iiinova — where your digital self keeps evolving

Romain Lacombe @rlacombe.bsky.social · 10/04/2025

Excited to speak at the @stanforddata.bsky.social Sustainability Data Science conference today! I’ll be presenting ClimateX, our work on LLM confidence calibration in the climate science domain. 📄 Paper: arxiv.org/abs/2311.17107 💻 Code: github.com/rlacombe/Cli...

010

Romain Lacombe @rlacombe.bsky.social · 05/04/2025

That’s a bit of relief at least! 🐚🦪

010

Romain Lacombe @rlacombe.bsky.social · 04/04/2025

Poster time! Fantastic workshop on experimental design in AI for Science today. Paper: arxiv.org/abs/2503.17368 Workshop: edai4science.github.io

030

Romain Lacombe @rlacombe.bsky.social · 03/04/2025

Nobel laureate talk time! With David Baker at the Stanford experimental design in AI for Science today.

010

Romain Lacombe @rlacombe.bsky.social · 02/04/2025

Pierre Clostermann during WWII

010

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

Excited to be presenting this work at the @stanfordbiosci.bsky.social workshop on Experimental Design in AI for Science next week! More info on the event here: edai4science.github.io

edai4science.github.io

Experimental Design: AI for Science

100

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

These results highlight the limits of the (extremely successful) evolutionary approach to protein structure prediction. We will likely need physics-informed models to reach accurate 3D structure predictions on rare PTMs and out-of-domain proteins beyond evolutionary priors.

100

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

No model achieved crosslink distances near the experimentally observed 1.8Å sulfur-to-α-carbon bond length, with average performance at 10-12Å. Most models incorrectly predicted disulfide bonds instead of thioether crosslinks, highlighting their reliance on evolutionary priors.

100

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

We evaluate 6 leading protein structure predictors—AlphaFold 2 and 3, Boltz-1, ESMFold, OmegaFold, and RoseTTAFold 2—on the 10 known sactipeptides. All models exhibit limited performance, with an average GDT-TS of only 11.5% for known sactipeptides and 12.6% for unknown ones.

100

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

This helps us probe how deep learning models generalize beyond evolutionary priors. We introduce a new, zero-shot benchmark for structure models: measuring how the 3D conformations predicted for sactipeptides match the geometry imposed by their known thioether crosslinks.

100

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

These non-canonical crosslinks are challenging and underrepresented in structural datasets. Crucially, *only 5* of the 10 known sactipeptides have a resolved 3d conformation with a PDB entry. *But* their 2d crosslink structure is known, which constrains possible 3d geometries!

100

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

Evolution-based protein structure prediction models have revolutionized structural biology but struggle with rare post-translational modifications (PTMs). We evaluate them on sactipeptides, a rare class of 10 known peptides with unique sulfur-to-α-carbon thioether crosslinks.

100

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

🔔🔔 New bioML paper alert! 🔔🔔 We evaluate evolutionary protein structure prediction models using the geometric structure of non-canonical crosslinks from a rare class of post-translationally modified peptides. A thread. 🧵

110

Romain Lacombe @rlacombe.bsky.social · 27/03/2025

"But you're so much more handsome than that!" – my beloved wife, when shown a Ghiblified selfie of ourselves. Right answer!! I'm swooning! 🥰🥰🥰

020

Romain Lacombe @rlacombe.bsky.social · 25/03/2025

This is much deeper than it sounds at first. Absolute nugget of wisdom.

000

Romain Lacombe @rlacombe.bsky.social · 25/03/2025

Oooh that looks beautiful! When can we pre-order?

100

Romain Lacombe @rlacombe.bsky.social · 25/03/2025

Anecdotally, @dcrainmaker.com has a treasure trove of GPS accuracy tests in land and marine conditions for most watches and trackers on the market.

020

Romain Lacombe @rlacombe.bsky.social · 22/02/2025

LLMs are Farenheit 451 for digitized human knowledge. There’s something comforting knowing that the ghost of every written word will live on, for as long as we can preserve open weights, and the GPUs to bring them to life.

010

Romain Lacombe @rlacombe.bsky.social · 04/01/2025

Chemical engineering rules everything around us. ⚗️ AI 🤖, bio 🧬, chips 💻, energy 🔋, fertilizers 🚜, materials ⛏️, space 🛰️ ... All the technologies shaping the world sit downstream from chemistry fundamentals. 👩‍🔬 Hold on to your periodic table and let's go! 🚀

080

Romain Lacombe @rlacombe.bsky.social · 23/12/2024

Few things are more heart wrenching than a walk by a wild, pristine, remote beach at the confines of civilization… …only to find billions of tiny speckles of microplastics all over the sand, which we *know* will end up in our bodies by way of the food chain. Very sad.

0100

Romain Lacombe @rlacombe.bsky.social · 15/12/2024

This guy is *such* a jerk. Ugh! Only good thing is it’s out in the open for anyone to see.

