DCPM-ADMET: fusion of dual-component pre-trained model and molecular fingerprints to enhance drug ADMET properties prediction Authors: Leilei Zhang, Yuchen Zeng, Yue Qi, Kaili Jiang, Xiaofei Zhou, Lu Liang, Jianping Lin Published: 2026 ABSTRACT AbstractThe absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of drugs are critical to their efficacy and safety in clinical trials; however, traditional machine learning methods have limited generalization ability in ADMET prediction due to insufficient data. To address this issue, we developed DCPM-ADMET, an innovative pre-trained model with higher accuracy, whose architecture employs a two-component system, an XLNet-based module for capturing the deep semantics of molecular sequences, and a specialized RNN-based component (GRU) designed to encode global molecular property descriptors into high-dimensional latent representations for robust property extraction. By further incorporating ECFP fingerprints to capture local substructures, the model outperforms traditional methods and most pre-trained models in prediction accuracy on multiple benchmark datasets for molecular properties; additionally, we fine-tuned it on a self-constructed database containing 465,470 entries covering 97 ADMET properties, and by integrating these 97 prediction models and 36 computational properties, we further developed a free online ADMET prediction tool with 133 endpoints (available at http://admet.bioai-global.com/), which is designed to assist researchers in conducting comprehensive molecular ADMET predictions. Scientific contribution The development of DCPM-ADMET provides a robust and effective framework for molecular property prediction in computational pharmacology. Our architecture innovatively employs a dual-component system: an XLNet-based module for deep capture of molecular sequence semantics, a multi-task GRU module for joint SMILES translation and physicochemical property descriptor regression. Furthermore, we incorporate ECFP fingerprints to achieve exhaustive substructural feature encoding. Leveraging this multimodal fusion strategy, DCPM-ADMET exhibits superior predictive performance across diverse molecular property benchmark datasets, outperforming both traditional fingerprinting methods and state-of-the-art pre-trained models. Subsequently, we fine-tuned the model on a self-developed proprietary database—currently the largest of its category—comprising 465,470 entries that cover 97 ADMET endpoints (including 43 regression tasks, the highest number reported to date). By integrating the 97 resultant prediction models with 36 computed physicochemical properties, we have developed and made publicly available a free, high-throughput online ADMET prediction tool with 133 endpoints which is poised to serve as a novel and valuable alternative for guiding early-stage drug discovery and safety assessment.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13321-026-01244-z.Keywords: Pretraining, ADMET, XLNet, RNN, Molecular fingerprints Introduction Drug development is a complex, time-consuming, and expensive process. Traditional drug development methods often require years and substantial financial investment to bring a new drug to market [1]. The pivotal success of clinical trials significantly shapes the course of drug development, holding paramount importance for pharmaceutical investors and patients alike. Nevertheless, about 40–60% of drug candidates succumb to unfavorable ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) characteristics during clinical trials [2]. By anticipating these ADMET properties, researchers can pinpoint and eradicate unfavorable pharmacokinetic traits in prospective compounds, expediting the developmental process while curbing time and financial expenses. This proactive strategy also facilitates the early recognition of potentially efficacious compounds during screening, prioritizing them for subsequent advancements and refining the selection of candidates with the highest likelihood of success as novel medications. Furthermore, it underpins the determination of optimal dosing, optimal routes of administration, and minimizes the threat of adverse reactions, enhancing both the safety and efficacy of drug candidates [3, 4]. In recent years, the