We built UniRep, a wide mLSTM that was trained on UniRef50 sequences to perform next amino acid prediction. In so doing, it learned a vector representation of proteins (2/8)pic.twitter.com/aUN3Iy8Upj
U tweetove putem weba ili aplikacija drugih proizvođača možete dodati podatke o lokaciji, kao što su grad ili točna lokacija. Povijest lokacija tweetova uvijek možete izbrisati. Saznajte više
We built UniRep, a wide mLSTM that was trained on UniRef50 sequences to perform next amino acid prediction. In so doing, it learned a vector representation of proteins (2/8)pic.twitter.com/aUN3Iy8Upj
UniRep vector space is semantically rich
. It at least includes organismal and structural level information, and is good for finding functional homologs that share little seq similarity. Similar to the sentiment neuron in NLP, UniRep has a pretty wild helix-sheet neuron! (3/8)pic.twitter.com/0L4IuiW0HC
Simple supervised models trained "on top" of UniRep reach SOTA performance on several diverse protein stability and quantitative function datasets. This indirectly suggests UniRep encodes fundamental protein features that underpin stability and function. (4/8)pic.twitter.com/irxU5IfPux
Putting it all together, UniRep shows promise for protein engineering by enabling rapid discovery of distant, diverse, and high functioning protein variants, and does so with orders of magnitude gain in efficiency!
(5/8)pic.twitter.com/ZiSRTh9Cqd
Currently, we're improving UniRep and seeing what it can do for protein eng. Also, so many new possibilities when proteins can be represented as vectors: deep generative design, fast semantic protein comparison, and semi-supervised learning of structure to name a few!
(6/8)
Finally, I can't speak highly enough of Ethan (@EthanAlley) and Grigory (@grigonomics) 
. These guys are destined for great things in protein design and biosecurity. Go follow them! Also, we can't thank @geochurch & @MoAlQuraishi enough for their stellar mentorship.
(7/8)
Last but not least, codeee: https://github.com/churchlab/UniRep … 
.. and some hashtags: #proteinengineering #MachineLearning #UniRep
FIN (8/8)
1/2 Hi Surge, great work! I'm not from the field so I have very basic questions. My wife has a proteomics table with a lot of uncharacterized proteins as the most up and downregulated, so we are interested in obtaining more information from them, specially because they do not...
2/2 ...share known homologs. Is it possible using the UniRep for that? What is the output?
Really interesting. Such cool directions for protein computational research
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.