Generating Novel Antibiotics with SyntheMol-RL

Instructions for generating antibiotic candidates for Staphylococcus aureus using SyntheMol-RL from the paper SyntheMol-RL: a flexible reinforcement learning framework for designing novel and synthesizable antibiotics.

This includes instructions for processing antibiotics data, training antibacterial activity prediction models, generating molecules with SyntheMol, and selecting candidates. Assumes relevant data has already been downloaded ( see docs/README.md).

Data
Build bioactivity prediction models
Generate molecules with SyntheMol-RL
Screen molecules with Chemprop-RDKit
Select generated molecules
Select screened molecules
Visualize selected molecules

Data

Process S. aureus training data

The training data consists of molecules tested for inhibitory activity against Staphylococcus aureus. The following command processes the data to compute binary activity labels based on the inhibition values based on the mean (of two replicates) normalized 16-hour optical density (OD) values.

python scripts/data/process_data.py \
    --data_paths rl/data/s_aureus/s_aureus_raw.csv \
    --value_column "16h Normalized Mean" \
    --save_path rl/data/s_aureus/s_aureus.csv \
    --save_hits_path rl/data/s_aureus/s_aureus_hits.csv \
    --activity_column s_aureus_activity

Output:

Data size = 10,716
Mean activity = 0.8938596024636059
Std activity = 0.3028107500370331
Activity threshold of mean - 2 std = 0.2882381023895396
Number of hits = 1,149
Number of non-hits = 9,567

Full data size = 10,716
Data size after dropping non-conflicting duplicates = 10,660
Data size after dropping conflicting duplicates = 10,658

Final data size = 10,658
Number of hits = 1,137
Number of non-hits = 9,521

Process ChEMBL antibacterials

Download lists of known antibiotic-related compounds from ChEMBL using the following search terms. For each, click the CSV download button, unzip the downloaded file, and rename the CSV file appropriately.

The results of the queries on November 8, 2023, are available in the Google Drive folder in the data/chembl subfolder.

The files are:

chembl_antibacterial.csv with 611 molecules (590 with SMILES)
chembl_antibiotic.csv with 636 molecules (591 with SMILES)

Merge these two files to form a single collection of antibiotic-related compounds.

python scripts/data/merge_chembl_downloads.py \
    --data_paths rl/data/chembl/chembl_antibacterial.csv rl/data/chembl/chembl_antibiotic.csv \
    --labels antibacterial antibiotic \
    --save_path rl/data/chembl/chembl.csv

The file chembl.csv contains 1,007 molecules.

Solubility data

Aqueous solubility data was obtained from ADMET-AI, which preprocessed data from the Therapeutics Data Commons. This dataset contains 9,982 molecules with aqueous solubility measurements in units of log mol/L. The data is saved to data/solubility/solubility.csv.

Build bioactivity prediction models

Here, we build three binary classification bioactivity prediction models to predict antibiotic activity against S. aureus and to predict aqueous solubility. The three models are:

Chemprop: a graph neural network model
Chemprop-RDKit: a graph neural network model augmented with 200 RDKit features
MLP-RDKit: a multilayer perceptron using 200 RDKit features

Compute RDKit features

Pre-compute the 200 RDKit features for the training data and the building blocks.

Training data

chemfunc save_fingerprints \
    --data_path rl/data/s_aureus/s_aureus.csv \
    --fingerprint_type rdkit \
    --save_path rl/data/s_aureus/s_aureus.npz

Time: 1 minute, 28 seconds with an 8-core machine.

Solubility data

chemfunc save_fingerprints \
    --data_path rl/data/solubility/solubility.csv \
    --fingerprint_type rdkit \
    --save_path rl/data/solubility/solubility.npz

Time: 2 minutes, 0 seconds with an 8-core machine.

Building blocks

for CHEMICAL_SPACE in real wuxi
do
chemfunc save_fingerprints \
    --data_path rl/data/${CHEMICAL_SPACE}/building_blocks.csv \
    --fingerprint_type rdkit \
    --save_path rl/data/${CHEMICAL_SPACE}/building_blocks.npz
done

Time: REAL = 10 minutes, 7 seconds; WuXi = 2 minutes, 3 seconds with an 8-core machine.

