Admetica

Predicting pharmacokinetic properties (ADMET: Absorption, Distribution, Metabolism, Excretion, and Toxicity) is essential for drug design. The earlier these properties can be reliably predicted, the better. While every company in the pharmaceutical industry requires this capability, there are currently no optimal solutions. Commercial tools are pricey, and public calculators have many limitations. Neither allows for easy customization.

Admetica addresses these challenges by offering a comprehensive set of pre-built predictive models, publicly available datasets, pipelines, and notebooks for training and evaluating models. It also provides tools for visual exploration of results—all under the permissive MIT license. It's a "batteries included" solution that is:

• Accurate: Built-in comparison with existing solutions ensures reliability.
• Open-source: The source code and data are freely available under the MIT license.
• Simple to use: CLI and REST APIs enable easy integration into existing workflows.
• Configurable: Customize settings or integrate your proprietary datasets into the pipeline.
• Reproducible: All scripts, notebooks, and training pipelines are included.
• Fast: Designed for high performance across datasets of all sizes.
• Interpretable: Tools for intuitive visual exploration of results.

Admetica is an open-source initiative with collaborators from academia, biotech startups, and big pharma. Interested in collaborating? Contact us via email.

Table of Contents

• Usage
• Integration with Datagrok
• Predictive models: see absorption, distribution, metabolism, excretion, toxicity
• Novartis ADMET predictions
  • Comparison
  • Enhancing Admetica with Novartis data
• Evaluation of free online ADMET tools: see PPB, Half-Life, CYP3A4-Substrate, HIA, Summary
• References

Usage

CLI tool

Installation

pip install admetica==1.4.1

By default, the pip installation will include all necessary dependencies for making ADMET predictions.

Predicting

Through the command-line interface:

admetica_predict \
    --dataset-path data.csv \
    --smiles-column smiles \
    --properties Caco2,PPBR \
    --save-path predictions.csv

This command assumes the presence of a file named data.csv with SMILES strings in the column smiles. In addition, you should specify the properties to be calculated (e.g. Caco2). The predictions will be saved to predictions.csv.

All models available in the repository are included and can be used.

Web server

To simplify running Admetica locally as a web server, you can use the provided setup.sh script. This script automates the setup by building the Docker image, running the container, and launching the Swagger UI documentation page in your browser.

Steps:

1. Ensure Docker is installed and running on your system.

2. Run the setup.sh script from the admetica_web folder:

   ./setup.sh

This script will:

• Build a Docker image named admetica.
• Stop and remove any existing container named admetica_container.
• Run a new container, exposing Admetica’s API on port 8080.
• Open the API documentation at http://localhost:8080/apidocs in your default browser.

The setup process should take about 2-3 minutes. If automatic URL opening is unsupported, manually open http://localhost:8080/apidocs.

Data

In order to train a model or obtain predictions, you must provide data containing molecules (as SMILES strings) and known target values.

The data used in this research and its overview can be found in the Datasets folder.

Training

You can create a model on your own using chemprop module, or use publicly available models that are located in the Models folder.

Integration with Datagrok

Predicting properties is just the beginning of the journey. To make a system truly usable, we need to expose it to chemists and medicinal chemists via the UI, without the need to run Docker containers or use CLI. Additionally, we want to easily interpret results in the context of the project (where you can specify desirable property ranges, etc).

To address that, we have developed an MIT-licensed Datagrok Admetica Plugin that allows scientists to calculate ADMET properties on demand. You can also visually assess results using a combination of color coding and a widget that fits in a grid cell and visualizes all properties at once. Additionally, the plugin enhances other Datagrok applications. For instance, Hit Triage for triaging molecular campaigns and Hit Design for collaborative, computation-augmented drug design.

Note that while both Admetica and Admetica Plugin are open-source, the Datagrok platform is proprietary. It is free for personal use, for academia, and non-profit research. Claim your license here.

Available predictive models

Currently, we have a total of 23 predictive models developed for Absorption, Distribution, Metabolism, Excretion and Toxicity.

