No one remembers authorising this firmware update

Very pleased to have contributed to this concise, digestible overview of AI structure prediction! I think this is a great resource for newbies and experienced AlphaFold users alike. Happy Folding!

Finally, we apologise to colleagues whose excellent work we could not include because of space limitations. Thank you to the community whose discoveries made these ten lessons possible! 🙏

Lesson 9 - Structural Modeling Can Guide Protein Engineering From prediction to design! 🛠️ Interfaces can guide changes that transfer recognition, restore function or create new connections. Combined with binder design, they can help build new immune receptor specificities.

Lesson 8 — When Predictions Deviate from Canonical Models What if the model is telling us something unexpected? Unusual assemblies or repeated failures can hint at different biology. Interpret outliers using existing evidence, then test them. 🔍

Lesson 7 - Capture Touch-and-Go Complexes Brief encounters can have big consequences! AI models can help probe contacts that evade biochemical detection, suggesting how partners meet and which residues to test in the lab. 🧪

Lesson 6 - AI Can Reveal Novel Protein–Protein Interactions From screening entire proteomes to identifying short binding motifs, modelling can uncover new partners and explain how they interact, generating testable hypotheses and guiding experimental discovery.

Lesson 5 - You Can Never Benchmark Enough! A convincing model is a start. How does it compare with what we know? Include known interactors, non-interactors and solved structures as controls. Revisit these controls as tools change. 🧪

Lesson 4 — Structural Similarity Can Reveal Hidden Relationships Proteins with no detectable sequence similarity can share a fold. Comparing predicted structures across families or proteomes can uncover connections that sequence searches miss. 🔍

Lesson 3 — Model Your Proteins with Other Molecules Proteins have company! Lipids, nucleic acids, ions and ligands shape their biology. Including relevant partners can reveal interfaces and improve regions that are poorly modelled in isolation.

On tοοls and metrics 🧰 Aside from the lessons, we put together a table of what's currently available and what each tool and metric is actually for. Reproduced below. Links in the reply.

Lesson 2 — Sample Multiple Seeds 🎲 How much depends on the seed? Repeat predictions to see which features hold and which vary. Consistency strengthens a hypothesis. Record seeds, model version, weights release and database version so others can reproduce the analysis.

Lesson 1 — Don’t Ignore Confidence Metrics 🔍 A model can be uncertain overall yet reliable at the interface that matters. Read local and interface scores alongside global confidence to judge which parts support your biological hypothesis.