Train models

For each model type, train 10 models using 10-fold cross-validation.

Chemprop for S. aureus

chemprop_train \
    --data_path rl/data/s_aureus/s_aureus.csv \
    --dataset_type classification \
    --target_column s_aureus_activity \
    --num_folds 10 \
    --split_type cv \
    --metric prc-auc \
    --extra_metrics auc \
    --save_dir rl/models/s_aureus_chemprop \
    --save_preds \
    --quiet

Chemprop for solubility

chemprop_train \
    --data_path rl/data/solubility/solubility.csv \
    --dataset_type regression \
    --target_column solubility \
    --num_folds 10 \
    --split_type cv \
    --metric mae \
    --extra_metrics r2 \
    --save_dir rl/models/solubility_chemprop \
    --save_preds \
    --quiet

Chemprop-RDKit for S. aureus

chemprop_train \
    --data_path rl/data/s_aureus/s_aureus.csv \
    --dataset_type classification \
    --target_column s_aureus_activity \
    --features_path rl/data/s_aureus/s_aureus.npz \
    --no_features_scaling \
    --num_folds 10 \
    --split_type cv \
    --metric prc-auc \
    --extra_metrics auc \
    --save_dir rl/models/s_aureus_chemprop_rdkit \
    --save_preds \
    --quiet

Chemprop-RDKit for solubility

chemprop_train \
    --data_path rl/data/solubility/solubility.csv \
    --dataset_type regression \
    --target_column solubility \
    --features_path rl/data/solubility/solubility.npz \
    --no_features_scaling \
    --num_folds 10 \
    --split_type cv \
    --metric mae \
    --extra_metrics r2 \
    --save_dir rl/models/solubility_chemprop_rdkit \
    --save_preds \
    --quiet

MLP-RDKit for S. aureus

chemprop_train \
    --data_path rl/data/s_aureus/s_aureus.csv \
    --dataset_type classification \
    --target_column s_aureus_activity \
    --features_path rl/data/s_aureus/s_aureus.npz \
    --no_features_scaling \
    --features_only \
    --num_folds 10 \
    --split_type cv \
    --metric prc-auc \
    --extra_metrics auc \
    --save_dir rl/models/s_aureus_mlp_rdkit \
    --save_preds \
    --quiet

MLP-RDKit for solubility

chemprop_train \
    --data_path rl/data/solubility/solubility.csv \
    --dataset_type regression \
    --target_column solubility \
    --features_path rl/data/solubility/solubility.npz \
    --no_features_scaling \
    --features_only \
    --num_folds 10 \
    --split_type cv \
    --metric mae \
    --extra_metrics r2 \
    --save_dir rl/models/solubility_mlp_rdkit \
    --save_preds \
    --quiet

Results for S. aureus (10-fold cross-validation, 8-core, 1-GPU machine):

Model	ROC-AUC	PRC-AUC	Time
Chemprop	0.862 +/- 0.013	0.527 +/- 0.034	68m, 34s
Chemprop-RDKit	0.875 +/- 0.013	0.570 +/- 0.050	79m, 38s
MLP-RDKit	0.873 +/- 0.018	0.553 +/- 0.040	60m, 09s

Results for solubility (10-fold cross-validation, 8-core, 1-GPU machine):

Model	MAE	R^2	Time
Chemprop	0.693 +/- 0.027	0.806 +/- 0.021	35m, 59s
Chemprop-RDKit	0.656 +/- 0.023	0.822 +/- 0.021	36m, 59s
MLP-RDKit	0.688 +/- 0.020	0.817 +/- 0.013	24m, 22s

The Chemprop-RDKit models are the best for both S. aureus and solubility. Therefore, those models are used to evaluate building blocks and molecules.

