AI in Drug Development: Limitations and Real-World Applications for Scientists
Artificial intelligence accelerates target and molecule discovery in drug development, but doesn't replace experts due to data quality and interpretability challenges. AlphaFold 2 predicts protein structures, with 45 AI-generated molecules in clinical trials, yet heterogeneous datasets and lack of universal descriptors require deep chemistry and biology expertise. In clinical research, AI assists with hybrid models, but simple regression models remain preferable for valid predictions.
Drug Development Stages and the Role of Data
The drug creation process includes target selection, active structure discovery, in vitro and in vivo testing, clinical phases, and regulatory assessment. Costs grow exponentially, with most candidates eliminated early.
Key challenges:
- Identifying pathogenic mechanisms and binding sites
- Assessing pharmacological properties of molecules
- Interpreting data from diverse experiments
AI addresses some tasks: AlphaFold 2 generates 3D protein structures from amino acid sequences, reducing costs for cryo-EM and X-ray crystallography. The 2024 Nobel Prize to Hassabis and Jumper confirmed this breakthrough.
AI Challenges in Molecule Discovery
Chemical structures aren't directly input into ML models. Descriptors are used: physicochemical (logP, molecular weight), graph representations, or Morgan fingerprints.
| Descriptor Type | Advantages | Limitations |
|---------------|-------------|-------------|
| 1D/2D (SMILES, ECFP) | Simple computation | Ignores 3D conformations |
| 3D Graph-based | Accounts for spatial geometry | High computational complexity |
| Physicochemical | Interpretability | Incomplete property coverage |
High dimensionality (thousands of features) requires large datasets unavailable outside big pharma. Bioactivity depends on conditions: cell types, receptor density. Heterogeneous data leads to artifacts without expert preprocessing.
Successes exist: By 2025, 45 AI candidates in clinical trials. FDA received >500 AI submissions from 2016-2023, establishing 10 principles: multidisciplinary approach, risk assessment, data documentation.
AI in Clinical Models: Interpretability vs. Power
For side effect assessment, logistic regression is used: predictors include dose, age, gender; outcome is event probability. The model is interpretable, confidence intervals calculable.
Neural networks lose this: black box nature, no explanations, unreliable with small early-phase samples. Example: models ignoring biology produce absurd predictions, as shown in original Figure 3.
Physiologically-based models (PK/PD) use differential equations for processes: cell division, enzyme inhibition. They provide not just predictions but also graphs and virtual studies.
Require cross-functional teams: mathematicians + biologists + clinicians verify assumptions and pathophysiology.
Hybrid Approaches and Future Directions
The compromise is hybrid models: differential equations + neural networks as approximators of unknown dependencies. Currently pilot projects but promising.
AI already assists: generates analysis code (Sonata Software). AI agents for building models from scratch are an active research area.
Key takeaways:
- AlphaFold 2 revolutionized structure prediction but not target validation
- 45 AI molecules in clinical trials, but success depends on expert descriptor interpretation
- Simple models preferred in clinical settings due to interpretability
- Hybrids combine physiology and ML for reliable predictions
- Regulators require documentation and risk assessment for AI
AI is a tool that enhances scientists but doesn't replace them: expertise in chemistry, biology, and common sense remain essential for valid conclusions.
— Editorial Team
No comments yet.