Drug repurposing identifies new therapeutic uses for existing drugs beyond their original medical indications. Existing safety profiles and pharmacological knowledge can reduce the time, cost, and risk of developing alternative treatments. section 7.1.3
A computational repurposing workflow can:
- Represent compounds as molecular fingerprints, such as 2,048-bit Morgan fingerprints generated from SMILES strings. section 7.1
- Use dimensionality reduction to visualize high-dimensional screening data and expose groups of structurally or pharmacologically similar compounds. section 7.1 section 7.4
- Train a self-organizing map, label each node by the plurality activity class of mapped training compounds, and treat nearby unseen compounds as repurposing candidates. section 7.1.3
- Evaluate candidates against retrospective hits and prioritize predicted hits for further study rather than treating model assignments as clinical evidence. section 7.1.3
From antihypertensive to alopecia treatment
Minoxidil was developed as an antihypertensive and later repurposed after observations showed that it promoted hair growth, leading to treatment for alopecia. section 7.1.3
SOM screening example
A SOM node containing three active, two inactive, and one unknown training compound receives the active label by majority vote; unseen compounds assigned there become hits for further investigation. section 7.1.3
Listing 7.3 Apply & Evaluate SOM for Drug Repurposing
def majority_label(labels):
return np.argmax(np.bincount(labels))
win_map = som.win_map(fingerprints, return_indices=True)
node_labels = {}
for position, instances in win_map.items():
instance_labels = [data['label'].iloc[idx] for idx in instances]
node_labels[position] = \ #A
majority_label(instance_labels) #A
file_path = \ #B
'data/ch07/raw/CH07_drug_repurposing_hub.csv' #B
drh_data = pd.read_csv(file_path)
drh_data = drh_data.dropna(subset=['smiles'])
retrospective_hits = [ #C
'atpenin a5', 'bictegravir', 'nelfinavir', #C
'prulifloxacin', 'tegobuvir', 'tinostamustine', #C
] #C
drh_data['Morgan_Fingerprint'] = drh_data['smiles'].apply(
generate_morgan_fingerprint)
drh_data = drh_data.dropna(subset=['Morgan_Fingerprint'])
unseen_fingerprints = np.array(
[list(fp) for fp in drh_data['Morgan_Fingerprint']])
predicted_activities = [] #D
for fingerprint in unseen_fingerprints:
winner_node = som.winner(fingerprint)
predicted_activity = node_labels.get( #E
winner_node, None) #E
predicted_activities.append(predicted_activity)
drh_data['predicted_labels'] = predicted_activities
drh_data['predicted_activities'] =
label_names.get)
predicted_drh_hits = set(
drh_data[drh_data['predicted_activities'] == 'Active']['pert_iname'])
drh_compounds = set(drh_data['pert_iname'])
# Retrospective hits present in the DRH, and those the SOM flagged as active.
retrospective_hits_in_drh = drh_compounds & set(retrospective_hits)
captured_drh_hits = predicted_drh_hits & retrospective_hits_in_drh
recall = len(captured_drh_hits) / len(retrospective_hits_in_drh)
print(f"Recall on retrospective hits: {recall:.2%}")
Output:
Recall on retrospective hits: 100.00%Inspect the implementation pattern for labeling SOM nodes, featurizing Drug Repurposing Hub compounds, and evaluating retrospective hits. section 7.1.3

