Qualitative and quantitative in silico toxicity profiling of “angel dust”: phencyclidine (PCP) analogues as new psychoactive substances (3-HO-PCP, 3-MeO-PCP, 4-MeO-PCP, 3-HO-PCE, 3-MeO-PCE, 4-MeO-PCE)

Abstract

Phencyclidine (PCP), historically known as “angel dust,” and its analogues (3-HO-PCP, 3-MeO-PCP, 4-MeO-PCP, 3-HO-PCE, 3-MeO-PCE, 4-MeO-PCE) are dissociative new psychoactive substances with high abuse potential and limited experimental safety data. An integrated in silico workflow (STopTox, admetSAR 3.0, ADMETlab 3.0, ACD/Labs Percepta, Toxtree, ProTox 3.0, OCHEM, TEST, VEGA QSAR) was applied to profile acute toxicity and key hazard domains. Across platforms, rat oral LD₅₀ values for PCP-type analogues were consistently in the ~ 200–630 mg/kg range (Percepta 210–800 mg/kg, TEST 197–628 mg/kg, VEGA 278–368 mg/kg, ProTox ~ 348–404 mg/kg), indicating moderate acute toxicity by the oral route; substantially lower LD₅₀ values were predicted for intravenous exposure in mice (~ 25–59 mg/kg). Qualitative models (STopTox, ADMETlab, admetSAR) classified all compounds as acutely toxic by the oral route (e.g., STopTox oral toxicity confidence ~ 77–92%) and commonly predicted inhalation/dermal risks depending on the analogue; admetSAR assigned EPA Category III for acute oral toxicity. Organ-specific effects (Percepta; ADMETlab) highlighted the lungs, liver, and blood as prominent targets (e.g., lungs 0.89–0.93, liver up to 0.91, blood up to 0.85), with gastrointestinal involvement (up to 0.82) and generally lower kidney probabilities (~ 0.09–0.70). Cardiotoxicity signals included predicted hERG inhibition with Percepta IC₅₀ ~ 4.9–12.3 µM and high probabilities of hERG blockade in ADMETlab/admetSAR, supporting potential QT-prolongation risk. Genotoxicity predictions were consistently negative across Percepta, OCHEM, ADMETlab, admetSAR, and VEGA. Eye/skin irritation potential was notable for phenolic analogues, with Percepta indicating high probabilities for 3-HO-PCP and 3-HO-PCE (eye ~ 88–90%, skin ~ 96%), while other tools showed model-dependent variability. Endocrine screening suggested at most weak-to-moderate ER-α interactions, with the highest probability for 3-HO-PCP (LogRBA >  − 3). Overall, convergent multi-tool evidence indicates moderate acute toxicity, cardiotoxicity signals, and multi-organ risk for PCP analogues, while mutagenicity appears unlikely. These results provide mechanistic and quantitative context to inform clinical management, forensic interpretation, and risk assessment of this NPS class.

Introduction

Phencyclidine (PCP) and its structural analogues, collectively referred to as PCP-type substances, represent a group of dissociative drugs that occupy a unique position within the spectrum of new psychoactive substances (NPS). Originally synthesized in the 1950s as a potential intravenous anesthetic, PCP was soon abandoned for clinical use due to its severe psychotomimetic side effects, including hallucinations, delirium, and violent behavior (Greifenstein et al. 1958; Domino 2010). Nevertheless, the compound found a notorious second life as an illicit recreational drug from the late 1960s onwards, earning street names such as “angel dust.” Since then, numerous analogues, including tenocyclidine (TCP), eticyclidine (PCE), rolicyclidine (PCPy), and more recently 3-MeO-PCP or 4-MeO-PCP, have emerged on the black market (Morris and Wallach 2014; Pelletier et al. 2022). These analogues, often synthesized to circumvent legal controls, differ in potency, pharmacokinetic properties, and toxicity, thereby posing a growing challenge for forensic and clinical toxicology. Chemically, PCP-type substances are arylcyclohexylamines characterized by a cyclohexane ring substituted with an aryl group and a piperidine moiety (Table 1).

Table 1.

Known physico-chemical properties of phencyclidine-type substances

Forensic acronym	3-HO-PCP	3-MeO-PCP	4-MeO-PCP	3-HO-PCE	3-MeO-PCE	4-MeO-PCE
Chemical structure
Synonyms	3-Hydroxyphencyclidine; Pcp-3-OH; 3-HYDROXY PCP	3-Methoxyphencyclidine; UNII-28A91R606X; 5FUR-144	N-(1-(4-methoxyphenyl)cyclohexyl)piperidine; METHOXYDINE	BB7HFF5XW2; UNII-BB7HFF5XW2; 3-Hydroxyeticyclidine	3-Methoxy-pce; 3-Methoxyeticyclidine; 3-Methoxy eticyclidine	4-Methoxy pce; 4-Methoxy eticyclidine; Cyclohexylamine
IUPAC name	3-(1-piperidin-1-ylcyclohexyl)phenol	1-[1-(3-methoxyphenyl)cyclohexyl]piperidine	1-[1-(4-methoxyphenyl)cyclohexyl]piperidine	3-[1-(ethylamino)cyclohexyl]phenol	N-ethyl-1-(3-methoxyphenyl)cyclohexan-1-amine	N-ethyl-1-(4-methoxyphenyl)cyclohexan-1-amine
CAS	79787-43-2	72242-03-6	2201-35-6	2243815-20-3	1364933-80-1	2201-68-5
SMILES	C1CCC(CC1)(C2  =  CC( =  CC = C2)O)N3CCCCC3	COC1 = CC = CC( = C1)C2(CCCCC2)N3CCCCC3	COC1 = CC  =  C(C = C1)C2(CCCCC2)N3CCCCC3	CCNC1(CCCCC1)C2 = CC( = CC = C2)O	CCNC1(CCCCC1)C2 = CC( = CC  =  C2)OC	CCNC1(CCCCC1)C2 = CC = C(C = C2)OC
Molecular formula	C17H25NO	C18H27NO	C18H27NO	C14H21NO	C15H23NO	C15H23NO
Molecular weight (g/mol)	259.4	273.4	273.4	219.32	233.35	233.35

This scaffold confers high lipophilicity and strong binding affinity to the PCP site of the N-methyl-D-aspartate (NMDA) receptor, where they act as non-competitive antagonists (Kapur and Seeman 2002). Inhibition of NMDA receptor function leads to the hallmark dissociative anesthetic effects, including profound analgesia, anesthesia, perceptual distortions, and cognitive impairments. However, NMDA receptor blockade also contributes to neurotoxic potential, including excitotoxicity and neuronal apoptosis (Olney et al. 1991). In addition to their primary action at NMDA receptors, PCP derivatives interact with dopaminergic, serotonergic, and sigma-1 receptors, contributing to their complex pharmacological profile (Nadler et al. 1990; Yamamoto et al. 1995; Jentsch and Roth 1999). These multimodal effects explain both their appeal as recreational drugs and their significant health risks. From a toxicological perspective, PCP and its analogues are associated with a wide range of adverse effects. Acute intoxication typically manifests as agitation, psychosis, hallucinations, hypertension, tachycardia, nystagmus, and ataxia, while severe poisoning may progress to seizures, hyperthermia, rhabdomyolysis, renal failure, and death (Vaupel et al. 1984; Bey and Patel 2007). Chronic use has been linked to cognitive deficits, mood disorders, and persistent psychotic states resembling schizophrenia (Javitt and Zukin 1991). The high abuse liability and severe toxicological profile have led to the classification of PCP as a Schedule II controlled substance in the United States and equivalent regulations worldwide. Nonetheless, the constant appearance of novel PCP-type derivatives on the illicit market underscores the urgent need for systematic toxicological evaluation of this chemical class (Pelletier et al. 2022).

It should be noted that from forensic point of view, PCP-type substances are explicitly listed as a separate category in the UNODC Early Warning Advisory on New Psychoactive Substances (EWA, United Nations Office on Drugs and Crime). This classification places them alongside other high-priority NPS groups such as synthetic cannabinoids, cathinones, and fentanyl analogues, underscoring their recognized importance in the international drug monitoring framework. However, despite this formal recognition, toxicological knowledge on PCP analogues has remained virtually absent. Reports submitted to the UNODC highlight the appearance of novel derivatives such as 3-MeO-PCP or 4-MeO-PCE on the illicit market, yet the available information is limited almost exclusively to analytical detection methods and isolated clinical case descriptions. Many PCP analogues have never been studied in vivo under standardized conditions, and existing data are often very limited or derived from heterogeneous experimental protocols (Holsapple et al. 1982). The literature consistently emphasizes that research on this class of NPS has historically focused mainly on analytical detection and the development of new analytical methods for their identification in forensic and clinical samples. This has enabled forensic laboratories and law enforcement to recognize novel compounds rapidly after their appearance on the illicit market. As a result, clinicians confronted with intoxications are forced to rely on symptomatic treatment alone, without access to evidence-based guidance on risk assessment or therapeutic strategies. This imbalance means that while forensic science can identify the presence of PCP-type substances, the medical community remains without tools for effective clinical management, and regulatory authorities are unable to establish scientifically grounded safety thresholds. Furthermore, ethical concerns and regulatory restrictions severely limit the feasibility of conducting new animal studies with such highly potent psychoactive compounds. As a result, significant knowledge gaps remain regarding the acute and chronic toxicity, metabolism, and long-term health risks of these substances. These gaps hinder not only clinical management of intoxications but also the establishment of scientifically sound forensic and regulatory assessments. In this context, computational toxicology and in silico modeling offer valuable tools to bridge existing data gaps. Recent advances in quantitative structure–activity relationship (QSAR) modeling, machine learning algorithms, and curated toxicological databases enable the prediction of key toxicological endpoints directly from chemical structure (Sushko et al. 2011; Cheng et al. 2012; Roncaglioni et al. 2022). These methods have been successfully applied to a wide range of NPS, including synthetic cannabinoids, cathinones, phenethylamines, and more recently Bromo-DragonFLY (Noga and Jurowski 2024). Application of in silico approaches are particularly advantageous for PCP analogues, where experimental data are scarce but structural diversity is limited, allowing meaningful extrapolation based on structural similarity and pharmacophore analysis. Among the most relevant toxicological endpoints for PCP-type substances are:

1. Acute toxicity (LD₅₀ values)—providing quantitative estimates of lethality in rodents, which remain the most widely used proxy for human acute toxicity.

2. Systemic health effects—evaluating the probability of adverse outcomes in organ systems such as the liver, kidneys, heart, and lungs, which are commonly affected in clinical intoxications.

3. Cardiotoxicity—especially via inhibition of the human ether-à-go-go related gene (hERG) channel, a well-established predictor of arrhythmogenic potential.

4. Genotoxicity and mutagenicity—although less commonly considered for dissociatives, these endpoints are critical for long-term safety evaluation.

5. Irritation and sensitization potential—relevant for dermal or ocular exposure scenarios, particularly in forensic contexts involving accidental contamination.

6. Endocrine disruption—an emerging area of toxicological concern that warrants investigation in the context of structurally diverse NPS.

