From Algorithm to Medicine: How AI Is Accelerating Drug Discovery

Drug discovery used to be a decades-long, billion-dollar race; generative AI for molecules and autonomous labs are turning it into a software problem.

October 28, 20266 min read

From Algorithm to Medicine: How AI Is Accelerating Drug Discovery

Pharmacology Enters the Realm of Algorithms

For decades, developing a drug was a marathon: 10 to 15 years of research, clinical trial failure rates above 90%, and costs that frequently exceeded $2 billion per approved drug. That model is starting to crack. Artificial intelligence has burst forcefully into the most uncertain phase of the process —the discovery of candidate molecules— and results are beginning to materialize in real trials, not just in scientific articles.

3D model of a protein structure displayed as a rotating model on a holographic screen, with candidate molecules floating around it — Protein structure prediction: the task AI solved first.

The turning point came with protein structural models. After the success of AlphaFold, which predicted the structure of more than 200 million proteins and earned Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry, the industry accepted that AI could solve biological problems that previously required years of crystallography. Now the question is no longer whether AI can contribute to drug discovery, but how much it can shorten each of its stages.

How a Data-Driven Laboratory Works

New systems combine three capabilities. First, generative models capable of proposing novel molecules with desired properties —affinity for a specific target, solubility, low toxicity— starting from natural language descriptions. Second, structure and protein-ligand binding prediction, which allows thousands of candidates to be discarded before synthesizing even one. And third, automated laboratory robotics, which closes the loop: the algorithm proposes, the robot synthesizes and tests, and the result feeds back into the model.

Robotic laboratory arm handling multi-well sample plates in an automated facility — Autonomous labs run thousands of assays per week, nonstop.

This approach, known as "closed-loop," has already produced the first generation of AI-generated clinical candidates that entered human trials in record time: in some cases, periods of 12 to 18 months from target selection to clinical candidacy were reported, compared to the usual 4–5 years with conventional methods.

AI does not replace the medicinal chemist or the pharmacologist: it frees them from mass screening to focus on decisions that require scientific judgment.

The Fronts Moving Fastest

The areas where AI is showing the greatest impact are:

Antibiotics against superbacteria: screening libraries of millions of compounds has made it possible to identify new antibacterial structures against resistant pathogens, a field that had been stagnant for decades.
Personalized oncology: models that analyze each patient's tumor molecular profile to prioritize drug combinations with a higher probability of response.
"Undruggable" targets: proteins long considered intractable —misfolded, with no obvious pockets— are now targeted with tailor-made molecules, such as algorithmically optimized proteolytic degraders.
Rapid-design vaccines: platforms that learned to model antigens during the pandemic have consolidated as tools for responding to new outbreaks.

Synthetic Biology, the Second Act of the Revolution

The advance is not limited to small molecules. Researchers at Stanford University and institutions such as SLAC and Arc Institute have introduced Boltz-2, an open-source biomolecular model that, in addition to predicting structures, estimates binding affinity —the property that determines whether a molecule has therapeutic potential or not—. Competing on that front, the AlphaFold 3 consortium and the open-source version Chai-Open Science are bringing equivalent tools to thousands of university labs without the resources of major pharmaceutical companies, democratizing a type of research previously reserved for a few.

Automating Science: Autonomous Labs

Where all of this converges is the autonomous lab. Facilities with robotic arms operating 24/7 run thousands of experimental assays per week, while an AI "scientific agent" plans the next experiment based on what it has learned. Coscientist, developed at Carnegie Mellon, demonstrated, for example, that it could design a chemical synthesis of ibuprofen from a request formulated in natural language. These systems do not make fewer mistakes than a human: they simply make them faster, and at that speed, making mistakes cheaply is a method.

What Pharmacological AI Still Lacks

However, it is worth tempering the enthusiasm. No clinical candidate generated entirely by AI has yet completed phase III of a trial, and the timelines AI cuts short (discovery) are a fraction of the path to market (clinical validation). There are also still insufficiently resolved problems:

False positives: generative models have proposed molecules that are theoretically synthesizable but useless or unstable in practice, which has motivated additional validation controls.
Historically biased data: large pharmacological datasets come disproportionately from populations and diseases of the wealthy world, and AI inherits these biases.
Interpretability: in healthcare, knowing that the model works is not enough: researchers and regulatory agencies demand to understand what makes a molecule work.

A Decade That Will Define 21st-Century Pharmacology

The United States, the European Union, and China have incorporated specific guidelines for the use of AI in pharmaceutical R&D, while major pharmaceutical companies are multiplying their agreements with AI laboratories: most large pharmaceutical companies now have alliances with companies specializing in AI-assisted discovery worth billions. The bet is double-edged: the same algorithms that search for new antibiotics could, in hostile hands, design dangerous compounds. That is why software biosecurity has been accompanying this debate since day one.

What is beyond doubt is that the funnel of drug discovery is widening at its base: there are more candidates, faster and cheaper. The next decade will tell whether that abundance translates into drugs that reach patients. That is, ultimately, the only metric that matters.