FutureHouse 等发布 OXtal：从 2D 分子图预测有机晶体结构
OXtal: Generative Molecular Crystal Structure Prediction
FutureHouse 与 Caltech、牛津大学、Mila、AITHYRA 合作发布 OXtal，一个 100M 参数的全原子扩散模型，可直接从 2D 分子图预测可实验实现的 3D 有机晶体堆积结构。传统方法依赖大规模 CPU 物理模拟搜索，最近一次 CSP Blind Test 中预测 7 个目标就消耗 46M CPU 核心小时，而 OXtal 用生成式推理在单块 GPU 上数秒内给出预测。模型不使用硬编码对称性，而是在超过 600,000 个实验解析结构上训练并配合新的裁剪算法，以软对称性作为归纳偏置学习分子局部堆积规律。
正文
当前显示原文，中文全文译文尚未就绪。
OXtal predicted molecular crystal structures (gray) compared against ground-truth crystal structures (green).
In 1988, the editor of Nature famously called the inability to predict crystal structures from chemical composition a "continuing scandal" in the physical sciences. Nearly forty years later, Crystal Structure Prediction (CSP) remains a grand challenge in computational chemistry, with high stakes for everything from drug stability to semiconductor performance.
Today, we are excited to highlight the work of FutureHouse AI-for-Science Post-doctoral Fellow Chenghao Liu, who, alongside a collaborative team from Caltech, University of Oxford, Mila, and AITHYRA, is changing how we approach this problem.
Introducing OXtal
OXtal is a new 100M parameter all-atom diffusion model that predicts experimentally realizable 3D organic crystal packings directly from a set of 2D molecular graphs.
Traditionally, predicting these structures is a brute-force process. It relies on running physics-based simulations on large CPU clusters to search for stable states. In the most recent CSP Blind Test, predicting just seven targets consumed 46M CPU core hours . OXtal replaces this prohibitively expensive search with a generative inference process, producing accurate predictions in seconds on a single GPU.
Scaling with Soft Symmetries
Rather than relying on hard-coded symmetries often used in this problem, OXtal uses scale and data augmentation instead—i.e., soft symmetries as an inductive bias. By training on over 600,000 experimentally resolved structures with a novel cropping algorithm, the model learns the local "packing puzzle" of how molecules sit next to one another. By solving for this local neighborhood, the global crystal structure emerges naturally.
This work represents a significant step toward high-throughput screening for pharmaceuticals and organic semiconductors, offering a high-speed filter that can sit upstream of traditional, expensive physics-based validation.
For more details, please visit the links below.