Alex Rubinsteyn

Going to be in Paris next week, anyone want to meet up to talk {ML, cancer genomics, immunology, pMHCs, TCRs, &c}?

De novo design and structure of a peptide–centric TCR mimic binding module T cell receptor (TCR) mimics offer a promising platform for tumor-specific targeting of peptide–major histocompatibility complex (pMHC) in cancer immunotherapy. In this study, we designed a de novo α-...

Cool burst of research, I think a lot of cancers will eventually get downgraded to inconveniences through pMHC directed therapies (or compensatory NK therapies to clean up MHC null escape)

www.science.org/doi/10.1126/...
www.science.org/doi/10.1126/...
www.science.org/doi/10.1126/...

Comfy-ness monotonically decreasing

Cool! What species is this?

Important caveat: in bacteria! More complicated in eukaryotes

A short translational ramp determines the efficiency of protein synthesis - Nature Communications Several factors contribute to the efficiency of protein expression. Here the authors show that the identity of amino acids encoded by codons at position 3–5 significantly impact translation efficiency...

TIL: A/T content of first few codons in a protein dramatically impacts translation efficiency

"A short translational ramp determines the efficiency of protein synthesis" (2019)

www.nature.com/articles/s41...

Functional analysis of the antigen binding sites on the... : AIDS hich recognize the specific antigen artificially may have tremendous potential in anti-MTB/HIV coinfection. We have previously successfully identified a MTB Ag85B199–207 and HIV-1 Env120–128 peptide-b...

Or even furin cleavage site : SGSG : P2A like in:

journals.lww.com/aidsonline/f...

Systematic identification and characterization of eukaryotic and viral 2A peptide-bond-skipping sequences Rao et al. identified thousands of previously unknown 2A peptides across both viruses and eukaryotes using an HMMER analysis. The authors further identified a unique class of 2A peptides, class B, who...

Going deep on 2A peptides:

www.cell.com/cell-reports...

(Trying to pick between GSG:T2A, GSG:P2A, and furin site:GSG:T2A)

The key parameter is the “eccentricity” of the orbit. 0 would mean a perfect circle orbiting the sun (Earth is close at 0.017). Between 0 and 1 means more and more stretched-out elliptical, and >1 is a not-coming-back hyperbolic path.

‘Oumuamua was 1.2, Borisov was 3.357. The new object is >6!

I’ve been using both o3 and Sonnet 4 for many different research questions (with and without Research mode), finally cancelled my Anthropic subscription last night.

o3 is consistently more accurate and more useful

Scientists retrieve 80-million-year-old dinosaur protein in 'milestone' paper Two independent papers help open door to a molecular science of dinosaur paleontology

Previously: Dino paleo-proteomics finds collagen peptides in bones (www.science.org/content/arti...)

Now: 20Mya rhino enamel full of MS identifiable peptides from many proteins (www.nature.com/articles/s41...)

Next: Dino enamel mass spec!

Interesting, any idea why?

Did anyone anticipate the NIH & FDA turning against animal research?

(I think it’s ethically great, scientifically shaky, and very surprising)

It’s a nice essay but hard to take it too seriously until Anthropic is kitting out automated labs with liquid handling robots

Dario Amodei — Machines of Loving Grace How AI Could Transform the World for the Better

Thanks for sharing - I’ve spent the evening reading Amodei’s essay. Really interesting - I liked the positive spin to it (at some parts).
Dropping the link here in case others are interested:
www.darioamodei.com/essay/machin...

Jobs Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

Friend asked me if I want to work at Anthropic & I was curious bc of Amodei’s Machines of Loving Grace essay.

Checked their job openings (anthropic.com/jobs) and there's not even a category that automated biology work could fall under.

Maybe next year?

We'll know that AI companies are shifting "AI will cure cancer" from vague hype to actual ambition when they start quietly spending a fortune on oligo and peptide synthesis.

They’re all links and I have yet to open one that’s not very topical to my query

Out of curiosity, where do those two links go?

(I wonder if I miss times where it mis-describes a paper but still links to something useful)

Huh, l use o3 + Deep Research almost daily and have yet to get a fake reference. It’s been great at finding important papers in a research niche for focused lit reviews. My main problem with is how credulous it is wrt claimed findings but as long as I read everything it’s citing, it’s great

Also would love to have co-admins for the feeds!

The canvas UI is still a bit clumsy, any way to make it feel more integrated with the rest of the UI (& make JSON edit a bit more hidden / niche)?

By popular demand, I moved #MLSky, #ImmunoSky, and #GeneSky over to @graze.social

Let me know if any problems popped up in the migration. Also, since Graze seems like a nicer platform for feed maintenance, I'll try to be more active -- I'm sure I missed some suggested improvements in the past

I’ll move it over today but wish there was a way to have multiple people manage it

Is there any way to switch it?

A Unifying Grammar for DNA Variants A unified framework and toolkit for standardizing DNA variants at scale. Part 1 of the neoantigen discovery with long-read sequencing series.

pirlblog.substack.com/p/variant-st...

A Unifying Grammar for DNA Variants A unified framework and toolkit for standardizing DNA variants at scale. Part 1 of the neoantigen discovery with long-read sequencing series.

When @benjamingvincent.bsky.social and I started PIRL one of our hypotheses was that neoantigen vaccines mostly don’t work bc SNVs create low value (self-like) targets. What other mutational antigens could we use instead? @jslee.bsky.social been working on making those legible…

This is masterful trolling.

Impressions from the AI trenches (aka whatever recent things I've tried across models):

o3 better at thinking, Sonnet 4 better at coding.

Not sure what to make of that.

Also ChatGPT Deep Research still infinitely better than Claude Research, o3-pro useless.

Gave Claude and o3 a detailed & annoying edit to a big blob of Python plotting code but accidentally pasted in the wrong code. They both obliged & tried to make the edit anyways, generating huge piles of incoherent slop.

Seeking: an AI that says "no way is this going to work"