From Prediction to Prescription
When DeepMind released AlphaFold 2 in 2021, it solved a 50-year-old problem in biology: predicting a protein’s 3D structure from its amino acid sequence. Scientists celebrated, papers piled up, and the AlphaFold Protein Structure Database grew to cover nearly every known protein. But predicting static structures is only part of the story. Proteins move, flex, and interact with other molecules. AlphaFold 3, released in late 2025, predicts these dynamic interactions with startling accuracy, and the drug industry has taken notice.
How AlphaFold 3 Works Differently
AlphaFold 2 predicted the folded shape of individual proteins. AlphaFold 3 goes further: it models how proteins bind to DNA, RNA, small molecules, and other proteins. This matters because most drugs work by fitting into a protein’s binding site like a key into a lock. If you can accurately simulate that binding computationally, you can screen millions of potential drug molecules in silico before ever touching a test tube. The new model uses a diffusion-based architecture, similar to the generative AI systems behind image creators like DALL-E, but applied to molecular geometry. Understanding the science behind mindfulness meditation and its impact on mental health provides background on how AI is reshaping scientific research.
Real Drug Candidates Are Emerging
Isomorphic Labs, DeepMind’s drug discovery spin-off, has partnered with Eli Lilly and Novartis on programs targeting previously “undruggable” proteins. In Q1 2026, they announced preclinical candidates for a notoriously difficult oncology target, KRAS G12D, a mutant protein involved in about 25% of all cancers. Traditional methods had failed to find molecules that bind tightly enough. AlphaFold 3’s binding predictions identified a novel molecular scaffold that showed nanomolar potency in cell assays, a strong early signal.
The Open Science Debate
AlphaFold 2 was open-sourced, giving every researcher on the planet access. AlphaFold 3 is different. DeepMind initially restricted access, offering it through a limited API and keeping the model weights proprietary. The backlash from the scientific community was swift. Open-science advocates argued that publicly funded research built on AlphaFold predictions would become dependent on a private company’s gatekeeping. DeepMind eventually released a more open version, but the tension between commercial incentives and scientific openness continues. Alzheimer’s Disease: Latest Research on Causes, Treatments, and Prevention examines related questions about technology access and equity.
Limits and Honest Caveats
AlphaFold 3 is not a magic wand. Its predictions are excellent for well-studied protein families but weaker for disordered regions and membrane proteins. Binding affinity predictions still have error margins that can mean the difference between a drug that works and one that does not. Wet-lab validation remains essential. The most honest assessment is that AlphaFold 3 dramatically accelerates the early stages of drug discovery, potentially cutting years and billions of dollars from the process, without eliminating the need for clinical trials and regulatory review.
What Comes Next
Several biotech startups are building on AlphaFold 3’s architecture to tackle specific disease areas. Canadian company Recursion Pharmaceuticals is using the technology combined with its own massive biological dataset to hunt for treatments for rare diseases. Academic labs at the University of Toronto and McGill are applying it to antibiotic resistance, trying to find new drugs that can defeat superbugs. The speed of iteration is unprecedented: what used to take a PhD student three years can now be done in three weeks. The bottleneck has shifted from computation to clinical testing.
