And this is why it's important to have controls in computational sciences... Even things as basic as numpy can be silently broken and give you the wrong results... 😬 (For those that are curious this was the default lambdalabs cloud instance setup)
reinstalling libopenblas/numpy fixed the problem! sudo apt-get -y autoremove --purge 'libopenblas*' sudo apt-get -y install libopenblas-base pip install --upgrade --force-reinstall numpy thanks to Yaroslav Bulatov for pointing out that numpy could be using different backend versions.
This appears computer specific... doing same test on a different CPU, all is fine 😬 This is a little concerning...
No, it was just a random thought 🤔 as I was walking my dog 🐕 this morning. I was just thinking, the really "smart" plagiarizers would just do a better job at paraphrasing, and a simple text matching method would not be able to detect this.
I guess the technical term is "Paraphrasing plagiarism" meaning rephrasing someone else’s ideas to present them as your own. These are the instances I'm more worried about ... And likely to survive the recent witch-hunt. 😅
Plagiarism comes in many forms, some more severe than others—from rephrasing someone’s ideas without acknowledgement to stealing a whole essay. These are
It's all about biasing AF to sample the alt state... sometimes multiple methods will work for a different reasons, it doesn't mean one is better than the other. Ironically, "cf-random" is invoking AF's internal clustering... 😅 so it is NOT a good "control" to test if clustering helps or not. (2/2)
I dont think there is a single method that works for all cases. It really depends on the MSA. If there are pockets of seqs that prefer one state vs other, then af-cluster can help. If both states are equally preferred and there is coevolutio signal for both, subsampling or col masking can help (1/2)
Not sure yet... Still waiting on the room preparation. 😅
But back to our paper... we explore this question in the SI, and actually take advantage of the fact that single sequence is sometimes enough to design a new protein where a few point mutations switch to the alternative conformation. 🤓 (8/8) www.nature.com/articles/s41...
Besides the examples where it's sometimes possible to predict from single sequence. There have been other examples where it was not possible. If it is memorization, why would AF memorize one example and not another? 🤔 twitter.com/GMondaSilva/... (7/8)
