AI drug discovery is the use of AI models to find or design the molecule for a new medicine. These models do two main things. They predict the 3D shape of a protein from its sequence, and they generate new molecules that would fit that protein. So discovery is moving from screening molecules that already exist to designing molecules that do not exist yet. It is also called AI-driven drug design or generative chemistry.
Here is why the 3D shape matters. A medicine usually works by sticking to one protein in the body, called the target, and blocking it. To design a molecule that sticks, chemists need to know the target's shape. Measuring that shape in a laboratory can take months or years. Since about 2020, AI models such as AlphaFold can predict the shape of most proteins in minutes. Many of these models are foundation models, trained on the shapes and sequences of hundreds of thousands of proteins.
A generative model then proposes molecules. It has learned from millions of known molecules which arrangements of atoms are stable and can be made. It is asked for molecules that fit the target's pocket, that do not stick to other proteins and that a chemist can actually make. Other models predict each molecule's properties, such as whether it would be toxic.
Consider a small biotech company in Gurugram that wants a medicine for a drug-resistant fungal infection. The fungal protein it wants to block has never been measured. The AI model predicts its shape. A generative model then proposes 50,000 new molecules for its pocket. The physics-based molecular modelling software checks how tightly each one would fit. A second model removes molecules that look toxic or hard to make. The chemists choose 40 molecules and make them. In this kind of work, the results from those 40 go back into the model, which proposes a better set. Each round takes a few weeks.
Some companies apply AI to a different source: the chemistry of plants. Plants make thousands of molecules that have never been studied. These companies measure the molecules in plant extracts with mass spectrometry, and use AI to identify which ones could become medicines.
The models depend on clean biological data, prepared on a bioinformatics data platform. Some laboratories connect their instruments, through lab instrument integration, so that the model can choose and run the next experiment itself. This is called a self-driving lab. The chosen molecules then go to in silico drug modelling, which predicts how they would behave in the body.
The first medicines designed with AI are now being tested in people. None has yet clearly shown that AI-designed molecules succeed more often in trials than molecules found the old way.