Drug discovery AI occupies an uncomfortable middle ground between transformative headlines and incremental reality. AlphaFold genuinely changed computational structural biology; AI-designed molecules have entered clinical trials; and yet no AI-designed drug has been approved as of early 2026. This post is the honest map of where AI helps in pharma R&D and where the 12-year cycle remains stubborn.
Target identification
Disease-target association. AI mines literature, omics data, phenotype databases to connect diseases with potential drug targets. Surfaces candidates that human literature review might miss.
Literature mining. Pharma has access to millions of papers, patents, internal reports. AI-based retrieval surfaces relevant prior work faster than manual review. Helps avoid pursuing targets others have already failed at.
Omics analysis. Genomics, proteomics, transcriptomics — AI finds patterns in high-dimensional biological data. Specialized models (not general LLMs) dominate; this space has existed pre-LLM and is only modestly changed.
Molecule design
Property prediction. Given a molecular structure, predict properties (toxicity, binding affinity, solubility, bioavailability). Accelerates screening. Replaces or augments traditional QSAR models.
Generative chemistry. AI models propose novel molecular structures meeting specified criteria. Output requires extensive human evaluation; many proposals are synthetically infeasible.
Retrosynthesis planning. Given a target molecule, AI proposes synthesis routes. Speeds up the chemistry planning; chemists still execute.
Structural biology
AlphaFold (DeepMind). Protein structure prediction with accuracy approaching experimental crystallography for many proteins. Genuine paradigm shift; widely used across pharma.
AlphaFold 3 and successors. Protein-ligand interactions, complexes. Still advancing rapidly. Each generation expands what's computable.
Docking and binding affinity. AI-augmented docking tools improve over classical methods. Physical simulation still essential for high-confidence predictions.
Cryo-EM image processing. AI substantially speeds structure determination from cryo-electron microscopy images. Operational workflow benefit in structural biology labs.
Clinical reality
AI-designed molecules entering trials. Yes, this is happening. Several biotechs have candidates in Phase 1/2 trials identified or designed with AI.
Approved AI-designed drugs. Not yet as of early 2026. The first will matter; companies claiming dramatic speedups need approved drugs to prove it.
Clinical attrition. Most drugs fail in trials. AI improves the probability of identifying viable targets but doesn't transform the base rate of clinical success. A candidate that looked great in silico still faces the same human biology in Phase 3.
Honest assessment
AI accelerates the early stages. Target identification, hit discovery, lead optimization — all meaningfully faster with modern AI. Saves 6-18 months out of a 12-year cycle.
AI does not fix clinical trial biology. Phase 2 and Phase 3 trials are human experiments that take time. AI doesn't speed up recruitment without large additional interventions, doesn't shorten endpoints, doesn't change regulatory review.
Companies claiming 10x speedup. Evaluate with skepticism. Ask: what drugs have they shipped? The proof will be in approvals, not pipeline announcements.
Economic model shifts
Discovery-phase productivity. More candidates through early stages per dollar. Portfolio strategy: cast wider net, pursue more parallel programs.
Specialized AI biotechs vs pharma. AI-focused biotechs (Recursion, Insitro, Relay Therapeutics, many others) specialize in AI-driven discovery. Traditional pharma partners or acquires for capability. Hybrid model stabilizing.
Data is the moat. Best AI teams without quality training data underperform. Pharma's proprietary clinical and experimental data matters more than model architecture.
Where the field is heading
Multimodal foundation models for biology. Training on combined genomics, proteomics, imaging, literature data. Early-stage research; will mature through 2026-2028.
Active learning loops. AI proposes molecules, lab synthesizes and tests, AI learns from results. Closing the experimental loop; some teams have robot-augmented labs specifically for this.
Regulatory-focused AI. Submission preparation, trial design against regulatory precedent, pharmacovigilance. See pharma regulatory AI post.