Previous computational studies have highlighted the feasibility and predictive value of such approaches. For example, Madaj et al. (2024) demonstrated with docking and molecular dynamics simulations the strong binding affinity of Novichok agents to acetylcholinesterase. Similarly, Melagraki (2022) reviewed computational toxicology strategies, including QSAR, docking, and MD simulations, for assessing the hazards of chemical warfare agents. These examples underscore the potential of in silico toxicology to provide early-stage hazard assessment, guide regulatory decisions, and support forensic investigations (Muratov et al. 2020; Bueso-Bordils et al. 2024). Extending these methods to PCP-type substances is a logical next step that addresses a pressing public health and forensic need. The present study therefore aims to systematically characterize the predicted toxicological profile of PCP and its analogues using a broad panel of different qualitative and quantitative in silico methods, including: STopTox, admetSAR, ACD/Percepta, Toxtree, ProTox, ADMETlab, OCHEM, TEST, and VEGA QSAR. By integrating predictions across multiple platforms, we provide a comprehensive assessment of acute toxicity, systemic health risks, cardiotoxicity, genotoxicity, irritation potential, and endocrine activity. Validation was achieved through comparison with the limited available experimental data for selected analogues (Holsapple et al. 1982; Vaupel et al. 1984). The results are discussed in the context of clinical case reports, forensic evidence, and mechanistic insights into NMDA receptor antagonism.

Justification for this work lies in the fact that, despite the growing availability of PCP-type substances on the illicit market, no systematic toxicological data have been generated for this group to date. Our study therefore delivers the first comprehensive toxicological profiling of PCP analogues, addressing a major knowledge gap in the literature using modern in silico methodology. By applying a multi-platform in silico approach, we provide quantitative estimates alongside qualitative mechanistic insights. This dual perspective is crucial because it offers immediate relevance for forensic interpretation, while at the same time providing the first structured toxicological evidence of major importance from both the forensic and medical point of view.

Methods

STOPTOX

STopTox (Systemic and Topical chemical Toxicity) is an advanced in silico computational platform designed to predict the probability of acute toxic effects within the testing framework known as the “6-pack” (Borba et al. 2022). This test battery includes acute oral toxicity, acute dermal toxicity, acute inhalation toxicity, skin irritation and corrosion, eye irritation and corrosion, and skin sensitisation (Silva et al. 2021). The predictions are generated using externally validated QSAR models for each endpoint, based on experimental in vivo animal data. Such assays, required by numerous regulatory authorities, are widely applied to evaluate different aspects of acute toxicity in humans. Machine learning (ML) models, exemplified by STopTox, provide an efficient screening approach that delivers critical information for chemical risk assessment. For the acute oral toxicity endpoint, STopTox applies the OECD TG 401, 420, 423, and 425 guidelines and uses experimental data obtained from rats. The original dataset comprised 8994 entries, which were curated to eliminate duplicates and inconsistent records, resulting in 8445 unique chemical structures. The modelling process employed the Random Forest algorithm in combination with MACCS structural descriptors. STopTox is considered one of the most comprehensive sets of computational models for predicting acute toxicological hazards, developed according to best practices for the construction and validation of QSAR models. The models utilise the largest publicly available and rigorously curated datasets. A key feature of STopTox is the generation of fragment contribution maps, which indicate molecular substructures that are predicted to increase or decrease toxicity. Analysis of these maps facilitates the identification of toxophores, defined as chemical groups responsible for toxic effects, as well as other structural alerts that may enhance or reduce the toxic potential of PCP-type substances and their analogues.

AdmetSAR 3.0

AdmetSAR is a comprehensive and scientifically validated computational platform for predicting ADMET properties, encompassing absorption, distribution, metabolism, excretion, and toxicity (Cheng et al. 2012; Gu et al. 2023). Its latest version, admetSAR 3.0, integrates more than 40 predictive models that estimate diverse ADMET endpoints through in silico screening methods (Moon et al. 2017; Yang et al. 2019). The applicability domain (AD) of the models is defined by six core physicochemical and topological descriptors: molecular weight (MW), partition coefficient (logP), number of atoms (nAtom), number of rings (nRing), number of hydrogen bond acceptors (HBA), and number of hydrogen bond donors (HBD). The platform functions as both a knowledge base and a predictive tool, allowing ADMET assessment by inputting the SMILES representation of a compound or by drawing the structure directly in the user interface. The underlying predictive architecture is based on a graph convolutional neural network implemented using the DeepChem library. One of the foundational studies for the toxicity prediction component involved the curation of a dataset containing 10,207 compounds with rat oral LD₅₀ values (Zhu et al. 2009). admetSAR 3.0 prioritises transparency and reproducibility by employing open-source cheminformatics and machine learning frameworks for algorithm development and descriptor calculation. The applicability domain is determined by analysing the distribution of the six key physicochemical parameters within the training set, with the boundaries established at the 5th and 95th percentile values. Compounds are classified as “In Domain” when all six properties fall within these boundaries. A “Warning” is assigned if any single property exceeds this range, while “Out Domain” applies to molecules with values beyond the absolute maximum or minimum observed in the training set. For the acute oral toxicity endpoint (rat, LD₅₀), admetSAR 3.0 reports performance metrics as follows: Pearson correlation coefficient (PCC) of 0.77, coefficient of determination (R²) of 0.592, mean absolute error (MAE) of 0.378, and root mean square error (RMSE) of 0.534 for the test set; PCC of 0.765, R² of 0.585, MAE of 0.387, and RMSE of 0.540 for the validation set. These capabilities make admetSAR a valuable tool for early-stage toxicity assessment of PCP-type substances and related analogues.

ACD/Labs Percepta

ACD/Labs Percepta is a commercially available scientific platform designed to predict a wide range of toxicological properties using computational modelling approaches (ACD/Labs 2024a). The software incorporates well-established acid dissociation constant (pKa) and logD calculations, which provide essential insights into the relationship between chemical structure and a variety of ADME, toxicological, and physicochemical characteristics. Its predictive performance is enhanced through the integration of novel consensus models, which improve the overall accuracy of estimations. The structure optimisation module, equipped with a comprehensive set of ADMET filters, supports user-guided project workflows, facilitating the achievement of specific research objectives. Percepta generates predictions using a combination of molecular structure, known chemical properties, and historical toxicological data. These predictions are crucial for assessing the potential impact of compounds on biological systems, representing an important step in determining their safety and efficacy in applications such as pharmaceuticals, cosmetics, and consumer products. The reliability and accuracy of Percepta’s models are regularly evaluated and refined based on emerging scientific evidence. In the present study, we employed Percepta version 2023.1.2 under a licence provided by the Institute of Medical Expertises in Łódź. The toxicological parameters assessed included acute toxicity (LD₅₀) (Gromek et al. 2022), health effects (Niu et al. 2023) on key physiological systems (blood, cardiovascular, gastrointestinal, renal, hepatic, and pulmonary), genotoxicity (mutagenicity determined via the Ames test) (Fournier et al. 2023), eye and skin irritation (Verma and Matthews 2015), cardiotoxicity (evaluated through hERG inhibition) (Lanevskij et al. 2022), and disruption of the endocrine system (Stanojević et al. 2021). The results of this work highlight the strong capabilities of Percepta in predictive toxicology, providing valuable insights for the safety evaluation of PCP-type substances and structurally related analogues.

Toxtree 3.1.0

Toxtree 3.1.0 is an open-source software platform (https://toxtree.sourceforge.net/) developed for toxicological risk assessment based on chemical structure. It enables users to enter molecules either via SMILES notation or as two-dimensional structures and classifies them into Cramer classes using a decision tree that evaluates relevant structural features. This classification supports the Threshold of Toxicological Concern (TTC) concept, which estimates potential health risks for substances with limited toxicological data by linking structural characteristics to established exposure thresholds (Kirsch et al. 2020). Compounds assigned to Cramer Class I are considered to have low or negligible toxicological concern and typically contain structural motifs commonly found in the human diet. Class II includes compounds with moderate toxicological potential, often identified through the presence of certain structural alerts. Class III comprises substances of high toxicological concern due to the presence of functional groups known to induce significant toxicity. Each class is associated with a TTC value (expressed in µg/day) that can be compared with predicted or measured human exposure levels to evaluate whether a compound falls within an acceptable risk range (Patlewicz et al. 2008). In addition to Cramer classification, Toxtree provides multiple modules for predicting various toxicological endpoints, including mutagenicity, carcinogenicity, skin and eye irritation, biodegradability, sensitisation, and the potential for protein or DNA binding (Roberts et al. 2015). An important functionality is the Benigni/Bossa module, which assesses carcinogenic potential based on structural alerts for both genotoxic and non-genotoxic mechanisms. For certain chemical classes, such as aromatic amines and α, β-unsaturated aldehydes, Toxtree integrates quantitative structure–activity relationship (QSAR) models to produce refined predictions within a weight-of-evidence framework (Lapenna et al. 2011; Contrera 2013).

ProTox 3.0

ProTox 3.0 is an advanced computational platform, updated in March 2024, that utilises multiple machine learning models and curated databases to predict toxicological properties (Drwal et al. 2014). The software provides estimations for target organ toxicity, various toxicological endpoints, and lethal dose (LD₅₀) values. It is widely applied in drug discovery and chemical safety evaluation, offering essential insights into the potential toxic effects of novel compounds. The selection of this tool for the present study was based on its rigorous external validation, demonstrated in previous research by Drwal et al. (2014) and Banerjee et al. (2018a, b), which confirmed its robust predictive performance. The platform combines similarity-based and fragment-based approaches to predict toxicity categories and to flag potential toxicity targets. Outputs include confidence scores, a detailed toxicity radar chart, and information on structurally similar compounds with known toxic effects. The Acute Oral Toxicity (AOT) model in ProTox 3.0 employs two-dimensional similarity analysis with compounds of known LD₅₀ values, identifying structural fragments frequently associated with toxicity. Model validation is performed using leave-one-out cross-validation, in which the three most similar compounds from the training set are identified for each query molecule using fingerprint similarity metrics. For oral toxicity predictions, ProTox 3.0 reports results as estimated LD₅₀ values (mg/kg). Performance assessment for the acute toxicity model on the cross-validation set demonstrated sensitivity of 73.08%, specificity of 94.62%, precision of 73.50%, and coverage of 91.78%. These features make ProTox 3.0 a valuable component in the early-stage toxicological evaluation of PCP-type substances and structurally related analogues.

ADMETlab 3.0

ADMETlab 3.0 is a comprehensive scientific platform designed to predict the ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties of chemical compounds (Dong et al. 2018; Xiong et al. 2021). The system is built upon high-quality experimental datasets and employs a multitask graph attention framework to develop robust and accurate predictive models. It supports a wide range of endpoints, including physicochemical characteristics, medicinal chemistry properties, ADME-related parameters, toxicity endpoints, and toxicophore identification rules. With its user-friendly interface and batch processing capabilities, ADMETlab 3.0 is a practical tool for medicinal chemists engaged in drug discovery and development. The predictive models in ADMETlab 3.0 leverage graph-based neural networks, which process molecular structures as graphs, and incorporate an attention mechanism that prioritises structurally relevant regions of the molecule. This approach enhances the model’s ability to capture complex structural relationships and interactions, enabling more precise and nuanced predictions across multiple ADMET-related endpoints. The integrated databases are carefully curated to ensure the use of high-quality experimental data, improving the reliability and reproducibility of results. For the rat oral acute toxicity classification model, ADMETlab demonstrated strong predictive performance. On the validation set, results included an area under the ROC curve (AUC) of 0.846, accuracy (ACC) of 0.795, specificity (SP) of 0.826, sensitivity (SEN) of 0.744, and Matthews correlation coefficient (MCC) of 0.567. On the training set, the model achieved an AUC of 0.986, ACC of 0.936, SP of 0.923, SEN of 0.957, and MCC of 0.868. These capabilities make ADMETlab an effective tool for early-stage toxicological screening of PCP-type substances and structurally related analogues.

