Atoms, the smallest tangible building blocks of the physical world. A step closer to designing directly at the atomic level.
Huge congrats to Weikun and the team! Amazing to see the atomic universe we once dreamed of becoming a reality. 🚀
Natural language is becoming a practical interface for protein design. Recent work such as Prompteina shows how text can guide the generation of protein sequences and all-atom structures.
Here we wanted to explore a more unified approach: can we make language, sequence, and structure native to one generative language modeling framework?
We're excited to share Trion-1, our native tri-modal diffusion language model for proteins, developed at GenSI, Tsinghua AIR, and Shanghai AI Laboratory.
Prompteina adds text conditioning to a pretrained all-atom protein generator. In contrast, Trion-1 jointly models natural-language, amino-acid, and quantized backbone structure tokens with a shared diffusion language model (dLLM).
As a 4B language model, Trion-1 enables six any-to-any capabilities cross its internalized modalities: folding, inverse folding, sequence and structure understanding, and text-conditioned sequence and backbone design.
Two highlights:
· On Folding, but as an LM, Trion-1 reaches TM Score of 0.86 on the PDB test split.
· Sequence understanding: Trion-1 reaches 62.16% on GeneralQA, closely matching Claude Opus 4.7's 62.37%.
The broader goal is a protein-native language model that lets us move naturally between describing a function, generating a sequence, and modeling its structure.
More on the motivation, design, and results are elaborated in the following thread. Also, please check out our manuscript: arxiv.org/abs/2610.31215 @haozhou_ai @hello_gensi
WeiQu retweeted
We can finally talk about it:
We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company.
We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.
Agent for protein design 🧐
We spent the last decade scaling model intelligence. I believe the next frontier is scaling the loop between intelligence and experience.
AI's center of gravity is shifting from generating content to completing work. This is not merely a product shift. It changes the research problem.
A system that completes work cannot be judged by a single answer. It must operate in a real environment, use tools, read state, take multi-step actions, learn from feedback, recover from mistakes, and carry useful experience across tasks.
Thinking backward from this endgame, I see three foundational capabilities for large-scale agent systems:
1. Low-latency foundation models. In an agent loop, latency determines how much search, verification, and action the system can afford.
2. Feedback-driven self-improvement. An agent should not repeat the same failure across thousands of tasks. It should turn trajectories and outcomes into better decisions.
3. Learnable long-horizon memory. A longer context window is not memory. The system must learn what to retain, compress, retrieve, and reuse.
These are not independent checkboxes. Speed enables wider search. Feedback makes search improve. Memory makes improvement persist.
Together they create a compounding loop: intelligence generates experience, and experience improves intelligence.
My bet is that the next scaling curve in AI will come not only from larger models, but from faster action, more precise learning, and more durable memory.
We scaled parameters. Now we need to scale closed-loop intelligence. 🧵
1
65
WeiQu retweeted
It’s getting hard to keep track of all the new open-source cofolding models coming out. To help everyone stay up to date, I'm excited to share cofoldarena.ai, a PDB-synced leaderboard of open models, updated weekly. All predictions are available to view & download! (1/4)
WeiQu retweeted
This is not a retreat from hard scientific problems. It is a shift in who solves them: from humans using AI tools to solve one problem at a time, to foundation models autonomously pursuing dozens through deep, long-horizon reasoning.
That is why, in early 2025, we at GenSI (@hello_gensi) moved beyond generative AI for Science and went all-in on LLMs for Science. DeepMind’s move and John Jumper’s choice clearly suggest that this consensus is beginning to form, and that the inflection point is getting close.
No human scientist can reason 24/7—we need sleep, food, and emotional recovery—but an AGI scientist could. Combined with human curiosity, judgment, and ambition, that capacity could create a new kind of scientific partnership. Together, humans and AGI could explore deeper layers of the unknown and push the frontier of discovery into territory neither could reach alone.
Science MUST be where this transition becomes visible first. The next step: towards AGI for science.
FT exclusive: Google DeepMind has dismantled the team behind AlphaFold, its Nobel Prize-winning AI system for predicting protein structures, as the company overhauls the research strategy that made it one of the world’s leading AI laboratories ft.trib.al/khCoAdr
Readers added context they thought people might want to know
Google DeepMind's VP of Science states the AlphaFold team continues its work on hard problems in biology and has not been dismantled or pivoted. x.com/pushmeet/statu…
WeiQu retweeted
Excited to share our now published manuscript @CellCellPress highlighting advances in modeling the evolution of protein-protein interactions with MSA Pairformer. Big thanks to @ZhidianZ, @oliviatang116, @eunbelivable, Milot Mirdita, Martin Steinegger, and @sokrypton !🧵 (1/10)
cell.com/cell/fulltext/S0092…
WeiQu retweeted
😎Binder #5 from PallAtom v2 was designed using a head-and-tail binding strategy that simultaneously optimizes for target-binding affinity and E3 inhibitory potency. We look forward to conducting additional E3 activity inhibition assays to further validate its functional efficacy
RBX1 Binder Design Competition results are out!
