MolX: A Geometric Foundation Model for Protein–Ligand Modelling

This model is a graph Transformer foundation model that jointly learns geometric and chemical representations of protein pockets and ligands from large-scale 3D structural data. MolX integrates over 3 million protein pockets and 5 million molecules, representing both entities as E(3)-equivariant graphs that preserve spatial geometry and chemical context.

Pretrain:

1. Prepare the data. To pretrain MolX, users are supposed to download the pretrain datasets including MoltextNet, Pcqm4m-v2, and PDB pocket from .

2. Place files. Place the downloaded pretrained data in the corresponding paths in examplex/property_prediction/dataset.

3. Prepare the environment. Here we export our anaconda environment as the file "env.yaml". You can use the command:

    conda env create -f env.yaml
    conda activate MOLX

   to get the same environment. Also, we use one A100 GPU with Ubuntu 18.04 to conduct pretraining.

   Besides, we highly recommond to install openbabel (2.3.2) (https://openbabel.org/wiki/Main_Page) and preprocess the mol2 files.

   apt install openbabel

4. Run the training script.

   sh multiloss.sh

Finetuning:

Please run the corresponding shell files in examples/property_prediction to finetune different datasets. For example, to finetune MolX on Molecule Glue dataset, run the following script.

sh mg.sh