Skip to content

Latest commit

 

History

History
383 lines (309 loc) · 14.5 KB

antibiotics.md

File metadata and controls

383 lines (309 loc) · 14.5 KB
Raw

Generating Novel Antibiotics with SyntheMol

Instructions for generating antibiotic candidates for Acinetobacter baumannii. 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).

Process antibiotics training data

The antibiotics training data consists of three libraries of molecules tested against the Gram-negative bacterium Acinetobacter baumannii. Here, we merge the three libraries into a single file and determine which molecules are hits (i.e., active against A. baumannii).

Merge the three libraries into a single file and determine which molecules are hits.

python scripts/data/process_data.py \
    --data_paths data/Data/1_training_data/library_1.csv data/Data/1_training_data/library_2.csv data/Data/1_training_data/library_3.csv \
    --save_path data/Data/1_training_data/antibiotics.csv \
    --save_hits_path data/Data/1_training_data/antibiotics_hits.csv \
    --activity_column antibiotic_activity

Output:

library_1
Data size = 2,371
Mean activity = 0.9759337199493885
Std activity = 0.2673771821337539
Activity threshold of mean - 2 std = 0.4411793556818807
Number of hits = 130
Number of non-hits = 2,241

library_2
Data size = 6,680
Mean activity = 0.9701813713618264
Std activity = 0.17623388953330232
Activity threshold of mean - 2 std = 0.6177135922952217
Number of hits = 294
Number of non-hits = 6,386

library_3
Data size = 5,376
Mean activity = 0.9980530249533108
Std activity = 0.13303517604898704
Activity threshold of mean - 2 std = 0.7319826728553367
Number of hits = 112
Number of non-hits = 5,264

Full data size = 14,427
Data size after dropping non-conflicting duplicates = 13,594
Data size after dropping conflicting duplicates = 13,524

Final data size = 13,524
Number of hits = 470
Number of non-hits = 13,054

Note: Newer versions of RDKit canonicalize SMILES differently, which may result in different deduplication results. The results above were generated with RDKit version 2022.3.4.

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 August 8, 2022, are available in the Google Drive folder in the data/Data/2_chembl subfolder.

The files are:

  • chembl_antibacterial.csv with 604 molecules (589 with SMILES)
  • chembl_antibiotic.csv with 636 molecules (587 with SMILES)

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

python scripts/data/merge_chembl_downloads.py \
    --data_paths data/Data/2_chembl/chembl_antibacterial.csv data/Data/2_chembl/chembl_antibiotic.csv \
    --labels antibacterial antibiotic \
    --save_path data/Data/2_chembl/chembl.csv

The file chembl.csv contains 1,005 molecules.

Build bioactivity prediction models

Here, we build three binary classification bioactivity prediction models to predict antibiotic activity against A. baumannii. The three models are:

  1. Chemprop: a graph neural network model
  2. Chemprop-RDKit: a graph neural network model augmented with 200 RDKit features
  3. Random forest: a random forest model using 200 RDKit features

Train models

For each model type, we trained 10 models using 10-fold cross-validation on the training data. Each ensemble of 10 models took less than 90 minutes to train on a 16-core CPU machine.

Chemprop

python scripts/models/train.py \
    --data_path data/Data/1_training_data/antibiotics.csv \
    --save_dir data/Models/antibiotic_chemprop \
    --dataset_type classification \
    --model_type chemprop \
    --property_column antibiotic_activity \
    --num_models 10

Chemprop-RDKit

python scripts/models/train.py \
    --data_path data/Data/1_training_data/antibiotics.csv \
    --save_dir data/Models/antibiotic_chemprop_rdkit \
    --model_type chemprop \
    --dataset_type classification \
    --fingerprint_type rdkit \
    --property_column antibiotic_activity \
    --num_models 10

Random forest

python scripts/models/train.py \
    --data_path data/Data/1_training_data/antibiotics.csv \
    --save_dir data/Models/antibiotic_random_forest \
    --model_type random_forest \
    --dataset_type classification \
    --fingerprint_type rdkit \
    --property_column antibiotic_activity \
    --num_models 10
Model ROC-AUC PRC-AUC
Chemprop 0.803 +/- 0.036 0.354 +/- 0.086
Chemprop-RDKit 0.827 +/- 0.028 0.388 +/- 0.078
Random forest 0.835 +/- 0.035 0.401 +/- 0.099

Compute model scores for building blocks

In order to speed up the generative model, we pre-compute the scores of all building blocks for each model.

