Cambridge-Yale-MIT team releases Zatom-2, a single transformer pretrained for generative modeling of molecules, materials and proteins
rishabh16_ · x · 2026-10-11
A team from Cambridge, Yale, MIT and Lawrence Berkeley National Laboratory released Zatom-2 (arXiv:2610.11454), a generative model pretrained with multitask learning on atomistic data.
- A single multiscale Transformer covers small molecules, periodic materials and proteins via atom- and residue-level tokenization
- Generation uses flow matching, with interactive 3D visualizations of denoising trajectories (MolStar) on the project page
- Fully open-sourced: project page, GitHub monorepo, Hugging Face checkpoints, plus zevals, an evaluation suite for generation and structure prediction
- The paper reports joint-training results and scaling curves across domains
More from Research
- Variational Inference Explained: Turning Intractable Posteriors into Optimization — goyal__pramod · 2026-10-12
- RSIArena Round 1: GPT-6 Astra wins as Grok 4.7 places top-2 with 70% GPU budget unused — my_cat_can_code · 2026-10-12
- Billion AI scientists will tackle aging, predicts AI commentator — Dr_Singularity · 2026-10-12
- Sakana AI demonstrates a working recursive self-improvement loop with MASS — ResultBackground2450 · 2026-10-12
- LeCun vs. Dietterich: Should We Prefer Models Expressed by Short Formulas? — tdietterich · 2026-10-12
- Anshul Kundaje advises young engineer to present bio-lab replication results with more humility — anshulkundaje · 2026-10-12