Basic Concepts¶

This tutorial is an introduction to medicinal chemistry molecular filtering. It will show that applying those type of filters and rules systematically and blindly is often not a good idea. While powerful, such filtering technics must always be carefully assesed and prototyped before using systematically and at large scale.

In [1]:

import matplotlib.pyplot as plt
import matplotlib.colors
import seaborn as sns

from rdkit.Chem import PandasTools
import datamol as dm
import pandas as pd

import medchem as mc

import matplotlib.pyplot as plt import matplotlib.colors import seaborn as sns from rdkit.Chem import PandasTools import datamol as dm import pandas as pd import medchem as mc

Marketed drugs does not always pass medchem filters¶

The most obvious example for not blindly applying medicinal filters, alerts and rules systematically is when looking at the already approved and marketed drugs. The below example shows that for many common filters, a large proportion of approved drugs does not pass them.

That being said it's important to keep in mind that all the approved drugs has gone into a very long and lengthy development process in which their structures have been further optimized and finetuned. During that process it's very common to include or exclude particular features that early during a drug pipeline could have been seen as an unwanted feature but turned out to be a beneficial one during the late stages of a program.

In [2]:

# Load ChEMBL approved and marketed drugs
data = pd.read_parquet("./data/chembl_approved_drugs.parquet")

# Build mol objects
data["mol"] = data["smiles"].apply(dm.to_mol)

# Apply basic rules
data["rule_of_five"] = data["smiles"].apply(mc.rules.basic_rules.rule_of_five)
data["rule_of_ghose"] = data["smiles"].apply(mc.rules.basic_rules.rule_of_ghose)
data["rule_of_veber"] = data["smiles"].apply(mc.rules.basic_rules.rule_of_veber)
data["rule_of_zinc"] = data["smiles"].apply(mc.rules.basic_rules.rule_of_zinc)

# Apply some default medchem filters
data["alerts_BMS"] = mc.functional.alert_filter(
    mols=data["mol"].tolist(),
    alerts=["BMS"],
    n_jobs=-1,
    progress=True,
    return_idx=False,
)
data["alerts_PAINS"] = mc.functional.alert_filter(
    mols=data["mol"].tolist(),
    alerts=["PAINS"],
    n_jobs=-1,
    progress=True,
    return_idx=False,
)

data["alerts_SureChEMBL"] = mc.functional.alert_filter(
    mols=data["mol"].tolist(),
    alerts=["SureChEMBL"],
    n_jobs=-1,
    progress=True,
    return_idx=False,
)

data["filters_NIBR"] = mc.functional.nibr_filter(
    mols=data["mol"].tolist(),
    n_jobs=-1,
    progress=True,
    return_idx=False,
)

data["filter_complexity"] = mc.functional.complexity_filter(
    mols=data["mol"].tolist(),
    complexity_metric="bertz",
    threshold_stats_file="zinc_15_available",
    n_jobs=-1,
    progress=True,
    return_idx=False,
)

data["filter_bredt"] = mc.functional.bredt_filter(
    mols=data["mol"].tolist(),
    n_jobs=-1,
    progress=True,
    return_idx=False,
)

data["filter_molecular_graph"] = mc.functional.molecular_graph_filter(
    mols=data["mol"].tolist(),
    max_severity=5,
    n_jobs=-1,
    progress=True,
    return_idx=False,
)

data["filter_lilly_demerit"] = mc.functional.lilly_demerit_filter(
    mols=data["mol"].tolist(),
    n_jobs=-1,
    progress=True,
    return_idx=False,
)

