AlphaFold3 is useful. Here's where it still fails us.

AlphaFold3 expanded coverage to small molecules, DNA, and RNA alongside proteins. For a lot of structural biology workflows, this matters.

But I’ve been running into the same failure modes repeatedly and wanted to write them down.

Where it actually breaks

Disordered regions. AF3 still doesn’t handle intrinsically disordered proteins well. pLDDT scores flag this, but people downstream often ignore them and treat low-confidence predictions as real structure.

Novel folds. If your protein is genuinely unlike anything in the training set, the predictions degrade fast. For environmental samples with unusual sequences, I’ve had to fall back to comparative modelling for sanity checking.

Multimeric complexes with no homologue. The interface prediction is better than AF2, but it’s not magic. Large assemblies with no structural precedent are still mostly guesswork.

What I actually use it for

Screening candidates before wet lab. It saves time at the prioritisation stage — not as a ground truth. If I treat a prediction as a hypothesis rather than a result, it holds up.

The colabfold implementation with MMseqs2 makes this fast enough to run on a list of candidates without a GPU cluster.

---

Still useful. Just be honest about what it is.