Hugging Science

immaculate community energy last night with the people behind: > EquiformerV3 > The Well > NVIDIA Atlas > Meta OMat ✨✨✨

tomorrow! sf! come hang! luma.com/2562pz76

🤗 Hosting a happy hour to meet the people building the future of open science in SF on Thursday! Wanna come? luma.com/2562pz76

We taught a DNA model to learn its own tokenization. It learned the genetic code with no supervision. And outperforms Evo 2's architecture with 3x faster inference. Great work with Arnav (@arnavshah0), Victor (@victor_ljz), Parsa (@Radii2323), Brandon (@fluorane), Sukjun (@sukjun_hwang), Bo Wang (@BoWang87), Patrick Hsu (@pdhsu), Hani Goodarzi (@genophoria) and Albert Gu (@_albertgu) 🔥

We released Gemma 4 last week, and seeing the community's response has been amazing! 🚀 Honored to lead the vision efforts in which we made huge performance leaps from Gemma 3, I wanted to help you make the most of the new capabilities. Deep dive 🧵

Welcome to Hugging Science!

The next OpenADMET blind challenge is about to launch. This one is will predict Pregnane-X Receptor (PXR) induction. PXR is a nuclear hormone receptor and master regulator of drug-metabolizing enzymes and transporters. This challenge comes with a PXR dataset of more than *11,000* compounds sourced from two Enamine libraries (Discovery Diversity 10 set and FDA Approved Drugs set) profiled through a rigorous multi-step assay flow reminiscent of an on-target drug discovery program.

I’ve been at a small conference this week, one where the AI people have been presenting early in the week and the domain science people will be presenting later in the week. At the end of the talks last night, the conversation turned very doomer with all the AI people talking about how well Claude Code or Codex can do hill-climbing AI research and how we (the AI people) are maybe all about to lose our jobs! The domain science people expressed their shock at this attitude because, though Claude Code can be let loose to complete lots of banal hill-climbing AI research projects, basically no experimental science is hill-climbing or even metric driven. Most scientific fields are about much more taste-driven exploration that is incredibly difficult to make metrics for or to parameterize, and this misunderstanding from the AI community is one of the most damaging things to the realization of great science with AI. Seems like we’re actually pretty far from having AI models do that… Over the summer, @evijit and I wrote about this (and some other things hindering AI for science) at a bit more length, and today that work is out in Patterns! So, if you care about these problems and the real challenges in bringing AI to science in the real work, I recommend giving it a read!

Simply adding Gaussian noise to LLMs (one step—no iterations, no learning rate, no gradients) and ensembling them can achieve performance comparable to or even better than standard GRPO/PPO on math reasoning, coding, writing, and chemistry tasks. We call this algorithm RandOpt. To verify that this is not limited to specific models, we tested it on Qwen, Llama, OLMo3, and VLMs. What's behind this? We find that in the Gaussian search neighborhood around pretrained LLMs, diverse task experts are densely distributed — a regime we term Neural Thickets. Paper: arxiv.org/pdf/2603.12228 Code: github.com/sunrainyg/Rand… Website: thickets.mit.edu

New OpenFold3 preview out! (OF3p2) It closes the gap to AlphaFold3 for most modalities. Most critically, we're releasing everything, including training sets & configs, making OF3p2 the only current AF3-based model that is functionally trainable & reproducible from scratch🧵1/9

Absolutely killer new blog from one of the OGs in open protein engineering, @amelie_iska. This is the only piece of writing I've come across that explains to someone in non-specialist language why you need each of these models and what role they all have to play in making sure that your protein design pipeline isn't actually splitting out garbage that you'll only discover after spending tens of thousands on wetlab validation. She does a deep dive on how to build a full-stack protein design sequence model with all the appropriate filters so that we, in her words, "do not merely reconstruct plausible sequences, but actively sample diverse, high-value protein variants under a multi-fidelity oracle stack." This blog covers all the big models that are used in industry for these kinds of generative pipelines (i.e., SPURS, BioEmu, UMA, GraphKcat, KcatNet, and MMKat) and why you need them all.

💊💊💊 @Ginkgo just dropped GDPx4 💊💊💊 29.9 MILLION rows of DRUG-seq data, perfect for benchmarking + large-scale perturbation modeling critical for anyone who can't afford their own high-throughput lab a little more in 🧵

Zero code to protein pipeline now on @huggingscience 🤗 As a part of the PDW hackathon, the organizers built inference spaces for: 🧬 RFDiffusion 🧬 RosettaFold3 🧬 BoltzGen2 (+ soon to be MCP servers)

Hey twitter! I'm releasing the LLM Evaluation Guidebook v2! Updated, nicer to read, interactive graphics, etc! huggingface.co/spaces/OpenEva… After this, I'm off: I'm taking a sabbatical to go hike with my dogs :D (back @huggingface in Dec *2026*) See you all next year!

thanks so much to everyone who came out to the @huggingscience bar takeover last night in San Diego amazing community with epic science coming out of it if you have any photos/videos, plz drop in the thread 🧫🧬🔭🔬💛

At NeurIPS next week. AI × Science afterparty. 800+ people on registration. See you there!

I would say my biggest takesway from spending the last 8 months singlemindedly studying bioML is that understanding the biology actually is important

Sakana AI will be at #NeurIPS2025! 🐟🐟🐟 We’re pleased to announce that Sakana AI is co-hosting “AI for Science: Algorithms to Atoms” panel event alongside NeurIPS. Join our CEO, @hardmaru, and other industry leaders as we explore the future of AI-driven scientific discovery. • Featured panelists: Yann LeCun (Meta), Bill Dally (NVIDIA), Anima Anandkumar (Caltech), David Ha (Sakana AI), Max Welling (CuspAI) • In collaboration with: NVIDIA, CuspAI, Sakana AI, and Samsung Next. • Dec 5 | 3:30pm PT. If you'll be at NeurIPS San Diego, here is the link to join: luma.com/AI-for-Science…

Very cool! The world needs a lot more labs like @grocklin 's!

happy to see people enjoying the article, and thanks @cgeorgiaw for sharing! Go follow @huggingscience!🔬