Against BindCraft's dominant performance, PallAtom stands out as the strongest Diffusion-based Model.
🎉Congratulations to my old crew!
😎Binder #5 from PallAtom v2 was designed using a head-and-tail binding strategy that simultaneously optimizes for target-binding affinity and E3 inhibitory potency. We look forward to conducting additional E3 activity inhibition assays to further validate its functional efficacy
1
1
160
WeiQu retweeted
Introducing PAR: We’ve been exploring autoregressive protein generation lately — generating protein structures in an autoregressive language modeling style is surprisingly fun.
🧬 Protein Autoregressive Modeling (PAR), ByteDance Seed
We introduce PAR, the first multi-scale AR framework for protein backbone generation.
Autoregressive modeling of 3D protein structures has long been considered difficult.
Here’s how we make it possible with PAR. 👇
par-protein.github.io/
WeiQu retweeted
🧬 Protein Autoregressive Modeling (PAR), ByteDance Seed
We introduce PAR, the first multi-scale AR framework for protein backbone generation.
Autoregressive modeling of 3D protein structures has long been considered difficult.
Here’s how we make it possible with PAR. 👇
par-protein.github.io/
WeiQu retweeted
Excited to share STAR-MD: a scalable autoregressive diffusion model that generates stable, high-quality protein MD trajectories at microsecond timescales, where existing methods fail catastrophically.
Accepted at ICLR 2026! Links and details in the thread 🧵👇
1/7
WeiQu retweeted
We’re thrilled to announce our collaboration with @EliLillyandCo.
We’ll be deploying Chai’s AI models to power biologics discovery across a broad set of drug programs at Lilly.
This is a major step forward in pharma’s adoption of AI-native R&D.
WeiQu retweeted
Announcing a new @GPU_MODE kernel writing competition: our first featuring both NVIDIA and AMD hardware!
The first problem will be the Triangle Multiplication operator essential to the AlphaFold 🧬 models! It's a particularly tricky problem with no good public implementation!
WeiQu retweeted
Benchmarking Sequence-Based and AlphaFold-Based Methods for pMHC-II Binding Core Prediction: Distinct Strengths and Consensus Approaches
1. This study benchmarks four recent methods for pMHC-II binding core prediction: two sequence-based (NetMHCIIpan, DeepMHCII) and two AlphaFold-based (AF2-FT, AF3). AF3 achieved the highest positive recall (0.86), while NetMHCIIpan excelled in predicting non-binders, with the highest negative recall (0.93).
2. The study demonstrates that AlphaFold-based methods are excellent at predicting positive binders but tend to misclassify non-binding peptides. NetMHCIIpan, in contrast, is highly precise in identifying non-binders but struggles with positive recall, missing true binders.
3. A consensus approach combining AlphaFold-based methods’ strengths for positive prediction and NetMHCIIpan’s non-binder filtering improves overall accuracy and precision. This consensus approach achieves an F1 score of 0.93 with perfect recall, highlighting the potential for improved pMHC-II binding core predictions.
4. The study also identifies challenges with longer peptides and chimeric MHCs, both of which reduce prediction accuracy, especially for sequence-based methods.
5. Future improvements should focus on enhancing AlphaFold-based models’ ability to filter out non-binders and refining NetMHCIIpan’s binder prediction capabilities to strike a better balance between precision and recall.
📜Paper: doi.org/10.1101/2024.10.06.6…
WeiQu retweeted
Improved deep learning prediction of antigen–antibody interactions @PNASNews
1/ This paper presents a breakthrough in predicting antigen–antibody interactions using deep learning models, particularly focusing on the spike protein of SARS-CoV-2. It evaluates the AF2Complex model's accuracy in predicting the structure of antibody-antigen complexes.
2/ One key innovation is the use of B cell sequencing data to improve prediction accuracy. By employing known RBD-binding antibodies, the AF2Complex model significantly outperforms previous approaches, achieving a top prediction accuracy of 61% on a benchmark of 36 antibodies.
3/ The study demonstrates that the deep learning approach can distinguish true antigen-binding antibodies from randomly selected antibodies with remarkable precision. For example, the model identifies true positives with a precision of 90% while maintaining a recall of 25%.
4/ This deep learning method offers a robust tool for predicting antibody-antigen interactions without requiring experimental structural data, making it highly valuable for vaccine design and therapeutic antibody development, especially in cases like COVID-19.
5/ The study highlights the AF2Complex model's capability to detect antibodies that target diverse epitopes, proving its utility for complex immune-related tasks and expanding the potential applications of deep learning in immunology and drug development.
💻Code: github.com/FreshAirTonight/a…
📜Paper: pnas.org/doi/10.1073/pnas.24…
WeiQu retweeted
1/X Pallatom is an end-to-end all-atom protein generation model that directly learns the distribution of atomic coordinates. We believe optimizing the P(all-atom) may open a new way for de novo protein design.
code(ing):github.com/levinthal/Pallato…
bioxiv:biorxiv.org/content/10.1101/…





