landscape of computational simulation software for ADMET property prediction has undergone remarkable advancements, achieving notable strides in enhancing prediction accuracy [5]. To further refine this intricate endeavor, artificial intelligence technologies have emerged as pivotal players in diverse ADMET computational frameworks, yielding an array of outstanding prediction tools, including ADMETlab, admetSAR, Interpretable-ADMET, HelixADMET, among others [6–10]. Nevertheless, the application of current models confronts myriad challenges, encompassing poor data quality, limited generalization capabilities, interpretability gaps, data scarcity and uncertainties, thereby restricting model accuracy and reliability. These hurdles underscore the urgency for continued research and optimization. The algorithmic evolution in drug ADMET prediction has witnessed profound transitions, evolving from traditional machine learning through deep learning to the advent of pre-trained models. Initially, researchers harnessed traditional methods like linear regression, support vector machines, and random forests, relying on manual feature engineering to capture molecular attributes [11]. However, these approaches often grappled with the intricacies of ADMET prediction. As computational prowess and data availability soared, deep learning techniques emerged, automating feature learning and demonstrating exceptional performance in ADMET prediction tasks [12]. Nevertheless, their hunger for massive datasets and challenges in generalization amidst data scarcity remain obstacles. Recently, pre-trained models have emerged as promising avenues for ADMET prediction. By pre-training on extensive datasets, these models acquire rich feature representations that are subsequently fine-tuned for specific ADMET tasks, significantly bolstering generalization and prediction precision. They not only adapt seamlessly to diverse ADMET prediction scenarios but also address data scarcity and uncertainty more adeptly, marking the cutting edge of algorithmic developments in this field [13]. In the present work, the model underwent an extensive unsupervised pre-training regimen utilizing a vast small molecule dataset. The pre-training employs a dual-component architecture to discern intricate and abstract feature representations from different aspects, which we refer to as the Dual-component Pre-trained Model (DCPM). One component, based on the XLNet architecture, is used to learn the semantic differences in molecular sequences, while the other, utilizing an RNN-based architecture, captures complex representations at the level of molecular properties. Subsequently, DCPM was trained and evaluated across a broad spectrum of downstream tasks, encompassing various challenges within the MoleculeNet benchmark dataset, such as predicting Caco-2 Permeability, BBB Penetration, and identifying CYP450 1A2 Inhibitors. Its competitive performance and robust generalization capabilities across these tasks validated its accuracy and reliability. Building upon this success, we further extended DCPM's reach by subjecting it to rigorous training and testing within 97 downstream ADMET tasks. We conducted a comparison with the traditional molecular fingerprint ECFP and our method demonstrated superiority in nearly 69.07% of the tasks. In addition, compared with existing tools, our model covers the largest number of ADMET regression tasks, which are designed to provide more accurate prediction results. Simultaneously, our overall dataset is also the largest, containing over 460,000 entries, thereby laying a solid foundation for a more holistic and comprehensive approach to drug ADMET prediction. Materials and methods Data preparation Pre-training data The DCPM-ADMET training process incorporates diverse datasets, encompassing chemical molecule collections for pre-training and targeted ADMET datasets for constructing predictive models. The pre-training phase leverages PubChem [14], a comprehensive bioactivity database housing millions of compound records spanning pharmaceuticals, natural products, and chemicals. From this repository, we extracted SMILES and InChI representations of chemical molecules, subjecting them to a rigorous data preparation pipeline: Data standardization: