Simon Kozlov

Bro resurrected NASA simulator in a weekend with Claude Code and used data from the real space launches to find the parameters to run it, and turns out it still works really damn well.

bes-dev.bsky.social@bes-dev.bsky.social

How was your weekend? I spent mine launching the Space Shuttle. With NASA code from 1970 that nobody could run for 50 years.

science.orgToward life with a 19–amino acid alphabet through generative artificial intelligence designBecause all known living organisms are made from at least 20 canonical amino acids, the feasibility of life using a more simplified alphabet remains unclear. In this work, we leveraged computational d...

Yours truly is a proper scientist now! TL;DR: we used AI to redesign parts of essential cell machinery with only 19 canonical amino acids instead of 20. Why? Great thread by @harriswang.bsky.social provides more context and details. Let me talk a bit about the AI design part of this. 1/

Harris Wang@harriswang.bsky.social

1/ Excited to share our new paper in Science @science.org: “Toward life with a 19-amino acid alphabet through generative artificial intelligence design.” A great collab w/ Sergey's group @sokrypton.org at MIT @columbiasysbio.bsky.social science.org/doi/10.1126/... 🦠🧬🛠️🖥️💥

No single model was able to solve every gene, it took all of them + developing novel methods. And even then, there are genes no model was able to solve - we had to use mutagenesis on last 3 stragglers. No model thinks these solutions are good, but they work in vivo. 9/

We're still far from solving biology with AI, but the progress is astonishing: median number of designs to solve a gene was around 4. Huge thanks to @sokrypton.org @harriswang.bsky.social for the mentorship and support, and all collaborators! Liyuan Liu, Charlotte Rochereau, @gogurtecho.bsky.social

Harris Wang@harriswang.bsky.social

But we still haven't solved all. For the rest, we expand the design space: allowing to change not just ILE residues, but allow compensating mutations elsewhere. We masked ILE neighbors in 3D space and repeated the process. Among several new working designs, one is crazy. 7/

Here pLM redesigns most of the helix and introduces other mutations. What makes it work? I wish I knew. Understanding the mechanics of this solution alone requires an experimental effort comparable to this whole project, if not more. Which brings me to Feynman's famous quote. 8/

Harris Wang@harriswang.bsky.social

Or we can find designs AlphaFold considers to produce the same structure. Or better - sequences both models consider a good match! In this work, we used a new method to optimize sequences by inverting AlphaFold with a bias towards ProteinMPNN logits x.com/sim0nsays/st... 5/

x.comx.com

This solves a bunch more and produces several viable solutions not explored by nature! 6/

Harris Wang@harriswang.bsky.social

Next, you can feed sequences into pLM like ESM, mask ILEs and take max predicted AA other than ILE! pLMs store evolutionary information so we're modeling what else evolution could place there. We do it with different methods developed by Charlotte Rochereau, solve 25% more. 3/

But say there's a super-conserved position where nature didn't explore anything else. How can we go beyond that? Well, maybe we can use structure as a prior. We run ProteinMPNN with a known structure for still-not-solved proteins, mask ILEs, see if it makes different designs. 4/

Harris Wang@harriswang.bsky.social

@gauravbhardwaj.bsky.social

So, you want to redesign the ribosome to not use a specific AA, and you chose isoleucine. How would you go about it? First, baseline: just try replacing it with most similar AA, evolutionary data says it's valine. Turns out this alone works in 36% of the cases 2/

Simon Kozlov reposted

Gaurav Bhardwaj

23 May 2025

Great to have our manuscript with @sokrypton.org 's lab describing AfCyDesign finally out in @natcomms.bsky.social . Structure prediction, sequence redesign, de novo hallucination of cyclic peptides, and some binder design examples in this version. rdcu.be/em0vA

rdcu.beCyclic peptide structure prediction and design using AlphaFold2Nature Communications - AfCycDesign: Cyclic offset to the relative positional encoding in AlphaFold2 enables accurate structure prediction, sequence redesign, and de novo hallucination of cyclic...

Simon Kozlov reposted

Eric Columbus

@ericcolumbus.bsky.social

16 Apr 2025

Oh wow great @haleaziz.bsky.social piece on a 2018 case in which Trump’s ICE screwed up and erroneously deported a guy. AND THEY BROUGHT HIM BACK because why wouldn’t they? We’re in a much darker place now. Free link: www.nytimes.com/2025/04/15/u...

Simon Kozlov reposted

Yehlin Cho

@yehlincho.bsky.social

8 Apr 2025

Excited to share our preprint “BoltzDesign1: Inverting All-Atom Structure Prediction Model for Generalized Biomolecular Binder Design” — a collaboration with @martinpacesa.bsky.social, @Zhidian Zhang, @Bruno E. Correia, and @sokrypton.org 🧬 Code will be released in a couple weeks

Simon Kozlov reposted

Axel Visel

@axelvisel.bsky.social

4 Dec 2024

BOOSTER, a newly identified orphan gene, supercharges photosynthesis and, when overexpressed, boosts height of poplars by up to 200% (at least in greenhouse) To make it even more interesting, part of the gene seems derived from a poplar fungus-farming ant 🐜 Paper: www.cell.com/developmenta...

youtu.beDNA Break Repair by Homologous Recombination (2024) Drew Berry wehi.tvYouTube video by WEHImovies

5 Dec 2024

Biology is amazing

Drew Berry wehi.tv@drewberry.bsky.social

Delighted to publish my new molecular animation: DNA Break Repair by Homologous Recombination youtu.be/Xe-83tBcxhs

Simon Kozlov reposted

Harris Wang

@harriswang.bsky.social

30 Nov 2024

💡💥Here’s a modest proposal I’ve been pondering for a while on how to change the PEER REVIEW process and start more community-driven SCIENCE FUNDING. Lots of feedback/discussion welcomed! Other variants of the idea might exist already—not claiming originality. (a long 🧵) (1/9)

29 Nov 2024

Fully accepting nothing may work out is a surprisingly effective way to work something out. Here’s the catch: knowing this fights against acceptance.

25 Nov 2024

I would love for a social network to exist where questions like this get a meaningful discussion. I don’t think Twitter had it even if it’s best days, with a few exceptions when tweet randomly goes viral. Can we make it work here? That would take the whole community to a next level.

Akos Nyerges@akosnyerges.bsky.social

Friends, does anyone know the per-base error frequency of the new Illumina MiSeq i100 Series, with its 2 x 300 nt XLEAP-SBS chemistry? Is it better or worse than the old MiSeq v3 chip's error frequency (based on your sequencing runs)? #NGS #Illumina #MiSeq

25 Nov 2024

Holy shit, @akosnyerges.bsky.social got a meaningful reply to a highly technical question. We are so back! bsky.app/profile/geno...

Gary Schroth@genomics-guy.bsky.social

I am no longer affiliated with the company, but I can confirm that the 600c kits have much better error rates compared to the old MiSeq kits. Same thing applies the to 600c kits for NextSeq 2000. That MiSeq chemistry a decade old, which is a long time in NGS!

Simon Kozlov reposted

Jason Kelly

@jasonrkelly.bsky.social

20 Nov 2024

Made a biotech founders starter pack, DM me to be added! go.bsky.app/LWm8uPE