In silico ADMET: Supporting new drug discovery

Rule-based approaches

One of the earliest forms of in silico ADMET prediction involves the application of empirical rules derived from known drug properties. The most prominent example is Lipinski’s rule of five, which outlines criteria for oral bioavailability, including molecular weight ≤500 Da, logP ≤5, ≤5 hydrogen bond donors, and ≤10 hydrogen bond acceptors. Other notable filters include Veber’s rule (≤10 rotatable bonds, polar surface area ≤140 Ų) and Egan’s egg, which predicts gastrointestinal absorption based on polarity and lipophilicity.^5,9

While Lipinski’s rule of five remains valuable for small, orally administered molecules, an increasing number of contemporary drug candidates fall into the beyond rule-of-five (bRo5) chemical space, including macrocycles, cyclic peptides, and large, flexible molecules. These compounds often violate classical physicochemical thresholds yet can exhibit acceptable permeability, oral bioavailability, and target engagement through mechanisms such as conformational flexibility and intramolecular hydrogen bonding.^8,9

To address these challenges, modern in silico ADMET approaches incorporate alternative descriptors and models tailored to bRo5 chemistry, including polar surface area efficiency, lipophilic permeability efficiency (LPE), 3D conformational analysis, and dynamic hydrogen-bonding patterns. AI-based models and molecular dynamics-assisted simulations are increasingly applied to better capture the complex absorption, distribution, and metabolic behavior of macrocycles, where classical rule-based filters are insufficient.^7,8

Strengths

These empirical rules offer a simple, interpretable, and rapid method for assessing drug-likeness. Their ease of use makes them highly valuable during the early stages of drug discovery, enabling the quick elimination of compounds that are unlikely to possess favourable pharmacokinetic properties. By applying such filters, researchers can efficiently narrow down large chemical libraries and focus on more promising candidates.^10,11

Limitations

Despite their utility, these rules have significant limitations. They often lack the precision needed for predicting complex ADMET endpoints, especially those involving metabolism, active transport, or toxicity profiles. Additionally, they are not suited for all drug types, such as biologics or compounds targeting the central nervous system, and may fail to capture the nuances of molecular interactions within biological systems.^11,12

QSAR and traditional ML models

QSAR models correlate molecular descriptors or fingerprints with specific ADMET properties, trained on experimental datasets. Traditional ML algorithms, such as random forests, support vector machines, and k-nearest neighbors, have enhanced QSAR modeling by capturing non-linear relationships. QSAR-based tools are commonly applied for endpoints such as logP, blood-brain barrier penetration, CYP450 inhibition, and hERG liability.^13,14

In QSAR and ML-based ADMET modeling, molecular representation plays a critical role in determining predictive performance. Structure-based molecular fingerprints, such as extended-connectivity fingerprints (ECFPs) and MACCS keys, are widely used to encode substructural features and chemical environments into fixed-length binary or count-based vectors suitable for ML algorithms. ECFPs, in particular, are effective in capturing local atomic neighborhoods and have demonstrated strong performance across multiple ADMET endpoints.^13,15

In addition to fingerprints, QSAR models frequently incorporate molecular descriptors, which can be broadly categorized into physicochemical descriptors (e.g., molecular weight, logP, topological polar surface area, hydrogen bond donors/acceptors) and topological or graph-based descriptors (e.g., connectivity indices, molecular complexity, and shape descriptors). Physicochemical descriptors are particularly informative for absorption and distribution-related endpoints, whereas topological descriptors often enhance predictions related to metabolism and toxicity by capturing global molecular architecture.^13,14

The optimal selection and combination of fingerprints and descriptors are highly endpoint-dependent, and hybrid representations that integrate physicochemical, topological, and fingerprint-based features are increasingly used to improve model robustness and generalizability.

