AI in Drug Discovery: From Lab to Patient in Record Time

Here’s a number that should make you uncomfortable: bringing a single new drug to market takes, on average, 10 to 15 years and costs roughly $2.6 billion. That’s not a typo. Billions. And for every drug that makes it through, dozens fail somewhere along the pipeline – burned through years of research, animal trials, and early human studies before someone realizes the molecule just doesn’t work the way they hoped. The pharmaceutical industry has been living with this brutal math for decades.

But something is shifting. Not gradually, not politely – fast. Artificial intelligence is rewriting the playbook for how we discover, design, and deliver new medicines. And the early results are hard to ignore.

The AlphaFold Moment

When DeepMind released AlphaFold2 in late 2020, structural biologists collectively lost their minds – and for good reason. The system could predict 3D protein structures with accuracy rivaling experimental methods that take months or years per protein. Within two years, AlphaFold had mapped the predicted structures of over 200 million proteins, essentially the entire known protein universe.

Why does that matter for drug discovery? Because understanding a protein’s shape is the first step to designing a molecule that can interact with it. Before AlphaFold, researchers might spend an entire PhD thesis solving a single protein structure through X-ray crystallography. Now, a graduate student can pull up a predicted structure during their morning coffee and start designing potential drug candidates before lunch.

That’s not hyperbole. It’s the kind of acceleration that changes who gets to participate in drug discovery and how quickly initial hypotheses can be tested.

AI-Designed Drugs Are Already in Clinical Trials

This isn’t theoretical anymore. Real companies are pushing AI-designed drug candidates through human trials right now.

Insilico Medicine made headlines when their AI-discovered molecule, INS018_055, reached Phase 2 clinical trials for idiopathic pulmonary fibrosis – a devastating lung disease with limited treatment options. The kicker? They went from target identification to a preclinical candidate in about 18 months. Traditional pharma would typically need four to six years for that stretch alone.

Recursion Pharmaceuticals is taking a different but equally ambitious approach. They’ve built a massive biological dataset – billions of images of cells treated with various compounds and genetic perturbations – and turned machine learning loose on it. Their pipeline now includes candidates for rare diseases, oncology, and inflammation, several of which are in clinical stages.

Then there’s Absci, using generative AI to design novel antibodies from scratch. Not tweaking existing ones – creating them. They reported designing and validating therapeutic antibodies in as little as six weeks, a process that conventionally takes 6 to 12 months.

Where AI Fits in the Pipeline

It helps to understand where AI is actually making a difference, because it’s not just one thing. The drug discovery pipeline has distinct phases, and machine learning is infiltrating nearly all of them.

• Target identification: ML models sift through genomics data, scientific literature, and patient records to identify which proteins or pathways are most likely to be involved in a disease. What used to require teams of biologists reading thousands of papers can now be narrowed to a ranked list of candidates in weeks.
• Molecular design: Generative models – think of them as the drug discovery equivalent of image generators – can propose novel molecular structures optimized for specific properties like binding affinity, solubility, and synthesizability. They explore chemical spaces far too vast for humans to search manually.
• Toxicity prediction: One of the biggest reasons drugs fail in clinical trials is unexpected toxicity. ML models trained on historical safety data can flag potential problems before a compound ever touches a living cell. Atomwise, for instance, uses deep learning to predict off-target binding that might cause dangerous side effects.
• Clinical trial optimization: AI is helping design smarter trials – identifying the right patient populations, predicting enrollment challenges, and even optimizing dosing regimens. Unlearn.AI creates “digital twins” of control-group patients, potentially reducing the number of people who need to receive placebos.

The Superpower Framing

I’ve talked to a few medicinal chemists about how AI is affecting their day-to-day work. The consensus is surprisingly consistent: AI hasn’t replaced chemists; it’s given them superpowers.

The experienced chemist still brings intuition that no model captures well – decades of understanding about what’s practical to synthesize, what’s likely to be metabolically stable, what will actually work in a living organism rather than just in a simulation. But the volume of possibilities they can now explore has expanded by orders of magnitude.

A senior researcher at a mid-size biotech told me their team used to evaluate maybe 200 compound variations per project. Now, with generative chemistry tools, they routinely assess 10,000 or more computationally before selecting which ones to make in the lab. The hit rate hasn’t changed dramatically – but the starting pool is so much larger that they’re finding better leads, faster.

The Regulatory Question

Here’s where things get complicated. The FDA, EMA, and other regulatory bodies weren’t built to evaluate drugs designed by algorithms. Their frameworks assume human-led research, documented reasoning, and interpretable decision-making at every step.

To their credit, regulators are adapting faster than you might expect. The FDA has been engaging with AI-driven drug developers through programs like its Innovative Science and Technology Approaches for New Drugs (ISTAND) pilot. In 2023 alone, the FDA received over 300 drug applications that included some AI/ML component – up from a handful just a few years earlier.

But there’s a tension. Many of the most powerful AI models are essentially black boxes. A graph neural network might predict that a molecule will bind to a target with high affinity, but explaining why it predicts that – in terms a regulatory scientist can evaluate – remains an open challenge. Interpretability tools are improving, but they’re not where they need to be for high-stakes medical decisions.

Some companies are getting around this by using AI for hypothesis generation but validating everything through traditional experimental methods. That’s probably the right approach for now, even if it means we’re not capturing the full speed advantage AI could theoretically offer.

What the Next Five Years Probably Look Like

I’m going to make some predictions, fully aware that predictions about AI timelines are almost always wrong in one direction or another.

By 2028, at least one AI-designed drug will receive full regulatory approval in a major market. Insilico Medicine and a few others are close enough that this feels more like a when-not-if situation. The real question is whether it’ll be a first-in-class drug for an unmet need or an incremental improvement for a condition with existing treatments.

Foundation models for biology will become as important as foundation models for language. Companies like Genentech (with its partnership with NVIDIA) and startups like Evozyne are building large-scale models trained on biological data. These models will understand cellular behavior the way GPT understands English – imperfectly, but usefully enough to transform workflows.

The cost of early-stage drug discovery will drop dramatically, but clinical trials will remain expensive. AI can’t (yet) make Phase 3 trials cheaper. The real savings come from killing bad candidates earlier, so money isn’t wasted on doomed compounds that limp through years of development before failing.

And here’s the outcome I’m most hopeful about: rare diseases will finally get serious attention. When it costs $2.6 billion to develop a drug, companies naturally focus on huge markets – diabetes, cancer, cardiovascular disease. But if AI can reduce early discovery costs by 50-80%, the economics of developing treatments for diseases affecting 10,000 or 50,000 people start to pencil out. That’s not a technology story. That’s a human story, and it’s the one that matters most.

The Bottom Line

We’re not in a world where you feed a disease name into a computer and a pill comes out the other end. We’re probably never going to be in that world. Biology is messy, clinical trials are hard, and the gap between “works in simulation” and “works in a human body” remains humbling. But AI is compressing timelines, expanding the search space, and making drug discovery less of a brute-force lottery. For the roughly 7,000 known rare diseases that currently have no treatment, that shift can’t come soon enough.