'We have developed OpenFold, a complete open-source reimplementation of AlphaFold2 that includes training code and data. By training OpenFold from scratch and matching the accuracy of AlphaFold2, we have demonstrated the reproducibility of the AlphaFold2 model for protein structure prediction. Furthermore, the OpenFold implementation introduces technical advances over AlphaFold2, including markedly faster prediction speed.'
https://www.biorxiv.org/content/10.1101/2022.11.20.517210v1?med=mas
@gaymanifold Sorry if there was a misunderstanding, but as the inverted commas indicate, this is a quote from the preprint in the link -it's not my work. That said, my two cents are that with an open source code we have a better change of gaining fundamental understanding beyond structural predictions. It's also important to note that for basic biochemistry, cell biology & physiology, these structures are also of great value, not just for engineering.
@cyrilpedia Nice! https://github.com/aqlaboratory/openfold even gives conda-based installation instructions, so OpenFold ought to be much easier to install than Google's AlphaFold2
@cyrilpedia thank you first of all for bringing something proprietary into public. I have a question. Do humans learn from these models? As in do they give us any insight into the physics/chemistry of proteins? Or do we not care at all about this and it's useful that we can do this so that we can engineer proteins etc?