PHARMACOKINETIC AND TOXICOLOGICAL EVALUATION OF NOVEL BIOACTIVE COMPOUNDS USING IN SILICO ADMET PREDICTION

Dr. Nathaniel J. Brooks
 Department of Pharmaceutical Chemistry,

Global Center for Biomedical Research, USA
nathaniel.brooks@example.org

 

Abstract.

 


Early pharmacokinetic and toxicological profiling is a central component of contemporary drug discovery because potent molecules may fail when exposure, metabolism, distribution, or safety properties are unfavorable. The present study describes an integrated in silico ADMET framework for prioritizing novel bioactive compounds before resource-intensive experimental testing. The workflow combines physicochemical and drug-likeness descriptors with predicted intestinal absorption, aqueous solubility, plasma protein binding, cytochrome P450 interactions, systemic clearance, hERG-related cardiotoxicity, Ames mutagenicity, and drug-induced liver injury liability. A structured literature-based methodology was established around widely used platforms and model classes including SwissADME, pkCSM, admetSAR, ADMETlab, ProTox-II, and contemporary deep-learning approaches. To demonstrate interpretation, an illustrative eight-compound dataset was evaluated with a consensus desirability score; these numerical outputs are methodological examples rather than experimentally measured or independently validated predictions. The integrated analysis shows how apparently favorable oral absorption can coexist with metabolic inhibition or toxicity liabilities, emphasizing the need for multi-endpoint rather than single-parameter optimization. Compounds with balanced lipophilicity, moderate polar surface area, acceptable solubility, lower CYP inhibition probability, and reduced hERG, mutagenicity, and hepatotoxicity signals were prioritized. In silico ADMET prediction can substantially improve early triage, but applicability domains, data quality, endpoint imbalance, model uncertainty, and inter-platform discordance must be considered. The proposed framework is suitable for hypothesis generation and lead prioritization and should be followed by orthogonal in vitro ADME and toxicology assays before translational decisions.


 

Keywords: ADMET; pharmacokinetics; toxicity; in silico prediction; SwissADME; pkCSM; hERG; cytochrome P450; drug-induced liver injury; lead optimization.

1. INTRODUCTION

Drug discovery increasingly treats pharmacokinetics and safety as design variables rather than as late-stage filters. A bioactive molecule can show high affinity for a therapeutic target yet remain unsuitable because it is poorly soluble, incompletely absorbed, rapidly cleared, strongly bound to plasma proteins, prone to clinically relevant drug-drug interactions, or associated with organ toxicity. The traditional sequence in which efficacy was established first and ADMET properties were assessed later contributed to avoidable attrition. Contemporary discovery therefore integrates absorption, distribution, metabolism, excretion, and toxicity information at the hit-to-lead and lead-optimization stages (1,5,6).

Simple physicochemical rules remain useful because many pharmacokinetic outcomes are constrained by molecular size, lipophilicity, hydrogen-bonding capacity, flexibility, and polar surface area. The rule-of-five framework linked excessive molecular weight, lipophilicity, hydrogen-bond donors, and hydrogen-bond acceptors with increased risk of poor absorption or permeability (1). Egan and colleagues developed multivariate models for passive intestinal absorption using lipophilicity and polar surface area (2), while Veber et al. highlighted molecular flexibility and polar surface area as important determinants of oral bioavailability (3). Solubility itself can be estimated from structure using statistical models such as ESOL, which demonstrated the practical value of combining logP, molecular weight, aromaticity, and rotatable-bond information (4).

