Roundup #3 (announcements, links, jobs)
716 words, 4 minute reading time
Announcements
Since the last roundup, two articles have been published. In order:
A walkthrough of a Baker-lab paper that designed a complex class of enzymes
But this was actually published on Asimov Press, that link is here.
A socratic dialogue over the utility of DNA language models (Part 1 of 2)
An essay that discusses the question that everyone has been asking for awhile: what’s the point of DNA language models? At least with regards to variant pathogenicity, I answer that question. Next part will come out soon and cover these models role in genome generation!
Thank you to everyone who reached out to me to be a guest writer! Got tons of really great pitches, hoping to publish some of them over the next few months. Also thank you to everyone who offered recommendations on people to use for podcast editing!
Links
I say this with a deep appreciation for how hard it is to take a strong stance online and deal with the vitriol sent your way: this article has an excellent hook but fails to deliver. I think there is a very interesting argument to make here about how, for example, CDR-only redesigns of existing targets fall short of the possibility of full de novo engineering (e.g. Fc engineering, which some AI biotechs do!) or the beginnings of full de novo engineering in the nanobody space, but neither is touched upon. Instead, headlines and soundbites by these companies (which, fairly, may overpromise!) are focused on as the opponent, instead of the very real research that is being done. My beef here really may have little to do with the author of this piece — they are reporting on the news, and the cultural sentiment being echoed here is indeed news. And it is not the job of news to search for capital-T truth, but the essay would’ve been so much stronger had it at least tried to do so.
Again, great title with a potentially amazing point (RFDiffusion doesn’t work that well for binder generation? Tell me more..), but the methods used are a bit lackluster. They only test 5 RFDiffusion-generated binders against 6 targets and find that most fail. Which is fully expected! RFDiffusion does not have a 20%< success rate for binder creation, it’s more on the order of 1%~ (potentially an order of magnitude lower, which is still far better than wet-lab tools for binder generation), which is explicitly detailed in the paper and known by users of the tool. The comments on the original thread may be interesting to read as well.
Interesting essay on reasoning model approaches from Andrew White (a founder of FutureHouse)
Essay on practical guide to biotech partnerships by Lucas Harrington (a founder of Mammoth Biosciences)
Jobs:
Contact me if you’d like to be posted here! If I’ve posted you here before, feel free to ask again. Moving forwards, I’ll just throw up a posting here if I see a job description that my reader-base might be interested in.
Dyno Therapeutics (where I work!) is hiring a machine learning scientist.
Excellent company, excellent culture, and excellent research. You should apply!
Plasmidsaurus is hiring across the board, but especially for wet-lab and software positions.
I think it’d be difficult for me to overstate how valuable of a company I think Plasmidsaurus already is and will become, both financially and in terms of net good for the world. Amongst every company I’ve written about, I strongly believe Plasmidsaurus will be the one that will have a book written about them in a decade. For a primer on what they do, here is an article I’ve written about them.
Pratyush Tiwary’s lab at U Maryland is hiring a postdoc in physics-driven AI for RNA and proteins.
Ideal for someone who wants to transition to industry!
Must be based in Switzerland!
Beautifully written post. Thanks for debunking and highlighting key issues in biotech and reporting facts in a fair unbiased approach.