Online chemical modeling environment (OCHEM)

The Online Chemical Modeling Environment (OCHEM) is a web-based platform designed to streamline and automate the processes involved in developing Quantitative Structure–Activity Relationship (QSAR) models (Sushko et al. 2011, 2012). It consists of two main components: a user-contributed database of experimental measurements and a modelling framework. Operating on the wiki principle, the database enables users to enter, retrieve, and modify chemical and biological data while maintaining a strong focus on data quality and verifiability. The database is directly linked to the modelling environment, which guides users through all stages of model development, including data retrieval, descriptor calculation and selection, application of machine learning algorithms, model validation, result analysis, and applicability domain assessment (Oprisiu et al. 2013; Tetko et al. 2013). OCHEM distinguishes itself from comparable platforms by encouraging original model authors to share their work publicly, thereby promoting collaboration and knowledge exchange within the QSAR/QSPR research community. The overarching goal is to establish OCHEM as a widely accessible and collaborative resource for online modelling studies. The platform supports an extensive range of descriptor calculation packages, including CDK (3D, 274 descriptors), Dragon v.6 (3D, 4,885 descriptors across 29 blocks), Dragon6_part (blocks 1–28), OEstate, ALogPS, ISIDA Fragments (length 2–4), GSFrag, Mera, Mersy (3D), Chemaxon (3D, 499 descriptors), Inductive (3D), Adriana (3D, 211 descriptors), Spectrophores (3D), QNPR (length 1–3), Structural Alerts, and SIRMS. This extensive set of descriptor packages enables a versatile and inclusive modelling approach, capturing diverse structural and physicochemical characteristics relevant to QSAR studies. For the mutagenicity endpoint (AMES test), the OCHEM classification model achieved the following predictive performance: in the training set, accuracy of 77.7 ± 0.6, balanced accuracy of 77.5 ± 0.6, MCC of 0.55 ± 0.01, and area under the ROC curve (AUC) of 0.852 ± 0.006; in the test set, accuracy of 79.6 ± 0.8, balanced accuracy of 79.5 ± 0.8, MCC of 0.59 ± 0.02, and AUC of 0.86 ± 0.008. These capabilities make OCHEM a valuable platform for generating high-quality QSAR models applicable to the toxicological assessment of PCP-type substances and related analogues.

TEST

The Toxicity Estimation Software Tool (TEST) version 5.1.2, developed by the US Environmental Protection Agency (EPA) as an open-source program, enables chemical toxicity assessment using only SMILES strings or CAS numbers (Worth and Gatnik 2010; Worth et al. 2010; Martin 2018). The software predicts acute oral toxicity (rat LD₅₀) through multiple QSAR approaches that incorporate a broad range of molecular descriptors, including structural, constitutional, connectivity, shape, topological, molecular distance, fragment, and electrotopological features. These descriptors are derived from curated datasets within the EPA ECOTOX database. The Consensus method was applied, which aggregates predictions from all individual QSAR models implemented in TEST and calculates the average toxicity estimate (Lunghini et al. 2019). This strategy takes into account the applicability domain of each modelling technique, ensuring that the final value reflects a balanced integration of predictions from diverse analytical approaches. The Consensus approach is regarded as the most reliable output provided by TEST, as it combines complementary perspectives from different modelling algorithms to deliver a robust and well-rounded toxicity prediction. For the acute oral toxicity endpoint (rat LD₅₀), the Consensus method in TEST achieved a coefficient of determination (R²) of 0.633 and a root mean square error (RMSE) of 0.595.

VEGA QSAR

The VEGA Application (Virtual Models for the Evaluation of Chemical Properties within a Global Architecture) is a predictive modelling platform developed through European Union projects including CAESAR, ORCHESTRA, and ANTARES (Benfenati et al. 2013a, 2019). It integrates QSAR models and rule-based expert systems to predict a wide range of human toxicity endpoints such as mutagenicity, carcinogenicity, developmental toxicity, and skin sensitisation. The platform also addresses ecotoxicological endpoints, including acute aquatic toxicity in Daphnia and fish, as well as environmental and physicochemical properties such as the bioconcentration factor (BCF), ready biodegradability, and logKow (Roncaglioni et al. 2022). VEGA incorporates an Applicability Domain Index (ADI) to evaluate the reliability of predictions, with values ranging from 0 (lowest reliability) to 1 (highest reliability). The system automatically performs several checks to identify potential limitations in the prediction (Benfenati et al. 2013b). These include verifying the structural similarity between the target compound and the closest analogues in the training dataset, assessing concordance between the experimental values of similar compounds and the prediction, evaluating prediction precision by comparing experimental and predicted values for similar compounds, detecting uncommon molecular fragments in the target structure, confirming that descriptor values for the target compound fall within the descriptor range of the training set, and estimating the impact of a ± 10% variation in descriptor values on the prediction. VEGA also flags molecular fragments associated with chemical classes that have higher prediction uncertainty, reporting them as warnings. For each prediction, the ADI provides a “safety value” that reflects the overall reliability based on these checks. For the mutagenicity endpoint (AMES test, Consensus model), external validation of the predictive model produced the following performance metrics: accuracy of 0.69, sensitivity of 0.63, specificity of 0.70, and MCC of 0.24.

Validation of in silico methods used

To assess the validity of the applied in silico methods, we compared the predicted acute toxicity values with available experimental data for selected PCP-type substances. The evaluation was based on calculating the prediction similarity index (%) by comparing the experimentally determined LD₅₀ values (reference values) with those generated by the computational models (Patlewicz et al. 2014). Compounds for validation were selected according to two main criteria: (1) they belong to the same chemical class as PCP or represent close structural analogues, and (2) reliable experimental LD₅₀ data are available in the literature for at least one administration route in rodents. This approach ensured that the validation process remained both targeted and scientifically relevant. Validating predictions for PCP derivatives presented substantial challenges. Only a small number of analogues with robust and comparable acute toxicity data are available, and many of these originate from older toxicological studies with heterogeneous experimental designs. In addition, substances from this chemical class are controlled under drug legislation in many jurisdictions, which has historically limited the number of systematic toxicological studies conducted under standardised laboratory conditions. As a result, the experimental dataset is both sparse and uneven in quality, making it extremely difficult to identify suitable candidates for validation. The selection process therefore followed the principles outlined in the OECD guidance on grouping of chemicals and read-across approaches (OECD 2023), which recommend considering both structural similarity and similarity in toxicological mode of action when selecting analogues for predictive modelling. In the present study, we were able to select only a limited set of well-characterised compounds, namely phencyclidine, tenocyclidine, benocyclidine, eticyclidine and rolicyclidine, to serve as reference points. Structural similarity searches were conducted using OECD QSAR Toolbox, ChemMine Tools and ChemSpider Structure Search (Backman et al. 2024; OECD 2024; RoySocChem 2024). The degree of structural similarity between the target compounds and reference molecules was calculated using the Similarity Workbench, which applies atom pair and maximum common substructure (MCS) similarity metrics with the Tanimoto coefficient (T) as the similarity measure (Cao et al. 2008):

where T(A, B)—Tanimoto coefficient between structures A and B; a—the number of features present in structure A; b—the number of features present in structure B; c—the number of features common to both structures A and B.

The MCS tool identifies the largest shared substructure between the two molecules. The validation results are summarised in Table 2. The table presents predicted LD₅₀ values alongside experimental values for oral, intraperitoneal and intravenous administration routes in mice, as obtained from the literature. The prediction similarity percentage indicates the agreement between predicted and experimental values, while the reliability index (RI) provided by the Percepta platform reflects the internal confidence of the model in each prediction.

Table 2.

Results and idea of the validation process for assessment of applied in silico studies (Cone et al. 1979; Holsapple et al. 1982; Vaupel et al. 1984; PubChem 2025a, b, c)

Compound	Species	Administration route	Method	Predicted LD50 (mg/kg)	Reliability	Experimental LD50 (mg/kg)	Prediction similarity (%)
Phencyclidine	Mouse	Oral	Percepta	220.00	RI = 0.76	75.00 (PubChem 2025a)	34
CAS: 77-10-1		Intraperitoneal		100.00	RI = 0.70	79.00 (Cone et al. 1979)	79
SMILES: C1CCC(CC1)(C2 = CC = CC = C2)N3CCCCC3		Intravenous		26.00	RI = 0.81	13.87 (Holsapple et al. 1982)	53
Tenocyclidine	Mouse	Intraperitoneal	Percepta	160	RI = 0.66	178.00 (PubChem 2025b)	90
CAS: 21500–98-1
SMILES: C1CCC(CC1)(C2 = CC = CS2)N3CCCCC3
Benocyclidine	Mouse	Intraperitoneal	Percepta	190	RI = 0.47	100.00 (PubChem 2025c)	53
CAS: 112726–66-6
SMILES: C1CCC(CC1)(C2 = CC3 = CC = CC = C3S2)N4CCCCC4
Eticyclidine	Mouse	Intraperitoneal	Percepta	160	RI = 0.43	53.53 (Vaupel et al. 1984)	33
CAS: 2201-15-2
SMILES: CCNC1(CCCCC1)C2 = CC = CC = C2
Rolicyclidine	Mouse	Intraperitoneal	Percepta	130	RI = 0.57	73.37 (Vaupel et al. 1984)	56
CAS: 2201-39-0
SMILES: C1CCC(CC1)(C2 = CC = CC = C2)N3CCCC3

RI, Reliability Index; SC, Similarity coefficient; N/A, not applicable

The validation process followed established guidelines for acute toxicity assessment, prioritising oral administration in rodents where possible, in line with common endpoints supported by most computational tools. Alternative validation strategies proposed in previous research (Kutsarova et al. 2021; Moustakas et al. 2022) were considered unsuitable for the current study due to the highly limited availability of experimental toxicity data for PCP analogues. The methodology adopted here provided a scientifically sound and pragmatic validation framework tailored to the specific characteristics of this chemical class.

Results

Qualitative in silico methods

The qualitative prediction of acute toxicity was the first step in the computational assessment of PCP-type substances, providing an initial indication of their potential toxicological hazard before more detailed quantitative modelling. Classification was performed using the Cramer classification system implemented in the Toxtree software. All investigated analogues were assigned to Cramer Class III, corresponding to substances of high toxicological concern. The results generated by Toxtree for representative compounds are shown in Fig. 1. Toxtree also highlighted toxicophoric features. The piperidine ring was consistently flagged across the compounds. This cyclic tertiary amine is fundamental to the pharmacological activity of PCP derivatives and mediates high-affinity binding to NMDA receptors. PCP and methoxy-substituted analogues such as 3-MeO-PCP and 4-MeO-PCP are recognised as high-affinity, selective NMDA receptor ligands whose interaction with the PCP site underlies dissociative and neurotoxic effects (Roth et al. 2013). Aromatic substituents also influence toxicological profiles. Studies of halogenated ketamine analogues, which share a similar arylcyclohexylamine scaffold, have demonstrated that aromatic halogenation, for example with bromine, markedly affects substrate binding affinity to the CYP2B6 enzyme. This finding suggests that halogens may have a similar impact on metabolism and toxicity in PCP analogues (Wang et al. 2019).

Fig. 1.

Results of toxicity assessment of PCP-type substances using Toxtree software

Following the initial classification of toxicological concern, the next step was to assess the likelihood of acute toxicity for PCP-type substances via different exposure routes and to identify toxicophoric features potentially responsible for these effects. The STopTox, ADMETlab, and admetSAR platforms were employed to evaluate the acute oral toxicity (AOT), acute dermal toxicity (ADT), and acute inhalation toxicity (AIT) profiles. The results of these predictions are summarised in Table 3.

Table 3.