The rapid expansion of public chemical and pharmacological data has enabled machine-learning models for endpoints that previously required costly experimental screening. pkCSM uses graph-based molecular signatures to estimate pharmacokinetic and toxicity properties (5), SwissADME combines physicochemical calculations and empirical prediction models for drug-likeness and pharmacokinetics (6), ProTox-II integrates similarity, fragment, pharmacophore, and machine-learning approaches across multiple toxicity endpoints (7), and admetSAR 2.0 provides web-based prediction and optimization of chemical ADMET properties (8). ADMETlab 2.0 further expanded endpoint coverage and introduced multi-task graph-attention models for broad ADMET evaluation (9). More recently, Deep-PK has illustrated the transition toward deep-learning systems that jointly predict pharmacokinetic and toxicity properties from molecular structure (10).

Although these developments are important, in silico ADMET predictions are not interchangeable with laboratory measurements. Performance varies by endpoint, chemical domain, class balance, assay definition, and data provenance. Toxicity outcomes such as hERG inhibition, Ames mutagenicity, and drug-induced liver injury are particularly sensitive to the representation of mechanistically diverse pathways. Molecular machine-learning benchmarks have also shown that strong average performance does not eliminate the effects of data scarcity, imbalance, scaffold shift, and uncertain applicability domains (11). Consequently, computational predictions are most defensible when used as a consensus prioritization layer that is linked to transparent uncertainty and followed by orthogonal experimental testing.

The aim of this study was to construct an OAK-compatible IMRAD framework for pharmacokinetic and toxicological evaluation of novel bioactive compounds using in silico ADMET prediction. Specific objectives were to: (i) define a reproducible sequence of physicochemical, pharmacokinetic, metabolic, and toxicological assessments; (ii) organize endpoints according to decision relevance; (iii) demonstrate multi-parameter interpretation using an illustrative compound set; and (iv) discuss limitations and experimental follow-up requirements. The central premise is that lead quality is multidimensional and should be optimized through a balance of exposure, metabolic stability, interaction liability, and safety rather than by maximizing a single predicted property.

2. MATERIALS AND METHODS

2.1. Study design and literature framework

The study was designed as a structured methodological analysis with an illustrative in silico case study. Peer-reviewed international literature was used to define the principal ADMET endpoints, prediction concepts, and interpretation criteria. Foundational property-based studies were combined with modern web-server and machine-learning publications. The methodological framework emphasized endpoints that commonly influence early medicinal-chemistry decisions: physicochemical suitability, solubility and intestinal absorption, distribution and plasma protein binding, CYP450-mediated metabolism, systemic clearance, cardiotoxicity, mutagenicity, and hepatotoxicity (1-15).

A total of 15 international references were selected to provide coverage of classical medicinal-chemistry rules, public prediction platforms, deep learning and molecular machine learning, and endpoint-specific safety prediction. The intent was not to rank software systems but to identify complementary strengths. Consensus interpretation was prioritized because different tools may use different datasets, molecular representations, label definitions, and decision thresholds.

2.2. Compound preparation and physicochemical profiling

A hypothetical set of eight novel bioactive compounds (BC-01 to BC-08) was used solely to demonstrate the reporting structure. Compound names and numerical model outputs are illustrative and must not be interpreted as experimentally measured data. In a real project, structures should be standardized before prediction by removing counterions, normalizing valence and aromaticity, resolving tautomeric or protonation states when appropriate, and documenting the canonical structure used for every platform. Such standardization reduces avoidable differences between tools and improves reproducibility.

The initial property panel included molecular weight, calculated lipophilicity (cLogP), topological polar surface area (tPSA), and estimated aqueous solubility. These descriptors were interpreted using the concepts established by Lipinski, Egan, Veber, and Delaney (1-4). Rather than applying rigid pass/fail criteria, physicochemical rules were treated as risk indicators because orally active drugs may legitimately fall outside conventional boundaries, particularly when active transport, prodrug strategies, or nontraditional chemical modalities are involved.

