iiinova — where your digital self keeps evolving

amelie-iska @amelie-iska.bsky.social · 20/06/2026

Trained a tiny ~17M parameter model for the OpenAI Parameter Golf Bits-per-Byte (BPB) Competition and beat the leading model (BPB<0.57, and the next best model was ~0.74BPB) huggingface.co/AmelieSchrei... huggingface.co/blog/AmelieS... codebase: github.com/amelie-iska/...

github.com

GitHub - amelie-iska/ToricGT

Contribute to amelie-iska/ToricGT development by creating an account on GitHub.

060

amelie-iska @amelie-iska.bsky.social · 17/01/2025

Yes, insurance companies really did cancel fire insurance for Californians 2 months before the LA fires. www.youtube.com/watch?v=5h1H...

youtube.com

California Wildfires Are An Insurance Nightmare

YouTube video by LegalEagle

000

Reposted by amelie-iska

Nuala Bishari @nuala.bsky.social · 11/01/2025

Did you know that court docs for name and gender changes are not confidential? You can google a trans person’s current name and find their old, dead name — and vice versa. In CA, my amazing bud Hazel is working to fix this, and she may just be successful. My latest for @sfchronicle.com Gift Link 🎁

sfchronicle.com

Changing names puts transgender Californians at risk. New protections could be on the way

A new transgender privacy bill in California would make official name changes confidential.

17224

Reposted by amelie-iska

Shannon Osaka @shannonosaka.bsky.social · 09/01/2025

Pretty amazing that nonprofit Watch Duty -- which says it has 150 volunteers working around the clock to track wildfires -- has become the most trusted source of information on LA fires. Operated by retired/volunteer firefighters, dispatchers, and reporters. Real people. www.watchduty.org

watchduty.org

Watch Duty - Wildfire Maps & Alerts

Watch Duty, a 501(c)(3) nonprofit, alerts you of nearby wildfires and firefighting efforts in real-time.

291809705

amelie-iska @amelie-iska.bsky.social · 09/01/2025

If you’re in LA, stay safe, leave if you’re told to. Sending good vibes.

010

amelie-iska @amelie-iska.bsky.social · 29/12/2024

Okay…density matrices/operators I can do.

010

amelie-iska @amelie-iska.bsky.social · 26/12/2024

So, DFT is pretty interesting 💀

050

amelie-iska @amelie-iska.bsky.social · 15/12/2024

If people find MSA subsampling useful, they will likely find MDGen very useful and they will definitely find BioEmu useful.

0102

Reposted by amelie-iska

Danielle L. Schmitt 🔬 @schmittwitt.bsky.social · 09/12/2024

UCLA Chem & Biochem is looking for an assistant professor who has "outstanding potential as a scientist, educator, & mentor to students from groups under-represented in the physical sciences". This is part of a cluster hire in physical sciences. @chemjobber.bsky.social recruit.apo.ucla.edu/JPF10039

recruit.apo.ucla.edu

Physical Sciences HSI Cluster Assistant Professor in Chemistry & Biochemistry

University of California, Los Angeles is hiring. Apply now!

01615

amelie-iska @amelie-iska.bsky.social · 09/12/2024

Immunotherapy for oncology New more effective & cheap anti-venoms Allosteric switch proteins New Alzheimer’s treatments De novo functional nanopores Vaccines And so much more from Baker Lab. Thank you for being such an inspiration and for forging a path forward. 🚀💜 www.youtube.com/live/HnT1VWz...

youtube.com

YouTube

Share your videos with friends, family, and the world

031

Reposted by amelie-iska

Roland Dunbrack 🏳️‍🌈 @rolanddunbrack.bsky.social · 07/12/2024

Thanks! Also check our update to CDR clustering where we modified the D score as max over all dihedral distances including omega, instead of average. Also electron density cutoff to detect peptide flips. Eliminated lots of small CDR clusters with poor density. www.biorxiv.org/content/10.1...

biorxiv.org

A penultimate classification of canonical antibody CDR conformations

Antibody complementarity determining regions (CDRs) are loops within antibodies responsible for engaging antigens during the immune response and in antibody therapeutics and laboratory reagents. Since...

121

amelie-iska @amelie-iska.bsky.social · 07/12/2024

Anthropic needs to get on fine tuning Claude with tree of thought reasoning and GFlowNets.

100

amelie-iska @amelie-iska.bsky.social · 06/12/2024

My apologies to those in love with classical molecular dynamics 😅

130

Reposted by amelie-iska

Frank Noe @franknoe.bsky.social · 06/12/2024

Super excited to preprint our work on developing a Biomolecular Emulator (BioEmu): Scalable emulation of protein equilibrium ensembles with generative deep learning from @msftresearch.bsky.social ch AI for Science. www.biorxiv.org/content/10.1...

21441147

Reposted by amelie-iska

Dr. Rebecca Calisi-Rodríguez @rebeccacalisi.bsky.social · 05/12/2024

When my daughter came out to me, she asked if being gay was “natural.” A classmate suggested otherwise… As a professor of animal behavior, I shared 2k+ examples from the animal kingdom! Those became a book, & today it’s still #1 top release and #3 best seller. #LoveWins #WildPride ❤️🏳️‍🌈 a.co/d/947x66x

Colorful book cover of Wild Pride, by Dr. Rebecca Calisi-Rodríguez, with label “Top New Release”

Colorful book cover of Wild Pride, by Dr. Rebecca Calisi-Rodríguez, showing the paperback version as #1 and kindle version as #2 in Top new releases.

Colorful book cover of Wild Pride, by Dr. Rebecca Calisi-Rodríguez, showing ranking as #3 best seller.