Chemprop

python scripts/models/predict.py \
    --data_path data/Data/4_real_space/building_blocks.csv \
    --save_path data/Models/antibiotic_chemprop/building_blocks.csv \
    --model_path data/Models/antibiotic_chemprop \
    --model_type chemprop \
    --average_preds

Chemprop-RDKit

python scripts/models/predict.py \
    --data_path data/Data/4_real_space/building_blocks.csv \
    --save_path data/Models/antibiotic_chemprop_rdkit/building_blocks.csv \
    --model_path data/Models/antibiotic_chemprop_rdkit \
    --model_type chemprop \
    --fingerprint_type rdkit \
    --average_preds

Random forest

python scripts/models/predict.py \
    --data_path data/Data/4_real_space/building_blocks.csv \
    --save_path data/Models/antibiotic_random_forest/building_blocks.csv \
    --model_path data/Models/antibiotic_random_forest \
    --model_type random_forest \
    --fingerprint_type rdkit \
    --average_preds

Generate molecules with SyntheMol

Here, we apply SyntheMol to generate molecules using a Monte Carlo tree search (MCTS). We limit the search space to molecules that can be formed using a single reaction (as opposed to multi-step reactions) to ensure easy, quick, and cheap chemical synthesis. We run SyntheMol three times, each time using a different property prediction model to score molecules and guide the search.

Chemprop

synthemol \
    --model_path data/Models/antibiotic_chemprop \
    --model_type chemprop \
    --building_blocks_path data/Models/antibiotic_chemprop/building_blocks.csv \
    --building_blocks_score_column chemprop_ensemble_preds \
    --building_blocks_id_column Reagent_ID \
    --reaction_to_building_blocks_path data/Data/4_real_space/reaction_to_building_blocks_filtered.pkl \
    --save_dir data/Data/6_generations_chemprop \
    --max_reactions 1 \
    --n_rollout 20000 \
    --replicate

Chemprop-RDKit

synthemol \
    --model_path data/Models/antibiotic_chemprop_rdkit \
    --model_type chemprop \
    --building_blocks_path data/Models/antibiotic_chemprop_rdkit/building_blocks.csv \
    --building_blocks_score_column chemprop_rdkit_ensemble_preds \
    --building_blocks_id_column Reagent_ID \
    --reaction_to_building_blocks_path data/Data/4_real_space/reaction_to_building_blocks_filtered.pkl \
    --save_dir data/Data/7_generations_chemprop_rdkit \
    --fingerprint_type rdkit \
    --max_reactions 1 \
    --n_rollout 20000 \
    --replicate

Random forest

synthemol \
    --model_path data/Models/antibiotic_random_forest \
    --model_type random_forest \
    --building_blocks_path data/Models/antibiotic_random_forest/building_blocks.csv \
    --building_blocks_score_column random_forest_rdkit_ensemble_preds \
    --building_blocks_id_column Reagent_ID \
    --reaction_to_building_blocks_path data/Data/4_real_space/reaction_to_building_blocks_filtered.pkl \
    --save_dir data/Data/8_generations_random_forest \
    --fingerprint_type rdkit \
    --max_reactions 1 \
    --n_rollout 20000 \
    --replicate

Filter generated molecules

Run three filtering steps to select molecules for experimental testing based on novelty, predicted bioactivity, and diversity.

Novelty

Filter for molecules that are structurally novel, i.e., dissimilar from the training hits and the ChEMBL antibacterials.

First, compute the nearest neighbor Tversky similarity between the generated molecules and the training hits or ChEMBL antibacterials to determine the novelty of each generated molecule. Tversky similarity is used to determine what proportion of the functional groups of known antibacterials are contained within the generated molecules.

Compute Tversky similarity to train hits.

for NAME in 6_generations_chemprop 7_generations_chemprop_rdkit 8_generations_random_forest
do
chemfunc nearest_neighbor \
    --data_path data/Data/${NAME}/molecules.csv \
    --reference_data_path data/Data/1_training_data/antibiotics_hits.csv \
    --reference_name antibiotics_hits \
    --metric tversky
done

Compute Tversky similarity to ChEMBL antibacterials.

for NAME in 6_generations_chemprop 7_generations_chemprop_rdkit 8_generations_random_forest
do
chemfunc nearest_neighbor \
    --data_path data/Data/${NAME}/molecules.csv \
    --reference_data_path data/Data/2_chembl/chembl.csv \
    --reference_name chembl \
    --metric tversky
done

Now, filter to only keep molecules with nearest neighbor Tverksy similarity to the train hits or ChEMBL antibacterials <= 0.5.