Table 1. Core in silico ADMET endpoints and their decision relevance

Endpoint

Biological meaning

Early discovery interpretation

Key refs

Solubility / logS

Dissolution and formulation feasibility

Very low solubility may limit exposure even with favorable permeability

(4,6,9)

HIA / permeability

Likelihood of intestinal uptake

Low predicted absorption can deprioritize oral development

(2,5,6,9)

PPB

Fraction bound to plasma proteins

High binding can alter free exposure and interaction risk

(5,9)

CYP inhibition / substrate status

Metabolic clearance and drug-drug interactions

Strong inhibition may increase interaction liability

(5,9,14)

Systemic clearance

Persistence of parent drug

Extremely high or low clearance can complicate dosing

(5,9,10)

hERG liability

Cardiac repolarization risk

High predicted blockade requires focused cardiac safety work

(7,9,12)

Ames mutagenicity

Genotoxicity hazard

Positive prediction triggers structural review and testing

(7,13)

DILI / hepatotoxicity

Potential liver injury

Elevated risk requires mechanistic and cellular follow-up

(7,9,15)

 

2.3. Pharmacokinetic prediction strategy

Absorption was evaluated conceptually through human intestinal absorption probability and permeability-related descriptors. Distribution was considered through plasma protein binding and, where relevant to a therapeutic hypothesis, blood-brain barrier penetration. Metabolic assessment focused on major CYP450 isoforms, especially CYP3A4, CYP2D6, CYP2C9, CYP2C19, and CYP1A2. CYP prediction is relevant because the enzyme system accounts for metabolism of a large fraction of xenobiotics, and in silico tools can assist with substrate, inhibitor, and site-of-metabolism hypotheses (14). Predicted systemic clearance was included as a summary pharmacokinetic endpoint, while recognizing that clearance is influenced by hepatic blood flow, enzyme/transporter activity, protein binding, renal elimination, and species-specific physiology.

Outputs from tools such as SwissADME, pkCSM, admetSAR 2.0, ADMETlab 2.0, and Deep-PK can be combined using a consensus logic (5,6,8-10). A high-confidence decision should not rely on a single probability without reviewing whether the compound is within the model's applicability domain. For example, classification probabilities close to a model threshold should be regarded as uncertain rather than as biological discontinuities.

2.4. Toxicological prediction strategy

The toxicity panel included predicted hERG-channel inhibition, Ames mutagenicity, and hepatotoxicity or drug-induced liver injury liability. hERG inhibition is a major computational safety endpoint because blockade of the cardiac potassium channel can delay repolarization and contribute to QT prolongation and ventricular arrhythmia (12). In silico hERG models can reduce early attrition but should be confirmed using dedicated ion-channel assays when a compound progresses.

Ames mutagenicity was included as a genotoxicity-related endpoint. Machine-learning studies have demonstrated that structural descriptors and learned representations can identify mutagenicity patterns, but performance remains dependent on data curation and class definitions (13). ProTox-II and related platforms extend toxicity prediction to hepatotoxicity, carcinogenicity, immunotoxicity, and pathway-level endpoints (7). Drug-induced liver injury remains especially difficult because it can reflect dose, reactive metabolites, mitochondrial toxicity, bile-acid disruption, immune mechanisms, and patient susceptibility; consequently, DILI prediction should be treated as a hazard signal rather than as a clinical diagnosis (15).

2.5. Integrated scoring and statistical presentation

For the illustrative dataset, each compound was summarized using a consensus desirability score ranging from 0 to 1. Higher intestinal absorption increased the score, whereas CYP3A4 inhibition probability, hERG liability, Ames mutagenicity, DILI probability, and excessive plasma protein binding reduced the score. The weighting scheme was selected for demonstration and is not a validated clinical or regulatory algorithm. In a real medicinal-chemistry program, weights should be adjusted to the therapeutic indication, route of administration, target tissue, expected dose, safety margin, and comparator profile.

Descriptive statistics and visualizations were used to compare compounds. Heatmaps emphasized correlated toxicological liabilities, radar plots illustrated multi-parameter balance, and bar charts provided a transparent representation of consensus prioritization. No inferential hypothesis test was applied because the case-study values were not experimental observations.

