High-Brightness Self-seeded X-ray Free Electron Laser to Precisely Map Macromolecular Structure


 We demonstrate a hard-X-ray self-seeded (HXRSS) free-electron laser (FEL) at Pohang Accelerator Laboratory with an unprecedented peak brightness (3.2 × 1035 photons/(s·mm2·mrad2·0.1%BW)). The self-seeded FEL generates hard X-ray pulses with improved spectral purity; the average pulse energy was 0.85 mJ at 9.7 keV, almost as high as in SASE mode; the bandwidth (0.19 eV) is about 1/70 as wide, the peak spectral brightness is 40 times higher than in self-amplified spontaneous emission (SASE) mode, and the stability is excellent with > 94% of shots exceeding the average SASE intensity. Using this self-seeded XFEL, we conducted serial femtosecond crystallography (SFX) experiments to map the structure of lysozyme protein; data-quality metrics such as Rsplit, multiplicity, and signal-to-noise ratio for the SFX were substantially increased. We precisely map out the structure of lysozyme protein with substantially better statistics for the diffraction data and significantly sharper electron density maps compared to maps obtained using SASE mode.


Introduction
The extreme peak brightness and ultrashort pulses provided by X-ray free-electron lasers (XFEL) [1][2][3][4][5] allow data collection from micrometersized protein crystals at room temperature (the functional temperature of their constituent molecules) while outrunning radiation damage. This 'diffraction-before-destruction approach' has been applied in serial femtosecond crystallography (SFX), which has revolutionized X-ray crystallography and has been considered an important tool to determine the structure of proteins that are di cult to crystallize [6][7][8][9] .
XFELs have noisy and spiky spectra because the devices exploit the self-ampli ed spontaneous emission (SASE) that starts from the electron beam shot noise. The uctuation of the noisy and spiky spectra of the XFEL can limit the data quality of SFX. Self-seeding is a promising approach to overcome the de ciencies of XFELs and to realize bright, fully-coherent FEL sources in the hard X-ray domain. The use of seeded FEL pulses with their higher reproducibility and 'cleaner' spectrum than SASE might accelerate convergence of the merged re ection intensities of the SFX data 10 . However, existing hard X-ray self-seeded (HXRSS) FELs have limited radiation pulse energy and spectral brightness, and inadequate stability. A previous study 11`o f the self-seeded XFEL for SFX did not show any improvement in the data-quality metrics of the SFX compared to SASE, in contrast to the expectation that the use of self-seeded pulses might result in SFX data of a superior quality to that collected using SASE pulses.
The idea of self-seeding 12 has been proposed to overcome the limitation of SASE FEL, and later a self-seeding scheme with a four-crystal monochromator in Bragg re ection geometry was proposed 13 . A more-compact self-seeding scheme 14 that uses a single-crystal monochromator was proposed for the hard X-ray region; this design exploits the phenomenon that the forward Bragg-diffracted monochromatic beam has a small delay time, which allows a very short (<5 m) chicane. The forward Bragg diffraction (FBD) through a thin diamond crystal produces a train of monochromatic wakes that trail the main X-ray pulse by a few tens of femtoseconds 15,16 . By using a magnetic chicane to detour the electron bunch (e-bunch) so that it and the wake overlap in time, the monochromatic seed signal can be ampli ed in the downstream undulators.
The rst successful demonstration of an HXRSS FEL using an FBD monochromator was performed out at the Linac Coherent Light Source (LCLS), produced 8.3-keV X-ray pulses with a bandwidth of 0.4-0.5 eV that was about 1/40-1/50 as wide as the SASE bandwidth 17 . The average pulse intensity of the HXRSS FEL pulses at the LCLS was 573 ± 290 µJ at 5.5 keV, and intensity uctuation was ~50% 18 . The average peak spectral intensity was 1.7 times larger than in SASE mode. In the self-seeding experiments at LCLS, including a soft X-ray self-seeding, the radiation spectrum often showed a pedestal-like distribution around the seeded frequency; this distribution limits spectral brightness 17,19,20 . The pedestals originate in longitudinal phase space modulations produced by microbunching instability (MBI) upstream of the undulators.
The problem of the large delays in Bragg-re ection monochromators 13 was overcome, and self-seeding using Bragg re ections was demonstrated at the SPring-8 Angstrom Compact Free Electron Laser (SACLA); the design used a channel-cut Si crystal monochromator with a tiny gap of 90 µm. 21 The seed, with a bandwidth of 1.3 eV (full-width at half-maximum, FWHM) at 9.85 keV, was ltered from the SASE radiation by the 111 Bragg re ections from the channel-cut crystal. The average pulse energy of the self-seeded XFEL at 9.85 keV was 450 µJ, vs. 780 µJ in SASE mode. The peak spectral intensity of the self-seeded XFEL pulses was six times higher than in SASE mode. However, the Xray pulses had a relatively large bandwidth of 3 eV (FWHM) because of the large bandwidth of the seed signal as well as an energy chirp in the e-bunch. Using the Si-220 Bragg re ections from the channel-cut crystal, the bandwidth was reduced to 0.6 eV (FWHM) at 9.0 keV with an average pulse energy of ~250 µJ 22 .
At PAL-XFEL we demonstrated an HXRSS FEL using an FBD monochromator ( Fig. 1) that has favorable source features compared to SASE mode: a high average peak spectral intensity exceeded that of SASE by a factor of 12 in self-seeded mode, pulse energy of 0.85 mJ at 9.7 keV; substantial improvements in the stability of self-seeding; and substantially suppressed pedestal effects. We also demonstrated that using the self-seeded XFEL with an unprecedented peak brightness and high stability, can substantially increase data-quality metrics of the SFX, such as R split , multiplicity, and signal-to-noise ratio (SNR), and yield signi cantly sharper electron density maps than those obtained using SASE mode.

