ADMETron: An AI-driven SaaS platform for comprehensive ADMET prediction and compound prioritisation

Technical Summary of ADMETron: An AI-driven SaaS Platform for Comprehensive ADMET Prediction

Problem Statement
Modern drug discovery requires the rapid and accurate assessment of Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties to prioritize chemical compounds and mitigate late-stage attrition. While various tools exist, there is a need for a unified environment that combines high-throughput virtual screening capabilities with intuitive, multidimensional visualization to support data-driven decision-making. Existing solutions often lack comprehensive endpoint coverage or fail to integrate complex predictive modeling with user-friendly structure-activity relationship (SAR) analysis in a single scalable platform.

Methodology
ADMETron is architected as a scalable SaaS platform integrating an interactive web interface with a robust predictive engine. The core modeling framework employs a hybrid approach:

• Feature Engineering: The system generates molecular embeddings using Recurrent Neural Networks (RNNs) trained on SMILES representations. These deep learning-derived features are combined with traditional physicochemical descriptors to create a comprehensive feature set.
• Predictive Modeling: The combined features are processed through Gradient Boosting Machines (GBMs) to generate predictions.
• Scope: The model is trained to predict 34 distinct ADMET endpoints, encompassing physicochemical properties, absorption rates, Cytochrome P450 (CYP450) interactions, hERG liability, and mutagenicity.
• Visualization: To interpret these multidimensional outputs, the platform utilizes an interactive radar graph-based framework. This allows users to perform real-time comparisons of multiple compounds against reference drugs across selected ADMET parameters.

Key Contributions

1. Unified Predictive Engine: The development of a single platform capable of handling both classification and regression tasks across a broad spectrum of 34 ADMET endpoints, bridging pharmacokinetic, toxicity, and medicinal chemistry properties.
2. Hybrid Modeling Architecture: The successful integration of RNN-derived molecular embeddings with physicochemical descriptors within a GBM framework to enhance predictive accuracy.
3. Interactive SAR Visualization: The introduction of a radar graph-based interface that facilitates the intuitive interpretation of complex molecular profiles, enabling rapid lead optimization and compound prioritization.
4. Scalability: The deployment of a system designed for high-throughput virtual screening and batch analysis, accessible via a web interface.

Results
The predictive performance of ADMETron was rigorously evaluated using benchmark datasets from the Therapeutics Data Commons (TDC). The evaluation demonstrated:

• Strong Performance: The platform exhibited robust accuracy across both classification and regression tasks.
• Generalizability: The models showed consistent performance across diverse chemical spaces.
• Comparative Advantage: When assessed against widely used online ADMET tools, ADMETron demonstrated broader endpoint coverage, successfully unifying pharmacokinetic, physicochemical, toxicity, and medicinal chemistry predictions within one environment.

Significance
The paper positions ADMETron as a comprehensive solution for ADMET assessment in drug discovery. By combining advanced AI-driven prediction with an interactive visualization framework, the platform aims to streamline the compound prioritization process. Its significance lies in its ability to support data-driven decision-making through a unified, scalable, and user-friendly environment, thereby addressing the need for efficient, high-throughput ADMET profiling in modern medicinal chemistry workflows.