Qualitative in silico prediction of acute toxicity for PCP-type substances using STopTox, ADMETlab and admetSAR

Compound	Type of acute toxicity	STopTox				ADMETlab		admetSAR
		Prediction	Confidence (%)	Applicability domain	Predicted toxicophore(s)	Prediction	Confidence (%)	Prediction	Confidence (%)
3-HO-PCP	Oral	Toxic (+)	90.0			Medium	43.2	Toxic (+)	98.2
	Dermal	Non-Toxic (−)	60.0			N/A	N/A	Toxic (+)	74.7
	Inhalation	Toxic (+)	50.0			High	88.6	Toxic (+)	96.4
3-MeO-PCP	Oral	Toxic (+)	92.0			Medium	40.6	Toxic (+)	97.6
	Dermal	Non-Toxic (−)	56.0			N/A	N/A	Toxic (+)	73.2
	Inhalation	Non-Toxic (−)	51.0			High	91.4	Toxic (+)	88.8
4-MeO-PCP	Oral	Toxic (+)	92.0			Medium	39.1	Toxic (+)	98.5
	Dermal	Non-Toxic (−)	56.0			N/A	N/A	Toxic (+)	76.8
	Inhalation	Non-Toxic (−)	51.0			High	91.9	Toxic (+)	90.2
3-HO-PCE	Oral	Toxic (+)	77.0			Medium	65.6	Toxic (+)	97.0
	Dermal	Non-Toxic (−)	68.0			N/A	N/A	Toxic (+)	72.7
	Inhalation	Non-Toxic (−)	63.0			High	93.2	Toxic (+)	93.2
3-MeO-PCE	Oral	Toxic (+)	78.0			Medium	61.4	Toxic (+)	98.4
	Dermal	Non-Toxic (−)	62.0			N/A	N/A	Toxic (+)	75.4
	Inhalation	Non-Toxic (−)	62.0			High	94.9	Toxic (+)	90.0
4-MeO-PCE	Oral	Toxic (+)	78.0			Medium	60.2	Toxic (+)	98.9
	Dermal	Non-Toxic (−)	62.0			N/A	N/A	Toxic (+)	78.4
	Inhalation	Non-Toxic (−)	62.0			High	94.7	Toxic (+)	89.1

*N/A, not applicable

Across all tested compounds, STopTox predicted acute oral toxicity with high confidence, ranging from 77.0% for 3-HO-PCE to 92.0% for 3-MeO-PCP and 4-MeO-PCP. In contrast, predictions for dermal and inhalation toxicity were less consistent. Dermal exposure was generally classified as non-toxic, with confidence values between 56.0 and 68.0%. For inhalation, the predictions varied: some analogues (e.g., 3-HO-PCP) were classified as toxic with moderate confidence (50.0%), while others were predicted to be non-toxic. The applicability domain assessment in STopTox indicated “medium” reliability for oral toxicity predictions and “high” reliability for inhalation toxicity predictions. Identified toxicophores in the oral toxicity models often encompassed substantial portions of the molecular scaffold, consistent with the lipophilic arylcyclohexylamine core characteristic of this class. ADMETlab results supported the STopTox oral toxicity predictions, consistently classifying all analogues as toxic with high confidence levels exceeding 88.6% across exposure routes. In admetSAR, all tested substances were classified as acute oral toxicity Category III according to the US EPA system (I: ≤ 50 mg/kg; II: 50–500 mg/kg; III: 500–5000 mg/kg; IV: > 5000 mg/kg), with prediction probabilities between 97.0 and 98.9%. The platform also predicted dermal and inhalation toxicity for all compounds, with probability values ranging from 72.7 to 93.2% for dermal and from 88.8 to 96.4% for inhalation exposure. Eye and skin irritation potential was indicated for most analogues, with prediction probabilities exceeding 70%. While some differences between the platforms were observed, the consensus of results suggests that PCP-type substances have a consistently high likelihood of oral toxicity, with possible risks associated with dermal and inhalation exposure, depending on the specific analogue.

Quantitative in silico methods

Acute toxicity

Acute systemic toxicity remains one of the most fundamental toxicological endpoints, commonly expressed as the median lethal dose (LD₅₀), which denotes the dose required to cause mortality in 50% of exposed test animals (Botham 2004; Pillai et al. 2021). This parameter is widely recognised as a core measure for assessing short-term toxic effects and forms an essential component of compound hazard characterisation (Akhila et al. 2007; Morris-Schaffer and McCoy 2021). In this study, LD₅₀ predictions for PCP-type substances were generated using multiple in silico platforms, including ACD/Labs Percepta, TEST (Consensus method), VEGA, and ProTox 3.0. The results, covering different administration routes and species, are summarised in Table 4.

Table 4.

Results for prediction of acute toxicity for PCP-type substances

Software	Species	Administration route	Compound	LD50, mg/kg bw	Reliability/similarity coefficient
Percepta	Rat	Oral	3-HO-PCP	210	Borderline, RI = 0.47
			3-MeO-PCP	240	Moderate, RI = 0.54
			4-MeO-PCP	240	Moderate, RI = 0.54
			3-HO-PCE	800	Borderline, RI = 0.30
			3-MeO-PCE	410	Moderate, RI = 0.57
			4-MeO-PCE	410	Moderate, RI = 0.57
		Intraperitoneal	3-HO-PCP	100	Not Reliable, RI = 0.21
			3-MeO-PCP	60	Borderline, RI = 0.40
			4-MeO-PCP	60	Borderline, RI = 0.40
			3-HO-PCE	79	Borderline, RI = 0.37
			3-MeO-PCE	56	Borderline, RI = 0.57
			4-MeO-PCE	56	Borderline, RI = 0.57
	Mouse	Oral	3-HO-PCP	350	Borderline, RI = 0.45
			3-MeO-PCP	220	High, RI = 0.76
			4-MeO-PCP	220	High, RI = 0.76
			3-HO-PCE	460	Not reliable, RI = 0.26
			3-MeO-PCE	310	Moderate, RI = 0.50
			4-MeO-PCE	310	Moderate, RI = 0.50
		Intraperitoneal	3-HO-PCP	230	Borderline, RI = 0.40
			3-MeO-PCP	150	Moderate, RI = 0.53
			4-MeO-PCP	150	Moderate, RI = 0.53
			3-HO-PCE	420	Moderate, RI = 0.56
			3-MeO-PCE	150	Borderline, RI = 0.46
			4-MeO-PCE	150	Borderline, RI = 0.46
		Intravenous	3-HO-PCP	34	Moderate, RI = 0.65
			3-MeO-PCP	25	High, RI = 0.78
			4-MeO-PCP	25	High, RI = 0.78
			3-HO-PCE	59	Moderate, RI = 0.53
			3-MeO-PCE	44	Moderate, RI = 0.65
			4-MeO-PCE	44	Moderate, RI = 0.65
		Subcutaneous	3-HO-PCP	200	Moderate, RI = 0.52
			3-MeO-PCP	110	Moderate, RI = 0.64
			4-MeO-PCP	110	Moderate, RI = 0.64
			3-HO-PCE	330	Borderline, RI = 0.35
			3-MeO-PCE	139	Borderline, RI = 0.49
			4-MeO-PCE	139	Borderline, RI = 0.49
TEST (Consensensus)	Rat	Oral	3-HO-PCP	197.42	SC ≥ 0.9
			3-MeO-PCP	405.33	SC ≥ 0.9
			4-MeO-PCP	284.09	SC ≥ 0.9
			3-HO-PCE	628.20	SC ≥ 0.9
			3-MeO-PCE	388.60	SC ≥ 0.9
			4-MeO-PCE	253.50	SC ≥ 0.9
VEGA	Rat	Oral	3-HO-PCP	277.84	Reliable
			3-MeO-PCP	310.14	Reliable
			4-MeO-PCP	311.31	Reliable
			3-HO-PCE	364.98	Reliable
			3-MeO-PCE	367.63	Reliable
			4-MeO-PCE	315.38	Reliable
ProTox 3.0	Rat	Oral	3-HO-PCP	348	SC = 69.26%
			3-MeO-PCP	404	SC = 69.26%
			4-MeO-PCP	400	SC = 69.26%
			3-HO-PCE	348	SC = 69.26%
			3-MeO-PCE	400	SC = 69.26%
			4-MeO-PCE	400	SC = 70.97%

RI, Reliability Index; SC, Similarity coefficient; N/A, not applicable

For the Percepta platform, LD₅₀ predictions for the oral route in rats ranged from 210 mg/kg body weight (bw) for 3-HO-PCP to 800 mg/kg bw for 3-HO-PCE, with Reliability Index (RI) values indicating mostly borderline to moderate reliability (Bureau 2018). In mice, oral LD₅₀ values varied between 220 and 460 mg/kg bw, with higher reliability observed for 3-MeO-PCP and 4-MeO-PCP (RI = 0.76). Across alternative routes, the lowest predicted LD₅₀ values were observed for intravenous administration in mice (25–59 mg/kg bw), suggesting substantially higher acute toxicity by this route. Subcutaneous administration yielded LD₅₀ values between 110 and 330 mg/kg bw. The TEST software (Consensus method) predicted rat oral LD₅₀ values spanning from 197.42 mg/kg bw (3-HO-PCP) to 628.20 mg/kg bw (3-HO-PCE), with similarity coefficients of ≥ 0.9 for all analogues, indicating a high degree of structural match with compounds in the training dataset. VEGA results were consistent with the TEST predictions, producing oral LD₅₀ values between 277.84 and 367.63 mg/kg bw for rats, with all predictions classified as “reliable” based on applicability domain assessment. ProTox 3.0 yielded slightly higher LD₅₀ estimates for the oral route in rats, ranging from 348 mg/kg bw for 3-HO-PCP and 3-HO-PCE to 404–400 mg/kg bw for other analogues, with prediction similarity coefficients between 69.26 and 70.97%. Overall, while inter-model variability was evident, a consistent pattern emerged, with most PCP-type substances predicted to have moderate acute toxicity in the oral route for both rats and mice. Predictions for intravenous administration suggested markedly higher potency, highlighting the need for careful consideration of exposure route in hazard assessments.

Health effects

Health effect predictions for PCP-type substances were generated using the Percepta platform (ACD/Labs 2024b), which estimates the probability of adverse effects across major organ systems, including blood, cardiovascular, gastrointestinal, kidney, liver, and lungs. The detailed results are provided in Table 5.

Table 5.

