XFEL diffraction measurements

Data collection. XFEL diffraction data were collected inside the LAMP (laser applications in materials processing) chamber (31) at the AMO endstation (32) of the LCLS. The particle beam exiting from the Uppsala aerosol injector was intersected with the x-ray beam. The LCLS generated x-ray pulses of 1 to 2 mJ at a photon energy of 800 eV (wavelength, 1.55 nm) with a pulse duration of 170 fs and a peak fluence of 0.02 mJ/μm² (14) at a repetition rate of 120 pulses/s. About 5% of the LCLS pulses were dumped (“BYKICK” mode) to continuously monitor the dark background. This means that the LCLS delivered effectively only about 114 pulses/s to the interaction region. Diffraction images were recorded synchronously with a pair of pnCCD area detector panels (33) operated in gain mode 5. The panels were placed at distances of 250 mm (TBSV data) and 370 mm (carboxysome and sucrose data). Each panel has a sensitive area of 76.8 mm × 38.4 mm with 1024 × 512 pixels. The direct beam and small-angle scattering passed through the gap between the panels. At a detector distance of 250 mm, the gap was 3.3 mm wide, and at a detector distance of 370 mm, it was 5.5 mm wide. Data were monitored live with the Hummingbird software package (34).

Data preprocessing. Diffraction data were preprocessed using the Hummingbird software package (34) and Psana (35). Configuration files (conf_preproc.py and conf_amol3416.py) can be downloaded from https://github.com/mhantke/electrospray_injection. The datasets that were used for analysis are listed in Table 2. Raw data were pedestal-subtracted using dark frames and rescaled to the unit of x-ray photons. Pedestal correction was followed by a three-step common mode subtraction procedure that was carried out for each panel individually, first for every quadrant (half panel), then for each fast, and finally for each slowly changing pixel dimension. Common mode is defined as the median pixel value of the selection of pixels that measure below 0.5 photons. For the faulty top-right quadrant, additionally ASIC (application-specific integrated circuit)–wise common mode subtractions were applied, first for the fast and then for the slowly changing pixel dimension. For certain runs (defined in amol3116_run_params.csv and amol3416_run_params.csv), all pixels of the inner one or two ASICs of the faulty quadrant were upscaled by a factor of 2. Detector geometry was applied by taking into account the relative position of the detector halves, the pnCCD readout timing issue for particular runs, and the column mismatch that was caused by a wiring error of the pnCCD chip. As hits, we selected those diffraction patterns that counted more than 3500 pixels measuring at least one photon and being located further than 200 pixels away from the center.

Data prediction. Diffraction data for carboxysomes, TBSV particles, and Rubisco proteins were simulated with the Condor software package (36). For Rubisco proteins, the electron density was estimated to be 0.43 Å⁻³ on the basis of a mass density of 1.35 g/cm³ and an atomic composition of H₈₆C₅₂N₁₃O₁₅S for proteins (3). The incident intensity was set to the measured peak fluence of 0.02 mJ/μm².

Image reconstruction. For retrieving the phase of selected carboxysome and TBSV diffraction patterns and reconstructing 2D projection images, we used the Hawk software package (37). Before phasing, the diffraction patterns were truncated at 0.5 photons and binned to 128 × 128 images. We used a binary mask excluding hot, saturated, and shadowed pixels.

The support was initialized with a static spherical mask of radius slightly larger than the expected particle size. The iterative phase retrieval was performed with 1000 iterations of the relaxed averaged alternating reflections algorithm (38) (TBSV hits) or the hybrid input-output (HIO) algorithm (39) (carboxysome hits), followed by 1000 iterations of the error reduction algorithm (39) in both cases, enforcing the projected electron densities to be real and positive. The final reconstruction is an average of 100 independent reconstructions with a random initial guess for the phases. To check for reproducibility of the reconstructions, we calculated phase retrieval transfer functions (PRTFs) (Fig. 4).

The dashed lines indicate the value e⁻¹, often used as threshold for judging the reproducibly of the retrieved phases.

Rubisco data analysis. Diffraction patterns were preprocessed as described above and then binned 16 × 16 pixels to improve the signal-to-noise ratio. Then, pixel values below the background floor of half a photon were set to zero to reduce the background from gas fluorescence and visible light, and all other values were rounded to the closest integer value. For every pixel, the variance and the mean value were calculated from the buffer run. Pixels for which the ratio of variance and mean value deviated by less than 0.3 from 1 were identified as good pixels because of the indication that their values followed Poisson statistics. Pixels that did not fall into this category were masked out. The mask was extended manually to exclude the halo of the direct beam and the edges of the detector quadrants. Last, images were background-corrected by subtraction of the median readout value for every pixel, respectively.