ADMET-vault: an interactive framework for real-time ADMET prediction and molecular optimization.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Drug discovery faces challenges due to poor absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiles of candidates in clinical trials.
• Early-stage ADMET profiling is crucial for successful drug development but often lacking in current workflows.

Purpose of the Study:

• To present ADMET-Vault, an interactive machine learning platform for early-stage ADMET prediction and molecular design.
• To integrate ADMET prediction into the drug design process, enabling real-time assessment and optimization of drug candidates.

Main Methods:

• ADMET-Vault integrates diverse molecular representations: physicochemical descriptors, molecular fingerprints, and graph neural network embeddings.
• The platform utilizes scaffold-based validation on 12 benchmark datasets from Therapeutics Data Commons.
• It combines predictive machine learning models with a live molecular editor for interactive design-predict-optimize cycles.

Main Results:

• ADMET-Vault demonstrated consistent predictive performance across multiple ADMET endpoints, with strong results for clearance, intestinal absorption, hepatotoxicity, and solubility.
• Regression tasks showed mean absolute errors from 0.28 to 7.68 and Spearman's rank correlation coefficients from 0.44 to 0.66.
• Classification tasks achieved area under the receiver operating characteristic curve values between 0.87 and 0.99.

Conclusions:

• ADMET-Vault effectively integrates predictive accuracy with interactive molecular editing, filling a critical gap in early drug discovery.
• The platform enables researchers to instantly assess the impact of structural changes on drug-like properties.
• This interactive approach facilitates a more efficient and optimized drug design process, potentially reducing late-stage failures.