SAMGPT

This is my final year project,works on inverse design of self-assembled-monolayers (Also other types of molecule if you want)

prop_finetuning
This file contains code for both structure-conditional and property-conditional molecular generation.
Compared with the original MolGPT, I introduced an adapter module and classifier-free guidance, which enable the model to be fine-tuned using a mixture of data (structures with or without property values (N/A)). This method allows the model to incorporate information from both the pre-trained structural features and the fine-tuned property values.

scaf_finetuning
This file contains codes for just strcture constrained generation.

salience_map This file contains codes for visualization of the generation. Highlighting next token with gradient colors to represent its probability to be generated.

Dataset I used.

Source	Dataset	Samples	Block Size (SMILES Len)	Maximum Scaffold Length
Zinc	MOSES	1.9 million	54 (train), 51 (validation)	48
ChEMBL	Guacamol	1.6 million	100 (train), 99 (validation)	100
10.1021/acs.jcim.6b00340	frontier orbitals	111,725 mols	148	115
PSC literatures	SAMs	200	202	123

Training steps

1. Before getting the pre-trained model weight, remember to check the vocabulary of you dataset. Vocabulary size mismatch might leads to failure!
2. I suggest first training the base model with MOSES and Guacamol with scaffold constrained. Each pre-trained has its different applications due to the distribution of SMILES length
3. Do scaffold finetuning (SAMs or frontier orbitals) or property finetuning (frontier orbitals).

Adapted model architecture

Script for finetuning.

nohup python ./prop_finetuning/cond_finetune.py 
  --run_name homo_lumo_Fine_tuning 
  --train_data_name full_prop_train_mix 
  --val_data_name full_prop_valid 
  --cond_props homo lumo 
  --pretrained_model_weight ./weights/scaffold_guacamol_all.pt 
  --output_ckpt homo_lumo_fine_tunning_mix.pt 
  --batch_size 200 
  --max_epochs 10 
  >> ./homo_lumo_Fine_tuning_mix.log 2>&1 &

Script for conditional generation

nohup python ./prop_finetuning/scaf_prop_generate.py 
  --model_weight ./homo_lumo_fine_tunning_mix.pt 
  --scaffold 
  --csv_name gen_homo_lumo_mix.csv 
  --cond_props homo lumo 
  --gen_size 1000 
  --vocab_size 143 
  --block_size 202 
  --batch_size 200 
  >> ./prop_homo_lumo_generated_fine_tune.log 2>&1 &

Model weight for reproducing

Google Drive: Model Weights

Property conditional generation results

Single property (LogP)

Multi-properties (LogP, TPSA)

Generated structures (-5.3 eV HOMO, -1.02 eV LUMO)