AI in drug discovery is one of the most hyped topics in life sciences. Some applications are now standard tools that genuinely accelerate the pipeline. Others remain aspirational despite glossy press releases. Knowing the difference matters when you’re deciding where to invest time or budget.
Where AI clearly delivers
Protein structure prediction
AlphaFold 2 and 3, RoseTTAFold, and ESMFold predict protein structures from sequence with experimental-quality accuracy in many cases. This has reshaped structure-based drug design, target validation, and protein engineering. The AlphaFold Protein Structure Database covers more than 200 million proteins.
Virtual screening
Deep learning models score billions of molecules against a target structure in hours. Tools like DiffDock, Equibind, and Glide ML predict binding poses; RFdiffusion designs novel binders from scratch. Pharma companies routinely run virtual screens of 10⁹–10¹⁰ molecules.
Generative chemistry
Models like REINVENT, Chemformer, and various diffusion models generate novel molecules with desired properties. Combined with synthesizability filters, these accelerate hit-to-lead optimization.
Property prediction
Machine learning predicts ADMET properties (absorption, distribution, metabolism, excretion, toxicity), solubility, and binding affinity reasonably well — especially for chemical scaffolds within the training distribution.
Target identification
Network-based AI integrates multi-omics data to prioritize disease-relevant targets. Open Targets and similar platforms now embed ML scoring routinely.
Where AI is still developing
Protein-ligand binding affinity
Predicting absolute binding affinity remains hard. Models do better on relative ranking within congeneric series than on de novo affinity prediction.
Pharmacokinetics from chemistry alone
PK predictions improve with more data, but in vivo behavior depends on route, dose, formulation, species, and individual variation in ways that current models capture incompletely.
Predicting clinical outcomes
Models predict drug-drug interactions and some safety signals, but predicting which compound will succeed in Phase III is still beyond reach.
“AI-discovered drugs” in patients
Several AI-originated compounds have entered clinical trials. Whether AI-discovered molecules outperform traditionally discovered ones in efficacy or speed-to-approval is still an open question.
Major tools and platforms
| Application | Notable tools |
|---|---|
| Structure prediction | AlphaFold, ESMFold, RoseTTAFold |
| Docking | DiffDock, Glide ML, Equibind |
| Generative chemistry | REINVENT, Chemformer, MolDQN |
| Property prediction | Chemprop, DeepChem, ADMET-AI |
| Protein design | RFdiffusion, ProteinMPNN, ESM-IF |
| Target ID | Open Targets, BenevolentAI |
Honest limitations
- Models are only as good as their training data. Public datasets are skewed toward kinase inhibitors, a few well-studied targets, and drug-like molecules
- Out-of-distribution prediction is unreliable. A model trained on FDA-approved drugs may extrapolate poorly to novel chemotypes
- Wet-lab validation remains essential. AI predictions must be tested experimentally — and failure rates are higher than the press releases suggest
- Reproducibility issues: Many published AI methods don’t generalize beyond the original benchmarks
Practical advice
- Use AlphaFold structures by default unless you have experimental structures — they’re often good enough for design
- Treat virtual screening hits as starting points, not final candidates
- For property prediction, prefer models trained on data similar to your chemical space
- Always validate experimentally before scaling — paper performance overestimates real-world performance
AI is now embedded throughout drug discovery, but it’s a tool, not a replacement for chemistry, biology, or clinical judgment. The biggest wins come from teams that combine AI methods with deep medicinal chemistry and biology expertise.