Add labels to test preds

When saving predictions, Chemprop does not save the labels. This command adds the labels to the predictions files (and also corrects the formatting of SMILES strings).

for MODEL in chemprop chemprop_rdkit mlp_rdkit
do
python scripts/data/add_labels_to_test_preds.py \
    --true_path rl/data/s_aureus/s_aureus.csv \
    --preds_path rl/models/s_aureus_${MODEL} \
    --value_column s_aureus_activity

python scripts/data/add_labels_to_test_preds.py \
    --true_path rl/data/solubility/solubility.csv \
    --preds_path rl/models/solubility_${MODEL} \
    --value_column solubility
done

Compute model scores for building blocks

After training, use the Chemprop-RDKit models to pre-compute scores of building blocks.

for CHEMICAL_SPACE in real wuxi
do
for MODEL in s_aureus solubility
do
chemprop_predict \
    --test_path rl/data/${CHEMICAL_SPACE}/building_blocks.csv \
    --smiles_column smiles \
    --preds_path rl/data/${CHEMICAL_SPACE}/building_blocks.csv \
    --checkpoint_dir rl/models/${MODEL}_chemprop_rdkit \
    --features_path rl/data/${CHEMICAL_SPACE}/building_blocks.npz \
    --no_features_scaling
done
done

Time with an 8-core, 1-GPU machine:

Model	Chemical Space	Time
S. aureus	REAL	20m, 11s
S. aureus	WuXi	2m, 28s
Solubility	REAL	20m, 06s
Solubility	WuXi	2m, 19s

Generate molecules with SyntheMol-RL

Generate molecules with SyntheMol-RL using RL models for S. aureus and solubility with dynamic multiparameter and exploring REAL & WuXi chemical spaces.

Note: Logically, the RL prediction types for the two models should be classification and regression. However, at the time this experiment was run, the code required a single model type for both models. Therefore, the models were trained as regression models and used as such in the generation.

RL Chemprop-RDKit

synthemol \
    --score_model_paths rl/models/s_aureus_chemprop_rdkit rl/models/solubility_chemprop_rdkit \
    --score_types chemprop chemprop \
    --score_fingerprint_types rdkit rdkit \
    --score_names 'S. aureus' 'Solubility' \
    --success_thresholds '>=0.5' '>=-4' \
    --chemical_spaces real wuxi \
    --building_blocks_paths rl/data/real/building_blocks.csv rl/data/wuxi/building_blocks.csv \
    --building_blocks_score_columns s_aureus_activity solubility \
    --reaction_to_building_blocks_paths rl/data/real/reaction_to_building_blocks.pkl rl/data/wuxi/reaction_to_building_blocks.pkl \
    --save_dir rl/generations/rl_chemprop_rdkit_s_aureus_solubility_dynamic_weights_real_wuxi \
    --n_rollout 10000 \
    --search_type rl \
    --rl_model_type chemprop \
    --rl_model_fingerprint_type rdkit \
    --rl_model_paths rl/models/s_aureus_chemprop_rdkit/fold_0/model_0/model.pt rl/models/solubility_chemprop_rdkit/fold_0/model_0/model.pt \
    --rl_prediction_types regression regression \
    --use_gpu \
    --num_workers 8 \
    --replicate_rl \
    --wandb_project_name synthemol_rl \
    --wandb_run_name rl_chemprop_rdkit_s_aureus_solubility_dynamic_weights_real_wuxi \
    --wandb_log

RL MLP-RDKit

synthemol \
    --score_model_paths rl/models/s_aureus_chemprop_rdkit rl/models/solubility_chemprop_rdkit \
    --score_types chemprop chemprop \
    --score_fingerprint_types rdkit rdkit \
    --score_names 'S. aureus' 'Solubility' \
    --success_thresholds '>=0.5' '>=-4' \
    --chemical_spaces real wuxi \
    --building_blocks_paths rl/data/real/building_blocks.csv rl/data/wuxi/building_blocks.csv \
    --building_blocks_score_columns s_aureus_activity solubility \
    --reaction_to_building_blocks_paths rl/data/real/reaction_to_building_blocks.pkl rl/data/wuxi/reaction_to_building_blocks.pkl \
    --save_dir rl/generations/rl_mlp_rdkit_s_aureus_solubility_dynamic_weights_real_wuxi \
    --n_rollout 10000 \
    --search_type rl \
    --rl_model_type mlp \
    --rl_model_fingerprint_type rdkit \
    --rl_model_paths rl/models/s_aureus_mlp_rdkit/fold_0/model_0/model.pt rl/models/solubility_mlp_rdkit/fold_0/model_0/model.pt \
    --rl_prediction_types regression regression \
    --replicate_rl \
    --wandb_project_name synthemol_rl \
    --wandb_run_name rl_mlp_rdkit_s_aureus_solubility_dynamic_weights_real_wuxi \
    --wandb_log

