De novo design

Drugs and materials consist of molecules, meaning specific arrangements of atoms. For a long time, such molecules were mainly searched for: in plants, in bacteria, in vast collections of already known substances. De novo design takes the opposite path. First you define what the molecule should be able to do, and only afterward design a fitting structure on the computer. “De novo” is Latin and means “anew” or “from the beginning.” The result is a proposal for a molecule that never existed in this form before, either in nature or in a laboratory.

Why one doesn’t simply keep searching

The number of theoretically possible molecules is unimaginably large. For small drug molecules it is estimated at roughly 10 to the power of 60. All known substances combined, by contrast, are a tiny fraction of that. Anyone who only searches through what already exists is thus ignoring almost all possibilities.

On top of that comes the time factor. A new drug traditionally takes ten to fifteen years to reach approval and often costs more than a billion euros. A large part of that is spent on candidates that turn out late in the process to be ineffective or toxic. If a computer already delivers better proposals beforehand, that saves years and money.

A second reason is independence from existing templates. For some tasks there simply is no molecule in nature that could be modified. Enzymes that break down a particular plastic have to be invented. This is exactly where the strength of the approach lies.

From target profile to finished molecule

At the beginning stands a goal. For drugs, this is often a protein in the body to which the active substance is meant to dock. The shape of this protein is known, and one knows where the binding site is located. It’s like a lock for which a matching key is sought. The difference: the key isn’t tried out from a keyring, but freshly filed.

This task is nowadays mostly taken over by generative AI models, meaning systems that generate new data rather than merely classifying existing data. They have learned from tens of thousands of known molecular structures which atomic arrangements are chemically stable at all. On this basis, they generate proposals that fit the desired binding site. Well-known tools of this kind are RFdiffusion and ProteinMPNN from the group of David Baker, who received a Nobel Prize in Chemistry in 2024.

Afterward comes a filtering step. Programs calculate for each proposal how tightly it is likely to bind and whether it can be manufactured. Out of thousands of designs, a few dozen remain. Only these are then actually synthesized and tested in the laboratory. The computer thus does not replace the experiment, it merely drastically narrows down the selection.

Where designed molecules already appear

In business news, you’ll mostly encounter the term in the context of biotech companies. Companies such as Isomorphic Labs, a spin-off of Google DeepMind, or Insilico Medicine are working on exactly this. The first drug candidates designed this way are already in clinical trials in humans. So far none has been approved, which shows: the design is the fast part, the testing remains slow.

Outside of medicine, the same principle is used for materials. Sought after, for example, are enzymes for breaking down plastic or new compounds for batteries. There too, one first formulates the desired property and then has candidates generated.

A common misconception is to confuse de novo design with drug screening. In screening, millions of existing substances are tested automatically. In de novo design, the candidate does not exist beforehand at all. Both methods complement each other, but only one of them expands the pool of possible molecules.