Atomic structure of a single large biomolecule from diffraction patterns of random orientations

The short and intense pulses of the new X-ray free electron lasers, now operational or under construction, may make possible diffraction experiments on single molecule-sized objects with high resolution, before radiation damage destroys the sample. In a single molecule imaging (SMI) experiment thousands of diffraction patterns of single molecules with random orientations are recorded. One of the most challenging problems of SMI is how to assemble these noisy patterns of unknown orientations into a consistent single set of diffraction data. Here we present a new method which can solve the orientation problem of SMI efficiently even for large biological molecules and in the presence of noise. We show on simulated diffraction patterns of a large protein molecule, how the orientations of the patterns can be found and the structure to atomic resolution can be solved. The concept of our algorithm could be also applied to experiments where images of an object are recorded in unknown orientations and/or positions like in cryoEM or tomography.

💡 Research Summary

The paper addresses one of the most critical bottlenecks in single‑molecule imaging (SMI) with X‑ray free‑electron lasers (XFELs): the determination of the unknown orientations of thousands to millions of extremely noisy diffraction patterns recorded from identical particles in random orientations. While previous approaches such as the Expectation‑Maximization (EMC) algorithm and the Generative Topographic Mapping (GTM) method have demonstrated that orientation can be recovered, they suffer from computational complexities that scale poorly (approximately R⁶–R⁸, where R = particle diameter / desired resolution).

The authors propose a new iterative orientation algorithm that dramatically reduces computational cost while retaining robustness to noise. The key steps are:

1. Initial 3‑D intensity guess – Start from a random 3‑D reciprocal‑space intensity distribution I(q).
2. Orientation grid – Construct a quasi‑uniform grid over the (Θ, Φ) sub‑space by subdividing the faces of an icosahedron. The number of grid points N_R scales as R². The third Euler angle Ψ (rotation about the incident beam) is treated separately because a rotation about the beam merely rotates the pattern.
3. Pattern extraction – For each grid orientation r, cut a spherical cap (the Ewald‑sphere section) from I(q) to obtain a synthetic 2‑D pattern I_rij. Because the rotated vectors do not fall exactly on the Cartesian grid, the nearest‑neighbor voxel is used, and a lookup table s_rij is pre‑computed.
4. Correlation evaluation – Compute the Pearson correlation between every measured pattern m_ij and every synthetic pattern I_rij after allowing for in‑plane rotations φ′ (equivalent to shifts in the azimuthal index). The correlation for a given (r, φ′) is obtained by averaging correlations over all resolution circles (constant θ).
5. FFT acceleration – The set of correlations for all φ′ can be evaluated simultaneously using the cross‑correlation theorem and fast Fourier transforms, reducing the naïve O(N_φ²) cost to O(N_φ log N_φ). The whole correlation matrix (size N_φ × N_R × N_M) is computed on GPUs via CUDA.
6. Orientation assignment – For each measured pattern, locate the (r_m, φ′_m) that yields the maximal correlation. Patterns with a correlation below a median‑based threshold are discarded for the current iteration to avoid contaminating the reconstruction with noise‑dominated data.
7. Reconstruction update – Rotate each accepted pattern to its assigned orientation, scale it by a factor that compensates for overall intensity differences, and deposit its values onto the 3‑D grid using the pre‑computed s_rij mapping. The new 3‑D intensity is obtained by averaging contributions from all accepted patterns.
8. Iteration and refinement – Repeat steps 2‑7 until the maximum correlation values plateau, indicating convergence. A final simplex search refines the Euler angles for each pattern to sub‑grid precision, after which the full set of original (unfiltered) patterns is used to build the final high‑resolution 3‑D intensity.

The authors provide a detailed scaling analysis. The number of voxels N_I in the 3‑D reciprocal space scales as R³, the number of orientation grid points as R², and the number of pixels per pattern as R². Consequently, the dominant computational effort (correlation matrix calculation) scales as O(R⁵ log R), a substantial improvement over the R⁶–R⁸ scaling of earlier methods.

To validate the method, the authors simulated 100 000 diffraction patterns of a 109 kDa periplasmic nitrate reductase (NapAB, PDB 3ML1). Each pattern was generated with random SO(3) orientations, atomic form factors derived from Cromer‑Mann coefficients, and Poisson noise corresponding to an average of ~10⁻² photons per Shannon‑Nyquist pixel at 1.9 Å resolution. Radial intensity variation was removed, and a mild Gaussian low‑pass filter (FWHM = 2 pixels) was applied. High‑angle data (θ ≤ 24°) were retained for final structure reconstruction, while the low‑signal outer region was excluded from orientation determination.

Running the algorithm on a GPU cluster, the correlation maxima rose sharply after roughly 10–15 iterations, and the final orientation assignments matched the ground‑truth rotations with sub‑degree accuracy. Using the refined orientations, the authors reconstructed the 3‑D reciprocal‑space intensity and applied standard phase‑retrieval techniques to obtain an electron‑density map at ~1.5 Å resolution, essentially recovering the atomic model of the protein.

The paper also discusses the broader applicability of the approach. Because the method relies only on the existence of a common intersection line (or plane) between any two patterns and on the similarity of nearby orientations, it can be transferred to cryo‑electron microscopy, where particles are imaged in random orientations on a detector, or to X‑ray tomography with unknown object positions. The use of a simple correlation metric, an efficiently sampled orientation grid, and GPU‑accelerated FFTs makes the algorithm both fast and scalable, opening the possibility of processing the massive data streams expected from next‑generation XFEL facilities.

In summary, the authors present a conceptually simple yet powerful orientation‑determination algorithm for SMI that (i) replaces the computationally heavy likelihood maximization of EMC with a Pearson‑correlation search, (ii) achieves O(R⁵ log R) scaling, (iii) converges rapidly with as few as 10⁵ noisy patterns for a 100 kDa protein, and (iv) yields atomic‑resolution structures when combined with established phase‑retrieval pipelines. This work represents a significant step toward practical, high‑throughput single‑particle XFEL imaging of large biomolecules.