Molecular docking and dynamics for ADMET

While docking and molecular dynamics are traditionally applied in target binding studies, they can also contribute to ADMET prediction. For example, docking can assess the likelihood of a compound binding to CYP enzymes, potentially predicting metabolic stability or drug-drug interactions. Similarly, interaction with efflux transporters (e.g., P-glycoprotein) can be explored.^16,17

AI and advanced ML models

Recent advances in deep learning, including graph convolutional networks and recurrent neural networks, have enabled models to learn directly from molecular structures without pre-defined descriptors. These AI models are capable of multi-endpoint prediction and may better generalize across diverse chemical spaces.^19,20

Examples include:

• ADMETlab 3.0 predicts >80 endpoints using AI/ML models

• DeepChem-based custom models, flexible frameworks for training bespoke models

• Emerging hybrid models that integrate physics-based features with AI predictions

The performance of AI-based ADMET models is commonly evaluated using standard classification and regression metrics. For binary endpoints such as CYP inhibition, hERG liability, or hepatotoxicity, area under the receiver operating characteristic curve (AUC) values reported in the literature typically range from 0.70 to 0.90, depending on endpoint complexity and data quality. Classification accuracy for well-characterized ADMET endpoints generally falls within the 65-85% range.^19,20

For continuous pharmacokinetic endpoints, including solubility, clearance, and volume of distribution, root mean square error (RMSE) values commonly reported are in the range of 0.3-0.7 log units, while coefficient of determination (R²) values often range from 0.5 to 0.8 in external validation datasets. Multi-task deep learning models frequently demonstrate improved average performance across related ADMET endpoints compared with single-task models.^18,19

Importantly, model performance varies substantially across endpoints and chemical space, underscoring the need for external validation and cautious interpretation of AI-generated predictions.

Validation strategies

To ensure predictive reliability, in silico ADMET models undergo several validation processes. Internal validation methods such as cross-validation, bootstrapping, and leave-one-out techniques are applied during model training to evaluate consistency and prevent overfitting. External validation involves testing the model on independent datasets, ideally containing chemical structures different from those used in training, to assess its predictive capability. Additionally, benchmarking studies compare the performance of different tools against experimental data or in vivo outcomes to evaluate their real-world applicability. However, despite these validation strategies, the availability of robust external datasets for many ADMET endpoints remains limited, posing challenges in evaluating the generalizability of these models [Figure 2].^26,27

Close

Figure 2:

Hybrid ADMET prediction model integrating AI & ML. ADMET: Absorption, Distribution, Metabolism, Excretion, and Toxicity, AI: Artificial intelligence, ML: Machine learning

Export to PPT

AI- and ML-based ADMET models are typically trained on large, curated datasets derived from publicly available and proprietary sources. Commonly used public databases include ChEMBL, which provides bioactivity and pharmacokinetic data; DrugBank, offering curated drug-related ADMET and target information; and Tox21, which supplies high-throughput toxicity screening data across multiple biological assays. Additional sources, such as PubChem BioAssay and eTOX, are frequently integrated to expand endpoint coverage.^24,25

While these datasets enable large-scale model training, they also introduce challenges related to data heterogeneity, experimental variability, and class imbalance. Consequently, the quality, diversity, and curation of training data remain critical determinants of AI model performance and generalizability.²⁶

Strengths of current tools

In silico tools offer a rapid and cost-effective approach for high-throughput ADMET profiling, allowing early-stage elimination of compounds with poor pharmacokinetic or toxicity profiles. This not only accelerates the drug discovery process but also significantly reduces associated expenses. Additionally, by prioritizing compounds with favorable ADMET characteristics, these tools help minimize the reliance on animal testing, thereby supporting the principles of the 3Rs: replacement, reduction, and refinement. Furthermore, the widespread availability of free, high-quality platforms such as SwissADME and pkCSM makes ADMET screening accessible to researchers across various institutions, enhancing the inclusivity and reach of early-stage drug development.²⁷

Limitations and challenges

In silico ADMET modeling faces several limitations that impact its predictive reliability and regulatory acceptance. One major concern is data quality and bias, as many models are built on legacy experimental datasets that may include errors, inconsistencies, or a bias toward certain chemical classes. Moreover, for rare endpoints such as idiosyncratic toxicity, data is often sparse, which diminishes the predictive power of the models. Another challenge lies in applicability domain restrictions; these models generally perform well only within the chemical space covered by their training data, and their accuracy tends to drop when applied to novel scaffolds or chemotypes.^27,28