MCTS

synthemol \
    --score_model_paths rl/models/s_aureus_chemprop_rdkit rl/models/solubility_chemprop_rdkit \
    --score_types chemprop chemprop \
    --score_fingerprint_types rdkit rdkit \
    --score_names 'S. aureus' 'Solubility' \
    --success_thresholds '>=0.5' '>=-4' \
    --chemical_spaces real wuxi \
    --building_blocks_paths rl/data/real/building_blocks.csv rl/data/wuxi/building_blocks.csv \
    --building_blocks_score_columns s_aureus_activity solubility \
    --reaction_to_building_blocks_paths rl/data/real/reaction_to_building_blocks.pkl rl/data/wuxi/reaction_to_building_blocks.pkl \
    --save_dir rl/generations/mcts_s_aureus_solubility_dynamic_weights_real_wuxi \
    --n_rollout 10000 \
    --search_type mcts \
    --replicate_rl \
    --wandb_project_name synthemol_rl \
    --wandb_run_name mcts_s_aureus_solubility_dynamic_weights_real_wuxi \
    --wandb_log

Screen molecules with Chemprop-RDKit

Chemprop-RDKit on random REAL 14 million and 100 and WuXi 7 million and 50.

for SPACE in real wuxi
do

if [ "${SPACE}" = "real" ]; then
  SIZES=("14m" "100")
else
  SIZES=("7m" "50")
fi

for SIZE in "${SIZES[@]}"
do
FILE_NAME="random_${SPACE}_${SIZE}"

chemfunc save_fingerprints \
    --data_path rl/data/${SPACE}/${FILE_NAME}.csv \
    --fingerprint_type rdkit \
    --save_path rl/data/${SPACE}/${FILE_NAME}.npz

for PROPERTY in s_aureus solubility
do
chemprop_predict \
    --test_path rl/data/${SPACE}/${FILE_NAME}.csv \
    --smiles_column smiles \
    --preds_path rl/screened/chemprop_rdkit_${FILE_NAME}_${PROPERTY}.csv \
    --checkpoint_dir rl/models/${PROPERTY}_chemprop_rdkit \
    --features_path rl/data/${SPACE}/${FILE_NAME}.npz \
    --no_features_scaling \
    --no_cache_mol
done

python -c "import pandas as pd
data = pd.read_csv('rl/screened/chemprop_rdkit_${FILE_NAME}_s_aureus.csv')
sol = pd.read_csv('rl/screened/chemprop_rdkit_${FILE_NAME}_solubility.csv')
data['solubility'] = sol['solubility']
data.to_csv('rl/screened/chemprop_rdkit_${FILE_NAME}.csv', index=False)"

done
done

Select generated molecules

Compute similarity to training hits and ChEMBL antibiotics.

for MODEL in rl_chemprop_rdkit rl_mlp_rdkit mcts
do
chemfunc nearest_neighbor \
    --data_path rl/generations/${MODEL}_s_aureus_solubility_dynamic_weights_real_wuxi/molecules.csv \
    --reference_data_path rl/data/s_aureus/s_aureus_hits.csv \
    --reference_name train_hits \
    --metric tversky

chemfunc nearest_neighbor \
    --data_path rl/generations/${MODEL}_s_aureus_solubility_dynamic_weights_real_wuxi/molecules.csv \
    --reference_data_path rl/data/chembl/chembl.csv \
    --reference_name chembl \
    --metric tversky
done