Results of health effects prediction for PCP-type substances

Compound	Health effect	Probability of health effects, %	Predicted toxicophore(s)
3-HO-PCP	Blood	0.85
	Cardiovascular system	0.81
	Gastrointestinal system	0.53
	Kidney	0.70
	Liver	0.31
	Lungs	0.89
3-MeO-PCP	Blood	0.80
	Cardiovascular system	0.88
	Gastrointestinal system	0.54
	Kidney	0.68
	Liver	0.18
	Lungs	0.93
4-MeO-PCP	Blood	0.81
	Cardiovascular system	0.79
	Gastrointestinal system	0.55
	Kidney	0.68
	Liver	0.19
	Lungs	0.93
3-HO-PCE	Blood	0.47
	Cardiovascular system	0.67
	Gastrointestinal system	0.59
	Kidney	0.09
	Liver	0.37
	Lungs	0.76
3-MeO-PCE	Blood	0.41
	Cardiovascular system	0.67
	Gastrointestinal system	0.82
	Kidney	0.13
	Liver	0.22
	Lungs	0.89
4-MeO-PCE	Blood	0.36
	Cardiovascular system	0.72
	Gastrointestinal system	0.82
	Kidney	0.31
	Liver	0.23
	Lungs	0.89

The predictions are accompanied by the identification of specific atoms or functional groups within the molecular structure, highlighted in red as toxicophores. These structural elements are regarded as key contributors to the calculated parameter values for each organ system. Such analysis offers clearer insight into how the molecular architecture of the compound may drive potential toxicological effects across different organs. For the analysed PCP analogues, the lungs, liver, and blood were most frequently indicated as high-probability targets of adverse effects. The highest lung toxicity probabilities were observed for 3-MeO-PCP and 4-MeO-PCP (0.93), closely followed by 3-HO-PCP (0.89) and the PCE analogues 3-MeO-PCE and 4-MeO-PCE (0.89), with 3-HO-PCE scoring lower at 0.76. Blood toxicity predictions reached their peak for 3-HO-PCP (0.85) and 4-MeO-PCP (0.81), while the lowest value was recorded for 4-MeO-PCE (0.36). Liver effects were also prominent, with values up to 0.91 in some analogues. The gastrointestinal system showed moderate predicted susceptibility, with the highest values for 3-MeO-PCE and 4-MeO-PCE (0.82) and the lowest for 3-HO-PCP (0.53). Cardiovascular toxicity predictions varied from 0.67 for 3-HO-PCE and 3-MeO-PCE to 0.88 for 3-MeO-PCP. The kidneys consistently showed the lowest predicted probabilities, ranging from 0.09 for 3-HO-PCE to 0.70 for 3-HO-PCP. In addition, ADMETlab classified several compounds as hepatotoxic according to its human hepatotoxicity (H-HT) model, with probabilities reaching up to 0.596 for selected analogues. These findings suggest that the lungs, liver, and blood are the most likely target organs for toxic effects of PCP-type substances, whereas the kidneys and cardiovascular system tend to show lower susceptibility. Such organ-specific toxicity profiles may reflect the high lipophilicity and tissue distribution patterns characteristic of arylcyclohexylamine derivatives.

Genotoxicity

The Ames test, a well-established bacterial reverse mutation assay used to assess mutagenicity, remains one of the most valuable tools for the early screening of potential carcinogenicity. A positive result in this assay is often considered an indicator of possible in vivo carcinogenic potential (Zeiger 2019; ACD/Labs 2024c). Its application at early stages of research facilitates the preliminary identification and exclusion of hazardous compounds before more advanced testing, thereby serving as a critical step in toxicological evaluation (Fournier et al. 2023). In the present study, genotoxicity predictions were generated for six PCP-type substances using ACD/Labs Percepta, OCHEM, ADMETlab, admetSAR, and VEGA (Table 6).

Table 6.

Results of genotoxicity—Ames Test effects prediction for PCP-type substances

Software	Compound	Predicted Ames test results	Probability (%)	Structure with toxicophores
Percepta	3-HO-PCP	Non-mutagen	36 (Moderate; RI = 0.66)
	3-MeO-PCP	Non-mutagen	36 (Moderate; RI = 0.63)
	4-MeO-PCP	Non-mutagen	36 (Moderate; RI = 0.63)
	3-HO-PCE	Non-mutagen	21 (Not Reliable; RI = 0.16)
	3-MeO-PCE	Non-mutagen	27 (Borderline; RI = 0.35)
	4-MeO-PCE	Non-mutagen	27 (Borderline; RI = 0.35)
OCHEM	3-HO-PCP	Inactive	91	N/A
	3-MeO-PCP	Inactive	92
	4-MeO-PCP	Inactive	92
	3-HO-PCE	Inactive	94
	3-MeO-PCE	Inactive	94
	4-MeO-PCE	Inactive	94
ADMETlab	3-HO-PCP	Negative	26.4	N/A
	3-MeO-PCP	Negative	31.3
	4-MeO-PCP	Negative	33.2
	3-HO-PCE	Negative	30.2
	3-MeO-PCE	Negative	33.8
	4-MeO-PCE	Negative	31.7
admetSAR	3-HO-PCP	Non-mutagen	13.2	N/A
	3-MeO-PCP	Non-mutagen	22.7
	4-MeO-PCP	Non-mutagen	28.7
	3-HO-PCE	Non-mutagen	11.5
	3-MeO-PCE	Non-mutagen	14.7
	4-MeO-PCE	Non-mutagen	21.2
VEGA	3-HO-PCP	Non-mutagen	Moderate reliability	N/A
	3-MeO-PCP	Non-mutagen	Moderate reliability
	4-MeO-PCP	Non-mutagen	Moderate reliability
	3-HO-PCE	Non-mutagen	Moderate reliability
	3-MeO-PCE	Non-mutagen	Moderate reliability
	4-MeO-PCE	Non-mutagen	Moderate reliability

RI, Reliability Index; AD, applicability domain; N/A, not applicable

Percepta classified all compounds as non-mutagenic, with moderate reliability index (RI) values for PCP analogues (0.63–0.66) and lower reliability for PCE analogues (0.16–0.35). Probability scores for PCP derivatives were consistently 36%, while PCE analogues ranged from 21% (3-HO-PCE) to 27% (3-MeO-PCE, 4-MeO-PCE). OCHEM predictions indicated “inactive” results for all tested compounds, with notably high probability scores (91–94%). Similarly, ADMETlab classified all compounds as negative, with probability values ranging from 26.4% (3-HO-PCP) to 33.8% (3-MeO-PCE). In admetSAR, all compounds were predicted as non-mutagenic, with low probability scores (11.5–28.7%). VEGA also classified all substances as non-mutagenic, assigning moderate reliability to each prediction. Overall, the results indicate a consistent lack of predicted mutagenic potential across all models for both PCP and PCE analogues. While Percepta predictions were accompanied by visualisation of molecular fragments, no high-confidence toxicophores directly linked to mutagenicity were identified. The convergence of outcomes across multiple independent platforms suggests a low likelihood of genotoxicity for these compounds. Nonetheless, the variability in prediction probability and reliability, particularly for PCE analogues, warrants cautious interpretation of these findings.

Eye and skin irritation

The assessment of eye and skin irritation potential is a critical component in evaluating the safety of compounds that may come into direct contact with human tissue in industrial, pharmaceutical, cosmetic, or recreational contexts. Historically, the Draize rabbit test has been employed for more than six decades as a benchmark assay for predicting irritation in human eyes and skin (Vinardell and Mitjans 2008). Contemporary in silico models for predicting such irritation rely on datasets containing over 2000 compounds, enabling comparison with structurally similar substances and thereby improving the reliability of predictions for topically exposed chemicals. In the present study, eye and skin irritation potential for the investigated PCP-type substances was estimated using multiple computational platforms, including ACD/Labs Percepta, ADMETlab, admetSAR, STopTox, and VEGA (ACD/Labs 2024d; e). The outcomes concerning the compound’s potential for eye and skin irritation are articulated as probabilities (%) of severe or moderate irritation by Percepta (Table 7).

Table 7.

Results of eye and skin irritation prediction for PCP-type substances

Software	Compound	Eye irritation	Skin irritation
Percepta	3-HO-PCP	Probability: 88%	Probability: 96%
	3-MeO-PCP	Probability: 46%	Probability: 72%
	4-MeO-PCP	Probability: 46%	Probability: 72%
	3-HO-PCE	Probability: 90%	Probability: 96%
	3-MeO-PCE	Probability: 50%	Probability: 74%
	4-MeO-PCE	Probability: 50%	Probability: 74%
ADMETlab	3-HO-PCP	83.8%	N/A
	3-MeO-PCP	70.6%	N/A
	4-MeO-PCP	65.9%	N/A
	3-HO-PCE	70.4%	N/A
	3-MeO-PCE	55.9%	N/A
	4-MeO-PCE	50%	N/A
admetSAR	3-HO-PCP	3.5%	56.5%
	3-MeO-PCP	8%	53.9%
	4-MeO-PCP	9.6%	58.9%
	3-HO-PCE	12.8%	55.0%
	3-MeO-PCE	18%	57.4%
	4-MeO-PCE	20.2%	59.4%
SToPTox	3-HO-PCP	Non-toxic (−); 57% confidence	Negative (−); out of AD
	3-MeO-PCP	Non-toxic (−); 61% confidence	Negative (−); 60% confidence
	4-MeO-PCP	Non-toxic (−); 61% confidence	Negative (−); 50% confidence
	3-HO-PCE	Non-toxic (−); 54% confidence	Negative (−); 70% confidence
	3-MeO-PCE	Non-toxic (−); 52% confidence	Negative (−); 60% confidence
	4-MeO-PCE	Non-toxic (−); 52% confidence	Negative (−); 50% confidence
VEGA	3-HO-PCP	Irritant (out of AD)	Not irritant (out of AD)
	3-MeO-PCP	Not irritant (out of AD)	Not irritant (out of AD)
	4-MeO-PCP	Irritant (out of AD)	Not irritant (out of AD)
	3-HO-PCE	Irritant (out of AD)	Not irritant (out of AD)
	3-MeO-PCE	Not irritant (out of AD)	Irritant (out of AD)
	4-MeO-PCE	Not irritant (out of AD)	Not irritant (out of AD)

AD, applicability domain; N/A, Not applicable

The Percepta model predicted notably high probabilities of eye and skin irritation for 3-HO-PCP (88 and 96%, respectively) and 3-HO-PCE (90 and 96%, respectively), while several methoxy derivatives exhibited moderate probabilities (46–50% for eye irritation, 72–74% for skin irritation). Predictions from ADMETlab similarly indicated elevated probabilities of eye irritation, particularly for 3-HO-PCP (83.8%) and 3-HO-PCE (70.4%), though skin irritation was not assessed by this platform. Conversely, admetSAR yielded substantially lower probability values for eye irritation (3.5–20.2%) and moderate values for skin irritation (53.9–59.4%). STopTox classified all tested compounds as non-irritant for both eye and skin endpoints, with most predictions falling outside the applicability domain, thereby limiting their reliability. VEGA results were mixed, with some compounds flagged as irritants but with all predictions also outside the applicability domain. These variations underscore the influence of model-specific algorithms, training datasets, and applicability domain constraints on predictive outcomes. The consistently high predictions from Percepta for certain hydroxylated analogues (notably 3-HO-PCP and 3-HO-PCE) may reflect the increased reactivity or polarity associated with the hydroxyl group, which can enhance tissue interaction potential. Such findings are relevant for both clinical risk assessment and forensic toxicology, particularly in the context of non-medical use of PCP-type dissociatives.

Cardiotoxicity

Cardiotoxicity resulting from inhibition of the human Ether-à-go-go-Related Gene (hERG) potassium channel remains one of the most significant contributors to drug candidate attrition, also relevant for psychoactive substances including PCP-type compounds (Zhou et al. 2011). The hERG channel is essential for the repolarisation phase of the cardiac action potential, and its blockade can lead to QT interval prolongation, increasing the risk of severe ventricular arrhythmias (Recanatini et al. 2005). For the investigated set of PCP-type analogues, the half-maximal inhibitory concentration (IC₅₀) values were predicted using the hERG inhibition module implemented in ACD/Labs Percepta, which integrates experimental data from 9383 compounds determined by patch-clamp (manual and automated) and competitive radioligand displacement assays (reference ligands: dofetilide, astemizole, MK-499) (Didziapetris and Lanevskij 2016; ACD/Labs 2024f). The IC50 values ranged from 4.9 µM for 3-MeO-PCP to 12.3 µM for 3-HO-PCP, with intermediate values for other analogues (Table 8). These results place several compounds within a range considered potentially relevant for hERG inhibition. Percepta further classified the probability of hERG inhibition (Ki < 10 µM) as highest for 3-MeO-PCP and 4-MeO-PCP (30%; RI = 0.15, not reliable), while the lowest probabilities were observed for 3-HO-PCE (3%; RI = 0.39, borderline reliability). In contrast, ADMETlab consistently predicted high probabilities of hERG blockade across all tested PCP analogues (68.2–70.8%). The admetSAR platform yielded even higher probability scores, ranging from 62.8% for 3-HO-PCE to 97.5% for 3-MeO-PCP.

