Introduction

This is the repository for our 3D-Transformer, which achieves state-of-the-art in many molecular representation learning problems. "3D-Transformer: Molecular Representation with Transformer in 3D Space (arXiv 2021)"
[arXiv]] [Code]

Intsallation

# Install packages
pip install pytorch scikit-learn mendeleev

Dataset

We test our model in three different domains: quantum chemistry, material science, and proteomics. You can download the raw dataset in the following links.

Quantum Chemistry

QM7 Dataset
Download (Official Website): http://quantum-machine.org/datasets/
Discription (DeepChem): https://deepchem.readthedocs.io/en/latest/api_reference/moleculenet.html#qm7-datasets

QM8 Dataset
Download (DeepChem): https://github.com/deepchem/deepchem/blob/master/deepchem/molnet/load_function/qm8_datasets.py
Discription (DeepChem): https://deepchem.readthedocs.io/en/latest/api_reference/moleculenet.html?highlight=qm7#qm8-datasets

QM9 Dataset Download (Atom3D): https://www.atom3d.ai/smp.html
Download (Official Website): https://ndownloader.figshare.com/files/3195389
Download (MPNN Supplement): https://drive.google.com/file/d/0Bzn36Iqm8hZscHFJcVh5aC1mZFU/view?resourcekey=0-86oyPL3e3l2ZTiRpwtPDBg
Download (Schnet): https://schnetpack.readthedocs.io/en/stable/tutorials/tutorial_02_qm9.html#Loading-the-data

GEOM-QM9 Dataset Download (Official Website): https://doi.org/10.7910/DVN/JNGTDF Tutorial of usage: https://github.com/learningmatter-mit/geom/blob/master/tutorials/01_loading_data.ipynb

Material Science

COREMOF
Download (Baidu Drive): https://pan.baidu.com/s/12N8gM8_TQ1mpBGx6gdkAog (password：l41s)
Reproduction of PointNet++: python coremof/reproduce/main_pn_coremof.py
Reproduction of MPNN: python coremof/reproduce/main_mpnn_coremof.py
Repredoction of SchNet: (1) load COREMOF python coremof/reproduce/main_sch_coremof.py
(2) run SchNet spk_run.py train schnet custom ../../coremof.db ./coremof --split 900 100 --property LCD --features 16 --batch_size 20 --cuda
(Note: official script of Schnet cannot be reproduced successfully due to the memory limitation.)

Proteomics

PDBbind
Atom3d: https://github.com/drorlab/atom3d
(1) download 'split-by-sequence-identity-30' dataset from https://www.atom3d.ai/
(2) install atom3D pip install atom3d
(3) preprocess the data by running python pdbbind/dataloader_pdb.py

Models

models/tr_spe: 3D-Transformer with Sinusoidal Position Encoding (SPE)
models/tr_cpe: 3D-Transformer with Convolutional Position Encoding (CPE)
models/tr_msa: 3D-Transformer with Multi-scale Self-attention (MSA)
models/tr_afps: 3D-Transformer with Attentive Farthest Point Sampling (AFPS)
models/tr_full: 3D-Transformer with CPE + MSA + AFPS

Quick Tour

After processing the dataset, it is time to establish the model. Suppose there are N types of atoms, and n downstream multi-tasks. If you only need to predict a single property, set n = 1. You can also speficy the number of attention heads, the number of encodes, the dimension size, the dropout rate, and etc, There we only adopt the defaults.

>>> import torch 
>>> from model.tr_spe import build_model
 
# initialize the model 
>>> model = build_model(N, n).cuda()

# take a 4-atom molecule for example
>>> x = torch.tensor([[1, 1, 6, 8]])
>>> pos = torch.tensor([[[7.356203877, 9.058198382, 3.255188164],
                         [5.990730587, 3.951633382, 9.784664946],
                         [1.048332315, 3.912215133, 9.827313903],
                         [2.492201352, 9.097616820, 3.297837121]]])
>>> mask = (x != 0).unsqueeze(1)
>>> out = model(x.long(), mask, pos)