OXtal

OXtal is a 100-million-parameter all-atom diffusion model designed to predict experimentally realizable three-dimensional organic crystal packings directly from two-dimensional molecular graphs. Developed through a collaboration between FutureHouse, Caltech, the University of Oxford, Mila, and AITHYRA, OXtal addresses one of the grand challenges in computational chemistry: Crystal Structure Prediction (CSP). The tool is aimed at researchers and organisations working in pharmaceuticals, drug stability, and organic semiconductor development who require fast, accurate crystal structure predictions at scale.

Crystal Structure Prediction has historically relied on brute-force physics-based simulations run across large CPU clusters. In the most recent CSP Blind Test, predicting just seven target structures consumed 46 million CPU core hours. OXtal replaces this prohibitively expensive search with a generative inference process, delivering accurate predictions in seconds on a single GPU — representing a fundamental shift in how the problem is approached.

Core Capabilities

• Predicts 3D organic crystal packings directly from 2D molecular graph inputs, without requiring physics-based simulation at inference time.
• Operates as a generative model, producing accurate crystal structure predictions in seconds on a single GPU.
• Trained on over 600,000 experimentally resolved crystal structures, giving the model broad coverage of real-world molecular packing arrangements.
• Uses a novel cropping algorithm to teach the model the local "packing puzzle" — how molecules sit next to one another — allowing the global crystal structure to emerge naturally from local neighbourhood solutions.
• Employs soft symmetries (scale and data augmentation) as an inductive bias rather than hard-coded symmetry constraints, enabling more flexible and generalisable learning.

Workflow and Approach

1. Input a set of 2D molecular graphs representing the target organic molecule or molecules.
2. OXtal's diffusion model generates candidate 3D crystal packings through a generative inference process.
3. The model resolves local molecular neighbourhood arrangements, from which the full global crystal structure is derived.
4. Predicted structures can be used as a high-speed upstream filter before committing to traditional, expensive physics-based validation pipelines.

Applications and Impact

• High-throughput screening for pharmaceutical compounds, particularly where crystal form affects drug stability and bioavailability.
• Accelerated materials discovery for organic semiconductors, where crystal packing directly influences electronic performance.
• Serves as a fast pre-screening layer that can be placed upstream of conventional CPU-intensive CSP workflows, dramatically reducing computational cost.

OXtal is the work of FutureHouse AI-for-Science Independent Post-doctoral Fellow Chenghao Liu and collaborators from Caltech, the University of Oxford, Imperial College London, Mila, AITHYRA, and Synteny. Full technical details are available via the project page at oxtal.github.io and the accompanying preprint on arXiv (arXiv:2512.06987). FutureHouse is a registered 501(c)(3) nonprofit organisation.