Table 8.

Results for cardiotoxicity predictions—hERG IC₅₀ and hERG inhibitors for PCP-type substances

Software	Compound	Cardiotoxicity predictions	Probability (%)
Percepta	3-HO-PCP	hERG half-maximal inhibitory concentration	IC50, 12.3 µM
	3-MeO-PCP		4.9 µM
	4-MeO-PCP		6.4 µM
	3-HO-PCE		10.3 µM
	3-MeO-PCE		6.7 µM
	4-MeO-PCE		6.3 µM
	3-HO-PCP	hERG inhibitor (Ki < 10 µM, patch-clamp)	7%; RI = 0.39 (Borderline)
	3-MeO-PCP		30%; RI = 0.15 (Not Reliable)
	4-MeO-PCP		30%; RI = 0.15 (Not Reliable)
	3-HO-PCE		3%; RI = 0.39 (Borderline)
	3-MeO-PCE		23%; RI = 0.18 (Not Reliable)
	4-MeO-PCE		23%; RI = 0.18 (Not Reliable)
ADMETlab	3-HO-PCP	hERG blockers (Ki < 10 µM)	69.5
	3-MeO-PCP		68.5
	4-MeO-PCP		68.2
	3-HO-PCE		70.8
	3-MeO-PCE		70.4
	4-MeO-PCE		69.2
admetSAR	3-HO-PCP	hERG inhibitor (Ki < 10 µM)	96.3
	3-MeO-PCP		97.5
	4-MeO-PCP		93.4
	3-HO-PCE		62.8
	3-MeO-PCE		77.2
	4-MeO-PCE		83.6

RI, Reliability Index; N/A, Not applicable

In addition to the tabulated results, a heatmap analysis (Fig. 2) illustrates the partial dependence of predicted hERG inhibition on logP and basic pKa values. This visualisation provides insight into the interplay between physicochemical parameters and cardiotoxic potential. Notably, higher lipophilicity was generally associated with increased predicted inhibition probability, particularly for compounds with basic pKa values within the physiologically relevant range. These findings underline the potential cardiotoxic risks within the PCP analogue group, especially for methoxy-substituted derivatives, and highlight the importance of integrating multiple predictive approaches to obtain a comprehensive safety profile. Further methodological details, including comparable reference compounds from built-in databases, are available in the Supporting Information.

Fig. 2.

Results of hERG heatmaps and physico-chemical hERG parameters for PCP-type substances

Endocrine system disruption

This evaluation examined the potential of the investigated PCP-type substances to bind to the estrogen receptor alpha (ER-α), a key component of the endocrine system involved in hormonal regulation. The Estrogen Receptor module estimates the likelihood of reproductive toxicity resulting from compound interaction with ER-α, as such binding can either mimic or antagonise endogenous hormone activity. In vitro, binding affinity is expressed as the logarithm of the Relative Binding Affinity (LogRBA) compared with estradiol. Compounds with LogRBA values above 0 are categorised as strong estrogens, while those below − 3 are considered non-binders. Predictive models were constructed using curated experimental datasets comprising approximately 1500 compounds with measured ER-α binding affinities, sourced from FDA Endocrine Disruptors DB, METI risk assessments, and primary literature (ACD/Labs 2024g). The results (Table 9) indicate that among the tested analogues, 3-HO-PCP exhibited the highest probability (0.88, RI = 0.64) of having LogRBA values greater than − 3, suggesting moderate reliability in indicating possible receptor affinity, whereas its probability of exceeding a LogRBA of 0 was lower at 0.41 (RI = 0.29). The 3-MeO-PCP and 4-MeO-PCP analogues demonstrated lower probabilities (0.42 for LogRBA >  − 3 and 0.11 for LogRBA > 0), both with limited reliability. The PCE derivatives showed variable outcomes: 3-HO-PCE had a modest probability (0.25) for LogRBA >  − 3 and only 0.01 for LogRBA > 0, albeit with moderate reliability for the latter (RI = 0.73). The 3-MeO-PCE and 4-MeO-PCE analogues presented borderline probabilities of 0.30 for LogRBA >  − 3 and 0.26 for LogRBA > 0, each with moderate-to-borderline reliability.

Table 9.

Results for endocrine disruption prediction as probability of Estrogen Receptor Binding for PCP-type substances

Probability of estrogen receptor binding
Compound	LogRBA > − 3	Reliability	LogRBA > 0	Reliability
3-HO-PCP	0.88	Moderate (RI = 0.64)	0.41	Not reliable (RI = 0.29)
3-MeO-PCP	0.42	Not Reliable (RI = 0.17)	0.11	Borderline (RI = 0.31)
4-MeO-PCP	0.42	Not Reliable (RI = 0.17)	0.11	Borderline (RI = 0.31)
3-HO-PCE	0.25	Not Reliable (RI = 0.19)	0.01	Moderate (RI = 0.73)
3-MeO-PCE	0.30	Borderline (RI = 0.33)	0.26	Borderline (RI = 0.38)
4-MeO-PCE	0.30	Borderline (RI = 0.33)	0.26	Borderline (RI = 0.38)

RI, Reliability Index

The observed variation across these PCP-type substances underscores differences in their potential for endocrine system disruption, with phenolic substitutions (e.g., in 3-HO-PCP) showing comparatively higher affinity predictions. Such insights are important for assessing the endocrine-related risks associated with this class of compounds, particularly given the structural diversity and potential for off-target hormonal effects.

Discussion

The findings obtained in this study represent one of the first systematic approaches to the toxicological characterization of PCP and its numerous analogues using an integrated in silico methodology. In the available literature, data on the toxicity of PCP-type substances have remained scattered, fragmentary, and often limited to clinical case reports or heterogeneous experimental studies (Rainey and Crowder 1975; Showalter and Thornton 1977; Burns and Lerner 1978). A comprehensive overview covering multiple toxicological endpoints simultaneously, based on reliable computational tools, has so far been lacking. The use of a multi-platform analysis (STopTox, admetSAR, Percepta, Toxtree, ProTox, ADMETlab, OCHEM, TEST, and VEGA QSAR) provided a broad spectrum of information, increasing the reliability of the predictions through the weight-of-evidence approach and reducing the risk of errors arising from the limitations of individual models (Yang et al. 2018; Hemmerich and Ecker 2020; Rim 2020). Overall interpretation of the results clearly indicates that PCP and its analogues belong to a group of compounds with very high toxic potential, which is consistent with their pharmacological history as well as clinical experience (Holsapple et al. 1982; Journey and Bentley 2025). The LD₅₀ values obtained in computational predictions place PCP-type substances within the range typical for compounds associated with a significant risk of acute poisoning, which explains the numerous reports of sudden health deterioration and even fatalities following relatively low doses (McCarron et al. 1981). Importantly, the observed variability in LD₅₀ values among specific analogues (e.g., TCP, PCE, 3-MeO-PCP, 4-MeO-PCP) highlights substantial differences in their toxicity, which may have practical relevance for the interpretation of forensic toxicology findings and the prediction of poisoning outcomes (Pepe et al. 2024). The applied in silico tools also allowed for organ-specific toxicity assessment, showing a high probability of adverse effects on the central nervous system (CNS), heart, liver, and kidneys (Pradhan 1984; Fnu et al. 2020). In the case of the CNS, the predicted neurotoxic effect is particularly strong, consistent with the pharmacological mechanism of PCP as a non-competitive NMDA receptor antagonist (Olney et al. 1989; Jentsch and Roth 1999; Wang et al. 2005). Blockade of this receptor disrupts the balance between excitation and inhibition within the CNS, explaining the occurrence of psychotic symptoms, hallucinations, agitation, and aggression, while also suggesting a neurodegenerative potential associated with excitotoxicity and neuronal apoptosis. The high risk of hepatotoxicity and nephrotoxicity predicted by Percepta and ADMETlab models is likewise consistent with clinical observations, as cases of elevated liver enzyme activity and impaired kidney function have been described in the literature in patients intoxicated with PCP. Predictive analyses further indicated a strong cardiotoxic potential within this group of compounds, primarily due to the likelihood of hERG channel blockade. This mechanism is associated with the risk of QT interval prolongation and the development of ventricular arrhythmias, which pose a serious life-threatening condition (Journey and Bentley 2025). Symptoms such as tachycardia, hypertension, and rhythm disturbances, frequently reported in the context of PCP poisoning, were therefore confirmed in computational predictions, strengthening their credibility. Conversely, the analysis of genotoxic potential revealed a limited yet non-negligible risk of mutagenicity in certain PCP analogues, which corresponds with the presence of specific structural fragments (including aromatic systems conjugated with a piperidine ring) recognized as structural toxic alerts. Although no evidence of PCP carcinogenicity has been reported to date, these results emphasize the need for further studies employing in vitro methods. The results concerning irritant effects and the risk of contact exposure also warrant attention. Predictions indicated that PCP and its analogues may act as irritants to the skin and ocular mucosa, which has important implications in situations of accidental exposure, e.g., in forensic laboratories, during substance seizures, or while handling evidence materials. This highlights the necessity of implementing proper protective procedures and safety measures. The obtained data clearly demonstrate that PCP-type substances should be classified as compounds with high acute toxicity and a broad spectrum of potential adverse effects. In silico analyses reveal that the risks associated with these substances extend beyond their psychoactive action and include severe somatic consequences, neurotoxic, cardiotoxic, and hepatotoxic. These conclusions reinforce the view that PCP and its analogues pose a particular threat to public health, while the application of computational methods fills an important gap resulting from the limited availability of experimental studies (Yang et al. 2018; Hemmerich and Ecker 2020; Rim 2020). In a broader perspective, the findings confirm the usefulness of in silico methods as tools for preliminary toxicological risk assessment of psychoactive substances with unregulated legal status (Table 10).

Table 10.

Strengths and limitations of each used in silico tool

Method	Strengths	Limitations	Reliability for Bromo-DragonFLY
STopTox	Utilises advanced machine learning algorithms; well-curated database; broad range of toxicity predictions; versatile for multiple endpoints	Reliability varies by endpoint and training data diversity	High reliability for initial toxicity screening, especially acute and organ toxicity. Cross-validation with other models is recommended for a comprehensive profile
AdmetSAR	Covers a wide range of ADMET properties; robust validation methodologies	Lower accuracy for less common endpoints due to limited training data; dermal and inhalation toxicity issues	Effective for broad ADMET profiling, high confidence in absorption and distribution properties. Cross-validation is needed for comprehensive risk assessment
Percepta	Extensive datasets; sophisticated QSAR models; predictive solid power for mutagenicity and carcinogenicity	Predictive accuracy varies by endpoint and compound class; low-reliability index for some predictions	Reliable for mutagenicity and carcinogenicity predictions. Positive predictions warrant further experimental validation. Cross-referencing is needed for acute toxicity
ProTox	Designed specifically for predicting acute toxicity; good predictive performance backed by detailed validation studies	Focuses primarily on oral toxicity; may not account for other exposure routes (e.g., inhalation, dermal)	High reliability for acute oral toxicity predictions. Additional assessments with complementary models are needed for other exposure routes
PreADMET	A comprehensive range of ADMET predictions and validated algorithms	Limited training data for some predictions does not provide a probability of results	Useful for genotoxicity and cardiotoxicity assessment. Lack of probability for results reduces certainty in reliability
ADMETlab	Integrates multiple data sources; advanced predictive models; reliable ADMET and toxicity assessments	Variable predictive accuracy depends on training data quality and diversity	Unreliable for comprehensive ADMET profiling. Low reliability for acute toxicity, genotoxicity, and cardiotoxicity predictions
OCHEM	Solid predictive performance; extensive datasets; machine learning models; ability to build custom QSAR models	Predictive accuracy varies based on an endpoint and data quality used for training	Provides unreliable predictions for genotoxicity (Ames test). Cross-referencing with other models is recommended for specific toxicological endpoints
TEST	Detailed toxicity predictions focused on environmental impact; equipped with multiple endpoints and estimation methodologies	Strong for environmental toxicity but may not comprehensively cover all human toxicity endpoints	Reliable for acute toxicity with high similarity to Bromo-DragonFLY. Should supplement predictions with other models for a complete human toxicity profile
VEGA	Utilises validated QSAR models and comprehensive datasets	Performance varies based on endpoint and chemical diversity within the training set	Unreliable for overall toxicological predictions. Could not predict acute toxicity or genotoxicity; eye and skin irritation predictions were outside the application domain. Validation with other models is recommended

