In December 2018, AlphaFold placed first in CASP13 — the biennial protein structure prediction competition. Its performance was good but not revolutionary: it predicted structures with ~60% accuracy on the most difficult targets, a modest improvement over previous methods. The DeepMind team went back to work.

Eighteen months later, at CASP14 in November 2020, AlphaFold 2 achieved near-atomic accuracy (~90% GDT) on the hardest targets. The problem of protein structure prediction — which biology had considered unsolvable for 50 years — was effectively solved.

The field of computational biology has not been the same since.

The AI-in-Biology Timeline

The growth curve is exponential. From one AI-discovered molecule entering trials in 2019 (Insilico's DDR1 inhibitor) to an estimated 200 in 2026, the field has crossed from proof-of-concept to pipeline reality. But the numbers tell only part of the story.

The Nobel That Changed Everything

When the 2024 Nobel Prize in Chemistry was awarded jointly to David Baker ("for computational protein design") and Demis Hassabis and John Jumper (DeepMind, "for protein structure prediction via AlphaFold"), it marked the moment AI-in-biology went from interesting subfield to central pillar of the discipline.

The Nobel had two effects. First, it validated AlphaFold's approach as a genuine scientific revolution — not hype, not a clever algorithm, but a fundamental shift in how biology is done. Second, it triggered an influx of talent and capital. Isomorphic Labs raised a $2.1 billion Series B in January 2026 — the largest private biotech round of the year.

The Three Waves of AI Drug Discovery

Understanding where AI drug discovery stands in 2026 requires understanding three distinct technology waves that are running concurrently.

Wave 1: Target Identification (2017-2023)

The first wave used machine learning to analyze genomic, proteomic, and literature data to find better drug targets. Companies like BenevolentAI and Recursion built knowledge graphs from millions of scientific papers, then used ML to prioritize novel target-disease pairs.

What it achieved: More targets, faster. BenevolentAI's platform identified a target for ALS (PIKFYVE) that was not obvious from literature alone. The drug is now in Phase 1.

What it didn't achieve: Faster drugs. Target identification was never the rate-limiting step in drug discovery — clinical trials were. AI narrowed the search but didn't shorten the timeline.

Wave 2: Molecule Generation (2020-2025)

The second wave used generative AI — first GANs, then diffusion models, then protein language models — to design molecules and proteins for specific targets.

What it achieved: Insilico Medicine's ISM001-055, designed for idiopathic pulmonary fibrosis, went from target to clinic in 18 months — 3-5x faster than traditional drug discovery. The drug showed positive Phase 2a data in 2023 and entered Phase 2b in 2025. If approved, it will be the first end-to-end AI-discovered drug.

What it didn't achieve: Perfect molecules. AI-generated molecules still need extensive optimization — pharmacokinetics, toxicity, manufacturability. The AI finds the starting point, not the finish line.

Wave 3: Protein Design (2022-2026)

The third wave, still in its early stages, uses AI to design entirely new proteins that don't exist in nature.

What it achieved: RFdiffusion (Baker lab, 2022) can design protein binders for any target structure. ProteinMPNN can design amino acid sequences for any protein backbone. ESM3 (EvolutionaryScale, 2024) designed a novel fluorescent protein with no sequence homology to known fluorescent proteins — the equivalent of an AI generating a functional new primary color.

What it didn't achieve: General-purpose enzyme design. Designing proteins that catalyze specific chemical reactions — enzymes — is much harder than designing structural binders. The field can now make proteins that bind to anything. It cannot yet make proteins that do anything on demand.

The First Approval: When and Who

No AI-discovered drug has been approved by the FDA as of May 2026. That statement sounds damning until you check the timeline: the first AI-discovered molecules entered Phase 1 in 2020. The average clinical development timeline is 7-10 years. We are exactly where we should be.

The most likely candidate for "first AI-discovered drug approved" is Insilico Medicine's ISM001-055 for idiopathic pulmonary fibrosis. It has positive Phase 2a data, is in Phase 2b, and IPF has accelerated approval pathways. If it succeeds, approval could come in 2028-2029.

The second candidate is Recursion's REC-994 for cerebral cavernous malformation (CCM), a rare genetic disorder with no approved therapies. REC-994's Phase 2 readout is expected in the second half of 2026. If positive, a Phase 3 trial would take 2-3 years.

The first AI-discovered drug approval will not be a breakthrough in the sense of curing a previously untreatable disease. It will be a validation of process — proof that AI-designed molecules can pass the same regulatory scrutiny as traditionally discovered drugs. That validation matters far more than the specific drug.

What History Teaches

The history of AI in biology from 2018-2026 contains a warning. Every wave has been overhyped at its peak — AlphaFold would "solve drug discovery" (it didn't), generative AI would "replace medicinal chemists" (it hasn't), protein design would "make any protein you want" (it can't, not yet).

The reality is more interesting. AI hasn't replaced any part of the drug discovery process. It has accelerated specific steps — target identification by 2-5x, hit finding by 10-100x, lead optimization by 3-10x. But clinical trials still take 7-10 years. The bottleneck has moved, not disappeared.

The companies that understand this — Insilico, Recursion/Exscientia, Schrödinger, Isomorphic Labs — are surviving the transition from hype to reality. The ones that promised more than acceleration could deliver — BenevolentAI (market cap from $2B to $200M), closed-down AI drug discovery startups — did not.

The lesson is not that AI in drug discovery was overhyped. It's that drug discovery is really hard, and AI made it somewhat less hard. That is not a revolution. It is something more durable: an improvement in the productivity of a critical industry.

---

Data Sources: Insilico Medicine press releases (2023-2026), DeepMind AlphaFold publications (Nature, 2018, 2021, 2024), Recursion Pharmaceuticals SEC filings, EvolutionaryScale ESM3 preprint (bioRxiv, 2024), Baker lab RFdiffusion and ProteinMPNN publications (Science, 2022-2023), ClinicalTrials.gov AI/ML interventional trials, Nobel Prize Committee press release (Chemistry 2024).

About the Author: Martin DAVILA is a bioeconomy analyst and the founder of Bioinfometrics.