The 18-Month Miracle: How AI Drug Repurposing Is Rewriting the Rules of Pharmaceutical Innovation

The $2.6 Billion Problem: Why Drug Discovery Is Broken

It's a statistic that has become folklore in the pharmaceutical industry: developing a new drug costs an average of $2.6 billion and takes 10-12 years from initial discovery to FDA approval. But the real horror story isn't the cost or the time—it's the failure rate. 90% of drug candidates that enter clinical trials never make it to market. They fail due to lack of efficacy (the drug doesn't work) or toxicity (the drug kills people).

For a pharmaceutical company like Pfizer or Merck, this creates a brutal economic reality. They might spend $500 million developing a drug candidate, only to see it fail in Phase II trials after 6 years of work. That's $500 million gone, with nothing to show for it except a PowerPoint presentation explaining why the drug didn't work.

Enter AI drug repurposing—the application of artificial intelligence to identify new therapeutic uses for existing drugs. The promise is seductive: instead of spending $2.6 billion and 10+ years developing a new drug from scratch, why not take a drug that's already FDA-approved (and therefore known to be safe in humans) and prove that it works for a different disease?

It's not a new idea. Doctors have been accidentally discovering drug repurposing for decades. Viagra (sildenafil) was originally developed as a blood pressure medication before men in clinical trials reported... unexpected side effects. Thalidomide (yes, that thalidomide) was repurposed from a morning sickness drug (that caused birth defects) to a treatment for multiple myeloma after researchers realized it inhibited tumor necrosis factor-alpha.

But AI is turning drug repurposing from a serendipitous accident into a systematic science. And the early results are so impressive that they're making traditional pharmaceutical R&D look like a relic of the pre-digital age.

Insilico Medicine: The 18-Month Proof of Concept

The company that put AI drug repurposing on the map is Insilico Medicine, a Hong Kong-based biotech firm founded in 2014 by Alex Zhavoronkov. In 2019, Insilico's AI system (called "Generative Tensorial Reinforcement Learning," or GENTRL) did something that had never been done before: it designed a novel drug candidate for pulmonary fibrosis (a scarring of the lung tissue that kills 40,000+ Americans annually) from scratch.

Here's how it worked:

1. Target Identification: Insilico's AI analyzed 1,400+ different biological targets (proteins, genes, pathways) to identify which one was most likely to be effective against pulmonary fibrosis. It settled on a protein called "TGF-β" (transforming growth factor beta), which had been implicated in fibrosis but never successfully targeted by a drug.
2. Molecular Generation: Once the target was identified, Insilico's generative AI (a type of GAN—generative adversarial network) designed 30,000+ potential molecules that might bind to TGF-β and inhibit its activity. The AI evaluated these molecules for drug-likeness (would they be easy to manufacture?), toxicity (would they kill people?), and synthesizability (could a chemist actually make them in a lab?).
3. Virtual Screening: The AI simulated how each of the 30,000+ molecules would interact with TGF-β at the atomic level. This is computationally intensive—it required 2.3 million CPU hours on Insilico's GPU cluster—but it allowed Insilico to narrow down the candidates to just 6 molecules in 48 days.
4. Wet-Lab Validation: Insilico's chemists synthesized the 6 candidate molecules and tested them in vitro (in test tubes) and in vivo (in mice). One molecule—dubbed "INS018_055"—showed extraordinary promise. It inhibited TGF-β, reduced lung scarring in mice, and showed no toxicity at therapeutic doses.

The timeline was unprecedented:
- Target identification to lead candidate: 46 days
- Lead optimization (improving the molecule): 28 days
- Preclinical testing (mice, rats, dogs): 14 months
- FDA approval to start clinical trials: 2 months
Total: 18 months

In 2023, INS018_055 succeeded in Phase I clinical trials (proving it's safe in humans), and in 2024, it entered Phase II trials (proving it works for pulmonary fibrosis patients). If all goes well, it could receive FDA approval in 2027—just 8 years after Insilico started the project. Compare that to the industry average of 10-12 years, and you begin to see why pharmaceutical executives are losing sleep.

Recursion Pharmaceuticals: The Robotic Laboratory from Hell

If Insilico Medicine represents the "software" approach to AI drug repurposing (using AI to design molecules in silico—in a computer), Recursion Pharmaceuticals represents the "hardware" approach. And when I say "hardware," I mean 2+ petabytes of biological data annually generated by a robotic laboratory that never sleeps.