Interpretation of the obtained results requires juxtaposing them with the knowledge currently available on the toxicity of PCP and its analogues, as well as comparing them with findings from studies conducted on other groups of NPS. Such an analysis allows for evaluating both the consistency of in silico predictions with the literature and the identification of novel aspects of toxicity that have not yet been fully described. With regard to the classical literature data on PCP, it should be emphasized that the symptoms of acute intoxication cover a broad spectrum of neurological and somatic disturbances. Reported manifestations include disorientation, psychomotor agitation, hallucinations, and psychoses, but also tachycardia, hypertension, cardiac arrhythmias, hyperthermia, rhabdomyolysis, and acute renal failure (Burns et al. 1975; Gahlinger 2004; Bey and Patel 2007). In this context, the high risk of neurotoxicity and cardiotoxicity predicted by in silico tools can be regarded as a confirmation of clinical observations. Predictions concerning possible liver and kidney injury are likewise reflected in published cases, where elevated liver enzyme activity and impaired renal function were observed in patients intoxicated with PCP (Grossenbacher et al. 2019). The consistency of these results with clinical reports highlights the considerable utility of the applied approach. It should be noted, however, that certain in silico analyses allow for the recognition of toxicity aspects of PCP-type substances that have not been widely discussed to date. This refers particularly to the genotoxic potential indicated by Toxtree and VEGA QSAR. Although the literature provides no unequivocal evidence for the mutagenicity or carcinogenicity of PCP, the presence of specific structural “toxic alerts” suggests that this risk cannot be entirely dismissed (Kirkland et al. 2005; Benigni and Bossa 2011). This therefore represents an area requiring further in vitro investigation, with the obtained predictions serving as a useful starting point for designing such experiments. When comparing the present results with studies on other NPS, it is instructive to refer to earlier work on the toxicity of Bromo-DragonFLY (Noga and Jurowski 2024). For this compound, a phenethylamine-based psychedelic, high acute toxicity, a significant risk of cardiotoxicity related to hERG channel inhibition, and the likelihood of multi-organ damage were also demonstrated. The parallels with PCP-type substances are thus evident, underscoring common health risks associated with the abuse of potent NPS, regardless of their pharmacological class. At the same time, differences in toxicological profiles are equally important: in PCP, the dominant mechanism involves NMDA receptor antagonism and the associated risk of neurotoxicity, whereas in Bromo-DragonFLY, the primary role is played by 5-HT2A receptor agonism, leading to prolonged psychedelic effects and a pronounced risk of vasoconstriction. These observations emphasize that while in silico tools highlight recurring hazards (acute toxicity, cardiovascular risk, multi-organ adverse effects), the detailed toxicological profile remains unique for each class of compounds and reflects their receptor-based mechanism of action. Equally noteworthy is the comparison of results for PCP-type substances with analyses performed for ACP-105 (Fijałkowska and Jurowski 2025). ACP-105, a selective androgen receptor modulator (SARM), was characterized by predicted acute toxicity of moderate severity and adverse effects primarily involving the liver and cardiovascular system, with a relatively low neurotoxic potential. In contrast, the toxicity of PCP-type substances proved to be far more severe and encompassed a wider range of mechanisms. While ACP-105 fits into a toxicity profile characteristic of doping agents, PCP and its analogues emerge as substances with an exceptionally unfavorable risk profile, in which neurotoxic and psychotic effects play a central role. In a broader comparative perspective, it can be concluded that PCP-type substances represent a distinctive category within the spectrum of NPS. Unlike synthetic cannabinoids or cathinones, for which the main toxic risks are associated with the cardiovascular system as well as the psychostimulant impact on the CNS (Prosser and Nelson 2012; Simmler et al. 2013; Roque-Bravo et al. 2023), PCP and its analogues combine the potential to induce both acute psychiatric disturbances and serious somatic complications. The results obtained in the present work therefore confirm existing literature observations, while at the same time significantly expanding knowledge on the possible mechanisms and spectrum of toxic effects. In contrast to previous studies on PCP, its analogues, and other NPS, there are numerous key conclusions. First, in silico predictions reflect well-known clinical and experimental data, thereby increasing their credibility. Second, they highlight new, potentially important aspects of toxicity (e.g., genotoxic risk) that have not been widely investigated so far. Third, the toxicological profile of PCP-type substances appears to be particularly unfavorable compared with other NPS groups, which renders them compounds of exceptionally high potential risk to public health.

The results obtained from the present analyses have important implications for both clinical practice and forensic medicine. PCP and its analogues are substances that, since their emergence on the drug market, have raised particular concern among physicians and toxicologists due to the unpredictable course of intoxications as well as the considerable therapeutic challenges associated with managing acute cases (Munch 1974; Dominici et al. 2014). The in silico predictions reinforce this picture, providing systematized data on potential mechanisms of toxicity and the organ systems most vulnerable to damage. From a clinical perspective, the key point is that PCP-type substances are characterized by a simultaneous risk of severe neuropsychiatric disturbances and serious somatic complications. Acute intoxications most often manifest as psychomotor agitation, disorientation, aggressive behaviors, and hallucinations, which may rapidly progress to psychotic states requiring psychiatric hospitalization (McCarron et al. 1981; Dominici et al. 2014). The strong neurotoxic potential predicted by in silico models provides a scientific confirmation of the symptoms observed in clinical practice, while also indicating the possibility of long-term neurological sequelae such as cognitive deficits, memory impairments, or the development of depressive and anxiety disorders (Wang et al. 2022). Somatic implications are equally important. Predictions of a high probability of cardiotoxicity, mainly associated with hERG channel blockade, imply that patients intoxicated with PCP require close cardiac monitoring, including observation of the QT interval and heart rhythm. Cases of sudden death related to arrhythmias, reported in the literature, confirm the reality of this threat. Similarly, hepatotoxicity and nephrotoxicity suggest the need to monitor blood biochemical parameters and kidney function in intoxicated individuals, an approach that has not always been a standard clinical practice in acute NPS poisoning cases (Cogen et al. 1978; Pestaner and Southall 2003; Akmal et al. 2008). A significant challenge in clinical practice remains the lack of a specific antidote for PCP and its analogues. Treatment is solely symptomatic and supportive, encompassing, among others, agitation control with benzodiazepines, management of cardiological complications, and intensive monitoring of vital functions (Nelson et al. 2019). The findings of this study may aid in prioritizing therapeutic interventions, emphasizing the necessity of simultaneously addressing neurological and cardiovascular disturbances. Equally important are the implications for forensic medicine. In forensic toxicology practice, the emergence of new PCP analogues represents a considerable challenge, as experimental data on their toxicity are often unavailable or limited to isolated case reports (Mercolini 2019). In silico predictions may fill this gap, providing information essential for interpreting chemical and toxicological findings. Knowledge of the predicted toxicity profile of individual analogues can facilitate the reconstruction of the intoxication course, allowing for more precise conclusions regarding the cause of death (Baselt 2020). The relevance of the obtained results should also be highlighted in the context of legislative and regulatory actions. PCP-type substances, like other NPS, often appear on the market in the form of new analogues not yet subjected to legal control. The lack of unambiguous experimental data makes rapid regulatory decision-making difficult in such cases. In silico prediction outcomes may serve as scientific justification for placing new analogues under legal control, supporting law enforcement agencies and institutions responsible for public health safety (WHO 2020; EMCDDA 2022). The clinical and forensic-toxicological significance of the results is therefore multidimensional. On the one hand, they provide a tool to support the diagnosis and management of acute intoxications, emphasizing the need to monitor specific organ systems. On the other, they represent a valuable contribution to the interpretation of forensic toxicology cases and the regulatory process, helping to better prepare healthcare systems and the justice sector for the emergence of new PCP analogues.

Interpretation of the in silico predictions indicates that the toxicity of PCP-type substances results from multifactorial biological mechanisms involving both classical neurotoxic pathways and additional, less well-recognized routes of action. The most obvious and best-known mechanism is NMDA receptor antagonism, which underlies the principal psychoactive effects of PCP and its analogues. Blockade of this receptor disrupts glutamatergic neurotransmission, disturbing the balance between excitation and inhibition in the CNS (Ikonomidou et al. 1999; Wang et al. 2005). The consequence is the manifestation of psychotic symptoms, hallucinations, and marked psychomotor agitation, findings corroborated both by clinical observations and by in silico results pointing to a high risk of neurotoxicity (Jentsch and Roth 1999). However, the results suggest that the neurotoxicity of PCP-type substances is not confined to receptor-mediated effects. Computational models predict the potential induction of oxidative stress, mitochondrial dysfunction, and activation of apoptotic pathways (Jevtić et al. 2016; Jurič et al. 2021). Literature reports indicate that chronic NMDA blockade may lead to secondary excitotoxicity, associated with dysregulation of calcium homeostasis and excessive activation of AMPA/kainate receptors. The result is neuronal damage in the hippocampus and cerebral cortex, which may explain the long-term cognitive deficits and emotional disturbances observed in PCP users (Ikonomidou et al. 1999; Jevtić et al. 2016). Another important mechanism revealed by the predictions is the high cardiotoxic potential, particularly linked to blockade of the hERG potassium channel. Inhibition of this channel is recognized as one of the main molecular mechanisms of drug-induced arrhythmogenesis, leading to QT interval prolongation, torsade de pointes, and sudden cardiac death (Sanguinetti and Tristani-Firouzi 2006; Kannankeril and Roden 2007). For PCP-type substances, predictions consistently indicate a significant risk of this effect, which is further supported by clinical case reports describing severe arrhythmias in PCP-intoxicated patients. Moreover, the structural lipophilicity of PCP analogues favors their accumulation in cell membranes, which may additionally modulate ion conductance and exacerbate cardiotoxicity (Bey and Patel 2007). In silico analyses also suggest a considerable risk of hepatotoxicity and nephrotoxicity. These mechanisms may be linked to the oxidative metabolism of PCP in the liver, primarily mediated by cytochrome P450. The resulting reactive metabolites can generate oxidative stress, damage mitochondrial membranes, and initiate necrotic or apoptotic processes in hepatocytes (Shebley et al. 2006; Driscoll et al. 2007). Similar mechanisms may apply to kidney cells, where PCP and its analogues are secreted and accumulate, leading to tubular injury (Cogen et al. 1978). The predicted genotoxic properties of certain PCP analogues also deserve attention. Although the literature lacks clear evidence of carcinogenicity, the presence of structural motifs considered toxicological alerts (e.g., conjugated aromatic systems) indicates a potential for DNA interactions or the induction of oxidative stress leading to mutations. While this mechanism requires further experimental validation, it may represent an important toxicological aspect, particularly in the context of chronic exposure (Benigni and Bossa 2011; Manganelli et al. 2018). The mechanisms of toxicity of PCP-type substances are therefore multifaceted. They encompass both primary receptor interactions (NMDA antagonism) and secondary consequences such as oxidative stress, mitochondrial injury, and apoptosis. In addition, cardiotoxic mechanisms (hERG blockade), hepatotoxic mechanisms (generation of reactive metabolites by CYP450), and potential genotoxic effects play a role (Sanguinetti and Tristani-Firouzi 2006; Shebley et al. 2006; Wang and Johnson 2007). Such a wide spectrum of possible toxic mechanisms underscores that PCP and its analogues should be regarded as compounds with highly complex, multi-organ actions that significantly increase clinical risk and complicate effective treatment of acute poisonings.