 

Figure 1. Integrated workflow for early in silico pharmacokinetic and toxicological evaluation.

3. RESULTS

3.1. Physicochemical and pharmacokinetic profile

The illustrative compound set occupied a conventional small-molecule property space, with molecular weights from 286.3 to 401.5 Da, cLogP values from 1.8 to 4.6, and tPSA values from 41 to 92 Ų. This range was intentionally selected to demonstrate how relatively modest changes in lipophilicity and polarity can alter the balance between solubility, absorption, protein binding, metabolic interactions, and safety. BC-04 and BC-07 represented the most lipophilic members of the series and also showed the least favorable estimated solubility. In contrast, BC-03 and BC-08 showed lower lipophilicity and higher tPSA with correspondingly more favorable illustrative solubility values.

Predicted intestinal absorption probabilities were high for most candidates, illustrating an important medicinal-chemistry principle: favorable absorption alone does not establish overall developability. The most lipophilic compounds also showed higher plasma protein binding and stronger CYP3A4 inhibition probabilities in the demonstration set. Thus, a compound may cross intestinal membranes efficiently while simultaneously developing liabilities related to low free fraction or metabolic drug-drug interactions. This type of trade-off supports the use of multi-endpoint optimization rather than isolated property filters.

Table 2. Illustrative physicochemical and pharmacokinetic outputs for the model compound set

Compound

MW

cLogP

tPSA

ESOL logS

HIA prob.

PPB

Clearance*

BC-01

312.4

2.1

68

-2.4

0.88

72%

7.8

BC-02

338.4

3.4

54

-3.1

0.93

91%

5.2

BC-03

286.3

1.8

92

-2.0

0.79

58%

9.4

BC-04

401.5

4.2

47

-4.0

0.96

96%

3.8

BC-05

355.4

2.7

76

-2.8

0.84

79%

6.5

BC-06

329.3

3.1

63

-3.0

0.91

87%

5.9

BC-07

378.5

4.6

41

-4.4

0.97

98%

3.1

BC-08

298.3

2.3

84

-2.2

0.82

65%

8.6

 

*Clearance values are illustrative arbitrary model outputs used only for methodological demonstration; units and thresholds should be defined by the actual prediction platform in a real study.

3.2. Metabolic interaction and toxicity signals

The illustrative toxicity matrix differentiated candidates that appeared superficially similar on physicochemical criteria. BC-07 showed the highest combined liability, with elevated CYP3A4 inhibition, hERG risk, Ames mutagenicity, and DILI probabilities. BC-04 displayed a secondary cluster of risk, driven particularly by CYP3A4 inhibition and hERG liability. Conversely, BC-01 and BC-03 showed lower probabilities across all four modeled safety endpoints. BC-08 also retained a comparatively balanced profile despite a higher polar surface area.

The pattern demonstrates why toxicity prediction should be interpreted as a constellation rather than as a single flag. A high hERG probability does not imply mutagenicity, and a DILI signal does not necessarily arise from the same structural features as CYP inhibition. Nevertheless, correlated liabilities can emerge when lipophilic, aromatic, or highly protein-bound structures interact promiscuously with proteins and membranes. Such hypotheses require structural inspection, analog analysis, and experimental confirmation.

 

Figure 2. Illustrative probability matrix for selected metabolic and toxicological liabilities.