Filter by Tversky similarity to train hits.

for NAME in 6_generations_chemprop 7_generations_chemprop_rdkit 8_generations_random_forest
do
chemfunc filter_molecules \
    --data_path data/Data/${NAME}/molecules.csv \
    --save_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5.csv \
    --filter_column antibiotics_hits_tversky_nearest_neighbor_similarity \
    --max_value 0.5
done

Filter by Tversky similarity to ChEMBL antibacterials.

for NAME in 6_generations_chemprop 7_generations_chemprop_rdkit 8_generations_random_forest
do
chemfunc filter_molecules \
    --data_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5.csv \
    --save_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5_chembl_sim_below_0.5.csv \
    --filter_column chembl_tversky_nearest_neighbor_similarity \
    --max_value 0.5
done

Bioactivity

Filter for high-scoring molecules (i.e., predicted bioactivity) by only keeping molecules with the top 20% of model scores.

for NAME in 6_generations_chemprop 7_generations_chemprop_rdkit 8_generations_random_forest
do
chemfunc filter_molecules \
    --data_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5_chembl_sim_below_0.5.csv \
    --save_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5_chembl_sim_below_0.5_top_20_percent.csv \
    --filter_column score \
    --top_proportion 0.2
done

Diversity

Filter for diverse molecules by clustering molecules based on their Morgan fingerprint and only keeping the top scoring molecule from each cluster.

Cluster molecules based on their Morgan fingerprint.

for NAME in 6_generations_chemprop 7_generations_chemprop_rdkit 8_generations_random_forest
do
chemfunc cluster_molecules \
    --data_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5_chembl_sim_below_0.5_top_20_percent.csv \
    --num_clusters 50
done

Select the top scoring molecule from each cluster.

for NAME in 6_generations_chemprop 7_generations_chemprop_rdkit 8_generations_random_forest
do
chemfunc select_from_clusters \
    --data_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5_chembl_sim_below_0.5_top_20_percent.csv \
    --save_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5_chembl_sim_below_0.5_top_20_percent_selected_50.csv \
    --value_column score
done

We now have 50 molecules selected from each model's generations that meet our desired novelty, score, and diversity criteria.

Map molecules to REAL IDs

Map generated molecules to REAL IDs in the format expected by Enamine to enable a lookup in the Enamine REAL Space database.

for NAME in 6_generations_chemprop 7_generations_chemprop_rdkit 8_generations_random_forest
do
python scripts/data/map_generated_molecules_to_real_ids.py \
    --data_path data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5_chembl_sim_below_0.5_top_20_percent_selected_50.csv \
    --save_dir data/Data/${NAME}/molecules_antibiotics_hits_sim_below_0.5_chembl_sim_below_0.5_top_20_percent_selected_50_real_ids
done

Predict toxicity

Predict the toxicity of the synthesized molecules by training a toxicity prediction model on the ClinTox dataset from MoleculeNet. ClinTox is a clinical trial toxicity binary classification dataset with 1,478 molecules with 112 (7.58%) toxic.

Train toxicity model using the CT_TOX property of the ClinTox dataset. (Note that the ClinTox properties CT_TOX and FDA_APPROVED are nearly always identical, so we only use one.)

python scripts/models/train.py \
    --data_path data/Data/1_training_data/clintox.csv \
    --save_dir data/Models/clintox_chemprop_rdkit \
    --model_type chemprop \
    --dataset_type classification \
    --fingerprint_type rdkit \
    --property_column CT_TOX \
    --num_models 10

The model has an ROC-AUC of 0.881 +/- 0.045 and a PRC-AUC of 0.514 +/- 0.141 across 10-fold cross-validation.

Make toxicity predictions on the synthesized molecules. The list of successfully synthesized generated molecules is in the data/Data/8_synthesized subfolder of the Google Drive folder.

python scripts/models/predict.py \
    --data_path data/Data/8_synthesized/synthesized.csv \
    --model_path data/Models/clintox_chemprop_rdkit \
    --model_type chemprop \
    --fingerprint_type rdkit \
    --average_preds