Absorption

Classification models

Name	Model	Size	Specificity	Sensitivity	Accuracy	Balanced Accuracy	ROC AUC
Pgp-Inhibitor	Chemprop	1,275	0.916	0.863	0.888	0.889

Regression models

Name	Model	Size	MAE	RMSE	R2	Spearman	Observed vs. Predicted
Caco2	Chemprop	910	0.317	0.415	0.701	0.832
Lipophilicity	Chemprop	4200	0.399	0.596	0.748	0.881
Solubility	Chemprop	9982	0.714	1.089	0.788	0.897

Distribution

Regression models

Name	Model	Size	MAE	RMSE	R2	Spearman	Observed vs. Predicted
PPBR	Chemprop	2790	6.919	11.294	0.609	0.762

Metabolism

Classification models

Name	Model	Size	Specificity	Sensitivity	Accuracy	Balanced Accuracy	ROC AUC
CYP1A2-Inhibitor	Chemprop	13,239	0.873	0.866	0.87	0.869
CYP3A4-Inhibitor	Chemprop	12,997	0.815	0.842	0.826	0.829
CYP3A4-Substrate	Chemprop	1,149	0.569	0.779	0.718	0.674
CYP2C19-Inhibitor	Chemprop	13,427	0.819	0.830	0.824	0.825
CYP2C9-Inhibitor	Chemprop	12,881	0.830	0.819	0.826	0.824
CYP2C9-Substrate	Chemprop	899	0.728	0.757	0.738	0.742
CYP2D6-Inhibitor	Chemprop	11,127	0.866	0.751	0.843	0.808
CYP2D6-Substrate	Chemprop	941	0.749	0.769	0.753	0.759

Here is a line chart illustrating various metrics for each of the corresponding models.

Excretion

Regression models

Name	Model	Size	MAE	RMSE	R2	Spearman	Observed vs. Predicted
Clearance Hepatocyte	Chemprop	1213	34.103	47.144	0.086	0.485
Clearance Microsome	Chemprop	1102	26.715	39.201	0.216	0.576

Toxicity

Classification models

Name	Model	Size	Specificity	Sensitivity	Accuracy	Balanced Accuracy	ROC AUC
hERG	Chemprop	22,249	0.811	0.897	0.885	0.854

Regression models

Name	Model	Size	MAE	RMSE	R2	Spearman	Observed vs. Predicted
LD50	Chemprop	7282	0.437	0.609	0.596	0.745

Novartis ADMET predictions

Scientists from Novartis recently published a paper in Nature. The dataset comprises a total of 273,706 molecules, which includes 70,465 from ChEMBL, 199,972 from ZINC, and 3,269 from PROTAC-DB. Naturally, we were curious how their results compare to Admetica ones. Also, if the Novartis predictions are better (very likely since they have massive proprietary datasets), can we improve Admetica models by training them on the publicly available Novartis predictions?

TL/DR: The Novartis models performed well on certain properties, such as CYP2C9-inhibitor, CYP3A4-inhibitor, and Caco-2, but were less effective on the CYP2D6-inhibitor. By incorporating some predictions to the Admetica training dataset, we have improved some of the Admetica models. See details below, or jump to the summary.

Comparison of Admetica and Novartis models

Cytochrome P450

Pipeline

We generated test datasets using data from the ChEMBL database. To ensure the test set was representative and produced accurate results, the data was preprocessed using the following steps:

• Extracting common structures:
  We compared the Novartis and ChEMBL datasets to identify shared molecular structures. Additionally, we filtered out values that overlapped with the Admetica training set to prevent redundancy.

• Filtering the ChEMBL dataset:
  We removed duplicate entries, prioritizing those labeled as IC50. We also excluded rows with undesired types such as Drug metabolism, FC, Retention_time, T1/2, mechanism based inhibition, Stability etc.

• Processing values:
  We standardized key values for consistency.

  Type	Action
  IC50, AC50, KI, Potency	Converted to µM by dividing the values by 1000
  Inhibition	Classified as binary:
  	Greater than 50%: 1
  	Less than 50%: 0
  Other values	Left unchanged (already in µM)

Both the original ChEMBL dataset and the processed data are available in the comparison folder. Additionally, the folder contains a Jupyter notebook, preprocessing_pipeline.ipynb, which fully reproduces the preprocessing steps and obtained datasets.

