The Marques team at EMBL Grenoble has developed an AI-based training method to automate crystal screening and identification for macromolecular crystallography studies
The way structural biologists study the mysteries of protein folding has changed significantly over the past few decades, thanks to advances in technology. Several innovations in the field of macromolecular crystallography developed at EMBL Grenoble have almost completely automated tedious manual processes, thereby improving sample quality and accelerating data collection. However, some steps in the crystallization process still require manual intervention, one of which is crystal screening and identification.
The Marques team, which operates a high-throughput crystallization and fragment screening facility (HTX Lab) at EMBL Grenoble, addressed this problem by developing an AI-based training method called AXIS (AI-based Crystal Identification System). They also designed and integrated its local implementation, CRIMS-AXIS, into the Crystal Information Management System (CRIMS), a bespoke platform for managing all information and data related to crystallization experiments.
The method, recently described in the International Union of Crystallography Journal, is based on a lab-in-the-loop approach that combines large-scale computer vision models with expert input obtained through CRIMS software. We provide a simple and cost-effective AI-assisted automated screening system to speed up structure-based drug design.
What is polymer crystallography?
Structural biology aims to gain insight into biological processes by determining the three-dimensional structure of macromolecules at the atomic level. One way to accomplish this is through polymer crystallography. This involves bombarding a regularly structured array of molecules packed within a crystal with high-intensity X-rays.
To perform these experiments, large quantities of crystals must be generated and prepared for data collection. This is usually done in a dedicated facility like the HTX lab at EMBL Grenoble. Technicians test different solutions on protein samples to find the solution that produces the crystal structure.
lab in the loop
For the past few decades, HTX Labs has been working with EMBL Grenoble’s technology-focused team to automate manual processes. Thanks to robotics and software development, much of the pipeline is already automated and users can run experiments remotely via the CRIMS software. However, the crystal screening and identification steps were still performed manually by the user through CRIMS.
“On average, each screening generates 13,000 images, about 5% of which show crystals,” explained Aurelien Persons, an ARISE postdoctoral researcher on the Marques team that developed AXIS. “The idea was to free researchers from the tedious task of checking for crystals in images, while improving quality control at the same time.”
Personas, who has a background in computer science and experience in web development and applied AI, has developed a new training method that uses visible and ultraviolet images to predict the probability of crystal formation. He started with a large-scale Vision Transformer model (a type of model that is adapted to computer vision tasks and shares architectural principles with large-scale language models) and pre-trained it on millions of images collected from intranets. We then first trained using a relatively large general-purpose crystal structure dataset before specializing on local data from the CRIMS system.
Thus, CRIMS-AXIS was born, but it still needed refinement. Personas took a “lab-in-the-loop” approach to training, comparing machine learning predictions to manual input from hundreds of expert scientists conducting experiments in the HTX lab. “This is like the best of both worlds, where AI-based predictions are modified by expert scientists to help retrain and improve the system,” commented Jose Marquez, team leader and senior scientist who heads the HTX Institute. “Two rounds of lab-in-the-loop training significantly improved the accuracy of AXIS predictions.”
What is “Lab in the Loop”?
Lab-in-the-loop aims to continually modify models using iterative loops and user feedback to improve the experimentation process using machine learning and generative AI techniques.
This method accelerates the learning process by focusing on the differences between machine learning predictions and human scores. However, humans can also make mistakes. For each conflicting score, HTX experts had to decide whether the machine or the scientist was right. At each iteration, Personnaz added an expanded training dataset along with selected data to retrain the model. “While the first iteration showed obvious errors, the second iteration showed fewer ‘stupid’ mistakes,” Personas commented. “This shows that CRIMS-AXIS is making good progress, because the number of remaining cases is increasing.
Fully integrated into CRIMS software, CRIMS-AXIS identifies not only crystals but also needle-like or other crystal forms. This model has received positive feedback from users. “AXIS removes a critical bottleneck, especially in the context of large-scale crystallization screening, and unlocks the potential for higher levels of automation, which is important for both basic and translational research,” explained Sihyun Sung, staff scientist on the Marquez team and CRIMS-AXIS user.
This research benefits from support from the European Commission through the Fragment-Screen project coordinated by Instruct-ERIC, and the machine learning models are stored in a central repository and available to the scientific community, making them easy to integrate in other laboratories.
next step
Personas is currently working with colleagues at EMBL Grenoble to improve CRIMS-AXIS and upgrade the automated pipeline.
On the machine learning side, we are collaborating with Alana de Sousa, an astrophysicist specializing in AI research who is currently training in the Marquez team. They seek to apply “self-supervised learning” to CRIMS-AXIS by leveraging the large number of unlabeled crystallographic images generated over the years using the HTX platform. The objective is to limit the diversity of training images, as we attempt to pre-train the model using only unlabeled crystal structure images. This allows the model to “learn to understand” crystallographic images and potentially achieve better results in crystal identification. The researchers also plan to test whether it can be used for other tasks such as multiclass classification, crystal detection, and segmentation.
To move toward a fully automated crystallization pipeline, Personnaz is working with software engineer Jeremy Sinoir on the Papp team to integrate automated crystal collection into CRIMS. Currently, HTX operators must use software to select which crystals should be collected and prepared for diffraction data collection experiments and how. The “automated harvest” being developed by Personnaz and Sinoir has been integrated into the latest version of the harvesting machine, the CrystalDirect Harvester 4, and will soon be available on the HTX platform. The Marquez team is also extending the lab-in-the-loop approach to other steps in the crystallography process.
This project has received funding from the European Union’s Horizon 2020 research and innovation program under Marie Skłodowska and Curie grant agreement no. 945405.
/Open to the public. This material from the original organization/author may be of a contemporary nature and has been edited for clarity, style, and length. Mirage.News does not take any institutional position or position, and all views, positions, and conclusions expressed herein are those of the authors alone. Read the full text here.
