Roundup #2 (announcements, links, jobs)
700 words, 4 minutes reading time
Announcements
Since the last roundup, one podcast and one story have been released:
Podcast:
Story:
On a related note, you should subscribe!
Also, one more poster has been put up:
Also, I am actively looking for two things:
A podcast editor.
Currently, I handle most of the editing myself and outsource the clip-generation process to freelancers. These freelancers often aren’t great, and I’d love to be able to push the editing process outside of my to-do’s. This blog generates some amount of money which I am happy to use to pay a good freelance editor above market rate. But I don’t know of any! Would highly appreciate a referral to anybody!
Guest writers.
I think this field (biology x computation) has an awfully high number of curious scientific ideas that would benefit from more eyes — some of which I try to cover, but my backlog list is ever growing. There is also a rather high number of students/scientists who are looking for their next role or desire more eyes on them. Writing is a good way to do that, and I’d love to host that! I don’t know what this looks like: maybe full articles, maybe a compilation of 3-4 small ‘thesis statements of interesting problems/ideas. Reach out to me if this feels interesting!
Links
Mammoth Biosciences releases a version of Cas9 that is three times smaller, termed ‘NanoCas’, which, for the first time, allows for genetic engineering tooling to fit inside an AAV vector. In-vivo genetic engineering on the horizon?
Sarah Constantin’s post about why prior synbio companies have failed. Open request for a deeper dive into this stuff!
Entirely Alphafold2-based model of a membrane protein complex with 17 chains! Fairly, no wet lab experiments done to confirm the model, so one may be fully in their rights to toss this into the trash. But it’s still quite interesting to see the reasoning here, very ‘lets try this out in Alphafold, okay it doesnt look good, lets try this out instead…’.
Very similar performance amongst all recent protein-ligand folding models (AF3, Chai-1, Boltz-1, etc) w.r.t unseen protein-ligand. I’ve long been pessimistic on the idea that building better foundation models for these problems is a (computational) skill issue instead of a data issue (something that Demis at Deepmind takes the other side on). Now, fairly, all of these models are trained in nearly identical ways with reasonably similar architectures, as they were all fast followers to AF3, so this is perhaps an unfair conclusion. But even if we expand to the world of all ligand-docking models in general, as this paper did, we still don’t see much difference on PoseBusters between ligand-only models like DiffDock-L and AF3-esque models (attached below). I get that PoseBusters is also not a great benchmark, but I don’t know of a leaderboard that uses something better. Anyway, as always, I could be wrong about all this and I hold this opinion loosely.
Jobs
(Contact me if you’d like to be posted here!)
Machine Learning/Computational Biology Scientist at Kerna Labs
From the JD: Kerna Labs is building foundation models of RNA biology to discover and develop better genetic medicines. Our mission is to leverage advanced computational techniques and high-throughput biology to fundamentally change the way we do drug discovery and development.
Alongside the institutional investors, there are a great set of angels: Patrick Hsu (cofounder of the Arc Institute), Melissa J Moore (CSO of Moderna) and Jacob Kimmel (cofounder of NewLimit).
Senior Researcher and Senior Research Engineer at Microsoft
From the JD: for our lab in Cambridge, UK, or Amsterdam, NL we are seeking Machine Learning Researcher candidates to work at the intersection of machine learning, chemistry, and drug discovery. If you are passionate about this research area and believe you can make a lasting impact, we would love to receive your application.