- AI in Lab Coat
- Posts
- What Makes a Drug “AI-Developed”? Here's Where Things Stand in 2025
What Makes a Drug “AI-Developed”? Here's Where Things Stand in 2025
I'm Building a Database of AI Drug Candidates. It's More Confusing Than You'd Think.

Developing a new drug from an initial idea to market can take anywhere from 10 to 15 years and cost $1 to 2.8 billion, and still have a failure rate of around 90%. But now with the rise of new artificial intelligence (AI) drug development startups, timelines and costs are being cut dramatically. From identifying the right biological targets to designing molecules, predicting preclinical outcomes, optimizing clinical trials, and even assisting in regulatory review, AI tools are accelerating each phase of the drug development pipeline.
To be clear, no drug has been fully discovered, developed, and approved purely by AI, at least not yet. Ever since AI entered drug discovery, there’s been a rush to claim the title of “first AI-drug”. And as the excitement around this grows, the definition of an AI-drug has also blurred.
How much AI involvement counts? A drug designed completely from nothing by AI? An existing drug repurposed by AI predictions? How about one flagged for toxicity risk? Or does it include cases where AI sped up clinical trials through patient matching or trial monitoring?
What’s clear, though, is that AI is enhancing the drug discovery pipeline. Instead of overhauling the industry, it’s making traditional processes faster and more efficient. We’re not at the point where AI is churning out new drugs on demand, but it’s getting closer. The question is, how far has it actually come, and what’s next?
Speeding Up Drug Target Identification
Finding the root of a disease, whether it’s a protein, gene or some other biological target, is a critical first step. Instead of manually combing through massive genomic, proteomic and clinical datasets we can use AI to look for key patterns to spot potential drug targets. More importantly, machine learning models can identify outliers that would otherwise go unnoticed.
One analysis found that in 22% of drug projects, AI helped uncover connections that led to fewer dead-end experiments, allowing resources to be focused on promising leads instead.
We can see a great example of this during the COVID-19 outbreak, when researchers at BenevolentAI used an AI knowledge graph to quickly identify baricitinib, an already approved arthritis drug, as a potential treatment. Their AI model identified a relevant pathway in just days, something that would have taken months of manual research and lab testing. This fast-tracked identification led to clinical trials and was successfully approved for emergency use authorization.

