Twitxy Saved

Jarrid Rector-Brooks

Verified account This account is verified. Learn more

@jarridrb

PhD student @Mila_Quebec with Yoshua Bengio, currently visiting Frances Arnold's group @Caltech
Joined October 2023
682 Following 926 Followers 128 Posts

Latest posts 5

Pinned post

Jarrid Rector-Brooks Verified account @jarridrb
What if AI could invent enzymes that nature hasn’t seen? 👩‍🔬🧑‍🔬

Introducing 🪩 DISCO: Diffusion for Sequence-structure CO-design

14 rounds of directed evolution and over a year of wet lab work. That's what it took to engineer an enzyme for selective C(sp³)–H insertion, one of the most challenging transformations in organic chemistry.

DISCO surpasses this with a single plate. No pre-specified catalytic residues, no template, no theozyme, no inverse folding, just joint diffusion over protein sequence and structure.

📝 Blog: disco-design.github.io
📄 Paper: arxiv.org/abs/2604.05181
💻 Code: github.com/DISCO-design/D…
0:15
32 replies 244 reposts 1K likes 283K views View on X
Jarrid Rector-Brooks Verified account @jarridrb
For all those interested in crystal structures, check out the new OXtalign!

It computes RMSD_n, the standard packing-similarity metric from CCDC’s COMPACK. OXtalign delivers similar accuracy as COMPACK, it is up to 200x faster, fully deterministic, and comes with an Apache license. It reads experimental CIFs, predicted clusters without a unit cell, PDB/XYZ files, and ASE or pymatgen objects. Apache-2.0. Have fun, crystal people :)

github.com/OXtal/oxtalign
github.com GitHub - OXtal/oxtalign: Fast, open RMSD_n crystal-packing similarity for molecular crystals... Fast, open RMSD_n crystal-packing similarity for molecular crystals (OXtalign) - OXtal/oxtalign
Joey Bose Verified account @bose_joey Sep 21
🚨As part of the growing OXtal family for ML on Molecular Crystals

We are releasing OXtalign 🔮, a fast, open-source Python package for comparing crystal packings. Developed by the juggernaut of a researcher, ChengHao Liu (last author of OXtal)!

It computes RMSD_n, the standard packing-similarity metric from CCDC's COMPACK. OXtalign delivers similar accuracy to COMPACK but SIGNIFICANTLY FASTER!!!!

Features
- Speed: ⚡️ Up to 200x faster and fully deterministic
- License 🟢: Comes with a permissive Apache 2.0 license.
- Data formats 📂: Reads experimental CIFs, predicted clusters without a unit cell, PDB/XYZ files, and ASE or pymatgen objects.

Have fun, crystal people 🚀
Code: github.com/OXtal/oxtalign
View quoted post on X
Jarrid Rector-Brooks Verified account @jarridrb
This and the megascale-style dataset work from @KotaroTsuboyama and @grocklin et al is so cool! Very exciting to see this paper. 😃
Yehlin Cho Verified account @ChoYehlin May 21
🚀 Excited to share our new work: Absolute Stability Predictor!

📊: forms.gle/4ZnXZSnTBvaykk…

Built the MGnify Stability Dataset (1.8M+ measurements) and developed stability prediction models, together with @grocklin, @KotaroTsuboyama, @sokrypton, and teams.
View quoted post on X
0 replies 0 reposts 9 likes 1.1K views View on X
Jarrid Rector-Brooks Verified account @jarridrb
Go Fred! Love to see more cool diffusion LLM stuff. 😀
Fred Peng @pengzhangzhi1 May 13
How to Train Diffusion LLM more efficiently?

Our paper has an answer for you:
Don’t Retrain, Align: Adapting Autoregressive LMs to Diffusion LMs via Representation Alignment

Diffusion language models are becoming increasingly attractive: they support bidirectional generation, non-sequential decoding, and flexible editing.

But training them from scratch is expensive.

So a natural question is:

If we already have strong pretrained autoregressive LMs, do we really need to relearn all language representations for diffusion LMs?

We argue: probably not.

Our view is that AR→DLM conversion should not be treated as learning language from scratch again. Much of the semantic structure is already inside the AR model. What changes is the generation order and denoising behavior.

So instead of only continuing denoising training, we explicitly preserve the representation geometry of the AR model.

We introduce REPR-ALIGN: during masked diffusion training, we align the hidden states of the DLM to a frozen AR teacher of the same architecture, layer by layer, using cosine similarity.

No adapters.
No architectural changes beyond the attention mask.
Just representation alignment + masked denoising.
The result: up to 4× training acceleration in our setting, with especially strong gains in low-data regimes.
The main takeaway is simple:

Don’t retrain the representation space from scratch.
Align it, and let the model relearn the decoding path.

Paper: arxiv.org/abs/2605.06885
Code: github.com/pengzhangzhi/O…

Work done with an amazing undergrad @alexisfox and advisors @Anru_Zhang @AlexanderTong7
View quoted post on X
2 replies 0 reposts 8 likes 1.1K views View on X