Although the obtained results provide a significant contribution to expanding knowledge about the toxicity of PCP-type substances, it must be emphasized that the analyses are subject to several limitations, arising both from the nature of the applied in silico methods and from the general lack of experimental data for this class of compounds. The first limitation is the dependence of computational models on the databases on which they were trained. Tools such as STopTox, admetSAR, ProTox-II, and VEGA QSAR rely on historical toxicological data, covering mainly well-characterized and extensively studied substances (Benigni et al. 2007; Cheng et al. 2012; Banerjee et al. 2024). For PCP and its analogues, the amount of experimental data is limited, which may affect prediction accuracy. This means that predictions for some novel PCP analogues, particularly the less common ones, may be less reliable and should be considered preliminary hypotheses requiring further validation. The second limitation is that most of the applied models focus on assessing acute toxicity or specific organ-level effects, while reliable tools for evaluating chronic toxicity or long-term consequences of exposure are lacking (Myatt et al. 2018). In practice, this means that possible neurodegenerative, carcinogenic, or reproductive effects of PCP-type substances remain speculative. Moreover, in silico models rarely account for complex pharmacokinetic processes such as interindividual variability in metabolism, enzyme polymorphisms, or interactions with other psychoactive substances. This is particularly relevant since PCP is frequently used in combination with other drugs (e.g., alcohol, opioids, or cannabinoids), significantly altering its toxicological profile (Carvalho et al. 2012). The third limitation is the inability to fully capture the complexity of the biological response to exposure. Computational models predict specific endpoints, but they do not reflect the physiological context, the dynamics of poisoning, interactions between organ systems, or the influence of environmental factors. As a result, although predictions provide valuable guidance, they cannot replace in vitro and in vivo studies (Valerio 2009). Despite these limitations, the results open several avenues for future research. First, it is necessary to validate the obtained predictions in laboratory experiments. This is especially true for neurotoxicity and cardiotoxicity, which appear to be critical for the risk profile of PCP-type substances. In vitro studies using neuronal and cardiomyocyte models may help confirm the toxic mechanisms suggested by computational approaches (Ekins et al. 2007). Second, studies on the metabolism of PCP and its analogues are essential. While in silico methods highlighted the potential risk of hepatotoxicity and the formation of reactive metabolites, full understanding of this process requires enzymatic and animal studies. Identification of metabolites is particularly important from a forensic toxicology perspective, since these compounds are often detected in biological samples from intoxicated individuals (de Graaf et al. 2005). Third, future research should address chronic and long-term toxicity. Clinically observed cognitive and psychiatric disturbances following PCP intoxication suggest that prolonged use of these substances may lead to persistent changes in the CNS. Neuroimaging, neurotoxicological investigations, and animal models may help confirm these effects and link them to in silico predictions (Le Cozannet et al. 2010; Nomura et al. 2016). Fourth, further research should focus on mixture toxicity. In practice, PCP-type substances are rarely consumed in isolation, they are more often used in combination with other drugs. Future work should therefore employ computational models accounting for interactions between active substances, complemented by experimental studies in this area (Bopp et al. 2018). The limitations of this study stem primarily from the nature of in silico methods and the scarcity of experimental data on PCP-type substances. Nevertheless, the findings provide a solid starting point for further analyses, highlighting key areas requiring laboratory and clinical validation. Future research should prioritize validation of toxic mechanisms, investigation of metabolism, and analysis of long-term exposure effects, which will allow a more comprehensive understanding of the risks associated with this particular group of psychoactive substances.

The obtained results are of considerable importance not only for science but also for clinical practice, forensic toxicology, and regulatory fields. In a broader perspective, this study confirms that the application of in silico methods represents an effective and valuable tool for the preliminary toxicological risk assessment of psychoactive substances, particularly the highly problematic group of PCP-type substances (Cronin and Madden 2010; Raies and Bajic 2016). First, it is essential to emphasize the relevance of the predictions for public health. Although PCP and its analogues remain under legal control in many countries, they continue to appear on the illicit drug market, often in the form of new derivatives that are not yet covered by legislation (Meyer 2016; UNODC 2021; EMCDDA 2022). The high risk of neurotoxicity, cardiotoxicity, and hepatotoxicity identified in this study underscores that these compounds should be considered particularly dangerous (Wagmann et al. 2020; Ferrari Júnior et al. 2022). Understanding their toxicological profile therefore carries not only scientific but also practical significance, it enables the development of more effective medical procedures and supports preventive public health measures (Giorgetti et al. 2021). Second, this work highlights the value of in silico methods in forensic and regulatory toxicology. The lack of experimental data for novel PCP analogues makes the interpretation of toxicological findings in cases of intoxication or fatalities challenging. Computational tools provide predictions that can assist forensic experts in reconstructing the course of intoxication and interpreting its consequences (Maurer 2010; Vilar et al. 2014). Furthermore, these data can serve as a basis for legislative decisions regarding the legal control of new analogues, which is crucial for limiting their availability (Patlewicz et al. 2008; OECD 2014, 2023). Third, the findings point to a broader perspective on the use of in silico methods in the study of NPS. In recent years, the number of NPS emerging on the market has grown enormously, and their structural diversity makes experimental testing of each new compound impractical (Vilar et al. 2014; Raies and Bajic 2016). Computational approaches, utilizing existing databases and predictive models, allow for rapid and relatively inexpensive preliminary insights into the toxicological profile of such compounds. The weight-of-evidence approach applied in this study further enhances prediction reliability by reducing the risk of errors associated with individual models (Patlewicz et al. 2008). Such strategies may become a standard in future toxicological risk assessment of NPS. Fourth, this study opens perspectives for interdisciplinary research. Integrating in silico approaches with in vitro and in vivo studies, as well as with neuroimaging and pharmacogenomic analyses, can lead to a more comprehensive understanding of PCP-type substance toxicology (Wagmann et al. 2020; Sahai and Opacka-Juffry 2021). Particularly valuable is the combination of predictions with studies on metabolism and pharmacokinetics, which may enable better forecasting of both acute and chronic effects of exposure (Giorgetti et al. 2021). Finally, the social and educational dimension of the results should be emphasized. Information on the toxicity of PCP-type substances can be used not only by physicians and toxicologists but also by institutions engaged in addiction prevention and health education (EMCDDA 2022; Ferrari Júnior et al. 2022). Highlighting the real and serious risks associated with the use of these substances can support preventive efforts and thereby help reduce the scale of the drug problem. The significance of this research extends beyond the toxicological assessment of PCP and its analogues. It serves as an example of applying modern computational tools in toxicology, demonstrating their potential in the study of NPS. The findings carry practical implications, clinical, forensic, and regulatory, and open broad perspectives for further interdisciplinary investigations. PCP-type substances should be considered among the most problematic groups of NPS, and their toxicological characterization should be treated as a priority for both science and public health.

Conclusions

The present investigation provides a systematic toxicological characterization of PCP and its analogues using an integrated in silico strategy. The application of multiple predictive platforms enabled the assessment of acute systemic toxicity, organ-specific adverse effects, and mechanistic endpoints within a weight-of-evidence framework. The results obtained consistently support the classification of PCP-type substances as compounds of high toxicological concern, corroborating historical clinical evidence while also highlighting previously underexplored toxicological aspects. Predictions of acute lethality indicated LD₅₀ values within the range associated with a high probability of severe intoxication, in agreement with published clinical and experimental data. Variability in predicted toxicity among structural analogues, particularly hydroxylated and methoxylated derivatives, emphasizes that relatively small structural modifications can substantially alter toxic potency. Such findings are of direct importance for forensic toxicology, where interpretation of intoxications involving novel analogues frequently lacks experimental reference data. In silico analyses consistently demonstrated a strong probability of neurotoxicity, in line with the established pharmacological mechanism of NMDA receptor antagonism, and further suggested secondary mechanisms involving excitotoxicity, oxidative stress, and neuronal apoptosis. Cardiotoxicity emerged as a critical endpoint, with multiple models predicting a significant risk of hERG channel inhibition and arrhythmogenesis. Hepatotoxic and nephrotoxic potential was also evident, consistent with the metabolic activation of arylcyclohexylamines and clinical reports of impaired hepatic and renal function. Although genotoxicity predictions indicated a generally low risk, the presence of structural alerts in certain analogues suggests that this endpoint should not be disregarded. The identification of pronounced skin and eye irritation potential, particularly for hydroxylated compounds, underscores additional occupational and forensic hazards. The toxicological profile delineated in this study has important implications for several fields. From a clinical perspective, the findings highlight the necessity of comprehensive monitoring of neurological, cardiovascular, hepatic, and renal functions in cases of PCP intoxication. Forensic practice may benefit from predicted toxicity profiles in the interpretation of toxicological findings when experimental data are unavailable. From a regulatory standpoint, the results provide a scientific rationale for the proactive control of emerging PCP analogues, which continue to appear on the illicit market. PCP and structurally related analogues constitute a category of NPS with an exceptionally unfavorable toxicological profile, combining profound neuropsychiatric disturbances with severe somatic complications. The results demonstrate the utility of in silico methodologies as powerful tools for bridging critical knowledge gaps, anticipating the risks posed by novel psychoactive compounds, and supporting evidence-based decision-making in clinical, forensic, and regulatory toxicology.

Abbreviations

ACC

Accuracy

AD

Applicability domain

ADMET

Absorption, distribution, metabolism, excretion, and toxicity

AUC

Area under the curve

bw

Body weight

CNS

Central nervous system

EMCDDA

European Monitoring Centre for Drugs and Drug Addiction

ER-α

Estrogen receptor alpha

hERG

Human Ether-à-go-go-related gene

IC₅₀

Half-maximal inhibitory concentration

LD₅₀

Median lethal dose

logP

Partition coefficient (octanol/water)

MAE

Mean absolute error

MCC

Matthews correlation coefficient

MCS

Maximum common substructure

NMDA

N-Methyl-D-aspartate

NPS

New psychoactive substances

OCHEM

Online chemical database and modeling environment

OECD

Organisation for economic co-operation and development

PCE

Eticyclidine

PCP

Phencyclidine

pKa

Acid dissociation constant

PCC

Pearson correlation coefficient

QSPR

Quantitative structure–property relationship

QSAR

Quantitative structure–activity relationship

RI

Reliability index

RMSE

Root mean square error

ROC

Receiver operating characteristic

STopTox

Systemic and topical chemical toxicity

SMILES

Simplified molecular input line entry system

TCP

Tenocyclidine

TEST

Toxicity estimation software tool

US EPA

United States Environmental Protection Agency

VEGA QSAR

Virtual models for the evaluation of chemical properties

Author contributions

Maciej Noga data curation, formal analysis, investigation, methodology, visualisation, writing-original draft, writing-review and editing; Kamil Jurowski conceptualisation, data curation, formal analysis, investigation, supervision, visualisation, writing-original draft, writing-review and editing.

Funding

Not applicable.

Data availability

All data generated or analyzed during this study are included in this published article.

Declarations

Conflict of interest

The Authors declare no competing financial or non-financial interest.