AI in Drug Discovery: From AlphaFold 3 to the First Clinical Approvals

Modern drug development resembles searching for a needle in a vast probability space. The number of synthesizable small molecules is estimated at around 10⁶⁰; finding, in this immense chemical space, a molecule that binds reliably to a disease target, remains stable in the body, and is not toxic is — by classical methods — the work of years and enormous budgets. By a commonly cited industry estimate, only a small fraction of candidates that enter development become approved drugs; that high failure rate is the root cause of cost and timeline. AI is changing this equation on two fronts: understanding the structure of the target and designing the molecule that will fit it.

Solving Structure: The AlphaFold Revolution

Knowing the three-dimensional structure of the protein a drug will act on is like seeing the inside of a lock before designing a key to open it. For decades, such structures were determined by painstaking experimental methods — X-ray crystallography or cryo-electron microscopy — one at a time and over months of work.

DeepMind's solution to much of the protein-folding problem with AlphaFold 2 in 2020 was a turning point. In May 2024, DeepMind and its sister company Isomorphic Labs introduced AlphaFold 3. Where earlier versions modeled only proteins, AlphaFold 3 uses a new diffusion-based architecture to predict the complex interactions of proteins together with DNA, RNA, small molecules (ligands), ions, and antibodies.

This jump in accuracy translates directly into more reliable predictions of binding sites and interaction energies, accelerating the target validation and drug optimization steps. In November 2024, DeepMind released the academic code and model weights of AlphaFold 3 — with commercial use restrictions — to the research community.

Designing the Molecule: Generative Chemistry and De Novo Design

Knowing the target's structure is only half the equation. The other half is producing a molecule that fits the target from scratch. This is where generative AI models come in. Trained on millions of known molecules and interactions, they can "imagine" novel candidate molecules with desired properties (binding to a specific target, solubility, low toxicity).

On the protein-engineering side, the Baker Lab's RFdiffusion is a pioneer of this approach. The diffusion-based model can generate realistic protein backbones and the building blocks of functional motifs, enabling the design of experimentally validated binders and functional proteins. RFdiffusion2, introduced in 2025, can design enzyme active sites with atom-level functional-group constraints: in a benchmark, it produced scaffolds for all 41 active sites; the previous method succeeded for only 16, and the new model identified active enzyme candidates by testing fewer than 96 sequences.

Proof of Promise: The First Generative-AI Drug in the Clinic

The real test of all this technology is whether an AI-designed molecule actually helps patients. The most important milestone in this area came from Hong Kong-based Insilico Medicine.

The company's Rentosertib (code names INS018_055 / ISM001-055) is a first-in-class TNIK inhibitor designed for idiopathic pulmonary fibrosis (IPF) — a progressive and fatal scarring disease of the lungs. Both the target on which the drug acts and the molecule that binds to it were identified with the help of generative AI.

On 3 June 2025, the first industry clinical evidence of AI-based drug discovery was published in Nature Medicine. In a randomized, double-blind, placebo-controlled Phase IIa study at 29 centers in China, patients receiving 60 mg of Rentosertib daily showed a mean improvement of +98.4 mL in forced vital capacity (FVC, the marker of lung function), while the placebo group declined by a mean of −20.3 mL.

This result is regarded as the first randomized evidence that an AI-designed molecule can produce a meaningful biological effect not only in the lab but in real patients. Insilico announced in 2025 that it was preparing for a Phase IIb study to confirm efficacy in a larger patient group. It is important to remember that the primary aim of the trial was safety and that Phase IIa was conducted in a relatively small patient group; larger phase studies will be decisive for efficacy claims.

The Three Key Links in AI Drug Discovery

Limits and Realistic Expectations

Excitement is justified, but careful reading is essential. AI today markedly accelerates the preclinical stage — target identification and molecule design — but the longest, most expensive, and riskiest part of drug development, clinical phase trials, still requires the same rigorous, multi-year processes. Predictions from models such as AlphaFold 3 also do not replace experimental validation; they guide and prioritize it. Another important point is that the performance of these models can vary on rare or unprecedented structures and on flexible/disordered proteins.

Even so, the direction is clear: AI is turning drug discovery from "trial-and-error art" increasingly into "design engineering." Rentosertib's Phase IIa results show that this transformation is no longer a theoretical promise but a clinically tested reality.