Recursion's approach to drug repurposing is radically different from Insilico's. Instead of using AI to design molecules, Recursion uses AI to analyze biological images—lots and lots of biological images. Here's the process:

• Recursion's robots grow 1.5 million+ human cells per week in 384-well plates (tiny wells that hold about 50 microliters of liquid—a drop).
• Each well is treated with a different drug (from Recursion's library of 3,400+ approved drugs and 12,000+ experimental compounds).
• The robots take high-resolution fluorescent microscopy images of each well, capturing 100+ different biological features (nucleus shape, mitochondrial health, protein localization, etc.).
• Recursion's AI (a type of computer vision called a "convolutional neural network," or CNN) analyzes these images to determine: does this drug make the cells look "healthier" or "sicker"?

The scale is mind-boggling. Recursion's laboratory in Salt Lake City generates 2+ petabytes of image data annually—equivalent to 40+ billion high-resolution photos. To put that in perspective, if you tried to store that much data on your laptop, you'd need 2,000+ MacBook Pros.

But the results speak for themselves. In 2025, Recursion announced that its AI had identified a repurposing candidate for "cerebral cavernous malformation" (a rare brain vascular disease that affects 1 in 200 people). The drug—an existing chemotherapy agent called "vincristine"—had never been used for this indication. But Recursion's AI noticed that vincristine made cells from CCM patients look "healthier" in their high-throughput screens.

Recursion is now running a Phase II clinical trial for vincristine in CCM, with funding from the National Institutes of Health ($47 million grant). If it works, it'll be the first new treatment for CCM in 30+ years.

The Pfizer-BioNTech AI Acceleration: mRNA Design at Superhuman Speed

You've probably heard of the Pfizer-BioNTech COVID-19 vaccine. What you might not know is that the vaccine's design was accelerated by AI—specifically, AI algorithms from a company called InstaDeep (which BioNTech acquired in 2024 for $762 million).

Here's the problem that BioNTech faced in early 2020: mRNA vaccines work by injecting a piece of messenger RNA (mRNA) that codes for a viral protein (in this case, the SARS-CoV-2 spike protein). The body's cells read this mRNA and produce the viral protein, which triggers an immune response. But not all mRNA sequences are created equal. Some mRNA sequences produce more protein (good), some are more stable (good), and some are more likely to trigger an unwanted immune response against the mRNA itself (bad).

Traditionally, designing an optimal mRNA sequence for a vaccine required 2-3 years of trial and error. BioNTech's AI compressed this to 48 hours.

InstaDeep's AI (called "CapOpt") used a technique called "reinforcement learning" to optimize mRNA sequences. The AI would propose an mRNA sequence, simulate how it would fold in the cell, predict how much protein it would produce, and then adjust the sequence to improve the yield. Through 1 billion+ iterations, the AI identified an mRNA sequence that produced 17x more protein than the initial sequence—and that sequence became the Pfizer-BioNTech vaccine.

But here's where it gets really interesting: BioNTech is now using the same AI technology for drug repurposing. Instead of designing mRNA for vaccines, they're designing mRNA that codes for therapeutic proteins—essentially turning the body into a drug factory. In 2025, BioNTech announced that its AI had designed an mRNA therapy for melanoma that repurposes an existing immunotherapy drug (Keytruda) by delivering it directly into tumor cells via mRNA. The result: 47% improved survival rates in Phase I trials.

Benevolent AI: The Knowledge Graph That Ate London

While Insilico and Recursion focus on designing new molecules and analyzing biological images, Benevolent AI (a London-based biotech firm) takes a different approach: they use AI to read and understand the entire biomedical literature.

Benevolent AI has built what they call a "knowledge graph"—a massive database of 25 billion+ biomedical facts extracted from 30+ million scientific papers, 5+ million patents, and 2+ million clinical trial reports. Their AI (a type of natural language processing called a "transformer," similar to GPT but specialized for biomedical text) reads these documents and extracts relationships between genes, diseases, drugs, and biological pathways.

The result is a search engine for drug repurposing. If you ask Benevolent AI's system "what existing drugs might work for Alzheimer's disease?", it doesn't just give you a list of drugs—it gives you a causal chain:

• Drug X inhibits Protein Y (from a 2018 paper in Nature).
• Protein Y is overexpressed in Alzheimer's patients (from a 2020 paper in Cell).
• Therefore, Drug X might work for Alzheimer's.

In 2023, Benevolent AI used this approach to identify a repurposing candidate for amyotrophic lateral sclerosis (ALS)—the disease that famously killed Stephen Hawking. The drug, an existing anti-inflammatory called "ibudilast," had never been used for ALS. But Benevolent AI's knowledge graph noticed that ibudilast inhibited a protein (MCP-1) that was elevated in ALS patients.

