux:
export PATH=$PATH:<path/for/trids>/bin export LD_LIBRARY_PATH=$CONDA_PREFIX/lib/python3.12/site-packages/torch/lib:$CONDA_PREFIX/lib:$LD_LIBRARY_PATHon Windows:
set PATH=%CONDA_PREFIX%\Lib\site-packages\torch\lib;%CONDA_PREFIX%\Library\bin;%CONDA_PREFIX%\Library\lib;%PATH% -
If python outputs the correct information, installation has been done. Here, -g denotes the CUDA id.
The following packages need to be installed manually only if this project is recompiled:
-
cmake (version >= 3.18, < 4.0)
conda install cmake==3.22 -c conda-forge -
Nvidia CUDA (version >= 11.8)
version == 11.8:
conda install cuda==11.8.0 -c nvidia/label/cuda-11.8.0or version == 12.6:
conda install cuda==12.6.0 -c nvidia/label/cuda-12.6.0
Note: The version of PyTorch should be consistent with the version of CUDA.
If installed CUDA version == 11.8, and default GCC version >= 12, the following package must be installed as C++ compiler.
-
GCC (version <= 11.2, Optional).
conda install gcc_linux-64==11.2 gxx_linux-64==11.2
-
Creating an empty folder for compilation:
mkdir build && cd build -
CMake configuration
cmake ../cmake/linux -DCMAKE_INSTALL_PREFIX=<path/for/trids> -DBUILD_PYTHON=ON -
Compilation & Installation
make install -j 8
Note:
- -DBUILD_PYTHON for cmake should be ON, if the PyTrids will be installed later manually; default is OFF;
- -DCMAKE_INSTALL_PREFIX is folder path for installation, please replace <path/for/trids> with appropriate location; default is /usr/local
- -j 8 means that the code will be compiled with eight CPU cores, this number could be set manually based on your own device.
-
Ninja (version >= 1.0.0)
conda install ninja -c conda-forge
Note: Please download and install Microsoft Visual C++ Build Tools manually.
-
The recommend version == 2017 and no more higher than 2022. If its version == 2022, install the following package into Conda
conda install vs2022-win_64 -c conda-forge
cmake/windows/build.bat
trids -r <receptor_path> -l <ligand_path> [-k <refrence_ligand_path>] [OPTIONS]
Options:
-h,--help Print this help message and exit
-r,--receptor <pdb> Rigid part of the receptor [REQUIRED]
-l,--ligand <smi, mol2, sdf, pdb> Ligand [REQUIRED]
-k,--pocket <pt, pth, mol2, sdf, pdb> Reference profile for Binding site identification
--hts High throughput screening mode with higher virtual screening accuracy but lower binding pose accuracy
-o,--out <string> File path for outputing docking results, the format of molecules is based on file extension
-e,--stream <uint> [256] Max number of sampling for Monte carlo research
-d,--depth <uint> [32] Max depths for Monte carlo research
-t,--top <uint> [1] Record Number of N best conformers in output
--seed <uint> User defined random seed
-c,--cpu <uint> [1] Number of CPU cores
-g,--cuda <int> [0] Index of Nvidia CUDA Device. If set to -1, no CUDA device will be used
--score_only Only calculating the score of given conformation without sampling
-v,--verbose <0,1,2> [0] Verbose mode, print more information, 0: Error, 1: Warning, 2: Info
--config Read an ini file
Note:
- For C++ version, the compiled program trids is in the ./bin file. You have to set it to your $PATH manually.
- For Python version, you could run trids directly, as long as the Conda envrionment of "trids" is activated.
Users could use eigher CLI mode or GUI mode with TRIDS plugin inside the PyMOL window.
To make sure the correct usage of PyMOL plugin, users should run this program with two steps:
- runing "trisite" to extract the binding site from the interested protein
- runing "triscore" or "trids" with the binding site object generated by "trisite"
trisite receptor [, reference [, cutoff [, show ]]]
Options:
receptor = str: PyMOL selection as the receptor
reference = str: (optional) reference ligand for pocket definition
cutoff = float: pocket radius in Angstrom (default: 8.0)
show = 0/1: auto-visualize pockets in PyMOL (default: 1)
triscore pocket, ligand [, use_vina ]
Options:
pocket = str: PyMOL selection generated by trisite
ligand = str: PyMOL selection as the ligand
use_vina = 0/1: use Vina scoring (default: 0)
trids pocket, ligand [, top_n [, streams [, depth [, use_vina [, name ]]]]]
Options:
pocket = str: PyMOL selection generated by trisite
ligand = str: PyMOL selection as the ligand
top_n = int: number of top conformations to return (default: 1)
streams = int: number of parallel sampling tasks (default: 256)
depth = int: Monte Carlo search depth (default: 32)
use_vina = 0/1: use Vina scoring function instead of TRI (default: 0)
name = str: base name for loaded result objects (default: trids_dock)
tridev [dev, num]
Options:
dev = str: computing device: either "gpu" or "cpu"
num = int: If "gpu", it is gpu id, otherwise, it is number of cores.
- Shenzhen Bay Laboratory
- Westlake University
- Changping Laboratory
- Peking University
- XtalPi Technologies Co., Ltd