Abstract
The 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-channel system—including an XLNet-based module for capturing the semantics of molecular sequences, an RNN-based component for small molecule property extraction, and ECFP fingerprints for capturing molecular substructures—and after initial pre-training, the model outperforms traditional methods 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.
Introduction
In the high-stakes field of drug discovery, potency is only half the battle. Approximately 40–60% of drug candidates fail during clinical trials due to poor ADMET properties—instances where the body cannot safely process the drug or the drug fails to reach its target effectively. Predicting these properties early using machine learning is a primary objective for the industry, yet it faces a significant "data desert" problem. Traditional models often struggle to generalize because experimental ADMET data is scarce and noisy. Furthermore, many current AI models suffer from a "pre-training discrepancy," where the learning phase (e.g., masking tokens) does not align with real-world application (where data is complete).
Methodology: The Dual-Channel Architecture
The researchers introduced DCPM-ADMET (Dual-channel Pre-trained Model), which utilizes a sophisticated three-pronged approach to molecular understanding:
- Channel 1 (XLNet-based): Unlike standard BERT-like architectures that rely on masking, this channel uses XLNet’s permutation-based autoregressive strategy. This allows the model to learn bidirectional context without artificial masks, effectively solving the pre-training discrepancy.
- Channel 2 (RNN-based): This channel focuses on chemical logic through a Sequence-to-Sequence (Seq2Seq) translation from SMILES to InChI identifiers, while simultaneously regressing nine fundamental physicochemical descriptors.
- Feature Fusion: The model integrates these deep learning features with ECFP (Extended Connectivity Fingerprints). By explicitly encoding molecular substructures, the model ensures it retains critical information about chemical fragments that dictate toxicity and reactivity.
Data Scale and Benchmarking
The scale of data utilized in this study represents a significant advancement over previous efforts. The model underwent initial pre-training on a massive set of 111,295,885 data points from PubChem. For specific ADMET tasks, the researchers constructed a proprietary database of 465,470 entries covering 97 properties, aggregated from high-quality sources like ChEMBL and the Therapeutics Data Commons (TDC).
Key Results and Performance
DCPM-ADMET demonstrated superior predictive accuracy across multiple benchmark datasets. In the MoleculeNet benchmark, the model outperformed traditional ECFP-based methods in 7 out of 10 tasks. Specifically, in regression tasks such as Caco-2 permeability, the model achieved a Pearson Correlation Coefficient (PCC) of 0.82, indicating a high degree of alignment between predicted and experimental values.
Beyond benchmark metrics, the model showed real-world utility by correctly identifying the toxicity of 80% of a test set containing 45 drugs withdrawn from the market. For example, it flagged the drug Nefazodone as high-risk for liver injury, catching subtle toxicological signals that traditional screenings often miss.
Conclusion and Future Directions
The development of DCPM-ADMET addresses the fundamental limitation of poor generalization in computational pharmacology. By deploying their findings into a free online tool offering 133 endpoints—the highest number among comparable public platforms—the researchers have provided an essential resource for early-stage drug discovery. Future work will focus on extending the DCPM architecture to predict drug-drug interactions (DDIs) and moving from property prediction to the de novo design of molecules with optimized ADMET profiles.
Summary updated