Select the top 150 diverse, novel hit molecules. Hits are defined as S. aureus >= 0.5 and solubility >= -4. Novelty is defined as maximum 0.6 Tversky similarity to training hits and ChEMBL antibiotics. Diversity is defined as maximum 0.6 Tanimoto similarity to other selected molecules (maximum independent set). Final selection is the top 150 diverse, novel hits molecules sorted by S. aureus score.

for MODEL in rl_chemprop_rdkit rl_mlp_rdkit mcts
do
python scripts/data/select_molecules.py \
    --data_path rl/generations/${MODEL}_s_aureus_solubility_dynamic_weights_real_wuxi/molecules.csv \
    --save_molecules_path rl/selections/${MODEL}/molecules.csv \
    --save_analysis_path rl/selections/${MODEL}/analysis.csv \
    --score_columns "S. aureus" "Solubility" \
    --score_comparators ">=0.5" ">=-4" \
    --novelty_thresholds 0.6 0.6 \
    --similarity_threshold 0.6 \
    --select_num 150 \
    --sort_column "S. aureus" \
    --descending
done

Select screened molecules

Filter by hits, where hits are defined as S. aureus >= 0.5 and solubility >= -4.

for NAME in real_14m wuxi_7m
do
chemfunc filter_molecules \
    --data_path rl/screened/chemprop_rdkit_random_${NAME}.csv \
    --save_path rl/screened/chemprop_rdkit_random_${NAME}_hits.csv \
    --filter_column "s_aureus_activity" \
    --min_value 0.5

chemfunc filter_molecules \
    --data_path rl/screened/chemprop_rdkit_random_${NAME}_hits.csv \
    --save_path rl/screened/chemprop_rdkit_random_${NAME}_hits.csv \
    --filter_column "solubility" \
    --min_value -4
done

REAL Output:

Original data size = 14,000,000
Data size after filtering with min_value 0.5 = 3,702
Final data size = 3,702

Original data size = 3,702
Data size after filtering with min_value -4.0 = 853
Final data size = 853

Note: The REAL random 14M has 3,702 (0.026%) with S. aureus >= 0.5; 11,371,677 (81.23%) with solubility >= -4; and 853 (0.006%) with both.

WuXi Output:

Original data size = 7,000,000
Data size after filtering with min_value 0.5 = 105,055
Final data size = 105,055

Original data size = 105,055
Data size after filtering with min_value -4.0 = 437
Final data size = 437

Note: The WuXi random 7M has 105,055 (1.50%) with S. aureus >= 0.5; 1,868,378 (26.69%) with solubility >= -4; and 437 (0.006%) with both.

Combine REAL and WuXi hits.

python -c "import pandas as pd
pd.concat([
    pd.read_csv('rl/screened/chemprop_rdkit_random_real_14m_hits.csv').assign(chemical_space='real'),
    pd.read_csv('rl/screened/chemprop_rdkit_random_wuxi_7m_hits.csv').assign(chemical_space='wuxi')
]).to_csv('rl/screened/chemprop_rdkit_random_real_wuxi_21m_hits.csv', index=False)"

Compute similarity to training hits and ChEMBL antibiotics.

chemfunc nearest_neighbor \
    --data_path rl/screened/chemprop_rdkit_random_real_wuxi_21m_hits.csv \
    --reference_data_path rl/data/s_aureus/s_aureus_hits.csv \
    --reference_name train_hits \
    --metric tversky

chemfunc nearest_neighbor \
    --data_path rl/screened/chemprop_rdkit_random_real_wuxi_21m_hits.csv \
    --reference_data_path rl/data/chembl/chembl.csv \
    --reference_name chembl \
    --metric tversky

python scripts/data/select_molecules.py \
    --data_path rl/screened/chemprop_rdkit_random_real_wuxi_21m_hits.csv \
    --save_molecules_path rl/selections/chemprop_rdkit/molecules.csv \
    --save_analysis_path rl/selections/chemprop_rdkit/analysis.csv \
    --score_columns "s_aureus_activity" "solubility" \
    --score_comparators ">=0.5" ">=-4" \
    --novelty_thresholds 0.6 0.6 \
    --similarity_threshold 0.6 \
    --select_num 150 \
    --sort_column "s_aureus_activity" \
    --descending

