Appearance
ADMET: Why 90% of Drug Candidates Fail — and How AI Can Help
Drug discovery is one of the most complex and expensive processes in modern science. Despite advances in genomics, chemistry, and computational modeling, roughly 90% of all drug candidates fail before reaching the market. The majority of these failures can be traced back to one critical area: ADMET — Absorption, Distribution, Metabolism, Excretion, and Toxicity.
Understanding ADMET and how artificial intelligence (AI) is transforming its prediction and optimization is essential for the future of pharmaceutical innovation.
What Is ADMET?
ADMET refers to a group of pharmacokinetic and toxicological properties that determine whether a compound can become a safe and effective drug.
| Property | Description |
|---|---|
| Absorption | How well a drug is absorbed into the bloodstream after administration. |
| Distribution | How the drug spreads through the body’s tissues and organs. |
| Metabolism | How the body chemically transforms the drug, primarily in the liver. |
| Excretion | How the drug and its metabolites are eliminated (usually via kidneys or bile). |
| Toxicity | How harmful the drug or its metabolites are to the body. |
Even a compound with strong therapeutic potential can fail if it performs poorly in one of these categories.
Why 90% of Drug Candidates Fail
According to data from major pharmaceutical studies, more than half of clinical failures are due to unfavorable pharmacokinetics or toxicity. Here's a breakdown of where things go wrong:
- 40% fail due to poor pharmacokinetics (ADME-related issues)
- 30% fail due to toxicity
- 20% fail due to lack of efficacy
- 10% fail for other reasons (e.g., formulation or manufacturing)
Traditional ADMET testing relies on in vitro (cell-based) and in vivo (animal-based) experiments. These are slow, expensive, and often don’t accurately translate to human biology. Predicting human ADMET behavior remains one of the biggest bottlenecks in drug development.
How AI Transforms ADMET Prediction
AI is reshaping how researchers evaluate ADMET properties — not by replacing laboratory testing, but by optimizing it with predictive models that can identify potential failures before costly clinical trials begin.
1. Predictive Modeling
Machine learning (ML) models can learn from large databases of known compounds to predict ADMET properties of new molecules.
Examples:
- QSAR models (Quantitative Structure–Activity Relationship) use molecular features to estimate solubility, permeability, or toxicity.
- Graph Neural Networks (GNNs) model molecular structures directly, learning from atom-level relationships.
2. Toxicity Forecasting
AI can detect patterns in molecular structures associated with hepatotoxicity, cardiotoxicity, or mutagenicity, reducing the need for animal testing.
For example, models like DeepTox and Tox21 Challenge have shown that deep learning can outperform traditional methods in toxicity prediction.
3. Metabolism Pathway Simulation
AI can simulate cytochrome P450 enzyme interactions, predicting how drugs are metabolized — a key step in preventing harmful drug–drug interactions.
4. Multi-parameter Optimization
Modern AI systems can optimize efficacy, safety, and pharmacokinetics simultaneously, balancing trade-offs that would be impossible to explore experimentally.
The New ADMET Workflow With AI
A next-generation drug discovery pipeline might look like this:
- Data Integration – Gather molecular, biological, and clinical data from public and proprietary sources.
- AI-Based Screening – Use ML models to predict ADMET outcomes for thousands of virtual compounds.
- Ranking & Filtering – Eliminate molecules with poor absorption or predicted toxicity.
- Experimental Validation – Test only the top candidates in the lab.
- Iterative Refinement – Feed experimental results back into the model to improve accuracy.
This approach can reduce costs by up to 70% and development time by several years.
Real-World Applications
- Atomwise uses deep learning to screen billions of compounds for ADMET properties.
- Insilico Medicine combines generative AI with ADMET prediction to design safer molecules.
- DeepMind’s AlphaFold has accelerated the understanding of protein–ligand interactions, which indirectly improves ADMET modeling accuracy.
Challenges Ahead
While AI is powerful, challenges remain:
- Data quality: Many datasets are biased or incomplete.
- Interpretability: AI models often act as black boxes.
- Regulatory acceptance: Agencies like the FDA are still developing frameworks for AI-driven predictions.
Despite this, the trajectory is clear: AI is rapidly becoming indispensable in early-stage drug discovery and development.
Conclusion
ADMET has always been the gatekeeper between chemical possibility and medical reality. With 90% of candidates still failing, traditional methods alone are no longer enough.
AI offers a new paradigm — one where drug safety, efficacy, and pharmacokinetics can be predicted, optimized, and understood before human trials even begin. As models continue to improve, the dream of faster, cheaper, and safer drug discovery is becoming reality.
Written by Tobi Bück
Founder of Drug Design Hub – exploring AI, pharmacology, and the future of drug discovery.