How Generative AI is Compressing the Decades-Long Drug Discovery Timeline

A breakthrough Phase 2a clinical trial published in 2025 has demonstrated that a drug discovered and designed entirely by generative artificial intelligence can produce measurable therapeutic signals in patients with idiopathic pulmonary fibrosis. The journey of a new molecule from the laboratory bench to the pharmacy shelf traditionally mirrors the time it takes for a child to graduate from primary school. This process typically spans 10 to 17 years, with costs ranging from hundreds of millions to over two billion dollars per successful drug, while the vast majority of candidates fail during clinical trials.¹

Artificial intelligence is not a magic wand that completes this journey overnight, but two fundamental shifts have occurred in recent years. First, predicting the three-dimensional shape of proteins—an experimental task that once took years—can now be performed in minutes through computational modeling. Second, generative models are now capable of both proposing biological targets and designing the precise molecules required to bind to them. A 2025 study in Nature Medicine moved these laboratory promises into the realm of clinical reality by showing a measurable response in human patients.²

Which parts of drug development does AI actually accelerate?

Artificial intelligence accelerates the "pre-clinical" phase of drug development, specifically target identification and molecular scaffolding, but it does not yet bypass the lengthy human safety trials required by regulators. Traditionally, scientists had to conduct massive screening campaigns to determine which protein to target and which chemical structure would fit its binding pocket. Today, AI models utilize omics data and vast literature to generate target lists, while structure-based engines draw candidate molecules from scratch. This compresses the early discovery window from the traditional 2.5 to 4 years down to approximately 12 to 18 months.

The release of AlphaFold 2 in 2021 was a watershed moment, as it predicted protein structures with near-experimental accuracy based solely on amino acid sequences.³ For many projects, this inverted the workflow: rather than waiting for a physical crystal structure to be solved, researchers could establish a binding hypothesis within minutes. While this does not guarantee an approved drug, the difference between blind screening and designing a molecule for a visible "pocket" provides a significant head start in the early stages.

What are the clinical results of the AI-designed drug rentosertib?

Rentosertib (INS018_055), a small molecule designed by an AI platform to treat idiopathic pulmonary fibrosis (IPF), demonstrated safety and a positive signal in lung function during a randomized Phase 2a trial. IPF is a condition characterized by the progressive scarring of the lungs; current treatments only slow the decline rather than reversing it. Using a generative AI platform, researchers identified the kinase TNIK as an anti-fibrotic target and designed rentosertib to fit it. This process took roughly 18 months from target discovery to the nomination of a preclinical candidate.⁴

In June 2025, Nature Medicine published the results of a double-blind, placebo-controlled trial involving 71 patients across 22 centers. While the primary endpoint was safety—where adverse events were found to be similar across groups—secondary endpoints showed a notable signal. The Forced Vital Capacity (FVC), a measure of lung air volume, increased by an average of 98.4 ml in the highest dose group (60 mg once daily), whereas the placebo group saw a decline of 20.3 ml.²

This data does not imply that the fibrosis was reversed, as the sample size was small and the duration short. However, for the pharmaceutical industry, it represents a milestone: a candidate chosen and designed by generative AI has transitioned from a headline to a peer-reviewed Phase 2a result with a dose-dependent response in lung function. The researchers are now preparing for larger, longer-term Phase 3 trials, which are necessary for regulatory approval.

What are the current limitations of AI in pharmacology?

The limitations of AI in drug development involve the scale of clinical validation and the necessity of human-led biological testing. While AI can propose a target and a chemical scaffold, it cannot yet predict with 100% certainty how a molecule will be metabolized in the human liver or how it will affect long-term safety over years of use. Tasks such as synthesis, toxicology, clinical protocol design, and ethics committee navigation remain firmly within the domain of human experts and institutions.

Furthermore, while AI reduces the time spent in early discovery, it does not yet shorten the years required for Phase 3 clinical trials, which are the most expensive and time-consuming part of the process. The true promise of AI lies in efficiency: reducing the amount of capital wasted on incorrect targets and poor molecules, thereby bringing high-potential candidates to trial faster. The case of rentosertib is not a miracle, but it is a measurable shift in the tempo of medical innovation.

Notes and Sources

1. Wouters OJ, McKee M, Luyten J. Estimated Research and Development Investment Needed to Bring a New Medicine to Market, 2009-2018. JAMA. 2020.
2. Ren F et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nature Medicine. 2025.
3. Jumper J et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021.
4. Ren F et al. A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models. Nature Biotechnology. 2024.