Merge random molecules (no filtering).

mkdir rl/selections/random

python -c "import pandas as pd
pd.concat([
    pd.read_csv('rl/screened/chemprop_rdkit_random_real_100.csv').assign(chemical_space='real'),
    pd.read_csv('rl/screened/chemprop_rdkit_random_wuxi_50.csv').assign(chemical_space='wuxi')
]).to_csv('rl/selections/random/molecules.csv', index=False)"

Visualize selected molecules

Create images of the structures of the selected molecules.

for NAME in rl_chemprop_rdkit rl_mlp_rdkit mcts chemprop_rdkit random
do
chemfunc visualize_molecules \
    --data_path rl/selections/${NAME}/molecules.csv \
    --save_dir rl/selections/${NAME}
done

Plot t-SNE of the selected molecules vs training hits and ChEMBL antibiotics.

chemfunc plot_tsne \
    --data_paths rl/data/chembl/chembl.csv \
    rl/data/s_aureus/s_aureus_hits.csv \
    rl/selections/random/molecules.csv \
    rl/selections/chemprop_rdkit/molecules.csv \
    rl/selections/mcts/molecules.csv \
    rl/selections/rl_mlp_rdkit/molecules.csv \
    rl/selections/rl_chemprop_rdkit/molecules.csv \
    --data_names chembl train_hits random chemprop_rdkit mcts rl_mlp_rdkit rl_chemprop_rdkit \
    --save_path rl/plots/selections_tsne.pdf \
    --point_size 2000

Compute ADMET properties of selected molecules

Compute ADMET properties of selected molecules using ADMET-AI.

Note: This requires installing ADMET-AI via pip install admet_ai.

for NAME in rl_chemprop_rdkit rl_mlp_rdkit mcts chemprop_rdkit random
do
admet_predict \
    --data_path rl/selections/${NAME}/molecules.csv
done

Merge selections into a single file

Merge all selections into a single file.

python -c "import pandas as pd

# Combine all selections
data = pd.concat([
    pd.read_csv('rl/selections/rl_chemprop_rdkit/molecules.csv').assign(method='rl_chemprop_rdkit'),
    pd.read_csv('rl/selections/rl_mlp_rdkit/molecules.csv').assign(method='rl_mlp_rdkit'),
    pd.read_csv('rl/selections/mcts/molecules.csv').assign(method='mcts'),
    pd.read_csv('rl/selections/chemprop_rdkit/molecules.csv').assign(method='chemprop_rdkit'),
    pd.read_csv('rl/selections/random/molecules.csv').assign(method='random')
]).reset_index(drop=True)
print(data['method'].value_counts())

# Combine chemical space labels into one column
data['chemical_space'] = data['chemical_space'].combine_first(data['reaction_1_chemical_space'])

# Save
data.to_csv('rl/selections/all_molecules.csv', index=False)"

Separate selections by chemical space

Separate selections by chemical space.

python -c "import pandas as pd
data = pd.read_csv('rl/selections/all_molecules.csv')
for chemical_space in ['real', 'wuxi']:
    space_data = data[data['chemical_space'] == chemical_space]
    print(f'{chemical_space}: {len(space_data):,}')
    space_data.to_csv(f'rl/selections/{chemical_space}_molecules.csv', index=False)"

Select final candidates based on availability and ClinTox score

Select 50 molecules from each model from the available molecules ranked by ClinTox score (lowest to highest), except for random where random molecules are selected. Note that rl/selections/all_molecules.csv has been modified to include a column indicating which molecules are available based on quotes from Enamine and WuXi.

