Mass Spectrum Prediction from Molecular Fingerprint

This is an implementation of mass spectrum prediction from the Extended Connectivity Fingerprint (ECFP) and the Extended 3-Dimensional Fingerprint (E3FP) with a simple 5 layers' MLP model. The inputs are the SMILES strings of the molecules and the output is the MS2 level mass spectrum of the molecules.

Set up

Set up RDKit environment referring https://www.rdkit.org/docs/Install.html.

conda create -c conda-forge -n rdkit-env rdkit
conda activate rdkit-env 

Train & Test

The training and test datasets are split randomly with 9:1.

usage: main.py [-h] [--device DEVICE] [--num_mlp_layers NUM_MLP_LAYERS]
               [--drop_ratio DROP_RATIO] [--batch_size BATCH_SIZE]
               [--in_dim IN_DIM] [--emb_dim EMB_DIM] [--out_dim OUT_DIM]
               [--train_subset] [--epochs EPOCHS] [--num_workers NUM_WORKERS]
               [--radius RADIUS] [--fp_type FP_TYPE]
               [--train_data_path TRAIN_DATA_PATH]
               [--test_data_path TEST_DATA_PATH] [--data_type DATA_TYPE]
               [--log_dir LOG_DIR] [--checkpoint_path CHECKPOINT_PATH]
               [--resume_path RESUME_PATH]

GNN baselines on ogbgmol* data with Pytorch Geometrics

optional arguments:
  -h, --help            show this help message and exit
  --device DEVICE       which gpu to use if any (default: 0)
  --num_mlp_layers NUM_MLP_LAYERS
                        number of mlp layers (default: 6)
  --drop_ratio DROP_RATIO
                        dropout ratio (default: 0.2)
  --batch_size BATCH_SIZE
                        input batch size for training (default: 256)
  --in_dim IN_DIM       input dimensionality (default: 1024)
  --emb_dim EMB_DIM     embedding dimensionality (default: 1600)
  --out_dim OUT_DIM     output dimensionality (default: 2000)
  --train_subset
  --epochs EPOCHS       number of epochs to train (default: 200)
  --num_workers NUM_WORKERS
                        number of workers (default: 0)
  --radius RADIUS       radius (default: 2)
  --fp_type FP_TYPE     fingerprint type [2d | 3d] (default: 2d)
  --train_data_path TRAIN_DATA_PATH
                        path to training data
  --test_data_path TEST_DATA_PATH
                        path to test data
  --data_type DATA_TYPE
                        type of dataset (sdf or mgf)
  --log_dir LOG_DIR     tensorboard log directory
  --checkpoint_path CHECKPOINT_PATH
                        path to save checkpoint
  --resume_path RESUME_PATH
                        path to resume checkpoint

Command examples:

ECFP:

# NIST17 
# (positive) 
python main.py --train_data_path ./data/NIST17/train_single_nist_msms_posi.sdf \
	--test_data_path ./data/NIST17/test_single_nist_msms_posi.sdf \
	--data_type sdf \
	--log_dir ./logs \
	--checkpoint_path ./check_point/nist_posi.pt \
	--resume_path ./check_point/nist_posi.pt 
# (negative) 
python main.py --train_data_path ./data/NIST17/train_single_nist_msms_nega.sdf \
	--test_data_path ./data/NIST17/test_single_nist_msms_nega.sdf \
	--data_type sdf \
	--log_dir ./logs \
	--checkpoint_path ./check_point/nist_nega.pt \
	--resume_path ./check_point/nist_nega.pt
# GNPS 
# (positive) 
python main.py --out_dim 3000 --train_data_path ./data/GNPS/train_ALL_GNPS_posi_high.mgf \
	--test_data_path ./data/GNPS/test_ALL_GNPS_posi_high.mgf \
	--data_type mgf \
	--log_dir ./logs \
	--checkpoint_path ./check_point/gnps_posi.pt \
	--resume_path ./check_point/gnps_posi.pt 
# (negative) 
python main.py --out_dim 3000 --train_data_path ./data/GNPS/train_ALL_GNPS_nega_high.mgf \
	--test_data_path ./data/GNPS/test_ALL_GNPS_nega_high.mgf \
	--data_type mgf \
	--log_dir ./logs \
	--checkpoint_path ./check_point/gnps_nega.pt \
	--resume_path ./check_point/gnps_nega.pt

E3FP:

# NIST17 
# (positive) 
python main.py --fp_type 3d \
	--train_data_path ./data/NIST17/train_single_nist_msms_posi.sdf \
	--test_data_path ./data/NIST17/test_single_nist_msms_posi.sdf \
	--data_type sdf \
	--log_dir ./logs \
	--checkpoint_path ./check_point/nist_posi_3d.pt \
	--resume_path ./check_point/nist_posi_3d.pt 
# (negative) 
python main.py --fp_type 3d \
	--train_data_path ./data/NIST17/train_single_nist_msms_nega.sdf \
	--test_data_path ./data/NIST17/test_single_nist_msms_nega.sdf \
	--data_type sdf \
	--log_dir ./logs \
	--checkpoint_path ./check_point/nist_nega_3d.pt \
	--resume_path ./check_point/nist_nega_3d.pt
# GNPS 
# (positive) 
python main.py --fp_type 3d \
	--out_dim 3000 --train_data_path ./data/GNPS/train_posi_main_GNPS.mgf \
	--test_data_path ./data/GNPS/test_posi_main_GNPS.mgf \
	--data_type mgf \
	--log_dir ./logs \
	--checkpoint_path ./check_point/gnps_posi_3d.pt \
	--resume_path ./check_point/gnps_posi_3d.pt 
# (negative) 
python main.py --fp_type 3d \
	--out_dim 3000 --train_data_path ./data/GNPS/train_nega_main_GNPS.mgf \
	--test_data_path ./data/GNPS/test_nega_main_GNPS.mgf \
	--data_type mgf \
	--log_dir ./logs \
	--checkpoint_path ./check_point/gnps_nega_3d.pt \
	--resume_path ./check_point/gnps_nega_3d.pt

Performance

Dataset	Fingerprint	Accuracy (Cosine Similarity on Validation)
NIST17 (positive)	ECFP	0.4055
NIST17 (negative)	ECFP	0.3364
GNPS (positive)	ECFP	0.3629
GNPS (negative)	ECFP	0.5146
NIST17 (positive)	E3FP
NIST17 (negative)	E3FP
GNPS (positive)	E3FP
GNPS (negative)	E3FP