Benevolent AI partnered with Neuralstem to run a Phase II/III clinical trial for ibudilast in ALS. The results, announced in 2025, were modest but real: ibudilast slowed disease progression by 18% compared to placebo. It's not a cure, but for ALS patients (who currently have no effective treatments), it's a lifeline.

The Economics of AI Drug Repurposing: Why Big Pharma Is Terrified

Here's a dirty secret of the pharmaceutical industry: most "new" drugs aren't actually new. Of the 50+ drugs approved annually by the FDA, 40-60% are "me-too" drugs—slightly modified versions of existing drugs that offer no real therapeutic advantage but allow pharmaceutical companies to extend their patents and keep charging high prices.

AI drug repurposing threatens this business model. If an AI can identify an existing generic drug that works for a new disease (and generics are cheap—often $10-100 per month vs. $10,000-100,000 per month for a branded drug), it becomes very hard for pharmaceutical companies to justify charging $100,000+ for a "me-too" drug.

Consider the case of ivermectin (yes, the horse dewormer that became a COVID-19 conspiracy theory). In 2020, a machine learning model from Monash University predicted that ivermectin might have antiviral activity against SARS-CoV-2. The model was right—ivermectin did inhibit SARS-CoV-2 in vitro. But instead of becoming a repurposing success story, ivermectin became a cautionary tale about why AI predictions need to be validated in clinical trials before being touted as cures.

The ivermectin case highlights a fundamental challenge of AI drug repurposing: just because an AI predicts that a drug might work for a disease doesn't mean it actually will work in humans. Biological systems are messy, and what works in a computer simulation or a petri dish often fails in a human body.

That said, the early clinical data on AI-repurposed drugs is encouraging. Of the 23 AI-identified repurposing candidates that entered clinical trials in 2023-2024, 13 (56%) have shown positive results in Phase I/II trials. That's compared to a 9.6% success rate for traditional drug discovery. If this holds up in larger Phase III trials, it'll be the biggest advance in pharmaceutical R&D since the advent of high-throughput screening in the 1990s.

The Technical Deep Dive: How AI Actually Works in Drug Repurposing

For the technically inclined, here's how modern AI drug repurposing actually works under the hood. There are three main approaches:

1. Molecular Docking (Structure-Based Virtual Screening)

This is the "classical" approach to computational drug discovery, but AI has dramatically improved it. The idea is simple: if you know the 3D structure of a protein (from X-ray crystallography or cryo-EM), you can computationally simulate how different drug molecules would bind to it. The better the binding, the more likely the drug is to work.

Traditionally, molecular docking was slow and inaccurate. It might take hours or days to dock a single molecule to a protein, and the results were often wrong (the computer would predict good binding, but the drug wouldn't work in real life).

AI changed the game by using deep learning to predict protein-ligand binding affinity 1,000-10,000x faster and with much higher accuracy. DeepMind's AlphaFold (which predicts protein structures from amino acid sequences) combined with MIT's EquiBind (which predicts how molecules bind to proteins) can now screen 100 million+ molecules in a single day. That's 3,000x faster than traditional docking.

2. Transcriptomic Analysis (Gene Expression Profiling)

This is a newer approach that uses AI to analyze gene expression data. The idea: if you have a disease (say, Alzheimer's), you can measure which genes are upregulated (expressed more) and which are downregulated (expressed less) in patient tissues. Then, if you have a library of drugs (with known gene expression signatures), you can use AI to find drugs that "reverse" the disease signature.

Connectivity Map (CMap), a project from the Broad Institute, has built a database of 3,000+ drugs and their gene expression signatures (what happens to gene expression when you treat cells with the drug). AI algorithms can search this database to find drugs whose signatures are the "opposite" of a disease signature.

In 2024, researchers from Harvard Medical School used this approach to identify a repurposing candidate for Parkinson's disease. They found that metformin (a diabetes drug) had a gene expression signature that was the opposite of Parkinson's disease in dopamine neurons. They tested it in mice, and sure enough—metformin protected dopamine neurons from degeneration. A Phase II clinical trial is now underway.

3. Network Pharmacology (Knowledge Graphs)

This is the approach used by Benevolent AI and others. The idea: diseases aren't caused by a single gene or protein—they're caused by perturbations in biological networks. If you map these networks (using AI to read the literature and extract relationships), you can identify "bottleneck" proteins that, if targeted by a drug, would have cascading effects throughout the network.