# Load ChEMBL approved and marketed drugs data = pd.read_parquet("./data/chembl_approved_drugs.parquet") # Build mol objects data["mol"] = data["smiles"].apply(dm.to_mol) # Apply basic rules data["rule_of_five"] = data["smiles"].apply(mc.rules.basic_rules.rule_of_five) data["rule_of_ghose"] = data["smiles"].apply(mc.rules.basic_rules.rule_of_ghose) data["rule_of_veber"] = data["smiles"].apply(mc.rules.basic_rules.rule_of_veber) data["rule_of_zinc"] = data["smiles"].apply(mc.rules.basic_rules.rule_of_zinc) # Apply some default medchem filters data["alerts_BMS"] = mc.functional.alert_filter( mols=data["mol"].tolist(), alerts=["BMS"], n_jobs=-1, progress=True, return_idx=False, ) data["alerts_PAINS"] = mc.functional.alert_filter( mols=data["mol"].tolist(), alerts=["PAINS"], n_jobs=-1, progress=True, return_idx=False, ) data["alerts_SureChEMBL"] = mc.functional.alert_filter( mols=data["mol"].tolist(), alerts=["SureChEMBL"], n_jobs=-1, progress=True, return_idx=False, ) data["filters_NIBR"] = mc.functional.nibr_filter( mols=data["mol"].tolist(), n_jobs=-1, progress=True, return_idx=False, ) data["filter_complexity"] = mc.functional.complexity_filter( mols=data["mol"].tolist(), complexity_metric="bertz", threshold_stats_file="zinc_15_available", n_jobs=-1, progress=True, return_idx=False, ) data["filter_bredt"] = mc.functional.bredt_filter( mols=data["mol"].tolist(), n_jobs=-1, progress=True, return_idx=False, ) data["filter_molecular_graph"] = mc.functional.molecular_graph_filter( mols=data["mol"].tolist(), max_severity=5, n_jobs=-1, progress=True, return_idx=False, ) data["filter_lilly_demerit"] = mc.functional.lilly_demerit_filter( mols=data["mol"].tolist(), n_jobs=-1, progress=True, return_idx=False, )

[17:20:36] WARNING: not removing hydrogen atom without neighbors
[17:20:36] WARNING: not removing hydrogen atom without neighbors
[17:20:36] WARNING: not removing hydrogen atom without neighbors
[17:20:36] WARNING: not removing hydrogen atom without neighbors
[17:20:37] WARNING: not removing hydrogen atom without neighbors
[17:20:37] WARNING: not removing hydrogen atom without neighbors
[17:20:37] WARNING: not removing hydrogen atom without neighbors
[17:20:37] WARNING: not removing hydrogen atom without neighbors
[17:20:37] WARNING: not removing hydrogen atom without neighbors
[17:20:37] WARNING: not removing hydrogen atom without neighbors
[17:20:37] WARNING: not removing hydrogen atom without neighbors
[17:20:37] WARNING: not removing hydrogen atom without neighbors
[17:20:38] WARNING: not removing hydrogen atom without neighbors
[17:20:38] WARNING: not removing hydrogen atom without neighbors
[17:20:38] WARNING: not removing hydrogen atom without neighbors
[17:20:38] WARNING: not removing hydrogen atom without neighbors
[17:20:40] WARNING: not removing hydrogen atom without neighbors
[17:20:40] WARNING: not removing hydrogen atom without neighbors
[17:20:40] WARNING: not removing hydrogen atom without neighbors
[17:20:40] WARNING: not removing hydrogen atom without neighbors

Common alerts filtering:   0%|          | 0/2628 [00:00<?, ?it/s]

Common alerts filtering:   0%|          | 0/2628 [00:00<?, ?it/s]

Common alerts filtering:   0%|          | 0/2628 [00:00<?, ?it/s]

NIBR filtering:   0%|          | 0/2628 [00:00<?, ?it/s]

Complexity Evaluation:   0%|          | 0/2628 [00:00<?, ?it/s]

To mol:   0%|          | 0/2628 [00:00<?, ?it/s]

Filtering with catalogs:   0%|          | 0/26 [00:00<?, ?it/s]

Match:   0%|          | 0/2628 [00:00<?, ?it/s]

[17:21:08] WARNING: not removing hydrogen atom without neighbors
[17:21:08] WARNING: not removing hydrogen atom without neighbors
[17:21:08] WARNING: not removing hydrogen atom without neighbors
[17:21:08] WARNING: not removing hydrogen atom without neighbors