3A4

After performing the pipeline for ChEMBL 3A4, we obtained a dataset structured as follows:

Class	Number of entries
Inhibitor	549
Non-Inhibitor	239

We calculated CYP3A4-Inhibitor and assessed performance metrics for the Admetica and Novartis models, resulting in the following outcomes:

Metric	Admetica	Novartis
True Positives (TP)	134.0	159.0
True Negatives (TN)	368.0	352.0
False Positives (FP)	181.0	197.0
False Negatives (FN)	105.0	80.0
Sensitivity (Recall)	0.5607	0.6653
Specificity	0.6703	0.6412
Balanced Accuracy	0.6155	0.6532
AUC	0.6155	0.6532

2C9

After performing the pipeline for ChEMBL 2C9, we obtained a dataset structured as follows:

Class	Number of entries
Inhibitor	329
Non-Inhibitor	135

We calculated CYP2C9-Inhibitor and assessed performance metrics for the Admetica and Novartis models, resulting in the following outcomes:

Metric	Admetica	Novartis
True Positives (TP)	102.0	69.0
True Negatives (TN)	137.0	236.0
False Positives (FP)	192.0	93.0
False Negatives (FN)	33.0	66.0
Sensitivity (Recall)	0.7556	0.5111
Specificity	0.4164	0.7173
Balanced Accuracy	0.5860	0.6142
AUC	0.5860	0.6142

2D6

After performing the pipeline for ChEMBL 2D6, we obtained a dataset structured as follows:

Class	Number of entries
Inhibitor	444
Non-Inhibitor	195

We calculated CYP2D6-Inhibitor and assessed performance metrics for the Admetica and Novartis models, resulting in the following outcomes:

Metric	Admetica	Novartis
True Positives (TP)	88.0	55.0
True Negatives (TN)	329.0	403.0
False Positives (FP)	115.0	41.0
False Negatives (FN)	107.0	140.0
Sensitivity (Recall)	0.4513	0.2821
Specificity	0.7410	0.9077
Balanced Accuracy	0.5961	0.5949
AUC	0.5961	0.5949

The comparison is fully reproducible, and you can find the Jupyter notebook, comparison_cyp.ipynb, in the folder.

Caco-2 permeability

We generated test datasets using data from the supplementary material of the paper In Silico Prediction of Caco-2 Cell Permeability by a Classification QSAR Approach. The following preprocessing steps were applied:

• Identifying common structures:

  We compared the Novartis and Caco-2 dataset to identify shared molecular structures. Additionally, we filtered out

• values that overlapped with the Admetica training set to prevent redundancy.

• Unit normalization:

  To ensure consistent units across all datasets predicting Caco-2 permeability (including Novartis and Admetica), we

• applied a log10 transformation to the values.

After performing the preprocessing for Caco-2, we obtained a dataset that contains 34 structures.

We calculated Caco-2 and assessed performance metrics for the Admetica and Novartis models, resulting in the following outcomes:

Metric	Admetica	Novartis
MAE	0.411552	0.351543
MSE	0.286792	0.201841
RMSE	0.535530	0.449267
R²	0.319010	0.520728

Both the original Caco-2 dataset and the processed data are available in the comparison folder. Additionally, the folder contains a Jupyter notebook, comparison_caco2.ipynb, which fully reproduces the preprocessing steps, obtained datasets and metrics.

Results

During our model comparison, we discovered that the Novartis model outperformed the Admetica model for the targets CYP3A4, CYP2C9, and Caco2.

Model enhancement

After performing the comparison and seeing that in some cases the Novartis model is better, we consequently opted to continue training with the surrogate Novartis data used in our comparison.

Data preparation

Given the significant class imbalance in the data for CYP3A4 and CYP2C9, we implemented under-sampling techniques to reduce the risk of overfitting during the training process.

Property	Class distribution	Number of rows in final dataset
CYP3A4	0: 22618, 1: 22618	57781
CYP2C9	0: 3975, 1: 3975	20299

This table summarizes the class distributions and the row counts in the final dataset for each property.

For further details, the comparison folder contains a Jupyter notebook, undersampling.ipynb, that fully reproduces the process of obtaining the final datasets for training.

Using the same pipeline from our comparison, we assessed the performance metrics of the newly trained models. The results are summarized in the tables below.

CYP3A4-Inhibitor

Metric	Admetica (Baseline)	Admetica (Enhanced)	Novartis
True Positives (TP)	134.0	163.0	159.0
True Negatives (TN)	368.0	334.0	352.0
False Positives (FP)	181.0	215.0	197.0
False Negatives (FN)	105.0	76.0	80.0
Sensitivity (Recall)	0.5607	0.6820	0.6653
Specificity	0.6703	0.6084	0.6412
Balanced Accuracy	0.6155	0.6452	0.6532
AUC	0.6155	0.6452	0.6532