Table 3. Illustrative toxicological prediction outputs and consensus prioritization

Compound

CYP3A4

hERG

Ames

DILI

Score

Illustrative action

BC-01

0.21

0.12

0.08

0.18

0.87

Advance

BC-02

0.48

0.26

0.14

0.25

0.75

Advance

BC-03

0.18

0.10

0.11

0.16

0.86

Advance

BC-04

0.71

0.58

0.22

0.46

0.58

Deprioritize

BC-05

0.33

0.19

0.17

0.23

0.81

Advance

BC-06

0.44

0.31

0.13

0.29

0.76

Advance

BC-07

0.82

0.73

0.34

0.61

0.48

Deprioritize

BC-08

0.24

0.15

0.09

0.20

0.84

Advance

 

3.3. Multi-parameter prioritization

The radar representation showed that BC-01 and BC-03 achieved the most balanced profiles in the demonstration dataset. BC-07 had excellent predicted absorption but a poor overall shape because metabolic and safety liabilities offset that advantage. BC-05 showed an intermediate profile, illustrating a candidate that might remain chemically tractable if potency is strong and liabilities can be reduced by targeted structural modification. Importantly, the radar plot is not intended to imply equal biological importance of every axis; rather, it provides a visual prompt for medicinal-chemistry discussion.

The consensus score ranked BC-01 and BC-03 as the strongest illustrative candidates, followed by BC-08. BC-04 and BC-07 fell below the example advancement threshold. This result does not establish therapeutic superiority; it demonstrates how integrated ADMET triage can prevent a highly absorbed or potent candidate from being advanced when a cluster of safety risks is present. In a real program, the advancement threshold would be calibrated against reference drugs, assay uncertainty, intended dose, therapeutic window, and disease severity.

 

Figure 3. Multi-parameter ADMET desirability profiles for four representative illustrative compounds.

 

Figure 4. Illustrative integrated ADMET desirability scores; the dashed line is a nonvalidated demonstration threshold.

4. DISCUSSION

The present framework illustrates the central role of in silico ADMET prediction in modern lead optimization. Computational models can evaluate large numbers of compounds before synthesis or before extensive biological testing, allowing chemists to identify property trends, anticipate liabilities, and prioritize experiments. This is particularly valuable when a project contains many analogs but limited capacity for microsomal stability, Caco-2, transporter, ion-channel, genotoxicity, or organ-toxicity assays. The value is greatest when predictions are treated as decision support rather than as replacements for experimental evidence.

Physicochemical descriptors remain a practical first layer. Lipinski's rule-of-five, Egan's absorption model, Veber's flexibility and polar-surface-area analysis, and ESOL provide interpretable relationships between structure and developability (1-4). Their simplicity is also a limitation: they cannot represent active transport, enzyme kinetics, formulation effects, reactive metabolism, or tissue-specific toxicity. Therefore, a compound that violates a classical rule should not automatically be discarded, and a compound that passes all rules should not automatically be advanced.

Modern platforms add endpoint-specific machine learning. pkCSM demonstrated the usefulness of graph-based signatures for pharmacokinetic and toxicity prediction (5). SwissADME provides a broad set of descriptors and drug-likeness assessments in an accessible interface (6). ProTox-II integrates diverse computational concepts across acute and mechanistic toxicity endpoints (7), while admetSAR 2.0 and ADMETlab 2.0 expanded large-scale web-based ADMET screening (8,9). ADMETlab 2.0 is notable for using multi-task graph-attention models and a broad collection of classification and regression endpoints, reflecting the general movement toward shared learned representations across related pharmacokinetic and toxicity tasks (9). Deep-PK further demonstrates how deep learning can provide unified small-molecule pharmacokinetic and toxicity predictions (10).

Despite increasing sophistication, machine-learning performance depends on the data-generating process. MoleculeNet highlighted that representation learning can improve molecular prediction but also emphasized difficulties under data scarcity and class imbalance (11). These limitations are highly relevant to toxicity, where positive cases may be rare, labels may be inconsistent, and different experimental systems can yield discordant outcomes. In addition, scaffold split performance is often lower than random split performance because new chemical series differ from training data. Applicability-domain analysis and external validation therefore remain essential.

hERG liability illustrates the distinction between screening and safety assessment. Computational methods can recognize structural patterns associated with channel block and are valuable for early risk reduction, yet ion-channel electrophysiology and concentration-response measurements remain necessary for development decisions (12). Similarly, machine-learning models for Ames mutagenicity can identify structural correlates of genotoxicity, but a computational negative cannot substitute for a validated experimental test when regulatory genotoxicity assessment is required (13). CYP450 prediction also benefits from cross-tool comparison because enzyme inhibition, substrate status, and sites of metabolism are separate questions with different modeling challenges (14).