060

Romain Lacombe @rlacombe.bsky.social · 12/12/2024

1 failure per 1000 GPU•days is the going hardware MTF rate for large training runs these days. #NeurIPS2024

000

Romain Lacombe @rlacombe.bsky.social · 11/12/2024

Craziest talk of the day so far: you can recognize which room someone stands in with embeddings of firing rate maps of their place cells neurons! #NeurIPS2024

010

Romain Lacombe @rlacombe.bsky.social · 11/12/2024

Deep learning is chemical engineering for information: layers are process units, diffusion is transport, entropy is the loss function… Don’t believe me? Then why is the NeurIPS opening talk literally starting with the Haber Bosch process!?

030

Romain Lacombe @rlacombe.bsky.social · 11/12/2024

Made it to #NeurIPS2024! Here until Sunday, ping me for coffee, a walk and talk, or some Tim Hortons maple glazed donuts! 🇨🇦

020

Romain Lacombe @rlacombe.bsky.social · 08/12/2024

I used to add little incandescent light bulbs in my notes, for highlights.💡 But it’s time to decarbonize. 🔋 So now I add little LEDs! ⚡️

030

Romain Lacombe @rlacombe.bsky.social · 02/12/2024

Ooh yes tricky one. Thank you!

010

Romain Lacombe @rlacombe.bsky.social · 02/12/2024

Unbelievable! Directed airborne evolution 😀🎈

010

Reposted by Romain Lacombe

Rachel Thomas @math-rachel.bsky.social · 01/12/2024

An analogy of the differences between bulk RNA-seq, single-cell RNA-seq, and spatial transcriptomics #multiomics2024 10/

bulk RNA-seq: a big pile of different colored legs all mixed up
single-cell RNA-seq: small piles of Legos organized by color
spatial transcriptomics: image of the brain built out of Legos

0104

Romain Lacombe @rlacombe.bsky.social · 02/12/2024

Air pollution makes you dumb. iopscience.iop.org/article/10.1...

034

Romain Lacombe @rlacombe.bsky.social · 01/12/2024

Engineering is the ultimate LEGO set. Piece some unit processes together, add sensors and a PCB to control them, break a few molecular bonds and train a neural net to run the whole thing, and don’t forget your economics so you can keep going. Is anything more fun than this!? 😀

021

Romain Lacombe @rlacombe.bsky.social · 01/12/2024

Oh of course! It’s an elliptic attempt at capturing a few millenia of research in one tweet, so I certainly can’t claim it doesn’t lack nuance. :) What’s the faithfulness violation example?

110

Romain Lacombe @rlacombe.bsky.social · 01/12/2024

All I'm saying is they're all stories – the more believable ones, just like the ones less so.

100

Romain Lacombe @rlacombe.bsky.social · 01/12/2024

If you can actually disprove it then yes it's a false story! but it seems that it's not that easy for a lot of proposed mechanisms

100

Romain Lacombe @rlacombe.bsky.social · 30/11/2024

Ah great question. No more than 5-6, intuitions around this are hard to form. I mean that we often say “that’s correlation not causation” just because we find the mechanism absurd, or “that has to be causal” because we can’t imagine a good confounding factor. But it’s just stories in the end.

100

Romain Lacombe @rlacombe.bsky.social · 30/11/2024

But if you understand "why" then you have a story?

010

Romain Lacombe @rlacombe.bsky.social · 30/11/2024

Causality is correlation with a story. Change my mind.

340

Romain Lacombe @rlacombe.bsky.social · 30/11/2024

Tu es dur Frédéric, ça prend du temps à un conseiller d’écrire une note avec thèse antithèse et synthèse pour décider s’il faut rejoindre BlueSky. 🫠

010

Romain Lacombe @rlacombe.bsky.social · 29/11/2024

Carrément ! En famille dans le sud mais je te dis quand on monte !

010

Romain Lacombe @rlacombe.bsky.social · 28/11/2024

Yesterday's heavy fog lifted through the night, and it's a beautiful sunny morning in California. 🌴🇺🇸 Happy Thanksgiving everyone! 🦃🍠🥧🙏❤️

250

Reposted by Romain Lacombe

Oregon Zoo @oregonzoo.bsky.social · 28/11/2024

Happy Chompsgiving!

141420311

Romain Lacombe @rlacombe.bsky.social · 28/11/2024

France is spot on (I know because it’s mine too).

010

Romain Lacombe @rlacombe.bsky.social · 27/11/2024

"- Please generate a beautiful picture of a strange attractor. - Here's the visual representation of a strange attractor, blending intricate and dynamic patterns with a cosmic aesthetic. Let me know if you'd like any adjustments!" What do y'all think about ChatGPT's chaos theory art skills?

Visual representation of a strange attractor generated by ChatGPT, blending intricate and dynamic patterns with a cosmic aesthetic.

040

Romain Lacombe @rlacombe.bsky.social · 27/11/2024

Aw that’s incredibly nice of you to say! 🥹🙏 I return the compliment, follow Ben and get a front row seat to the latest in AI for science! ⤵️

030