CYP2C9-Inhibitor

Metric	Admetica (Baseline)	Admetica (Enhanced)	Novartis
True Positives (TP)	102.0	71.0	69.0
True Negatives (TN)	137.0	232.0	236.0
False Positives (FP)	192.0	97.0	93.0
False Negatives (FN)	33.0	64.0	66.0
Sensitivity (Recall)	0.7556	0.5259	0.5111
Specificity	0.4164	0.7052	0.7173
Balanced Accuracy	0.5860	0.6155	0.6142
AUC	0.5860	0.6155	0.6142

Caco-2

Metric	Admetica (Baseline)	Admetica (Enhanced)	Novartis
MAE	0.411552	0.364398	0.351543
MSE	0.286792	0.195037	0.201841
RMSE	0.535530	0.441630	0.449267
R²	0.319010	0.536883	0.520728

Summary

Metric	Δ (Enhanced - Baseline)	% Improvement
CYP3A4-Inhibitor
True Positives (TP)	+29.0 ↑	+21.6% ✅
True Negatives (TN)	-34.0 ↓	-9.2% ❌
False Positives (FP)	+34.0 ↑	+18.8% ❌
False Negatives (FN)	-29.0 ↓	-27.6% ✅
Sensitivity (Recall)	+0.1213 ↑	+21.6% ✅
Specificity	-0.0619 ↓	-9.2% ❌
Balanced Accuracy	+0.0297 ↑	+4.8% ✅
AUC	+0.0297 ↑	+4.8% ✅
CYP2C9-Inhibitor
True Positives (TP)	-31.0 ↓	-30.4% ❌
True Negatives (TN)	+95.0 ↑	+69.3% ✅
False Positives (FP)	-95.0 ↓	-49.5% ✅
False Negatives (FN)	+31.0 ↑	+94.0% ❌
Sensitivity (Recall)	-0.2297 ↓	-30.4% ❌
Specificity	+0.2888 ↑	+69.3% ✅
Balanced Accuracy	+0.0295 ↑	+5.0% ✅
AUC	+0.0295 ↑	+5.0% ✅
Caco-2
MAE	-0.047154 ↓	-11.4% ✅
MSE	-0.091755 ↓	-31.9% ✅
RMSE	-0.093900 ↓	-17.6% ✅
R²	+0.217873 ↑	+68.2% ✅

Where:

• ↑: Improvement
• ↓: Decline
• ✅: Positive Improvement
• ❌: Negative Improvement

Evaluation of free online ADMET tools

For our evaluation, we used 24 tyrosine kinase inhibitors (TKIs) and a comparison table from the supplementary materials of the study Evaluation of Free Online ADMET Tools for Academic or Small Biotech Environments. The table contained predictions for the 24 structures from various web services, including ADMETlab, admetSAR, SwissADME, and others. Since the table was slightly outdated, we updated the ADMETlab 2.0 predictions with those from ADMETlab 3.0.

Plasma Protein Binding

We compared plasma protein binding (PPB) predictions for 24 tyrosine kinase inhibitors (TKIs) using four tools: ADMETLab, admetSAR, preADMET, and Admetica. The plot below shows the results.

Model	MSE	MAE	RMSE
ADMETLab	37.42	3.25	6.12
admetSAR	148.04	6.71	12.17
Admetica	139.08	8.52	11.79
preADMET	265.00	13.00	16.28

ADMETLab consistently provided the most accurate predictions, achieving the lowest error metrics across all measures (MSE, MAE, RMSE). Admetica and admetSAR had similar performance, with Admetica exhibiting slightly better metrics overall, despite both having higher error rates compared to ADMETLab. PreADMET showed the least accuracy overall, as indicated by its significantly higher error metrics. Overall, ADMETLab remains the top choice for precise plasma protein binding predictions.

Half-Life

Among all online web services, only ADMETLab provides predictions for Half-Life. To assess the accuracy of these predictions, we performed a comparative analysis between ADMETLab and Admetica for 24 structures.

Model	MSE	MAE	RMSE
ADMETLab	1647.99	32.86	40.60
Admetica	1162.88	29.21	34.10

ADMETLab tends to underestimate Half-Life predictions compared to actual values, resulting in relatively high error metrics. In contrast, Admetica often overestimates Half-Life, but it has lower error metrics overall. This disparity highlights the importance of carefully evaluating the predictions from both tools, as they exhibit different tendencies in their predictions, leading to significant variability.

CYP3A4-Substrate

We compared CYP3A4 substrate predictions using five tools: ADMETLab, admetSAR, pkCSM, preADMET, and Admetica.