A Faster Way to Design and Test New Drugs
Once a target is identified, the next challenge is finding or designing a molecule that can actually act on it. Traditionally, chemists would screen tens of thousands of compounds in the lab, hoping to find a few promising “hits,” a process that is both tedious and costly. AI accelerates this by running virtual screening experiments, narrowing down potential hits in days instead of months.
Machine learning models can scan large chemical libraries in silico, predicting which compounds are most likely to bind effectively and have drug-like properties. Not only that, while screening for potential candidates, these models can also flag for potential toxicity early, again protecting researchers from pursuing false leads.
Major pharma companies like Bayer, Roche, and Pfizer are already using AI platforms to speed up hit discovery, particularly in cancer and heart disease research. But the evolution of these AI models isn’t just screening known compounds, eventually, they will learn to design entirely new ones.
Generative AI models can suggest novel chemical structures with desirable properties, refining them through quick iterations far faster than traditional chemistry. This approach dramatically speeds up optimization, where lead candidates can be algorithmically tested long before they reach the lab.
Take Insilico Medicine’s AI-designed fibrosis drug, for example. Their model generated novel fibrosis drug candidates in just 46 days, a process that would normally take 2 to 4 years. Similarly, Exscientia and Sumitomo Dainippon Pharma used AI to identify DSP-1181, a lead candidate for obsessive-compulsive disorder (OCD), in under 12 months.
These are fantastic examples of AI’s potential in drug design, but the technology is still maturing. Insilico’s GENTRL model, which designed the fibrosis candidates, was a proof-of-concept rather than a direct pipeline to a new drug. To my knowledge, none of those candidates progressed to clinical trials. Instead, the model was later incorporated into their commercially available Chemistry42 platform.
As for Exscientia’s DSP-1181, it entered a Phase I clinical trial in Japan in early 2020. But by 2022, development was discontinued. The company itself was later acquired and merged into Recursion Pharmaceuticals in 2024.
Even with these setbacks, AI in drug discovery is still pushing forward at an incredible pace. The models are getting better, the datasets are growing, and researchers are figuring out what actually works. Both Insilico Medicine and Recursion are still major players in AI-driven drug discovery, with plenty of candidates in the pipeline.
Making Clinical Trials Faster and Smarter
Having a drug candidate make it to clinical trials is a major achievement, but it’s also one of the biggest bottlenecks in the drug development pipeline. This stage typically takes six to seven years to complete. Recruiting the right patients, managing resources, and dealing with dropouts all add to the delays and costs.
Patient recruitment is where AI can make the most significant impact in expediting the trials. Finding patients who meet the strict eligibility criteria is the cause of most trial delays. Around 86% of trials miss their enrolment timelines, and nearly a third of Phase III trials fail due to enrolment issues.
AI can help with recruitment by mining electronic health records and other real-world data to find eligible patients faster. More than that, it can find sub-populations that are most likely to respond to treatment, improving trial design and increasing the chances of success.
AI is also being used to address patient adherence to the program. In a schizophrenia trial, an AI-powered mobile app (AiCure) used a smartphone camera to confirm medication intake, increasing adherence by 25%.
Another promising area being explored is synthetic control arms. Instead of recruiting a whole placebo group, AI attempts to use historical patient data to build a virtual one. This doesn’t mean completely removing entire control groups, but it could significantly reduce the number of participants needed to make one, speeding up trials and lowering costs.
The idea of replacing real control groups with past patient data sounds terrifying and maybe a little too futuristic, but it’s already happening. With massive datasets available, AI can accurately match historical patients to current trial participants, making synthetic control arms a legitimate alternative to traditional placebo groups.
AI will certainly not replace clinical trials; we’ll always need real human data to prove a drug works and probe for any potential adverse effects. But it’s making trials faster and more likely to succeed. And when you consider how much time and money is lost on failed trials, the value of AI shows up across the pipeline, not just at the discovery phase.
When is a Drug Truly an “AI-Drug”?
We started by noting how unclear the industry still is about what makes a drug AI-developed. We talked about baricitinib, probably one of the earliest and most cited examples of an AI-drug. Originally approved for arthritis, it was later identified by AI for emergency use in COVID-19 treatment. Here, AI didn’t design or invent the drug itself, but without AI’s insight, this new application might never have been spotted. Does this count as an “AI-developed” drug, or simply an AI-assisted repurposing?
Here’s another repurposing example: IGALMI (dexmedetomidine). BioXcel Therapeutics didn’t create a new molecule from scratch; instead, they used their AI platform to repurpose an existing drug (previously given via IV in hospitals for sedation) into a convenient oral film for treating agitation (associated with schizophrenia or bipolar disorder). It’s on the market now, FDA-approved in 2022. AI clearly accelerated this drug’s path to market, but once again, can we say it’s an AI-drug?
These examples show that the line between AI-assisted and AI-developed drugs isn’t clear yet. Maybe it never will be because, at some point soon, every new drug might be assisted by AI in some way. Given how quickly and deeply AI is embedding itself into every step of drug discovery (directly or indirectly), it’s entirely possible we’ll soon reach a stage where “AI-developed” drugs are just… drugs.
Still, AI’s impact is undeniable. As AI becomes further integrated into the drug discovery process, we’ll need a clearer definition of what an AI-drug really is. Not just for new drugs that materialize at the press of a button powered by algorithms but also for cases where AI plays an integral supporting role, like speeding up clinical trials or handling mountains of paperwork, basically making the whole drug development process less painful. Either way, the future of drug discovery looks a lot like AI, whether we officially call it that or not.
How Close Are We, Really?

My own research into AI-driven drugs has given me a clear look at exactly where things stand right now. So far, I’ve manually tracked 389 AI-driven drug candidates from 88 AI drug discovery companies, and that’s just the ones with active drug programs. I’m still discovering new companies regularly, so these numbers aren’t final. Plus, there are many other startups without disclosed candidates that aren’t yet in my database.
The reality is, despite hundreds of promising AI-generated leads, we haven’t seen a fully AI-discovered drug gain full regulatory approval, not yet anyway. But the sheer volume of candidates now progressing through clinical stages suggests it’s only a matter of time.
I’ll keep watching closely. This is an area of substantial personal interest to me, tracking each candidate’s progress, seeing which ones falter, and especially noting which will finally cross the finish line. While the drug discovery industry is still figuring out what counts as truly “AI-developed,” my database and this ongoing research will help me (and you) better understand AI’s real-world impact.
The first fully AI-designed drug approved for patients might still be around the corner, but at this rate, that corner is getting very close.
Access the public list of AI-Drug candidates database.
![]() | Got feedback or raised funds? Drop a comment—we’re always tracking the latest in AI healthcare. |
Reply