utilizing RDKit (version 2023.9.6), SMILES strings underwent standardization procedures, including bond order canonicalization and removal of redundant hydrogen atoms. Data standardization: utilizing RDKit (version 2023.9.6), SMILES strings underwent standardization procedures, including bond order canonicalization and removal of redundant hydrogen atoms. Missing & duplicate value filtering: molecules that failed standardization were deemed missing values, the missing and duplicate values were removed to ensure data uniqueness and validity. Missing & duplicate value filtering: molecules that failed standardization were deemed missing values, the missing and duplicate values were removed to ensure data uniqueness and validity. Property descriptor calculation: based on the standardized SMILES, we employed RDKit to compute nine crucial physicochemical properties for each molecule, serving as molecular property descriptors. These encompassed logP, maximal and minimal partial charges, valence electron count, hydrogen bond donor and acceptor numbers, Balaban's J value, molar refractivity, and topological polar surface area. Property descriptor calculation: based on the standardized SMILES, we employed RDKit to compute nine crucial physicochemical properties for each molecule, serving as molecular property descriptors. These encompassed logP, maximal and minimal partial charges, valence electron count, hydrogen bond donor and acceptor numbers, Balaban's J value, molar refractivity, and topological polar surface area. These processes yielded 111,295,885 data points for pre-training. ADMET datasets For the ADMET property datasets, we aggregated resources from multiple experimental compound databases, notably PubChem, ChEMBL [15] and TDC [16], and implemented a meticulous data processing strategy: Data standardization: SMILES representations were standardized using RDKit, and harmonizing data across datasets into consistent units like pEC50 and mol/L. Data standardization: SMILES representations were standardized using RDKit, and harmonizing data across datasets into consistent units like pEC50 and mol/L. Data cleaning & merging: a rigorous comparison of entries across similar datasets from different sources led to the removal of conflicting samples in classification sets and unreliable samples in regression sets exhibiting discrepancies exceeding an order of magnitude. Subsequently, datasets were merged and further cleaned. Data cleaning & merging: a rigorous comparison of entries across similar datasets from different sources led to the removal of conflicting samples in classification sets and unreliable samples in regression sets exhibiting discrepancies exceeding an order of magnitude. Subsequently, datasets were merged and further cleaned. Duplicate data elimination: following cleaning and merging, duplicate entries were purged, resulting in a refined ADMET dataset. Duplicate data elimination: following cleaning and merging, duplicate entries were purged, resulting in a refined ADMET dataset. This final dataset comprises 97 ADMET sets, segmented into 43 regression and 54 classification datasets, totaling 465,470 data entries. Model Architecture The DCPM-ADMET molecular property prediction framework encapsulates two essential components: a feature extraction module and a downstream task prediction module. The feature extraction module meticulously crafts deep and versatile vector representations from SMILES molecular sequences. These representations then underpin tailored modeling for diverse downstream tasks. Pre-training Inspired by Transformer-based models' exceptional performance, DCPM embraces pre-trained architectures to refine SMILES sequences. However, prevalent methods like MoLFormer [17] and KPGT [18], which adopt the Transformer encoder during pre-training with masked token recovery akin to BERT [19], confront a pre-training-fine-tuning disparity when real-world inputs lack mask tokens [20]. 1. XLNet [21], which combines autoregressive and permutation strategies, resolves this inconsistency while enabling bidirectional context analysis, a feat previously unattainable by autoregressive models. By leveraging XLNet's permutation technique and dual-stream attention, DCPM captures global SMILES sequence patterns through self-attention and distills chemical insights from bidirectionally permuted contexts, enriching feature representation. 