50%minimum gain on some interaction types the AlphaFold 3 paper reported over earlier toolsDesignRelease11.11.24
AlphaFold 3 code opens to academic researchers
Google DeepMind released the model code, with weights on request, six months after describing it in Nature.
Amino Weekly science desk11 Nov 20241 min read
What changed
When AlphaFold 3 was described in May 2024, scientists could only use it through a web server with limits on what they could submit, and the code was not published. That drew criticism, because results in a journal are supposed to be checkable. In November, DeepMind put the code online. The trained weights, which make it usable, are available to academic researchers who apply.
Why it matters for peptides
Earlier versions predicted the shape of single proteins. AlphaFold 3 predicts complexes: a protein with a peptide bound to it, with DNA, with a small drug molecule or with metal ions. That is closer to the questions peptide researchers ask: where will this hormone sit on its receptor, and will this designed binder fit?
The paper reported clear gains over earlier specialist tools for several kinds of interaction, by at least 50 percent on some benchmarks.
The limits
Running it locally lets labs screen many candidates and look closely at failures. But AlphaFold 3 can be confidently wrong, especially for flexible peptides, for molecules unlike its training data and for how tightly two things bind, which it does not predict directly. Commercial use stays with Isomorphic Labs, the DeepMind spin-off developing drugs. Open alternatives trained to match it appeared within months.
Source Abramson J. et al., Nature, May 2024; Google DeepMind code release, November 2024
Amino Weekly reports research. It is not medical advice and never gives doses.