Table of contents:

• Using DeepMol models
  • How to use
  • Visualization
  • Example
• Case studies for the DeepMol publication
  • AutoML experiments - TDC Commons
  • Comparison of DeepMol and QSARTuna
  • Benchmark computational resources and runtimes of DeepMol
  • Evaluate TDC commons benchmark datasets

Using DeepMol models

Models available so far:

Model Name	How to Call	Prediction Type
BBB (Blood-Brain Barrier)	BBB	Penetrates BBB (1) or does not penetrate BBB (0)
AMES Mutagenicity	AMES	Mutagenic (1) or not mutagenic (0)
Human plasma protein binding rate (PPBR)	PPBR	Rate of PPBR expressed in percentage
Volume of Distribution (VD) at steady state	VDss	Volume of Distribution expressed in liters per kilogram (L/kg)
Caco-2 (Cell Effective Permeability)	Caco2	Cell Effective Permeability (cm/s)
HIA (Human Intestinal Absorption)	HIA	Absorbed (1) or not absorbed (0)
Bioavailability	Bioavailability	Bioavailable (1) or not bioavailable (0)
Lipophilicity	Lipophilicity	Lipophilicity log-ratio
Solubility	Solubility	Solubility (log mol/L)
CYP P450 2C9 Inhibition	CYP2C9Inhibition	Inhibit (1) or does not inhibit (0)
CYP P450 3A4 Inhibition	CYP3A4Inhibition	Inhibit (1) or does not inhibit (0)
CYP2C9 Substrate	CYP2C9Substrate	Metabolized (1) or does not metabolize (0)
CYP2D6 Substrate	CYP2D6Substrate	Metabolized (1) or does not metabolize (0)
CYP3A4 Substrate	CYP3A4Substrate	Metabolized (1) or does not metabolize (0)
Hepatocyte Clearance	HepatocyteClearance	Drug hepatocyte clearance (uL.min-1.(10^6 cells)-1)
NPClassifier	NPClassifier	Pathway, Superclass, Class
Plants secondary metabolite precursors predictor	PlantsSMPrecursorPredictor	Precursor 1; Precursor 2
Microsome Clearance	MicrosomeClearance	Drug microsome clearance (mL.min-1.g-1)
LD50	LD50	LD50 (log(1/(mol/kg)))
hERG Blockers	hERGBlockers	hERG blocker (1) or not blocker (0)

How to use:

You can get information about each model just by instatiating the class.

from deepmol_models import NPClassifier

NPClassifier()

NPClassifier Overview

This model is a reimplementation of NPClassifier as described in the ACS journal publication.

All credits should be given to the authors. NPClassifier performs automated structural classification of natural products.

NPClassifier Details

Attribute	Value
Model Name	NPClassifier
Prediction Type	Pathway, Superclass, Class

Key Features

• Prediction of natural product Pathway, Superclass, and Class.
• Efficient performance on large datasets.

You can use them either individually or mixed together.

You can call one model individually, pass a CSV file and get the results in one dataframe:

from deepmol_models import BBB
results = BBB().predict_from_csv("dataset.csv", smiles_field="Drug", id_field="Drug_ID", output_file="predictions.csv")
results

ID	SMILES	BBB Penetration
0	OCC(S)CS	1.0
1	CCN+(C)c1cccc(O)c1	0.0
2	Nc1ncnc2c1ncn2[C@@H]1OC@HC@@H[C@@H]1O	1.0
3	CC(=O)OCC1=C(C(=O)O)N2C(=O)C@@H[...	0.0
4	CC1(C)S[C@@H]2[C@H](NC(=O)C@Hc3ccsc3...	0.0

Or pass SMILES strings and get the results in one dataframe:

from deepmol_models import BBB
results = BBB().predict_from_csv("dataset.csv", smiles_field="Drug", id_field="Drug_ID", output_file="predictions.csv")
results

ID	SMILES	BBB Penetration
1	CCN+(C)c1cccc(O)c1	0.0
2	Nc1ncnc2c1ncn2[C@@H]1OC@HC@@H[C@@H]1O	1.0

Complementarily, you can run several models:

from deepmol_models import BBB, PPBR, VDss, Caco2, HIA, Bioavailability, \
    Lipophilicity, Solubility, PlantsSMPrecursorPredictor, NPClassifier, MixedPredictor

# results = MixedPredictor([BBB(), Caco2(), CYP2D6Inhibition(), NPClassifier()]).predict_from_csv("test_molecules.csv", "Drug", "Drug_ID", output_file="predictions.csv")
results = MixedPredictor([BBB(), PPBR(), VDss(), Caco2(), 
                          HIA(), Bioavailability(), Lipophilicity(),
                          Solubility(), PlantsSMPrecursorPredictor(), NPClassifier()]).predict_from_csv("test_molecules.csv", smiles_field="Drug", id_field="Drug_ID", output_file="predictions.csv")
results

