News · 16 May 2024
AlphaFold 3 and the difference between a prediction and an explanation
On 8 May 2024 a team at Google DeepMind and Isomorphic Labs published AlphaFold 3 in Nature. Earlier versions predicted the structure of single proteins. AlphaFold 3 extends the approach to complexes: proteins interacting with DNA, RNA and small molecules such as drug candidates, using a diffusion-based architecture. The authors report substantial gains over existing methods for several interaction types.
It is an impressive piece of science. It is also a good case study for a question we ask in every project: what does a user need beyond the prediction?
Confidence is part of the output
One reason structure prediction has been usable by biologists is that the models report how sure they are. Per-residue confidence scores and estimates of expected positional error tell a researcher which parts of a predicted structure to trust and which to treat as a hypothesis. A structure without those numbers would be much less useful.
The same principle applies to clinical models. A classifier that says "malignant, 0.93" gives a clinician very little. One that also shows where in the image the evidence is, how typical this case is of the training data, and when the case lies outside what the model has seen, supports an actual decision.
Accuracy and understanding are different goals
A highly accurate model can still be a black box about why. In science, that matters because explanations generate hypotheses. In medicine, it matters because a clinician remains accountable for the decision and must be able to challenge the result.
There are practical ways to narrow the gap:
- Calibrated uncertainty, so that "0.93" really means about 93 times in 100.
- Evidence views, such as attribution maps or retrieved examples, designed with the people who will read them.
- Failure-mode documentation: which inputs the model handles badly.
- Independent validation on data from different sites and devices.
The takeaway
Foundation-scale models will keep improving, and biology is a field where that is welcome. As they move closer to clinical use, the work shifts from "can it predict?" to "can an expert interrogate it?" Building the second capability early is cheaper than retrofitting it.
Written in May 2024; details reflect the situation at that time.