In [8]:

filter_columns = [
    "rule_of_five",
    "rule_of_ghose",
    "rule_of_veber",
    "rule_of_zinc",
    "alerts_BMS",
    "alerts_PAINS",
    "alerts_SureChEMBL",
    "filters_NIBR",
    "filter_complexity",
    "filter_bredt",
    "filter_molecular_graph",
    "filter_lilly_demerit",
]

# Some sorting for a nice plot
data["n_filters_pass"] = data[filter_columns].sum(axis=1)
data = data.sort_values("first_approval", ascending=True)

# Plot

f, ax = plt.subplots(figsize=(14, 4), constrained_layout=True)

cmap = matplotlib.colors.ListedColormap(["#EF6262", "#1D5B79"], None)

a = sns.heatmap(
    data[filter_columns].T,
    annot=False,
    ax=ax,
    xticklabels=False,  # type: ignore
    yticklabels=True,  # type: ignore
    cbar=True,
    cmap=cmap,
)

ax.collections[0].colorbar.set_ticks([0.25, 0.75])
ax.collections[0].colorbar.set_ticklabels(["Don't Pass", "Pass"], fontsize=14)

ax.set_xlabel(f"Drug Molecules sorted by approval date from old to recent (n={len(data)})", fontsize=14)
ax.set_ylabel("Medchem Filters", fontsize=18)

# Add percentage of passing mols in the y labels
new_ylabels = []
for t in ax.yaxis.get_ticklabels():
    perc = data[t.get_text()].sum() / len(data) * 100
    new_ylabels.append(f"{t.get_text()} ({perc:.0f}%)")
_ = ax.yaxis.set_ticklabels(new_ylabels, fontsize=12)

filter_columns = [ "rule_of_five", "rule_of_ghose", "rule_of_veber", "rule_of_zinc", "alerts_BMS", "alerts_PAINS", "alerts_SureChEMBL", "filters_NIBR", "filter_complexity", "filter_bredt", "filter_molecular_graph", "filter_lilly_demerit", ] # Some sorting for a nice plot data["n_filters_pass"] = data[filter_columns].sum(axis=1) data = data.sort_values("first_approval", ascending=True) # Plot f, ax = plt.subplots(figsize=(14, 4), constrained_layout=True) cmap = matplotlib.colors.ListedColormap(["#EF6262", "#1D5B79"], None) a = sns.heatmap( data[filter_columns].T, annot=False, ax=ax, xticklabels=False, # type: ignore yticklabels=True, # type: ignore cbar=True, cmap=cmap, ) ax.collections[0].colorbar.set_ticks([0.25, 0.75]) ax.collections[0].colorbar.set_ticklabels(["Don't Pass", "Pass"], fontsize=14) ax.set_xlabel(f"Drug Molecules sorted by approval date from old to recent (n={len(data)})", fontsize=14) ax.set_ylabel("Medchem Filters", fontsize=18) # Add percentage of passing mols in the y labels new_ylabels = [] for t in ax.yaxis.get_ticklabels(): perc = data[t.get_text()].sum() / len(data) * 100 new_ylabels.append(f"{t.get_text()} ({perc:.0f}%)") _ = ax.yaxis.set_ticklabels(new_ylabels, fontsize=12)

Showcasing pro-drugs¶

Some drugs are designed as prodrug that will be only be made active once in the human body after being metabolized. Those compounds will tend to have unwanted molecular features that are often flagged by common alerts and rules.

Below we load a small dataset of 7 drug/prodrug pairs and apply some common medchem rules on them.