ID	SMILES	BBB Penetration	Human PPBR	VDss	Cell Effective Permeability	Human Intestinal Absorption	Bioavailability	Lipophilicity	Solubility	Precursors	Pathways	Superclass	Class
0	OCC(S)CS	1.0	64.832665	5.803529	-4.725497	0.0	0.0	0.290602	0.117866		Fatty acyls	Fatty alcohols	Fatty alcohols
1	CCN+(C)c1cccc(O)c1	0.0	32.882912	2.891243	-4.989814	0.0	0.0	0.271549	-0.606772	L-Lysine	Alkaloids	Tyrosine alkaloids	Phenylethylamines
2	Nc1ncnc2c1ncn2[C@@H]1OC@HC@@H[C@@H]1O	1.0	41.540812	2.722000	-6.059630	1.0	0.0	0.250526	-1.701435	Dimethylallyl diphosphate	Carbohydrates	Nucleosides	Purine nucleosides
3	CC(=O)OCC1=C(C(=O)O)N2C(=O)C@@H[...]	0.0	52.921825	0.329071	-5.473660	0.0	0.0	0.267842	-2.512460	Geranylgeranyl diphosphate; L-Alanine	Amino acids and Peptides	β-lactams	Cephalosporins

Visualization

You can use our API to access the bokeh representation of the chemical space and check some features of the molecules:

from deepmol_models import bokeh_plot
bokeh_plot(results, "Solubility", additional_labels=["Pathways", "Superclass", "Class"])

Example

You can find an example in example.

Case studies for the DeepMol publication

Installation

1. Clone the repository and move into the directory:

git clone https://github.com/BioSystemsUM/deepmol_case_studies.git
cd deepmol_case_studies

2. Create a conda environment and activate it:

conda create -n deepmol_case_studies python=3.10
conda activate deepmol_case_studies

3. Install the dependencies:

pip install -r requirements.txt
pip install --no-deps deepmol[all]==1.1.7

4. Install the package:

pip install .

AutoML experiments - TDC Commons

AutoML experiments can be found in here.

We used podman/docker for the experiments, the Dockerfile can be found in this repository.

The "run" file can be found in here.

Comparison of DeepMol and QSARTuna

AutoML experiments for comparison between DeepMol and QSARTuna can be found in here.

We used podman/docker for the experiments, the Dockerfile can be found in this repository.

The "run_automl_benchmark" file can be found in here.

Benchmark computational resources and runtimes of DeepMol

The scripts to evaluate computational resources and runtimes of each method in DeepMol are here.

All the dataframes with this information are available here.

Evaluate TDC commons benchmark datasets

To train and evaluate DeepMol models:

from dcs.evaluation import get_results

results = get_results(tdc_dataset_name="Bioavailability_Ma", pipeline="bioavailability_optimal")

The tdc_dataset_name parameter is used to download the TDC commons benchmark datasets. Available datasets:

• "AMES"
• "BBB_Martins"
• "Bioavailability_Ma"
• "Caco2_Wang"
• "Clearance_Hepatocyte_AZ"
• "Clearance_Microsome_AZ"
• "HIA_Hou"
• "Pgp_Broccatelli"
• "Solubility_AqSolDB"
• "Lipophilicity_AstraZeneca"
• "VDss_Lombardo"
• "CYP2C9_Veith"
• "CYP2D6_Veith"
• "CYP3A4_Veith"
• "CYP2C9_Substrate_CarbonMangels"
• "CYP2D6_Substrate_CarbonMangels"
• "CYP3A4_Substrate_CarbonMangels"
• "DILI"
• "Half_Life_Obach"
• "hERG"
• "LD50_Zhu"
• "PPBR_AZ"

While the pipeline parameter is for internal pipeline loading. The pipelines are listed according to the paper, where there are the pipelines created based on the first AutoML experiment and the ones that were further optimized (optimal). Available pipelines:

• "ames"
• "bbb"
• "bioavailability"
• "bioavailability_optimal"
• "caco"
• "clearance_hepatocyte"
• "clearance_microsome"
• "hia"
• "pgp"
• "solubility"
• "lipophilicity"
• "lipophilicity_optimal"
• "vdss"
• "cyp2c9"
• "cyp2d6"
• "cyp3a4"
• "cyp2c9_substrate"
• "cyp2d6_substrate"
• "cyp3a4_substrate"
• "dili"
• "half_life"
• "herg"
• "hia"
• "ld50"
• "ppbr"

If intended, the default pipelines (all but the optimal) for each dataset can be called as follows:

from dcs.evaluation import get_results

results = get_results(tdc_dataset_name="Bioavailability_Ma")