From Serendipity to Search
Drug discovery has historically been a painfully slow, expensive, and failure-prone process. The average new drug takes 10-15 years and costs $1-2.6 billion to bring to market. About 90% of drugs that enter clinical trials fail. Penicillin was discovered because Alexander Fleming left a petri dish out by accident. Viagra was originally developed as a heart medication. The industry has been looking for a better way for decades, and AI might finally be delivering one.
AlphaFold and the Protein Structure Revolution
The most visible AI success in biology is DeepMind’s AlphaFold. The first version (2018) won the Critical Assessment of Structure Prediction (CASP) competition. AlphaFold 2 (2020) essentially solved the 50-year-old protein folding problem — predicting a protein’s 3D structure from its amino acid sequence with accuracy competitive with experimental methods like X-ray crystallography. AlphaFold 3, released in 2024, extended this capability to predicting the structures of proteins interacting with DNA, RNA, small molecules, and ions — essentially modeling the molecular interactions that underlie most drug mechanisms.
The practical impact is significant but nuanced. Knowing a protein’s structure doesn’t automatically give you a drug. It gives you a target — a protein you want to inhibit, activate, or otherwise modulate. Drug discovery still requires finding a molecule that binds to that target with high specificity, can be manufactured at scale, survives digestion, reaches the right tissue, and doesn’t cause unacceptable side effects. AlphaFold narrows one stage of a multi-stage process. It doesn’t replace the process.
As of mid-2025, the AlphaFold Protein Structure Database contains over 200 million predicted protein structures, freely available to researchers worldwide. More than 1.8 million researchers have accessed it. The downstream impact on drug discovery is still unfolding, but early studies suggest structure-based drug design timelines have been compressed by months in some cases.
AI-Native Drug Discovery Companies
Several companies are trying to use AI not just for one step of the process but for end-to-end drug discovery:
Insilico Medicine, based in Hong Kong and New York, is perhaps the furthest along. Their lead AI-discovered drug candidate, INS018_055 for idiopathic pulmonary fibrosis, entered Phase II clinical trials in 2023 — the first AI-discovered drug to reach this stage. The entire process from target discovery to preclinical candidate took about 18 months and cost roughly $2.6 million, compared to industry averages of 4-5 years and tens of millions. Whether it proves effective in humans is the real test.
Recursion Pharmaceuticals (NASDAQ: RXRX) takes a different approach — high-throughput automated biology combined with machine learning. They run millions of cellular experiments per week, imaging cells treated with thousands of compounds, and use computer vision to detect subtle morphological changes. It’s essentially “phenotypic screening at scale, with AI pattern recognition.” Recursion went public in 2021 and acquired two AI drug discovery companies (Cyclica and Valence) in 2023.
Isomorphic Labs, DeepMind’s drug discovery spinout (led by Demis Hassabis), raised a reported $200+ million in 2024 and signed partnerships with Eli Lilly and Novartis worth up to $3 billion combined. The premise: apply AlphaFold’s structural biology insights and DeepMind’s general AI capabilities to the entire drug discovery pipeline.
Real FDA Approvals and Clinical Results
As of 2025, no drug discovered primarily by AI has received full FDA approval. Several AI-designed molecules are in clinical trials:
- Insilico’s INS018_055 (Phase II, idiopathic pulmonary fibrosis)
- BenevolentAI’s BEN-2293 (Phase II, atopic dermatitis, later discontinued due to lack of efficacy — an AI failure that illustrates the gap between computational success and clinical reality)
- Exscientia’s EXS-21546 (Phase I, oncology — note: Exscientia’s former CEO was fired in 2024 for inappropriate relationships, a reminder that AI companies still have human governance problems)
The attrition rate for AI-discovered drugs in clinical trials is still uncertain — the sample size is too small. Early data suggests the rate is comparable to traditional drug discovery, which means AI hasn’t yet solved the fundamental challenge: biology is complicated, and mice lie. (The industry adage: “We can cure cancer in mice. Humans are harder.”)
Limitations and Realistic Expectations
AI’s biggest contribution to drug discovery so far isn’t creating new drugs — it’s eliminating bad candidates faster. Virtual screening with AI models can predict which molecules are likely to be toxic, poorly absorbed, or metabolically unstable before anyone synthesizes them in a lab. This “fail fast, fail cheap” approach is genuinely valuable even if it doesn’t generate headlines.
The harder problems remain: understanding disease biology at a systems level (most diseases involve complex networks of interacting genes and proteins, not a single target), predicting clinical efficacy from preclinical data (the “valley of death” between lab and clinic), and dealing with the sheer combinatorial complexity of chemical space — there are an estimated 10^60 drug-like molecules, more than the number of atoms in the solar system. AI can search this space more efficiently than humans, but it can’t search all of it.
The most realistic near-term impact of AI in drug discovery: faster identification of drug candidates for well-understood targets, better prediction of toxicity and ADME (absorption, distribution, metabolism, excretion) properties, and expanded use of drug repurposing — finding new uses for existing drugs, where safety data already exists. AI won’t replace the painstaking, decade-long process of bringing a new drug to market, but it might make that process 20-30% faster and cheaper. In an industry where a single successful drug can generate billions in revenue, that’s worth pursuing.