1. XLNet [21], which combines autoregressive and permutation strategies, resolves this inconsistency while enabling bidirectional context analysis, a feat previously unattainable by autoregressive models. By leveraging XLNet's permutation technique and dual-stream attention, DCPM captures global SMILES sequence patterns through self-attention and distills chemical insights from bidirectionally permuted contexts, enriching feature representation. 2. In the realm of Natural Language Processing (NLP), XLNet leverages the Transformer-XL architecture to enhance its capability in managing extremely long texts. Nevertheless, the SMILES sequences of chemical molecules, characterized by their relatively shorter lengths and smaller vocabularies, are more susceptible to overfitting challenges. To mitigate this, we have omitted the long text processing feature and incorporated an additional RNN structure, drawing inspiration from the approach of Robin Winter [12]. This RNN conducts a Seq2Seq SMILES-to-InChI translation while regressing nine physicochemical descriptors. By integrating InChI's fundamental physicochemical insights and descriptor regression into the pre-training objectives, we augment task complexity, mitigating overfitting, and fostering deeper molecular knowledge extraction beyond mere sequence reconstruction. 2. In the realm of Natural Language Processing (NLP), XLNet leverages the Transformer-XL architecture to enhance its capability in managing extremely long texts. Nevertheless, the SMILES sequences of chemical molecules, characterized by their relatively shorter lengths and smaller vocabularies, are more susceptible to overfitting challenges. To mitigate this, we have omitted the long text processing feature and incorporated an additional RNN structure, drawing inspiration from the approach of Robin Winter [12]. This RNN conducts a Seq2Seq SMILES-to-InChI translation while regressing nine physicochemical descriptors. By integrating InChI's fundamental physicochemical insights and descriptor regression into the pre-training objectives, we augment task complexity, mitigating overfitting, and fostering deeper molecular knowledge extraction beyond mere sequence reconstruction. 3. Considering that tokenized SMILES sequences mainly consist of atomic tokens, lacking substructure fragment information of chemical molecules, we added a feature fusion layer to DCPM. This layer amalgamates XLNet and RNN hidden layers with Extended Connectivity Fingerprints (ECFP) [22], an atomic connectivity graph-based descriptor rich in topological and substructure information, augmenting the model's substructural comprehension. 3. Considering that tokenized SMILES sequences mainly consist of atomic tokens, lacking substructure fragment information of chemical molecules, we added a feature fusion layer to DCPM. This layer amalgamates XLNet and RNN hidden layers with Extended Connectivity Fingerprints (ECFP) [22], an atomic connectivity graph-based descriptor rich in topological and substructure information, augmenting the model's substructural comprehension. To verify the effectiveness of each component, we conducted a series of ablation experiments on benchmark datasets. The results confirmed that the multimodal fused features of XLNet_RNN_ECFP represent the optimal combination. Detailed results are provided in the Supplementary Information (SI): “Ablation Experiments on Components of DCPM-ADMET”. In Fig. 1a, the pre-training workflow is outlined: SMILES are tokenized, embedded, and processed by an Encoder amalgamating GRU cells and XLNet layers. Extracted features are fed into three decoders, optimizing across InChI reconstruction, property prediction, and masked structure tasks. To enhance reproducibility, the optimal hyperparameter settings of the pre-training model are explicitly reported in the Supplementary Data Table S1. Schematic workflow of the DCPM-ADMET framework. a Pre-training phase: The model captures comprehensive molecular representations by simultaneously learning to reconstruct InChI tokens, property descriptors, and masked molecular structures. b Fine-tuning phase: The pre-trained encoders