A promising drug can fail long before it reaches a patient. It's usually not because scientists lack an idea, but because they can't find a molecule that works safely in the human body. The US Food and Drug Administration notes that hundreds or even thousands of compounds may need to be made and tested during development, and that studying and testing a new drug before approval takes about eight and a half years. AI is helping researchers make some early decisions faster, but it hasn't removed the need for laboratory experiments or clinical trials.
Solving the 3D Puzzle of Proteins
Many medicines work by interacting with proteins involved in a disease. If researchers understand a protein's three-dimensional shape, they can investigate where a drug might bind and how that could affect the protein.
Scientists use experimental methods like X-ray crystallography and cryo-electron microscopy to find protein structures. AI systems such as AlphaFold can also predict likely structures just from amino acid sequences. The AlphaFold Protein Structure Database, built by EMBL-EBI and partners, gives researchers a huge collection of predicted protein structures.
Newer systems go even further. AlphaFold 3, described in a 2024 Nature paper, can predict structures of molecular complexes involving proteins, DNA, RNA, small molecules and other parts. These predictions help scientists form research hypotheses, but they aren't experimental proof. Models can make mistakes, and they may not capture all the shapes a protein takes inside living cells.
Designing Molecules with Generative AI
Once researchers find a possible disease target, they need to find or design a molecule that interacts with it. Generative AI models can suggest new chemical structures or tweak existing ones based on the properties scientists want.
This can narrow down a huge search space and help chemists decide which compounds are worth making. But a computer-generated molecule is only a candidate. Researchers still have to make it in the lab and test whether it binds as expected, affects the disease process and is safe. Reviews of generative AI in drug discovery talk about its potential, and also the challenges of producing molecules that are effective, stable and suitable for further development.
Checking Safety Before Human Testing
A compound that affects a disease-related protein can still cause unwanted effects somewhere else in the body. So researchers look at things like absorption, distribution, metabolism, elimination and toxicity, which are often grouped together as ADMET.
Machine learning can help estimate some of these properties and pick which compounds to test first in the lab. Still, laboratory studies and carefully designed clinical trials are needed to check safety in people.
Supporting Clinical Research
AI can also help with parts of clinical development, like analyzing medical data, planning trials and monitoring them. However, the evidence for real-world impact is still developing. A 2025 review of AI-driven clinical trials points out that only a few AI-designed drug candidates have entered human trials and that big scientific and practical challenges remain.
What Comes Next?
AI is becoming a useful tool across early drug discovery, from predicting protein structures to designing molecules and analyzing data. It doesn't guarantee a successful medicine, and it doesn't replace scientific judgment. Every promising candidate still has to pass experimental checks and clinical evaluation before it can reach patients.