What AI does in drug discovery

A drug discovery programme has to decide what to change in a biological system, find a molecule that produces the intended change and establish whether that change helps people. AI can contribute at several points. The label on a company's website rarely tells you which.

Start with the output. Did the model rank genes, predict a protein structure, propose a chemical structure or select the next experiment? Each output creates a different scientific claim. A model that ranks a target highly has proposed a hypothesis about biology. A model that proposes a molecule has supplied something chemists can attempt to make.

Finding and testing molecules

In the 2023 abaucin study, researchers measured the effects of thousands of compounds on Acinetobacter baumannii, trained a neural network on those measurements and used it to prioritise further compounds. They then tested the candidates. Laboratory work identified antibacterial activity and explored the mechanism; a mouse wound experiment tested activity in a living system.

That workflow ties a prediction to a defined measurement. The model learns relationships between molecular structures and the assay output. It can search far more candidate structures than a team can test physically, although the quality of the search still depends on the examples, measurements and chemical territory represented in its training data.

The earlier halicin work used deep learning to identify a candidate with antibacterial activity, followed by experimental tests. Both studies connect predicted antibacterial activity with laboratory testing. They do not establish a general ability to design rejuvenation therapies.

Generative design adds another task: proposing structures. A proposed structure still needs a feasible synthetic route, adequate purity and suitable physical properties. Chemists may need to alter it repeatedly after experiments reveal weaknesses. Those revisions belong in any account of what AI contributed.

Protein prediction and protein design

AlphaFold's 2021 study established substantial progress in predicting protein structures from sequence. A structural model can help researchers investigate a binding site or plan experiments. It leaves questions about activity, selectivity, exposure and the effect of manipulating that protein in a diseased human tissue.

RFdiffusion tackles protein design, including generating structures around a functional requirement. Its published experiments tested several classes of designed proteins. This is a different output from predicting the structure of a protein that already exists.

Consider a protein intended to bind a cellular receptor. A computational design can suggest a compatible surface. Binding measurements must establish whether the protein attaches; cellular experiments must establish what happens after attachment. The same receptor can have useful and harmful roles in different cells.

Reading an AI discovery claim

Look for the model's task, the data it used, the number of proposals tested and the experiments that determined which proposals survived. Human decisions matter too. Selecting a disease indication, rejecting an impractical molecule and redesigning an assay are contributions to discovery.

Claims about speed need a start and finish. Producing a candidate structure quickly, selecting a development candidate and completing a human trial are separate achievements. A fast computational step can improve a programme even when later experiments occupy most of its calendar.

For longevity research, the hardest inference often concerns the target. An age-associated protein can be a cause of damage, a response to damage or part of a repair process. A convincing programme explains why changing that target should improve a specified function, then tests the explanation.