A computer-generated model of a protein can look remarkably complete: coloured ribbons, intricate pockets and another molecule fitted neatly beside it. For researchers, such a prediction can offer a valuable starting point. For everyone reading a science headline, it can also create a misleading sense that the hardest questions have already been answered.
A predicted structure is not evidence that a treatment works in people. It can help researchers form and prioritise hypotheses, but those hypotheses still need testing. To understand an AI-assisted drug discovery story, first identify what the reported result actually demonstrates.
A structural prediction answers a particular question
Proteins are molecules whose shapes and interactions matter to biological processes. A structural prediction offers a model of how molecular components may be arranged. That can help researchers decide what to investigate next, including possible interactions worth testing.
AlphaFold 3 extends structural modelling to complexes involving proteins and other molecular components. However, EMBL-EBI's training material stresses that its output is a static structural prediction, not a simulation of a molecule's full dynamic behaviour in solution. A plausible predicted arrangement should not be interpreted as a complete account of what happens in a living system. EMBL-EBI on AlphaFold 3's limitations.
For a general reader, the distinction is between a proposed arrangement and an observed outcome. The first can guide an experiment. It cannot silently stand in for the second.
Confidence is attached to the prediction
Prediction software may provide confidence measures alongside a structure. Those measures need to be interpreted in relation to the question being asked. Confidence in one region of a model is not automatically confidence in an interaction between different molecular components.
EMBL-EBI explains that AlphaFold 3 offers several measures, including local confidence and measures used to assess the arrangement of complexes and interfaces. They support critical assessment of a prediction; they are not a single probability that a future medicine will succeed. EMBL-EBI's guide to assessing prediction quality.
Be especially careful when a headline places a percentage beside a sweeping claim. Ask what that number measures, how it was evaluated and whether it applies to the result being discussed. “High confidence” without an object leaves out the most useful information.
Read the story as a ladder of claims
Consider a hypothetical announcement about an AI-designed molecule. A colourful structural model might be accompanied by one of several very different findings:
| Reported finding | The next question to ask |
|---|---|
| The model predicts a promising arrangement | Has the prediction been checked experimentally? |
| A laboratory experiment detects the intended interaction | Does that interaction produce the desired effect in a relevant biological system? |
| An early study finds a potentially useful effect | What are the safety, dosing and reproducibility findings? |
| A clinical study reports a benefit | Who was studied, compared with what, and with which limitations? |
| A regulator approves a medicine | For which indication and population, and under what conditions? |
This is a reading framework, not a universal development sequence. Research often loops back, and different programmes use different methods. Its purpose is to stop a result at one level being described as if it established every level above it.
An experiment that rules out an attractive prediction can also be useful science. The value of modelling is not limited to producing a successful product; it can help make research choices more informed.
Drug development extends beyond the model
The US Food and Drug Administration describes a process encompassing discovery, preclinical research, clinical research, regulatory review and continuing safety monitoring after products reach the market. Its overview makes clear how much evidence gathering lies beyond an initial laboratory idea. That is the US regulatory framework; requirements and decisions elsewhere should be checked with the relevant authority. FDA overview of the drug development process.
When a company says AI accelerated discovery, ask which activity became faster. Generating candidate structures, selecting experiments and establishing a clinical benefit are different achievements. Saving time on one stage does not, by itself, establish the total time saved across a programme.
It is reasonable to be interested in a faster method while remaining uncertain about its eventual clinical impact. Those positions fit together.
Five questions that make a science headline clearer
Before describing an announcement as a medical breakthrough, look for answers to five questions:
- What exactly did the model predict or help researchers do?
- Which parts of the result were tested outside the model?
- What comparison supports the claim that the approach improved on existing methods?
- What important uncertainties do the researchers acknowledge?
- Does the evidence concern a research tool, an experimental candidate or an approved medicine?
A strong article should make those distinctions easy to find. It should also identify the underlying research and separate researchers' findings from a company's expectations.
AlphaFold's contribution becomes easier to appreciate when the claim is precise. A model can give scientists a better question to test. The next headline worth following is what happened when they tested it.
Sources & further reading
- EMBL-EBI on AlphaFold 3's limitations — checked 2026-09-29
- EMBL-EBI's guide to assessing prediction quality — checked 2026-09-29
- FDA overview of the drug development process — checked 2026-09-29