are frozen and integrated with ECFP features through a fusion layer to perform downstream molecular property predictions Fine-tuning For fine-tuning, depicted in Fig. 1b, the feature extraction layer is frozen, while the downstream prediction layer is trained. In this phase, the 512-dimensional structural embedding from XLNet, the 512-dimensional property embedding from the RNN, and the 512-dimensional ECFP fingerprints are concatenated into a final 1536-dimensional feature vector. For binary classification and regression tasks, we employ Random Forest to predict properties. The deep neural network applied in multi-task uses three hidden layers, optimizing key hyperparameters such as L2 regularization, dropout, and the number of neurons in each hidden layer. The Tree-structured Parzen Estimator (TPE) algorithm is used to identify the best hyperparameters across 50 evaluations. We employed early stopping and an 8:1:1 Bemis–Murcko scaffold-based train-validation-test split. Multiple evaluation metrics were utilized, with AUC as the primary indicator, and five independent runs were conducted to guarantee model robustness. Here, we systematically evaluated various fine-tuning strategies-specifically full-parameter fine-tuning versus parameter freezing (feature extraction)-alongside the selection of downstream models, including Random Forest (RF) and DNN (Deep Neural Network). For the DNN architecture, we further explored the differences between single-task and multi-task learning schemes. Our results indicate that parameter freezing (feature extraction) integrated with RF for single-task and DNN for multi-task learning achieves the best performance, aligning with the methodological strategies employed in prior research [23–26]. This optimized framework was adopted for all subsequent analyses, with full details provided in the Supplementary Information (SI): “Selection of Fine-Tuning Strategies and Model Architectures”. DCPM's prowess is validated against ECFP on MoleculeNet, demonstrating superiority across 10 subsets (7 classification and 3 regression tasks). Among the seven classification datasets, ClinTox, SIDER, Tox21, and ToxCast are multi-label classification tasks, while the other three classification datasets (BACE, BBBP, and HIV) and the regression datasets (ESOL, Lipophilicity and FreeSolv) are single-task. Ultimately, DCPM-ADMET is trained on a comprehensive ADMET dataset, encompassing 54 classification and 43 regression tasks, solidifying its standing as a useful tool for ADMET property prediction. Directly computable properties For enhanced usability, we have integrated 36 additional directly computable endpoints (see the Supplementary Data Table S2 for detailed values), encompassing the calculation of physicochemical properties, medicinal chemistry principles, and toxicophore rules using the RDKit and Scopy packages [27]. Results and discussion Results of DCPM-ADMET First, when compared with traditional fingerprint features, ECFP, across 10 benchmark datasets, DCPM outperformed traditional fingerprints in 7 out of 10 tasks (Tables 1-2, and Supplementary Data Tables S3-S5), demonstrating a clear advantage. This generalization is further confirmed by benchmarking against state-of-the-art pre-trained models, including KPGT [28]、KERMT [29]、MoleculeSTM-Graph [30]、MPG [31] and Mole-BERT [32] (Tables 1-2, and Supplementary Data Tables S3-S5). The results demonstrate that DCPM achieved the highest number of optimal performances across the benchmarks, securing top results in 5 out of 10 tasks (BBBP, HIV, ToxCast, ESOL, and FreeSolv). In comparison, KPGT, MPG, and Mole-BERT reached optimal performance in 3 (BACE, SIDER, and Tox21), 1 (ClinTox), and 1 (Lipophilicity) tasks, respectively. Table 1Performance comparison between DCPM and state-of-the-art (SOTA) methods on classification tasksMethodClassification (ROC-AUC↑)BACEBBBPHIVClinToxSIDERTox21ToxCastDCPM0.8253 ± 0.00630.7079 ± 0.00310.7884 ± 0.00240.9571 ± 0.00420.8311 ± 0.00200.8113 ± 0.00350.8318 ± 0.0009ECFP0.8287 ± 0.01050.6483 ± 0.00760.7448 ± 0.00460.9711 ± 0.00270.8339 ± 0.00160.7432 ± 0.00460.7847 ± 0.0012KERMT0.819 ± 0.00470.6807 ± 0.00780.7765 ± 0.00320.9716 ± 0.00280.8493 ± 0.00170.8193 ± 0.00570.8268 ± 0.0025KPGT0.8454 ± 0.00020.6911 ± 0.00010.7844 ± 0.00.9756 ± 0.00420.8502 ± 0.00340.8205 ± 0.00510.8279 ± 0.0119MoleculeSTM-Graph0.7432 ± 0.01470.6637 ± 0.00540.7068 ± 0.00330.9309 ± 0.00390.8370 ± 0.00190.7875 ± 0.00250.8070 ± 0.0014MPG0.7998 ± 0.01240.6965 ± 0.00790.7656 ± 0.01140.9789 ± 0.00140.8441 ± 0.00100.7871 ± 0.00310.8071 ± 0.0019Mole-BERT0.7997 ± 0.00380.6857 ± 0.00780.7725 ± 0.00660.9572 ± 0.00260.8238 ± 0.00100.8065 ± 0.00360.8088 ± 0.0008 Performance comparison between DCPM and state-of-the-art (SOTA) methods on classification tasks Table 2Performance comparison between DCPM and state-of-the-art (SOTA) methods on regression taskMethodRegression (RMSE↓)ESOLFreeSolvLipophilicityDCPM1.029 ± 0.00762.4621 ± 0.01760.9252 ± 0.0026ECFP1.9518 ± 0.00444.2936 ± 0.01171.0515 ± 0.0013KERMT1.1565 ± 0.00343.027 ± 0.02960.917 ± 0.0023KPGT1.5301 ± 0.03.6052 ± 0.01.0386 ± 0.0001MoleculeSTM-Graph1.468 ± 0.01283.6389 ± 0.01511.0377 ± 0.0008MPG1.4905 ± 0.00473.4328 ± 0.02881.0098 ± 0.0049Mole-BERT1.6142 ± 0.00523.119 ± 0.04690.9049 ± 0.0037 Performance comparison between DCPM and state-of-the-art (SOTA) methods on regression task To validate the practical utility of our approach, we developed a comprehensive modeling framework covering 97 ADMET property prediction tasks, enabling a holistic characterization of molecular ADMET profiles. Comparative analysis demonstrates that our proposed scheme outperforms traditional fingerprint-based methods in 67 tasks (comprising 39 classification and 28 regression tasks), achieving an overall success rate of 69.07%. Specifically, the average AUC for classification tasks improved from 0.812 to 0.831, while the average Pearson Correlation Coefficient (PCC) for regression tasks increased significantly from 0.661 to 0.701. Detailed performance metrics for all tasks are provided in Tables S6–S7 of the Supplementary Data. We conducted a comparative analysis of endpoint information and data size across several prominent ADMET prediction platforms, namely ADMETlab 3.0 [33], admetSAR 3.0 [34], ProTox3.0 [35], ADMETboost [36], toxCSM [37] and Interpretable-ADMET [9]. The comprehensive details of this comparison are outlined in Table 3. The results indicate that our platform encompasses the highest number of ADMET regression tasks. Furthermore, we possess the largest dataset, which serves as a robust foundation for model development. Overall, DCPM-ADMET provides more options for ADMET prediction. Table 3Comparison of DCPM-ADMET with other web-based platformsnamedata size(compounds)endpointclassification taskregression taskURLDCPM-ADMET465,4701335443http://admet.bioai-global.com/ADMETlab 3.0400,0001195918https://admetlab3.scbdd.comadmetSAR3.0370,0001199017http://lmmd.ecust.edu.cn/admetsar3/ProTox 3.0100,00061610https://tox.charite.de/protox3/ADMETboost100,00022139https://ai-druglab.smu.edu/admetInterpretable-ADMET250,729599028http://cadd.pharmacy.nankai.edu.cn/interpretableadmet/toxCSM43,23636315http://biosig.lab.uq.edu.au/toxcsm Comparison of DCPM-ADMET with other web-based platforms Implementation of DCPM-ADMET DCPM-ADMET presents a user-centric web application that seamlessly integrates pre-trained and 97 specialized ADMET models and 36 directly computable properties in the cloud. This platform enables users to effortlessly engage with ADMET prediction outcomes through a highly intuitive graphical interface, as exemplified in Fig. 2. DCPM-ADMET Web Application Prediction Workflow. The platform features a user-friendly interface for seamless interaction: (Left) User Input: Molecular structures can be submitted via SMILES strings. (Middle) Cloud Computing: Input data is processed in the background. (Right) Results Display: Prediction outcomes are presented through a comprehensive dashboard, featuring 2D/3D molecular visualizations, basic physicochemical properties (e.g., Molecular Weight, LogP), and detailed ADMET classification/regression results with visual indicators (green/red dots) to highlight potential toxicity or metabolic risks. A more detailed introduction is available through the Documents component In the DCPM-ADMET prediction module, users input molecular representations in SMILES format, and the system processes them via cloud-based RDKit to generate key molecular attributes, including molecular structure, molecular weight