17417

Reposted by amelie-iska

Diego del Alamo @delalamo.xyz · 05/12/2024

Improving structure prediction by padding MSAs with synthetic sequences. Potts model-designed seqs worsen prediction performance, so some higher-order coevolutionary interactions appear to be contributing to this boost openreview.net/pdf/fd516f23...

Table 1: Real-World MSA Challenges average pLDDT, LDDT and RMSD enhancement scores, averaged over 3 runs; avg. Det. represents average depth of MSA.

1182

Reposted by amelie-iska

Diego del Alamo @delalamo.xyz · 04/12/2024

More evidence that protein folding NNs have learned a crude approximation of an energy function, though the authors here suggest it is at the local level exclusively, because it is only able to make incremental refinements in the absence of coevolutionary data academic.oup.com/bioinformati...

Fig. 5. Impact of AlphaFold2 refinement in a homology modeling
pipeline. Scatter plots illustrate the change in TM-score (left) and lDDT
of Cα (right) after processing by AlphaFold2. Each point represents a
single target from the CAMEO dataset, colored by sequence identity.
The x-axis displays the scores of the templates compared to the target
structures, and the y-axis shows the scores after refinement with
AlphaFold2.

2257

amelie-iska @amelie-iska.bsky.social · 04/12/2024

Trying to get Baker Lab's new beta-pairing binder design technique going. Not quite there, but I'm getting close. Anyone have good adj and ss .pt prep code, or a bash script to run the model with on said .pt files?

193

amelie-iska @amelie-iska.bsky.social · 04/12/2024

I like collecting good ideas 😁 there are quite a few floating around in AI4Biochem these days. I love piecing together what has been done in the past 6 months and seeing what is possible in the next 6 months. It just gives me so much hope and excitement to be able to look into the near future. 🚀

020

amelie-iska @amelie-iska.bsky.social · 04/12/2024

This looks interesting. Reminds me of this… www.biorxiv.org/content/10.1...

biorxiv.org

020

Reposted by amelie-iska

Brian Naughton @btnaughton.bsky.social · 04/12/2024

This is a fantastic, bleeding-edge overview of the field. Semi-relatedly, anyone know what happened with AI2BMD? 10k atoms(!), "ab initio accuracy", published in Nature... It sounds great, but very few hits / github stars so I assume few people are using it.

281

Reposted by amelie-iska

Hannes Stark @hannes-stark.bsky.social · 01/12/2024

Who will be at NeurIPS and wants to meet? :) (generative models for biomolecules, ...) Also, in the Wednesday poster session, Bowen Jing and I will present our "Generative Modeling of Molecular Dynamics Trajectories" arxiv.org/abs/2409.178...

2602

Reposted by amelie-iska

Brian Naughton @btnaughton.bsky.social · 30/11/2024

New blogpost: The ABCs of Alphafold 3, Boltz and Chai-1 blog.booleanbiotech.com/alphafold3-b...

blog.booleanbiotech.com

Boolean Biotech

911536

Reposted by amelie-iska

Gabriele Corso @gcorso.bsky.social · 30/11/2024

Nice blogpost by @btnaughton.bsky.social about playing around with the different biomolecular structure prediction models AF3, Boltz-1, and Chai-1! My perspective? Current models are good, but future ones will be much better: let's build them together! 🚀

2172

Reposted by amelie-iska

Bird account @dialecticbio.bsky.social · 30/11/2024

Design of high specificity binders for peptide-MHC-I complexes www.biorxiv.org/content/10.1... New from the Baker lab

15910

Reposted by amelie-iska

Gabriele Corso @gcorso.bsky.social · 30/11/2024

If someone in the community is experienced with this, help is always super welcomed!

031

amelie-iska @amelie-iska.bsky.social · 29/11/2024

Testing out the new Boltz-1 chunking on a hemichannel…if it works, full gap junction is next. I’ll report back with results soon 😊

2152

Reposted by amelie-iska

Diego del Alamo @delalamo.xyz · 29/11/2024

Pretraining PLMs with both the standard MLM loss and a growth temperature loss dramatically improves prediction of thermostability, as well as identification of thermostabilizing point mutants. Lots of wet lab evaluation www.science.org/doi/10.1126/...

science.org

A general temperature-guided language model to design proteins of enhanced stability and activity

The PRIME model improves protein mutants’ key properties with a 30% success rate, using machine learning without prior data.

5268

amelie-iska @amelie-iska.bsky.social · 29/11/2024

More playing around with Chai-1 & Boltz-1

060

amelie-iska @amelie-iska.bsky.social · 28/11/2024

🧠 ^^This^^ is the moat…it’s not your model, it’s your ability to wield it and to pursuade it to conspire with other models to solve a problem.

120

amelie-iska @amelie-iska.bsky.social · 27/11/2024

Oh?! Really?! What an interesting development…I still like Boltz-1, but maybe if I get to know Chai-1 a little I’ll like it too.

3141

amelie-iska @amelie-iska.bsky.social · 27/11/2024

Something tells me trying to disrupt/modulate complicated pathways to inhibit cancer is a bit like using mechanist interpretability (ex. Golden Gate Claude) to get neural nets to do something in particular. It can be done, and sometimes must, but top down training of the system is much better.

000

Reposted by amelie-iska

Petter Brodin @petterbrodin.bsky.social · 27/11/2024

Thanks Carl for highlighting

2181

Reposted by amelie-iska

Milot Mirdita @milot.bsky.social · 27/11/2024

MMseqs2 Release 16 Highlights: GPU-accelerated search📄, ORF or new 6-frame translated search modes, contig taxonomy always keeps the longest ORF, bug fixes (reduced memory and higher sensitivity) and relicensed as MIT 📄 biorxiv.org/content/10.1... 💾 mmseqs.com and 🐍Bioconda 🖥️🧬🧶

011443