Additionally, the black-box nature of many advanced AI-driven models, particularly those using ML and deep learning techniques, limits their interpretability. This lack of transparency can hinder trust and complicate their integration into decision-making processes, especially in regulatory settings. Furthermore, most in silico tools are trained predominantly on human data, which may not translate well to preclinical species. They also often fail to incorporate important context-specific factors such as disease state or formulation effects, reducing their real-world applicability.^28,29

Regulatory hesitancy remains another significant barrier. Agencies such as the FDA and EMA currently regard in silico ADMET predictions as supplementary rather than primary evidence in drug development submissions. This scepticism is partly due to the absence of standardized validation criteria across different tools, which undermines confidence in their reliability and comparability. As a result, despite their potential, in silico models must overcome these challenges before gaining broader acceptance in regulatory frameworks.²⁹

Summary of validation outcomes in the literature

Studies benchmarking tools such as SwissADME, ADMETlab, and pkCSM have demonstrated good performance for simple physicochemical properties (e.g., logP, molecular weight, and GI absorption) but mixed accuracy for more complex endpoints such as CYP inhibition, hERG liability, and hepatotoxicity.

This underscores the importance of combining in silico predictions with experimental confirmation, particularly for critical safety assessments.

Integration of AI and deep learning

Recent advances in AI, particularly deep learning architectures such as graph neural networks and transformers, offer significant potential to enhance predictive performance across complex ADMET endpoints. These models can learn directly from molecular structures, which may help overcome the limitations associated with traditional descriptors and QSAR-based approaches.^6,28

Emerging trends in this area include multi-task learning, where models are trained to predict multiple ADMET endpoints simultaneously by leveraging shared underlying features. Transfer learning is also gaining traction, enabling the adaptation of models trained on large, publicly available datasets like ChEMBL to more specialized tasks that have limited data availability. Additionally, federated learning is being explored as a strategy to collaboratively build robust predictive models without the need to share proprietary or sensitive data, thereby addressing privacy and data security concerns.³⁰

Despite these advancements, a major challenge remains: the black-box nature of deep learning models. This lack of interpretability can hinder trust, widespread adoption, and regulatory approval. As a result, there is a growing emphasis on the development of explainable AI techniques to provide transparency into model decision-making processes and ensure responsible deployment in drug discovery.^29,30

Hybrid and multi-scale modeling

Combining physics-based methods (e.g., molecular docking, molecular dynamics) with data-driven AI predictions may provide complementary strengths. Such hybrid models can incorporate mechanistic insights while retaining the scalability and flexibility of ML.

Multi-scale modeling that links molecular, cellular, and systemic pharmacokinetics could further enhance predictive accuracy, particularly for complex endpoints such as organ-specific toxicity.¹⁹

Cloud-based and real-time ADMET screening platforms

Future in silico ADMET tools are likely to be integrated within cloud-based medicinal chemistry platforms, enabling real-time feedback during compound design and optimization. This seamless integration could facilitate iterative design cycles, with ADMET properties guiding lead selection at every stage.³¹

Open data and collaborative model development

The creation and curation of large, high-quality, and openly accessible ADMET datasets will be critical for advancing the field. Initiatives such as ChEMBL, PubChem, Tox21, and eTOX provide valuable resources, but further efforts are needed to expand coverage of rare endpoints and underrepresented chemical spaces.

Collaborative model development, including benchmarking challenges and open competitions, may help drive transparency, validation, and community consensus on best practices.²⁹

Towards regulatory integration

For in silico ADMET tools to gain wider regulatory acceptance, standardized validation frameworks and reporting guidelines are essential. The development of consensus criteria for model performance, applicability domain definition, and uncertainty quantification will support their inclusion in regulatory submissions as part of weight-of-evidence approaches.^6,27