Scalable 3D Reconstruction From Single Particle X-Ray Diffraction Images Based on Online Machine Learning (2312.14432v1)

Abstract: X-ray free-electron lasers (XFELs) offer unique capabilities for measuring the structure and dynamics of biomolecules, helping us understand the basic building blocks of life. Notably, high-repetition-rate XFELs enable single particle imaging (X-ray SPI) where individual, weakly scattering biomolecules are imaged under near-physiological conditions with the opportunity to access fleeting states that cannot be captured in cryogenic or crystallized conditions. Existing X-ray SPI reconstruction algorithms, which estimate the unknown orientation of a particle in each captured image as well as its shared 3D structure, are inadequate in handling the massive datasets generated by these emerging XFELs. Here, we introduce X-RAI, an online reconstruction framework that estimates the structure of a 3D macromolecule from large X-ray SPI datasets. X-RAI consists of a convolutional encoder, which amortizes pose estimation over large datasets, as well as a physics-based decoder, which employs an implicit neural representation to enable high-quality 3D reconstruction in an end-to-end, self-supervised manner. We demonstrate that X-RAI achieves state-of-the-art performance for small-scale datasets in simulation and challenging experimental settings and demonstrate its unprecedented ability to process large datasets containing millions of diffraction images in an online fashion. These abilities signify a paradigm shift in X-ray SPI towards real-time capture and reconstruction.

• The paper introduces X-RAI, combining a convolutional encoder with a physics-based decoder to achieve high-quality 3D reconstructions.
• It employs an amortized, self-supervised online learning approach to efficiently estimate particle poses from millions of XFEL images.
• Experimental validation on datasets like PR772 demonstrates X-RAI’s superior accuracy and scalability compared to traditional methods.

Introduction to Single Particle Imaging and Challenges

Single particle imaging (SPI) with X-ray free-electron lasers (XFELs) is a cutting-edge technique in the paper of biomolecules such as proteins and viruses, offering insights into their structures and functions. Unlike traditional methods, X-ray SPI can image individual particles in near-physiological conditions, capturing snapshots that might not be feasible using other methods. However, processing the massive datasets generated by high-repetition-rate XFELs presents a significant computational challenge. Conventional reconstruction algorithms are not adept at managing these large volumes of data due to their reliance on exhaustive search strategies to estimate unknown particle orientations.

Introducing X-RAI

To address these computational challenges, a novel approach called X-RAI has been introduced. It combines a convolutional encoder, which estimates the unknown orientation or pose of particles, with a physics-based decoder, to create high-quality 3D reconstructions from large X-ray SPI datasets. This method uses an amortized approach to pose estimation, which means it learns to predict poses across the entire dataset, thereby avoiding the need for time-intensive searches. A distinct feature of the decoder is its reliance on an implicit neural representation, offering a continuous 3D intensity model which is more suited for gradient-based optimization. X-RAI functions in a self-supervised, end-to-end manner. Moreover, it can be run online, processing data incrementally and allowing for real-time updates to the model, as opposed to other algorithms which require several passes over the entire dataset.

Online and Offline Reconstruction Capabilities

X-RAI's performance has been showcased in both simulated (offline) and on-the-fly (online) scenarios. Offline, it demonstrated superior reconstruction on small-scale datasets compared to previous techniques like M-TIP and Dragonfly. Online, X-RAI effectively processed millions of images sequentially, with its throughput unaffected by the dataset's size. This illustrates its potential for real-time reconstruction during actual SPI experiments.

Experimental Validation and Future Impact

Validation against large experimental datasets, such as the PR772 coliphage dataset, demonstrated both the accuracy and scalability of X-RAI. Additionally, its design accommodates soft enforcement of known symmetries in biomolecules during the reconstruction process. X-RAI's development signals a significant advancement in SPI by enabling efficient handling of the extensive data projected with the advent of next-generation XFEL sources. It promises to facilitate real-time analyses during experiments and could be crucial in decoding transient biomolecular states that are elusive to existing imaging techniques.

In conclusion, X-RAI is a transformative tool in the field of structural biology, suited for both the current and future landscape of X-ray SPI. It promises to unlock the full potential of emerging data-rich SPI experiments and could revolutionize our understanding of biomolecular dynamics.