For years we’ve been told that artificial intelligence will revolutionise drug discovery.
Recently, the CEO of Eli Lilly described how his company is investing heavily in AI, building enormous biological supercomputers and partnering with firms like NVIDIA.
But the reality is more complicated, and more interesting, than the hype.
AI didn’t suddenly arrive in drug discovery. It has been developing in research labs and biotech companies for decades, the public is only noticing now.
What we’re seeing today is simply the latest stage in a long technological evolution.
Phase 1: Computer Assisted Drug Design (1970s to 1990s)
Long before modern AI, scientists were already using computers to help design drugs.
This field became known as Computer Assisted Drug Design (CADD).
Researchers modelled how molecules might interact with proteins in the body. Instead of synthesising thousands of chemicals randomly, computers helped predict which compounds might bind to a biological target.
These techniques helped accelerate the development of targeted medicines, including early HIV protease inhibitors.
But the limitation was obvious.
Computers were only as good as the structural information scientists already had. If the protein structure wasn’t known, the models were largely guesswork.
Phase 2: High Throughput Screening (1990s to 2010)
The next revolution wasn’t really AI, it was automation.
Pharmaceutical companies built robotic systems capable of testing millions of compounds against biological targets.
This approach, known as high throughput screening, produced enormous datasets describing which molecules worked and which didn’t.
Ironically, this data explosion laid the foundation for the AI era.
Because machine learning needs one thing above all:
data.
Phase 3: Machine Learning Enters Drug Discovery (2010 to 2020)
During the 2010s, advances in machine learning allowed scientists to analyse chemical datasets far more effectively.
Instead of simply modelling molecular structures, algorithms could now:
predict binding affinity
estimate toxicity
forecast metabolism
rank candidate molecules
Companies such as Insilico Medicine, Exscientia and Recursion Pharmaceuticals began building AI systems trained directly on chemical and biological data.
Some of the first AI assisted drug candidates from these platforms have already entered clinical trials.
But the real breakthrough came from a completely different area of biology.
Phase 4: Foundation Models for Biology (2020 to Present)
In 2020, a system developed by DeepMind called AlphaFold solved one of the biggest problems in biology:
predicting how proteins fold.
Protein structure determines how drugs interact with biological targets. Before AlphaFold, determining these structures often required years of laboratory work.
AlphaFold made it possible to predict many structures in minutes.
This dramatically accelerated the earliest stages of drug discovery.
But it still only solves part of the puzzle.
The Real Problem: Biology Is Still Mostly Unknown
Even with modern AI tools, drug discovery remains incredibly difficult.
As Eli Lilly’s leadership has pointed out, scientists likely understand less than 15% of human biology.
A drug must successfully navigate an extraordinary number of biological hurdles:
absorption
distribution
metabolism
toxicity
off target effects
long term safety
This is why most drugs fail.
Historically, roughly 90% of drugs entering clinical trials never reach the market.
AI can help identify better candidates earlier. But it cannot remove biological uncertainty.
What AI Actually Changes
The real impact of AI isn’t replacing scientists.
It’s reducing the search space.
Instead of testing tens of thousands of molecules, researchers can use predictive models to focus on a much smaller set of promising candidates.
The workflow increasingly looks like this
1.AI proposes candidate molecules
2.robots synthesise them
3.automated laboratories test them
4.results feed back into the model
Each cycle improves the predictions.
Companies like Eli Lilly are now building continuous AI driven discovery systems combining supercomputers with automated laboratories.
Phase 5: Closed Loop Drug Discovery (2022 to Present)
The recent explosion of artificial intelligence has been driven by large language models (LLMs).
Examples include:
ChatGP
Gemini
Claude
These systems are trained on vast amounts of text and are extremely good at predicting language.
But language models do not understand chemistry or biology.
They learn patterns in human language, not the physical rules governing molecules.
Ask an LLM to write an essay and it performs brilliantly.
Ask it to design a drug molecule and it quickly runs into limitations.
The Rise of Scientific AI
Instead of relying on general language models, pharmaceutical companies are now building specialised AI systems trained directly on biological data.
These models learn from:
protein structures
molecular interactions
experimental screening data
pharmacology and toxicity results
AlphaFold’s ability to predict protein structures was a major step forward.
But it represents only one piece of the drug discovery puzzle.
The First AI-Designed Obesity Drug
Recently, one of the first obesity drugs designed with heavy AI assistance entered late stage clinical trials.
A biotechnology company called MindRank developed a drug known as MDR-001.
https://clinicaltrials.gov/study/NCT07110766
Unlike the injectable GLP-1 peptides dominating today’s market, MDR-001 is a small molecule GLP-1 receptor agonist designed to be taken orally.
AI systems helped generate and optimise the molecule by analysing massive datasets of chemical structures and biological activity.
The drug has now entered Phase III clinical trials, testing whether it can produce meaningful weight loss in humans.
If successful, it could represent one of the first medicines whose discovery was significantly accelerated by AI.
Why Obesity Drugs Are an Ideal Target for AI
The GLP-1 drug boom has created enormous incentives to discover better metabolic therapies.
Current blockbuster drugs such as:
Ozempic
Wegovy
Zepbound
have demonstrated that targeting metabolic hormones can produce dramatic weight loss.
But these drugs still have limitations:
injectable delivery
gastrointestinal side effects
complex manufacturing
AI systems are now being used to search for molecules that might deliver similar benefits with fewer drawbacks.
What the AI Drug Discovery Story Often Leaves Out
Most timelines make AI drug discovery sound like a straight line.
More data → better models → faster drugs.
Reality is more complicated.
The data quality crisis
Much of the historical data used in drug discovery is messy and inconsistent.
Companies are now generating new, standardised datasets using automated laboratories specifically designed for AI.
The ADMET problem
Designing a molecule that binds to a target is relatively easy.
Predicting how that molecule behaves inside a human body is much harder.
The clinical trial bottleneck
Even if AI designs a promising drug in days, proving safety and efficacy still takes years of human testing.
Discovery can accelerate.
Evidence cannot.
The IP battle
Courts have ruled that AI cannot be listed as a patent inventor.
So companies must still establish clear human inventorship for AI-assisted discoveries.
Why Peptides Are Perfect for AI Drug Design
Peptides sit in a sweet spot between chemistry and biology.
They are:
more complex than small molecules
simpler than full proteins
That makes them easier for AI systems to model.
Peptides are essentially sequences of amino acids biological “sentences”.
Example GLP-1 fragment
HAEGTFTSDVSSYLEGQAAKEFIAWLVKGR
Each letter represents an amino acid.
This structure is remarkably similar to language models.
Both are sequences of tokens.
Because of this similarity, AI systems can treat proteins and peptides almost like biological language.
The Bottom Line
AI is not about machines replacing scientists.
It’s about giving scientists more powerful tools to explore the enormous complexity of biology.
The combination of:
biological datasets
machine learning models
automated laboratories
supercomputing infrastructure
may gradually transform how medicines are discovered.
But even the pharmaceutical industry admits the biggest breakthroughs may still lie in the 2030s.
And when that future arrives, it may begin with molecules that look a lot like code.
Because peptides are essentially biological software.
And software is exactly the kind of system artificial intelligence excels at writing.
Comments