Model	Accuracy	Precision	Recall	F1 Score
ADMETLab	0.83	1.0	0.83	0.91
admetSAR	1.0	1.0	1.0	1.0
pkCSM	0.96	1.0	0.96	0.98
preADMET	0.46	1.0	0.46	0.63
Admetica	0.96	1.0	0.96	0.98

The results revealed that admetSAR was the best-performing model, demonstrating exceptional accuracy in predicting CYP3A4 substrates. Both pkCSM and Admetica also showed strong performance, making them reliable options for these predictions. ADMETLab provided decent results, but its performance was not as robust as the top three models. Conversely, preADMET exhibited the lowest performance, indicating that it is less reliable for CYP3A4 substrate predictions. Overall, we can confidently rely on admetSAR, pkCSM, and Admetica for accurate predictions in this area.

HIA

We compared models for predicting human intestinal absorption (HIA) using six free tools: ADMETLab, admetSAR, FAF-Drug4, pkCSM, SwissADME, and Admetica.

Model	Accuracy	Precision	Recall
ADMETLab	0.62	0.61	1.0
SwissADMET	0.62	0.63	0.86
admetSAR	0.58	0.58	1.0
FAF-Drug4	0.58	0.58	1.0
pkCSM	0.58	0.58	1.0
Admetica	0.58	0.58	1.0

ADMETLab and SwissADMET demonstrated the highest accuracy, effectively identifying absorbed compounds, while admetSAR, FAF-Drug4, pkCSM, and Admetica showed similar performance levels but lower precision. Overall, all models exhibited perfect recall, indicating their effectiveness in identifying absorbed compounds, though there is still room for improvement in minimizing false positives.

You can find both the original dataset and the processed data in the comparison folder. This folder also includes a Jupyter notebook, comparison_services.ipynb, that reproduces the steps taken to conduct the comparison.

Summary

Below is a summary table of the most notable metrics from the evaluation of free online ADMET tools using predictions for 24 tyrosine kinase inhibitors (TKIs).

Model	PPB (F1)	Half-Life (MAE)	CYP3A4 Substrate (F1)	HIA (Accuracy)
ADMETLab	0.91	32.86	0.83	0.62
admetSAR	1.00	N/A	1.00	0.58
Admetica	0.98	29.21	0.98	0.58
pkCSM	N/A	N/A	0.98	0.58
preADMET	N/A	N/A	0.63	0.58
FAF-Drug4	N/A	N/A	N/A	0.58
SwissADME	N/A	N/A	N/A	0.62

References

Our project is about improving and combining existing solutions, not reinventing the wheel. Here are some of the resources is the list of resources we've investigated:

1. ADMETlab: a platform for systematic ADMET evaluation based on a comprehensively collected ADMET database / Jie Dong, Ning-Ning Wang, Zhi-Jiang Yao та ін. // J Cheminform. – 2018. – https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6020094/.
2. Evaluation of Free Online ADMET Tools for Academic or Small Biotech Environments / Júlia Dulsat, Blanca López-Nieto, Roger Estrada-Tejedor, José I. Borrell // Molecules. – 2023. – https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9864198/.
3. Vishwesh Venkatraman. FP-ADMET: a compendium of fingerprint-based ADMET prediction models / Vishwesh Venkatraman // J Cheminform. – 2021. – https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8479898/.
4. Front Pharmacol. vNN Web Server for ADMET Predictions / Front Pharmacol // Front Pharmacol. – 2017. – https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5722789/.
5. ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties / Guoli Xiong, Zhenxing Wu, Jiacai Yi та ін. // Nucleic Acids Res. – 2021. – https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8262709/.
6. In silico Prediction of Chemical Ames Mutagenicity / Congying Xu, Feixiong Cheng, Lei Chen та ін. // J Cheminform. – 2012. – https://pubs.acs.org/doi/abs/10.1021/ci300400a.
7. Computational Models for Human and Animal Hepatotoxicity with a Global Application Scope / Denis Mulliner, Friedemann Schmidt, Manuela Stolte та ін. // Chem. Res. Toxicol.. – 2016. – https://pubs.acs.org/doi/10.1021/acs.chemrestox.5b00465.
8. ADMET Evaluation in Drug Discovery. 16. Predicting hERG Blockers by Combining Multiple Pharmacophores and Machine Learning Approaches / Shuangquan Wang, Huiyong Sun, Hui Liu та ін. // Mol. Pharmaceutics. – 2016. – https://pubs.acs.org/doi/10.1021/acs.molpharmaceut.6b00471.
9. Application of machine learning models for property prediction to targeted protein degraders