and other properties. ADMET predictions are provided for both regression and classification tasks. Regression results include predictions for critical aspects: basic properties, as well as absorption, distribution, metabolism, and toxicity. Drug-likeness is assessed based on thresholds derived from pharmacological research, marked as “√” (acceptable), “ × ” (unacceptable), or “warning” for intermediate values. For instance (Fig. 3b), Apixaban’s bioavailability is predicted at 46.48%, marked as “√”, indicating good bioavailability, while its plasma protein binding is 89.02%, marked as “warning”, suggesting reduced efficacy. Molecular structure and ADMET profile of apixaban. a 2D structure and SMILES. b Absorption and distribution predictions. c CYP450 metabolism predictions In classification tasks, probabilities for various ADMET categories are displayed, with values between 0–35% labeled as negative (“−”), 65–100% as positive (“ + ”), and 35–65% as “warning”. For example (Fig. 3c), Apixaban’s predictions for CYP450 2C9 inhibitor, CYP450 3A4 inhibitor, and CYP450 3A4 substrate are marked as “ + ”, indicating metabolism via these enzymes, consistent with its known metabolic pathways. Case Study We analyzed 45 withdrawn drugs from DrugBank [39] (version 5.1.12) and predicted their ADMET properties using DCPM-ADMET. These drugs were withdrawn due to potential toxicity, with hERG toxicity [40] and hepatotoxicity [41] being the main causes. Among the 45 drugs, 24 were withdrawn due to hERG toxicity, 16 due to hepatotoxicity, 2 due to carcinogenicity and genotoxicity, and 1 due to skin irritation. Prediction of classification Properties DCPM-ADMET accurately predicted potential toxicity in 36 out of 45 drugs, marking a success rate of approximately 80.0% (see Supplementary Data Table S8). To ensure unbiased results, we ensured no overlap between DrugBank molecules used in these tasks and our training sets. Molecular fingerprint similarity analysis using Tanimoto coefficients revealed low average similarities (0.06–0.164) between the 45 drugs and their training set counterparts, further validating the model's generalization. Among drugs with maximum similarity below 0.4, prediction accuracy remained consistent with the overall accuracy (87.5%), reinforcing DCPM-ADMET's robust generalization capabilities (see Supplementary Data Table S8). Furthermore, we compared our tool with the current mainstream ADMET web tools (ADMETlab 3.0 and admetSAR 3.0), and the results showed that we have certain advantages in hERG prediction. This demonstrated the necessity and practicality of DCPM-ADMET as another ADMET tool available for online use (see Supplementary Data Table S8). Identify adverse properties Thalidomide, initially synthesized by CIBA (predecessor of Novartis), was marketed in Europe in 1957 as an antiemetic for pregnant women, tragically leading to numerous birth defects. Our tool foresaw its genotoxicity (Fig. 4a), emphasizing the potential to avert such disasters with early predictions. Thalidomide & Nefazodone Prediction Insights Integrated Property Analysis For hERG inhibition and hepatotoxicity, DCPM-ADMET boasted prediction accuracies of 91.7% and 75.0%, respectively, underscoring its robust predictive capabilities. Specifically, Nefazodone has demonstrated the trustworthiness of DCPM-ADMET’s prediction. According to its description, Nefazodone has a very low incidence of liver damage, with a serious liver injury occurrence rate of 1 in 250,000 to 300,000 users annually. However, DCPM-ADMET is still capable of predicting its hepatotoxicity (Fig. 4b). On the other hand, Nefazodone was predicted as high risk for DILI (drug-induced liver injury) with DCPM-ADMET, prompting further investigation. Reports suggest that Nefazodone is oxidized by CYP3A4 (our prediction also shows it as a substrate for CYP3A4), and its active metabolites can bind with GSH to produce toxicity, potentially explaining the high DILI risk and the discrepancy in its hepatotoxicity prediction. This demonstrates the tool’s ability to accurately predict ADMET properties and guide our understanding of the intrinsic factors underlying drug properties. Comparative Prediction for Multiple Drugs DCPM-ADMET predicted that Fexofenadine (0.826) has a lower risk