Interpreting AlphaFold correctly: structure predictions need uncertainty information

Making a spatial hypothesis accessible

Proteins consist of chains of amino acids whose spatial arrangement is important to their biological role. AlphaFold uses sequence information to predict these structures. The jointly provided AlphaFold database makes these models available for scientific use and explains their limitations. For researchers, a plausible structure can provide a starting point for planning experiments or putting existing findings into context. It should remain clear what the visualisation shows: a computed structural hypothesis, not a direct image of the protein under all biological conditions. Proteins can be flexible, adopt different states and interact with other molecules. A single prediction therefore does not automatically tell us how the protein works in every environment. Its value lies in enabling more targeted questions, not in having already answered them all.

Local confidence is not global certainty

Among other things, EMBL-EBI explains pLDDT as a measure of local confidence in a predicted structure. This makes it possible to assess which regions the model predicts with relatively high confidence. Low scores can indicate uncertain predictions and require particularly careful interpretation in flexible or disordered regions. However, a high local score does not answer every question about the relative positions of larger structural components. Further information, such as predicted aligned error, becomes relevant for this. Nor is a confidence score a direct measurement of biological function. The widely used colour-coded visualisation makes it easier to find one's bearings, but can encourage viewers to overlook these distinctions at a glance. Specific analyses should therefore examine the underlying uncertainty measures and their meaning, rather than just the visual coherence of the three-dimensional model.

From prediction to verifiable research

A responsible workflow begins with a clear biological question. The next step is to check whether the relevant structural regions have been predicted reliably enough and whether the interaction under investigation is covered by the model being used at all. Existing experimental data provides an important comparison. Uncertain regions should not be silently treated as fixed spatial arrangements. In publications, the model version, provenance and relevant confidence information belong in the documentation. The prediction can then help to prioritise experiments or derive new hypotheses. It should not, however, serve on its own as proof of an effect, binding or a cause of disease. Precisely because the models are so visually compelling, this distinction matters: a useful research tool becomes more robust when its uncertainty remains visible and is deliberately incorporated into further investigation.

Keep exploring

• AI image descriptions: Useful when the purpose is clear
• Data quality: why provenance and selection matter more than sheer volume
• Detecting hallucinations: What contradictory answers reveal