In [69]:

# Load the dataset
data = pd.read_csv("./data/Drug_Prodrug_pairs.csv")

# Compute mol objects
data["Drug mol"] = data["Drug SMILES"].apply(dm.to_mol)
data["Prodrug mol"] = data["Prodrug SMILES"].apply(dm.to_mol)

# Reorder the columns
data = data[["Drug", "Drug mol", "Prodrug", "Prodrug mol"]]


# Apply a few medchem rules to the drugs and prodrugs
def _apply_rules(row):
    rule_names = [
        ("Ro5", "rule_of_five"),
        ("Ghose", "rule_of_ghose"),
        ("Veber", "rule_of_veber"),
        ("Zinc", "rule_of_zinc"),
    ]

    for rule_label, rule_name in rule_names:
        rule_fn = getattr(mc.rules.basic_rules, rule_name)

        # Apply the rule to the drug and prodrug
        drug_ok = rule_fn(dm.copy_mol(row["Drug mol"]))
        prodrug_ok = rule_fn(dm.copy_mol(row["Prodrug mol"]))

        row[rule_name] = f"{rule_label}\nDrug: {drug_ok}\nProdrug: {prodrug_ok}"

    return row


data = data.apply(_apply_rules, axis=1)

# Enable drawing in the dataframe
PandasTools.ChangeMoleculeRendering(data)

# Small hack to display new lines as well as RDKit drawings
HTML(data.to_html().replace("\\n", "<br>"))

# Load the dataset data = pd.read_csv("./data/Drug_Prodrug_pairs.csv") # Compute mol objects data["Drug mol"] = data["Drug SMILES"].apply(dm.to_mol) data["Prodrug mol"] = data["Prodrug SMILES"].apply(dm.to_mol) # Reorder the columns data = data[["Drug", "Drug mol", "Prodrug", "Prodrug mol"]] # Apply a few medchem rules to the drugs and prodrugs def _apply_rules(row): rule_names = [ ("Ro5", "rule_of_five"), ("Ghose", "rule_of_ghose"), ("Veber", "rule_of_veber"), ("Zinc", "rule_of_zinc"), ] for rule_label, rule_name in rule_names: rule_fn = getattr(mc.rules.basic_rules, rule_name) # Apply the rule to the drug and prodrug drug_ok = rule_fn(dm.copy_mol(row["Drug mol"])) prodrug_ok = rule_fn(dm.copy_mol(row["Prodrug mol"])) row[rule_name] = f"{rule_label}\nDrug: {drug_ok}\nProdrug: {prodrug_ok}" return row data = data.apply(_apply_rules, axis=1) # Enable drawing in the dataframe PandasTools.ChangeMoleculeRendering(data) # Small hack to display new lines as well as RDKit drawings HTML(data.to_html().replace("\\n", "
"))

Out[69]:

	Drug	Drug mol	Prodrug	Prodrug mol	rule_of_five	rule_of_ghose	rule_of_veber	rule_of_zinc
0	Gabapentin		Gabapentin enacarbil		Ro5 Drug: True Prodrug: True	Ghose Drug: False Prodrug: True	Veber Drug: True Prodrug: True	Zinc Drug: True Prodrug: True
1	Dabigatran		Dabigatran etexilate		Ro5 Drug: True Prodrug: False	Ghose Drug: False Prodrug: False	Veber Drug: False Prodrug: False	Zinc Drug: False Prodrug: False
2	Sofosbuvir		Sofosbuvir		Ro5 Drug: True Prodrug: False	Ghose Drug: False Prodrug: False	Veber Drug: False Prodrug: False	Zinc Drug: False Prodrug: False
3	Tedizolid		Tedizolid phosphate		Ro5 Drug: True Prodrug: False	Ghose Drug: True Prodrug: True	Veber Drug: True Prodrug: False	Zinc Drug: True Prodrug: False
4	Isavuconazole		Isavuconazonium		Ro5 Drug: True Prodrug: False	Ghose Drug: True Prodrug: False	Veber Drug: True Prodrug: False	Zinc Drug: True Prodrug: False
5	Aripiprazole		Aripiprazole lauroxil		Ro5 Drug: True Prodrug: False	Ghose Drug: True Prodrug: False	Veber Drug: True Prodrug: False	Zinc Drug: True Prodrug: False
6	ACT-333679		Selexipag		Ro5 Drug: True Prodrug: True	Ghose Drug: True Prodrug: False	Veber Drug: False Prodrug: False	Zinc Drug: True Prodrug: True

While this is not systematic, we observe that many drugs passes the rules while the associated prodrugs does not pass them.

---

-- The End :-)