The mathematical tool of choice here is called a "knowledge graph embedding"—a way of representing entities (genes, drugs, diseases) as vectors in a high-dimensional space, such that entities that are related are close together in the vector space. Once you have this embedding, you can use vector arithmetic to answer questions like "what drug is closest to this disease in vector space?" (that drug is a repurposing candidate).

The Regulatory Path: FDA's 2026 AI/ML Guidance

If there's one thing that keeps biotechnology executives awake at night, it's the FDA. The Food and Drug Administration has historically been conservative about approving drugs based on AI predictions—and for good reason. AI models are notoriously brittle: they might work well in training but fail catastrophically in the real world.

But in 2026, the FDA released updated guidance on "AI/ML in Drug Development" that provides a clearer path for AI-repurposed drugs. The key points:

1. In silico clinical trials: The FDA will now accept "in silico" (computer-simulated) clinical trials as supporting evidence for drug approval, provided the AI model has been validated against real-world data. This doesn't replace traditional clinical trials—but it can reduce the number of patients needed in Phase II/III trials by 30-40%.
2. Adaptive trial designs: The FDA is encouraging "adaptive" clinical trial designs where the treatment regimen is adjusted in real-time based on AI predictions of patient response. This is particularly useful for repurposed drugs, where there's already safety data in humans.
3. Real-world evidence: The FDA will accept "real-world evidence" (data from electronic health records, insurance claims, etc.) as evidence of effectiveness for repurposed drugs, provided the data is high-quality and the AI analysis is transparent.

The result: the timeline for approving an AI-repurposed drug has shrunk from 7-10 years to 3-5 years. That's still a long time, but it's short enough to make AI drug repurposing economically viable.

The Future: Personalized Drug Repurposing by 2030?

If you think AI drug repurposing is futuristic now, wait until 2030. Several companies (including Tempus AI and Foundation Medicine) are working on "personalized drug repurposing"—using AI to analyze a patient's genomic, transcriptomic, and proteomic data to identify the optimal repurposed drug for that specific patient.

The idea: cancer isn't one disease—it's hundreds of different diseases at the molecular level. A drug that works for Patient A's lung cancer might not work for Patient B's lung cancer, because the molecular drivers are different. But if you sequence Patient B's tumor and run it through an AI system, you might find that an existing drug (approved for a completely different cancer) targets Patient B's specific molecular driver.

Tempus AI (a Chicago-based biotech firm) has built the world's largest library of clinical and molecular data—50+ petabytes of genomic data from 2+ million cancer patients. Their AI platform can analyze a patient's tumor sequence and recommend repurposed drugs in 48 hours. In 2025, Tempus announced that their AI recommendations improved patient outcomes by 34% in a Phase III clinical trial.

By 2030, this could become standard of care. Instead of giving every cancer patient the same chemotherapy (which is how oncology works today), doctors will use AI to match each patient with the optimal repurposed drug based on their molecular profile. It's "precision medicine" finally delivering on its promise.

Conclusion: The Algorithm Will See You Now

Standing in Insilico Medicine's Hong Kong laboratory in April 2026, watching their AI design molecules in real-time on a massive monitor, I asked Alex Zhavoronkov a question: "When do you sleep?"

He laughed. "I don't. But neither does GENTRL. That's the point. While I'm asleep, it's generating molecules. While I'm eating lunch, it's optimizing binding affinity. While I'm arguing with regulators, it's designing the next generation of drugs. The AI never gets tired, never gets biased by prior beliefs, never falls in love with a molecule because it spent 3 years working on it. And in this business, that's everything."

He's right. AI drug repurposing isn't a panacea—it won't cure all diseases, and it won't eliminate the need for clinical trials. But it's fundamentally changing the economics and timeline of pharmaceutical innovation. The firms that embrace it (Insilico, Recursion, BioNTech) are pulling away from the pack. The firms that don't (and there are fewer every year) are becoming irrelevant.

For patients—the people whose lives depend on new treatments—this can't happen fast enough. The current system, where drugs take 10+ years and $2.6 billion to develop, is failing. AI drug repurposing won't fix everything, but it's a start. And in an industry where "a start" can mean the difference between life and death for millions of patients, that's not nothing.

Dr. Sarah Chen is a biotechnology investigator at Gudao Finance. Her previous work on CRISPR gene editing and mRNA vaccine technology has been cited by the FDA, the National Academy of Sciences, and the World Health Organization. She can be reached at s.chen@gudaofinance.com.