Intelligent automation finally bridges the last manual gap in high-throughput macromolecular crystallography

Macromolecular crystallography has long served as the mechanistic backbone of structural biology, providing atomic-resolution insights into proteins and nucleic acids that are essential to both fundamental discovery and structure-guided drug design. Over the past two decades, robotic handling, liquid dispensing automation, and synchrotron-coupled harvesting systems have dramatically compressed the protein-to-structure timeline, particularly for fragment and ligand screening campaigns that can generate tens of thousands of crystallization images in a single experiment. Yet one step has persistently resisted full automation: the visual identification of crystals from imaging plates. Manual inspection at this scale is not merely slow; it is inconsistent, cognitively taxing, and introduces subjective annotation errors that propagate through downstream decisions. To address this, researchers at the European Molecular Biology Laboratory in Grenoble developed AXIS, an AI-based Crystal Identification System that integrates the self-supervised DINOv2 vision transformer with state-of-the-art transfer learning and the MARCO dataset, the largest annotated collection of crystallization images. 

AXIS was trained on nearly 77,000 images consolidated from MARCO’s half-million-image repository into crystal and non-crystal classes, then fine-tuned on local facility data to capture the specific morphological and optical characteristics of each imaging environment. The system operates equally well on both visible-light and UV-fluorescence images, the latter providing intrinsic protein fluorescence signals that are particularly informative for confirming the proteinaceous nature of hits. Most significantly, AXIS implements a Lab-in-the-Loop methodology in which domain experts periodically review model outputs, correct errors, and resubmit curated annotations as a training signal, enabling the model to continuously specialize to local conditions rather than degrading over time. In practice, this iterative feedback architecture means AXIS not only matches human expert accuracy on large-scale annotation tasks but also actively corrects annotation errors introduced by human reviewers, removing a source of label noise that has undermined prior automated systems. This results in a deployable, facility-agnostic detection pipeline that removes the final manual bottleneck in high-throughput crystallography workflows, opening a credible path to fully automated protein-to-structure pipelines at scale for both academic and industrial structural biology.