of side effects compared to Terfenadine (0.868) (Fig. 5a), aligning with its safer profile and reduced risk of QT prolongation. In the case of IAP antagonists, DCPM-ADMET accurately predicted the impact of structural modifications on the ADMET properties of tolinapant (Fig. 5b). By incorporating a hydroxymethyl group, tolinapant's lipophilicity decreased (LogD: 2.12 vs. 2.58; LogP: 4.17 vs. 4.32), and its probability of being a CYP3A4 substrate (60.56%) was lower than that of AT-IAP (67.09%). These predictions were consistent with experimental data [42], validating DCPM-ADMET's effectiveness in drug optimization. Comparative ADMET prediction results of multiple drugs. a Comparison of side effect risks between fexofenadine and terfenadine; b Effects of structural modification on the ADMET properties of tolinapant Prediction of Regression Properties To assess DCPM-ADMET's performance further, we analyzed 71 drugs with Caco-2 permeability data from DrugBank. Caco-2 permeability gauges a drug's intestinal absorption, serving as the gold standard for in vitro predictions [43]. Our model achieved an RMSE of 0.47 and a PCC of 0.83 on these data (Fig. 6), signifying minimal error and strong correlation between predictions and experimental values. Molecules in the test set exhibited low mean similarity to the training set (0.045–0.16), highlighting the model's generalization ability (all data and prediction values see Supplementary Data Table S9). Prediction results for 71 drugs with Caco-2 values Conclusion In this study, we present DCPM-ADMET, an advanced tool for predicting drug ADMET properties using a novel pre-trained model architecture that combines the strengths of both RNN and Transformer models. By incorporating molecular substructure information through ECFP fingerprints, DCPM outperforms traditional molecular fingerprint methods and most pre-trained models, offering accurate and reliable predictions across a wide range of molecular properties. DCPM-ADMET is capable of predicting 133 ADMET endpoints, covering 54 classification tasks and 43 regression tasks. The model is trained on a high-quality dataset consisting of 465,470 entries associated with diverse ADMET profiles. Additionally, it integrates 36 computational descriptors linked to physicochemical characteristics, medicinal chemistry guidelines, and toxicophore rules. Benefiting from such comprehensive functionalities, DCPM-ADMET serves as a complementary online ADMET prediction platform and delivers a set of unique predictive endpoints. This tool enables the early recognition of drug candidates with poor pharmacokinetic behaviors, facilitating the structural optimization of molecules to further improve drug efficacy and safety. Looking forward, the methodologies employed in DCPM can be extended to other areas of drug discovery, such as predicting drug-drug interactions and optimizing molecular bioactivities. With the continuous availability of more data, the model's accuracy and generalizability can be further refined. Supplementary Information Supplementary Material 1. Supplementary Material 2. Acknowledgements We acknowledge the contributions of various individuals and organizations that have made this study possible. And the robotic AI-Scientist platform of the Chinese Academy of Sciences. Author contributions Leilei Zhang, Yuchen Zeng, Yue Qi and Kaili Jiang participated in project operations and article writing. Xiaofei Zhou participated in the project planning process. Lu Liang was involved in project operations and guidance. Jianping Lin participated in project planning and article writing. Funding This work was supported by the 2025 Tianjin Major Science and Technology Special Project in Biomedicine (25ZXSWSY00580) and the Tianjin Science and Technology Plan (25ZXZSSS00090). Data availability All the datasets and source code are publicly available through GitHub https://github.com/zhangzhangleilei/DCPM-ADMET Declarations Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Lu Liang, Email: lianglu@bioai-global.com. Jianping Lin, Email: jianpinglin@nankai.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Data Availability Statement All the datasets and source code are publicly available through GitHub https://github.com/zhangzhangleilei/DCPM-ADMET