ChatMol
September 27, 2024 ยท View on GitHub
Code and data for Interactive Molecular Discovery with Natural Language. Related operation guides will be gradually improved in the near future.
Requirement
Main packages: python==3.9.0; torch==1.13.0; transformers==4.28.0; pandas==1.5.2; numpy==1.23.5; nltk==3.7
Data Preprocess
Pre-training
Related codes and data instances are in predata/.
getmsk.py: Masked language modeling.getM2T.py: Experimental property prediction & SMILES-to-name.getT2M.py: Literature-to-SMILES mapping correlation.getrdkit.py: Spatial structure prediction.
Fine-tuning
Related codes and data instances are in data/.
- train/validation/test.txt are ChEBI-20k dataset.
- PCtrain/PCdev/PCtest.txt are PCdes dataset.
- Files under
ChEBI-dia/are the new ChEBI-dia dataset.
Model Training and Inferencing
We stronly recommend using GPUs for this stage. First create an empty folder save_model/.
For knowledgeable and versatile pre-training, the jbl data files have to be first put under predata/. Change the pth args to combine different pre-training tasks. For instance:
CUDA_VISIBLE_DEVICES=0 python multi_pretraining.py --version base --save_pth save_model/understanding_pre.pt --pth predata/mixmsk.jbl --pth2 predata/preM2T.jbl --pth3 predata/spat.jbl
For molecule understanding tuning, remember to adjust the evaluate step length for different downstream datasets. For instance:
CUDA_VISIBLE_DEVICES=0 python molecule_understanding --pth_train data/PCtrain.txt --pth_dev data/PCdev.txt --pth_test data/PCtest.py --eval_len 700 --init_checkpoint save_model/pre_understanding_base.pt --save_pth save_model/understanding_
For molecule generation tuning, first conduct the fine-tuning, and then separately generate the test result and do evaluation for the specific checkpoint. For instance:
CUDA_VISIBLE_DEVICES=0 python molecule_generation.py --save_pth save_model/generation_noplug_
cd evaluation/
python test.py --init_checkpoint ../save_model/generation_noplug_100000.pt
You can also download the reference checkpoint from our Google Drive and put them under save_model/. Run the demo.py and provide your input according to the hints:
Input 'describe' to trigger molecule understanding of the current molecule, otherwise to use molecule generation. Input 'END' to clear dialog memory.
user: It derives from a tryptamine.
system: choose one molecule to continue. 1. C1=CC=C2C(=C1)C(=CN2)CCN 2. CNCCC1=CNC2=CC=CC=C21 3. C1=CC=C2C(=C1)C(=CN2)CC(C(=O)O)N
user: 2
user: It is an organic phosphate, a tryptamine alkaloid and a secondary amino compound.
system: choose one molecule to continue. 1. C1=CC=C2C(=C1)C(=CN2)CCN 2. C1=CC=C2C(=C1)C(=CN2)CC(C(=O)O)N 3. CNCCC1=CNC2=CC=CC=C21
user: 3
user: describe the current molecule
system: The molecule is a tryptamine alkaloid and a member of tryptamines. It has a role as a metabolite. It derives from a tryptamine.
user: It has a role as a hallucinogen and a fungal metabolite.
system: choose one molecule to continue. 1. C1=CC=C2C(=C1)C(=CN2)CC(C(=O)O)N 2. C[C@H](CC1=CNC2=CC=CC=C21)N[C@@H](C)C(=O)N[C@@H](CC3=CC=CC=C3)C(=O)O 3. C[C@H](CC1=CNC2=CC=CC=C21)OP(=O)(O)O
user: 3
user: END
user: It derives from a formic acid.
system: choose one molecule to continue. 1. CC(C)(COP(=O)(O)OP(=O)(O)OC[C@@H]1[C@H]([C@H]([C@@H](O1)N2C=NC3=C(N=CN=C32)N)O)OP(=O)(O)O)[C@H](C(=O)NCCC(=O)NCCSC=O)O 2. C(=O)[O-] 3. CC(C)(COP(=O)(O)OP(=O)(O)OC[C@@H]1[C@H]([C@H]([C@@H](O1)N2C=NC3=C(N=CN=C32)N)O)OP(=O)(O)O)[C@H](C(=O)NCCC(=O)N)O
user: 2
user: It has a role as a buffer.
system: choose one molecule to continue. 1. C(=O)[O-] 2. C(C(=O)O)N(C(=O)N)CC(=O)O 3. C(C(=O)O)N(CC(=O)O)CC(=O)O
user: 1
user: The molecule is the potassium salt of formic acid.
system: choose one molecule to continue. 1. C(=O)[O-].[K+] 2. [C@@H]([C@H]([C@@H]([C@H](C(=O)[O-])O)O)O)([C@H](C(=O)[O-])O)O.[K+] 3. [C@@H]([C@H]([C@@H]([C@H](C(=O)[O-])O)O)O)O.[K+]
user: 1
user: describe it
system: The molecule is a potassium salt comprising equal numbers of potassium and formic acid ions. It has a role as a fertilizer, a food acidity regulator, a human metabolite and a Saccharomyces cerevisiae metabolite. It contains a formate.
user: END