Drug-induced liver injury is even more complex because toxicity may emerge through multiple mechanisms and can depend on exposure, metabolite formation, host genetics, immune status, and concomitant medication. Reviews of computational DILI modeling have emphasized the need to integrate mechanistic understanding and high-quality datasets and to recognize the limitations of structure-only models (15). For this reason, a predicted DILI probability should be treated as a hazard prioritization signal. Follow-up may include hepatocyte cytotoxicity, mitochondrial function, reactive metabolite trapping, bile-salt export pump studies, transcriptomic signatures, or mechanistic modeling depending on the suspected pathway.

The integrated approach also has a practical medicinal-chemistry advantage: it supports directional optimization. If a series shows high hERG and CYP3A4 risk together with excessive lipophilicity, reducing hydrophobic surface area or basicity may be explored. If solubility is poor while permeability is high, polarity or ionization can be tuned carefully to avoid overcorrection. If plasma protein binding is extreme, free drug exposure should be considered rather than total concentration alone. These decisions should be connected to potency and selectivity so that ADMET improvement does not eliminate the intended pharmacological effect.

A limitation of the current article is that the numerical case-study outputs are illustrative. They were constructed to demonstrate interpretation and reporting and were not generated by live queries to a defined prediction server. Therefore, they must not be cited as original experimental or computational discoveries. A journal submission claiming original compound-level findings should include exact structures or identifiers, software/version information, input settings, prediction date, endpoint units, platform-specific confidence values, and ideally cross-platform or experimental validation. This distinction is essential for reproducibility and scientific integrity.

For OAK-oriented scientific reporting, transparency is preferable to overstating prediction certainty. The Materials and Methods section should identify every model used; the Results section should separate measured from predicted values; tables should include endpoint units and probability definitions; and the Discussion should explicitly address uncertainty and applicability domains. Such reporting allows readers to understand what has been learned from the computational analysis and what remains to be established experimentally.

5. CONCLUSION

In silico ADMET prediction provides a powerful early-stage framework for integrating pharmacokinetic and toxicological considerations into the design of novel bioactive compounds. The most informative strategy is not to search for a molecule that is optimal on every isolated endpoint, but to identify candidates with a balanced profile compatible with the intended route, dose, tissue exposure, and therapeutic window. Physicochemical descriptors provide an interpretable first screen, while machine-learning platforms extend evaluation to absorption, protein binding, CYP interactions, clearance, cardiotoxicity, mutagenicity, and hepatotoxicity.

The illustrative analysis demonstrated that high predicted intestinal absorption may coexist with substantial metabolic or safety liabilities and that consensus scoring can make these trade-offs visible. However, model predictions are hypotheses generated from historical data and molecular representations; they are not substitutes for experimental ADME, toxicology, or clinical evidence. The recommended workflow is therefore computational triage followed by targeted in vitro confirmation, mechanistic investigation of flagged liabilities, and iterative medicinal-chemistry optimization. When applied within its applicability domain and reported transparently, in silico ADMET prediction can reduce unnecessary synthesis and testing, improve lead prioritization, and support a more efficient transition from bioactive molecules to experimentally validated drug candidates.

