I'm not a structural biologist, I'm an ML person who ended up embedded with a small biotech team for a project, which turned out to be the right vantage point to actually notice the gap between the AlphaFold3 headlines and what the wet-lab side of the team actually did with it. The pitch was "predicts protein structure plus how it binds small molecules and nucleic acids, not just the protein alone." The reality, from watching people who need the answer to be right use it, was more specific and more useful than the headline.

What the team actually trusted it for

Fold prediction for a novel protein with no close homolog in the training data — genuinely useful as a first-pass hypothesis, saves weeks of guessing at a starting point for crystallography. Nobody treated the output as ground truth; they treated it as a strong prior that narrows down where to point the actual experimental effort. That distinction matters more than it sounds like it should: the model is a hypothesis generator that happens to be right often enough to be worth the compute, not an oracle.

Where confidence dropped fast was ligand binding poses for anything chemically unusual — the confidence scores the model outputs were a genuinely useful signal here, and the team had learned, the hard way apparently, to just not trust predictions below a certain pLDDT-adjacent threshold rather than eyeballing whether a structure "looked right." That's a workflow lesson as much as a model lesson: the output includes a confidence signal, and ignoring it because the structure looks plausible is exactly how you get burned.

The actual bottleneck didn't move

Getting a fast, decent structural hypothesis was never the bottleneck for us. Validating it experimentally still takes the same number of weeks it always did.

That's the thing nobody puts in the press release. A faster hypothesis-generation step is genuinely valuable, but if your pipeline's bottleneck is wet-lab validation time — and for most real drug discovery work, it is — a better prediction model shifts where your team spends its early-stage time without shrinking the overall project timeline nearly as much as the headlines imply. It's a real improvement, applied to a part of the pipeline that wasn't the slowest part to begin with.

What I'd tell an ML person joining a biotech team

  • Learn what the confidence scores actually mean before trusting or dismissing a prediction — they're not decorative.
  • Ask where your team's actual bottleneck is before assuming a better model solves it. It's often not the modeling step.
  • The domain experts' skepticism about a "solved" problem is usually well-earned. Listen to it before getting excited about a benchmark number.