python -c "import pandas as pd

# Load molecules
data = pd.read_csv('rl/selections/all_molecules.csv')

# Get available molecules
available = data[data['available']]

# Select top 50 molecules by ClinTox score for each method (except random)
selected = []
for method in sorted(data['method'].unique()):
    method_data = available[available['method'] == method]

    if method == 'random':
        real_data = method_data[method_data['chemical_space'] == 'real']
        wuxi_data = method_data[method_data['chemical_space'] == 'wuxi']
        method_data = pd.concat([real_data.sample(33, random_state=0), wuxi_data.sample(17, random_state=0)])
    else:
        method_data = method_data.sort_values('ClinTox').head(50)

    selected.append(method_data)

selected = pd.concat(selected)

# Save selected molecules
selected.to_csv('rl/selections/all_molecules_available_clintox.csv', index=False)

# Save selected molecules by chemical space
for chemical_space in sorted(selected['chemical_space'].unique()):
    selected[selected['chemical_space'] == chemical_space].to_csv(
        f'rl/selections/{chemical_space}_molecules_available_clintox.csv',
        index=False
    )"

Calculate additional similarity metrics for novelty

In addition to the previously computed Tversky similarity, compute the Tanimoto and Maximum Common Substructure (MCS) similarity metrics for the final set of hits.

Tanimoto

chemfunc nearest_neighbor \
    --data_path rl/selections/final_hits.csv \
    --reference_data_path rl/data/s_aureus/s_aureus_hits.csv \
    --reference_name train_hits \
    --metric tanimoto

chemfunc nearest_neighbor \
    --data_path rl/selections/final_hits.csv \
    --reference_data_path rl/data/chembl/chembl.csv \
    --reference_name chembl \
    --metric tanimoto

Asymmetric MCS (normalized by reference molecule size, similar to Tversky normalization)

chemfunc nearest_neighbor \
    --data_path rl/selections/final_hits.csv \
    --reference_data_path rl/data/s_aureus/s_aureus_hits.csv \
    --reference_name train_hits_asymmetric \
    --metric mcs \
    --ring_matches_ring_only \
    --denominator mol_2

chemfunc nearest_neighbor \
    --data_path rl/selections/final_hits.csv \
    --reference_data_path rl/data/chembl/chembl.csv \
    --reference_name chembl_asymmetric \
    --metric mcs \
    --ring_matches_ring_only \
    --denominator mol_2

Symmetric MCS (average normalization by both molecule sizes, similar to Tanimoto normalization)

chemfunc nearest_neighbor \
    --data_path rl/selections/final_hits.csv \
    --reference_data_path rl/data/s_aureus/s_aureus_hits.csv \
    --reference_name train_hits_symmetric \
    --metric mcs \
    --ring_matches_ring_only \
    --denominator avg

chemfunc nearest_neighbor \
    --data_path rl/selections/final_hits.csv \
    --reference_data_path rl/data/chembl/chembl.csv \
    --reference_name chembl_symmetric \
    --metric mcs \
    --ring_matches_ring_only \
    --denominator avg

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Generating Novel Antibiotics with SyntheMol-RL

Data

Process S. aureus training data

Process ChEMBL antibacterials

Solubility data

Build bioactivity prediction models

Compute RDKit features

Train models

Add labels to test preds

Compute model scores for building blocks

Generate molecules with SyntheMol-RL

Screen molecules with Chemprop-RDKit

Select generated molecules

Select screened molecules

Visualize selected molecules

Compute ADMET properties of selected molecules

Merge selections into a single file

Separate selections by chemical space

Select final candidates based on availability and ClinTox score

Calculate additional similarity metrics for novelty

FilesExpand file tree

antibiotics.md

Latest commit

History

antibiotics.md

File metadata and controls

Generating Novel Antibiotics with SyntheMol-RL

Data

Process S. aureus training data

Process ChEMBL antibacterials

Solubility data

Build bioactivity prediction models

Compute RDKit features

Train models

Add labels to test preds

Compute model scores for building blocks

Generate molecules with SyntheMol-RL

Screen molecules with Chemprop-RDKit

Select generated molecules

Select screened molecules

Visualize selected molecules

Compute ADMET properties of selected molecules

Merge selections into a single file

Separate selections by chemical space

Select final candidates based on availability and ClinTox score

Calculate additional similarity metrics for novelty