REFERENCES

1. Lipinski, C. A., Lombardo, F., Dominy, B. W., & Feeney, P. J. (2001). Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings. Advanced Drug Delivery Reviews, 46(1-3), 3-26. https://doi.org/10.1016/S0169-409X(00)00129-0

2. Egan, W. J., Merz, K. M., Jr., & Baldwin, J. J. (2000). Prediction of drug absorption using multivariate statistics. Journal of Medicinal Chemistry, 43(21), 3867-3877. https://doi.org/10.1021/jm000292e

3. Veber, D. F., Johnson, S. R., Cheng, H. Y., Smith, B. R., Ward, K. W., & Kopple, K. D. (2002). Molecular properties that influence the oral bioavailability of drug candidates. Journal of Medicinal Chemistry, 45(12), 2615-2623. https://doi.org/10.1021/jm020017n

4. Delaney, J. S. (2004). ESOL: Estimating aqueous solubility directly from molecular structure. Journal of Chemical Information and Computer Sciences, 44(3), 1000-1005. https://doi.org/10.1021/ci034243x

5. Pires, D. E. V., Blundell, T. L., & Ascher, D. B. (2015). pkCSM: Predicting small-molecule pharmacokinetic and toxicity properties using graph-based signatures. Journal of Medicinal Chemistry, 58(9), 4066-4072. https://doi.org/10.1021/acs.jmedchem.5b00104

6. Daina, A., Michielin, O., & Zoete, V. (2017). SwissADME: A free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules. Scientific Reports, 7, 42717. https://doi.org/10.1038/srep42717

7. Banerjee, P., Eckert, A. O., Schrey, A. K., & Preissner, R. (2018). ProTox-II: A webserver for the prediction of toxicity of chemicals. Nucleic Acids Research, 46(W1), W257-W263. https://doi.org/10.1093/nar/gky318

8. Li, Y., Cheng, L., Liu, Z., Chen, X., Tang, Y., & Li, W. (2019). admetSAR 2.0: Web-service for prediction and optimization of chemical ADMET properties. Bioinformatics, 35(6), 1067-1069. https://doi.org/10.1093/bioinformatics/bty707

9. Xiong, G., Wu, Z., Yi, J., Fu, L., Yang, Z., Hsieh, C., Yin, M., Zeng, X., Wu, C., Lu, A., Chen, X., Hou, T., & Cao, D. (2021). ADMETlab 2.0: An integrated online platform for accurate and comprehensive predictions of ADMET properties. Nucleic Acids Research, 49(W1), W5-W14. https://doi.org/10.1093/nar/gkab255

10. Myung, Y., de Sá, A. G. C., et al. (2024). Deep-PK: Deep learning for small molecule pharmacokinetic and toxicity prediction. Nucleic Acids Research, 52(W1), W469-W475. https://doi.org/10.1093/nar/gkae254

11. Wu, Z., Ramsundar, B., Feinberg, E. N., Gomes, J., Geniesse, C., Pappu, A. S., Leswing, K., & Pande, V. (2018). MoleculeNet: A benchmark for molecular machine learning. Chemical Science, 9(2), 513-530. https://doi.org/10.1039/C7SC02664A

12. Shan, M., Jiang, C., & Qin, L. P. (2022). A review of computational methods in predicting hERG channel blockers. ChemistrySelect, 7(31), e202201221. https://doi.org/10.1002/slct.202201221

13. Chu, C. S. M., Simpson, J. D., O'Neill, P. M., & Berry, N. G. (2021). Machine learning - Predicting Ames mutagenicity of small molecules. Journal of Molecular Graphics and Modelling, 109, 108011. https://doi.org/10.1016/j.jmgm.2021.108011

14. Zhai, J., Man, V. H., Ji, B., Cai, L., & Wang, J. (2023). Comparison and summary of in silico prediction tools for CYP450-mediated drug metabolism. Drug Discovery Today, 28(10), 103728. https://doi.org/10.1016/j.drudis.2023.103728

15. Bassan, A., Alves, V. M., Amberg, A., et al. (2021). In silico approaches in organ toxicity hazard assessment: Current status and future needs in predicting liver toxicity. Computational Toxicology, 20, 100187. https://doi.org/10.1016/j.comtox.2021.100187