We could have this for biology if we hadnt collectively decided to spend ~10 Mio dollars per year on BioRender instead 😢
04.02.2026 17:09
👍 4
🔁 2
💬 1
📌 0
We could have this for biology if we hadnt collectively decided to spend ~10 Mio dollars per year on BioRender instead 😢
SVG is much harder to generate for LLMs and there is much less training data of complex images. Current state of the art is this work: starvector.github.io/rlrf/
Carbons with 5 bonds in AI generated images.
Curious to try it but it will probably still be unusable if you care about the things you want to illustrate. There is already some services that wrap NanoBananaPro for Figure generation and even the teaser images they show are wrong: For example grabstract.io
Zenodo probably isnt too bad, could also set up a small blog on GitHub pages with Rogue Scholar to get DOI or use OpenScienceFramework for non-preprints that can't be reviewed on Prereview
As a bonus, here's a video of ProteinEBM folding up the fast-folder NTL9, rendered in stunning 2D by py2Dmol from @sokrypton.org! We hope models like ProteinEBM can serve as a step toward solving the "real" protein folding problem.
Try the NanoBananaPro model with Gemini Think Mode. Performs much better but also opens new avenues for dishonest research bsky.app/profile/did:...
Yeah it can work really well as www.qedscience.com demonstrates (for wet lab biology)
Great. Would be nice to have a more direct comparison of the same pipeline on different hardware. If one has to specially optimize each model (TensorRT, BF16) its clearly an argument to stick with an x86 machine
Yes true. Also it's ARM which might create issues with dependencies. But CUDA ecosystem, large memory, low energy consumption are nice...
Maybe, someone from the NVidia Bio Team ( cc @kdidi.bsky.social @machine.learning.bio )can enlighten us if they got their hands on it for testing
Anyone from the Bio community already got their hands on a NVidia DGX Spark for structure prediction and Protein design workloads ?
And little sidenote that the versioning is becoming quite a mess: Now there is RFDiffusion,RFDiffusionAA, RFDiffusion2, RFDiffusion3 and RFDiffussion2-MI (last two preprinted weeks from each other)
They saw it is hard to achieve a specific binder that recognizes only the phosphorylated target and not the unphosphorylated. Sequence specificity was easy. I also find notable that the model doesn't use motifs common in natural proteins like SH2
Cool work to design phospho-tyrosine binder by the Baker lab. They show it works but success rate is low (<0.1%) and binders are not super strong (>500nM, likely related to cost of desolvating the phosphate) .
www.biorxiv.org/content/10.1...
We are hiring a PhD student 🎓
Work with me and Matthias Rüdt in our department on an exciting industry collaboration with Lonza combining high-throughput experiments and molecular modelling using AlphaFold and MolecularDynamics
📍 Sion, Valais, Switzerland
More info: www.jobup.ch/fr/emplois/d...
Interesting work to design antibodies by merging AlphaFold hallucination and an antibody specific language model to design de novo CDR regions in a given antibody framework
🖥️ github.com/SantiagoMill...
🗎 www.biorxiv.org/content/10.1...
Today we're dropping the "beta" tag from Adaptyv, launching our new website and announcing our $8M seed round.
When we started Adaptyv a few years ago, our core belief was: AI models for biology are only as good as the lab data they're trained on and the hypotheses they can test in the real world.
🚨To accommodate the addition of EuroMLSB, we have extended the submission deadline to October 1, 2025 11:59pm AoE.
Find information on paper guidelines at mlsb.io. Submissions will be made through CMT.
You asked and we listened... @workshopmlsb.bsky.social is excited to be expanding to Copenhagen, DK at @euripsconf.bsky.social 🎉
Two workshops (San Diego & Copenhagen) will run concurrently to support broader attendance. You can indicate your location preference(s) in the submission portal💫
Time to test their claim that diffusion is competitive with hallucination in generating realistic backbones of huge (>800 AA) proteins...
La-Proteina from NVidia to do co-design of fully atomistic protein structures (sequence + side-chains + backbone) for up to 800 residues is now open source with permissive licensing
🖥️ github.com/NVIDIA-Digit...
🗎 arxiv.org/abs/2507.09466
Ever since Twitter shut down 3rd party clients I sorely missed a client with "remember where I left off" feature. Finally, found a client for Bluesky to do this: @openvibe.social which also supports RSS.
Hoping to liberate scientific illustrations from eternal payments to them so that we all can communicate and iterate on ideas faster by building on openly available illustrations.
It's been quite frustrating to deal with BioRender's unwillingness to support open science (while publicly claiming that they do) and administrators in institutions who don't care while paying them $$$ to continue business as usual (globally >10k Mio $).
@pracheeac.bsky.social I'm working on creating a free and open alternative to BioRender for scientific illustrations. Many scientists support this effort but I've also encountered quite some fans who don't realize that they're supporting a defacto monopoly.
Stay tuned for details on the 6th edition of MLSB, officially happening this December in downtown San Diego, CA!
Unfortunately, the MLSB Workshop @ NeurIPS (@workshopmlsb.bsky.social) was rejected this year.
Feedback from the deciding committee indicates it was a coin flip decision, with 283 proposals & a number related to “computational biology”
More on the future of MLSB soon…
Exactly. Also for some countries like Germany where people have been complaining about the conditions for research for years (people finishing their thesis with unemployment money and other shenanigans) it's kinda a slap in the face that suddenly there is money to bring in new hires...
Not as pretty but this should give you a vector you can tune easily in a similar way with reduced color complexity (might be a bit slow for large proteins in contour mode) bioicons.com/pdb2vector/ And uses less CO2 😅
No. The other libraries have different names (e.g RCSB Mol*, PDBe Mol*). Mol* = mol* = molstar
