Love this project!! Congrats @samuelhking, @pdhsu and the @arcinstitute crew! my 2 cents: AI being used for biological design is best thought of as a translator.  Allowing us to speak in English and have it translated into DNA and vice versa.  We don't know how to design a phage out of parts but Evo 1/2 were trained by "reading" over 2 million phage genomes from nature and so it learned to "speak" phage DNA.  So then we could ask it to generate one -- just like you could ask ChatGPT to generate a poem for you in Chinese even if you don't speak it yourself. We had already trained AI models on the language of proteins with models like Alphafold and ESM and that worked well -- this paper shows we can do it at a higher level of complexity.  This AI model speaks multi-gene phage genomes, not just individual genes.  Very exciting demonstration and the work nicely proves it by actually making and testing designed phages. They work! There are two things that are obvious future directions in my opinion and that will ultimately be successful: (1) The model should be re-trained based on what it learns about the phages that were designed so it can get better at understanding what the human is asking for and translating that into DNA.  This "reinforcement learning" is similar to how Google taught AI models to play chess - you let the model play a game and then tell it if it won or lost.  Here you'd let the model design millions of phages, build them in the lab, and then tell it the performance of the different designs. (2) We should see if models trained on millions of bacterial genomes can enable us to build an AI-designed whole bacterial cell similar to what was done here for a phage. This will see if we can go from translating an English request into a DNA book (500,000 letters of DNA for the simplest bacteria) rather than the DNA poem of the phage (5,000 letters of DNA in a phage).  This would be a nation-scale scientific milestone as cells are the building blocks of all life and the US should make sure we get to it first. In order to do both (1) and (2) we need dramatic improvements in the efficiency of doing the actual wet lab biology needed to build DNA and test the performance of organisms.  It's indicative that they only built 302 phage designs and tested 16 designs -- that's because wet lab work is too slow and expensive.  The answer to that is lab automation -- I've been happy to see NSF investing $100M in AI-controllable, automated cloud labs and other efforts that will make US scientific infrastructure more efficient and industrial scale.  The White House AI action plan called out the need for these "cloud-enabled labs" too. Again, awesome work !!!

在@brianhie實驗室的一份新預印本中，團隊報告了可行的噬菌體基因組的首次生成設計。 利用Evo 1和Evo 2，他們生成了完整的基因組序列，結果得到了16種具有不同基因組結構的可行噬菌體。