AXIS: A Lab-in-the-Loop Machine Learning approach for generalized detection of macromolecular crystals
Authors:
- Aurelien PersonnazAPAurelien Personnaz
- Sihyun SungSSSihyun Sung
- Raphael BourgeasRBRaphael Bourgeas
- Sruthi SunniSSSruthi Sunni
- Florine DupeuxFDFlorine Dupeux
- Bukunmi AdediranBABukunmi Adediran
- Rosicler BarbosaRBRosicler Barbosa
- Anne-Sophie HummAHAnne-Sophie Humm
- Euan Colaco-OsorioECEuan Colaco-Osorio
- José Antonio MárquezJMJosé Antonio Márquez
Preprint
Abstract
Macromolecular crystallography provides mechanistic understanding of biological processes and can be applied in drug design. Nowadays, the use of robotic systems for crystal growth and diffraction analysis is widespread and high throughput protein-to-structure pipelines for ligand and fragment screening are revolutionizing the field. However, the identification of crystals is still largely carried out through manual inspection, sometimes involving tens of thousands of images, which represents a bottleneck in an otherwise highly automated process. Here we describe AXIS , an A I-based C rystal I dentification S ystem combining the DINOv2 computer vision model, state-of-the-art transfer learning and MARCO, the largest crystallization dataset available to date, for automated crystal detection. AXIS can operate both with visible and UV light images and integrates a Lab-In-The-Loop approach combining ML and expert inputs for continuous learning and specialization. AXIS enables automated annotation of large crystallization image datasets with performance and accuracy comparable to that of human experts and the Lab-In-The-Loop approach introduced here enables efficient adaptation to local conditions facilitating widespread application, which has been a major limitation to date. AXIS can help correct human errors in image annotation and removes critical bottlenecks, particularly in the context of extensive crystallization screens or high throughput applications like fragment and ligand screening unlocking the potential for higher levels of automation that are key both in fundamental and translational research. Appendix A. GraphicalKeywords
This publication has 45 references indexed in Scilit:
- The future of integrated structural biologyStructure, 2024
- The HZB F2X‐Facility—An Efficient Crystallographic Fragment Screening PlatformApplied Research, 2024
- Accurate structure prediction of biomolecular interactions with AlphaFold 3Nature, 2024
- Highly accurate protein structure prediction with AlphaFoldNature, 2021
- Fragment‐Based Discovery of Non‐bisphosphonate Binders of Trypanosoma brucei Farnesyl Pyrophosphate SynthaseChemBioChem, 2020
- Liquid application method for time-resolved analyses by serial synchrotron crystallographyNature Methods, 2019
- Protein crystallization analysis on the World Community GridJournal of Structural and Functional Genomics, 2010
- Protein crystallization: from purified protein to diffraction-quality crystalNature Methods, 2008
- Robotic nanolitre protein crystallisation at the MRC Laboratory of Molecular BiologyProgress in Biophysics and Molecular Biology, 2005
- Small is beautiful: protein micro-crystallographyNature Structural & Molecular Biology, 1998