Improving The Spectral Brightness Of Self-seeded Xfel
To increase spectral brightness, we used e-bunches with a higher charge (180 pC) and a longer duration (42 fs FWHM) than in the previous study 17,24 . The e-bunch energy is 8.538 GeV; the undulator parameter is 1.87; and the duration of the SASE FEL radiation pulse is 20 fs (FWHM), as measured using the cross-correlation method 23,24 . A laser heater (LH) was used to suppress MBI (Methods Section).
One prerequisite to generate X-rays that have narrow bandwidth (narrowband) and high spectral brightness for the HXRSS FELs is a narrowband seed 25 . For e cient seeding, the bandwidth of the monochromatic seed must match the FEL bandwidth, and the duration of the monochromatic wake should be comparable to or longer than that of the SASE signal for seeding. To generate such a seed, we use high-index Bragg re ections (ie., 33-3 or 115) with an FBD bandwidth of ~0.1 eV, instead of the 004 re ection with a ~0.3-eV bandwidth, as typically used in short-pulse modes 17,24 . Unlike the self-seeding in re ection geometry 21,22 , the path length delay of the seed pulse does not depend on the crystal index, so the high-index Bragg re ections, like 33-3 or 115, can be used to generate a seed bandwidth < 0.1 eV. For a better overlap with long e-bunch and a higher seed intensity, we use the 0 th -order wake of the FBD signal as the monochromatic seed instead of the 1 st -order wake that has been used in previous experiments ( Fig. 1f) (details in ref. [25] and Supplementary Fig. 1). The duration of the 0 th -order wake, in this case, is su ciently long (~65 fs) to allow for su cient delay of the e-bunch with respect to the SASE pulse, and full separation of the seed signal from the SASE background ( Supplementary Fig. 2).
Another prerequisite to increase spectral brightness is to suppress the pedestal-like distribution around the central seed frequency. This effect originates in the MBI induced by bunch compression, which creates detrimental sideband modulation of the e-bunch. Once the sidebands are generated, the electron oscillations are driven by the multiple-frequency ponderomotive potential. As a result, the e ciency of FEL generation at the carrier frequency is reduced, and the spectral quality is degraded by diversion of radiation power into sideband frequencies 26,27 . A laser heater can e ciently suppress MBI in both SXRSS 20 and SASE mode; 28-31 however, for the HXRSS, the improvement of spectral brightness by using a laser heater has not been investigated experimentally.
The LH can suppress the MBI and increase the peak spectral intensity in self-seeded mode ( Figure 2). The pedestal around the center peak is signi cantly reduced as slice energy spread increases, so the peak intensity increases (Fig. 2a). These results show that the sideband ampli cation due to the MBI is effectively suppressed, so the main peak is solely ampli ed. The fraction of FEL intensity enclosed within the bandwidth shows that spectral purity is signi cantly increased using the LH (Fig. 2b). As the slice energy spread increases, the peak intensity (solid red line) also increases (Fig. 2c); it reaches its maximum when the slice energy spread is ~27 keV. The optimal slice energy spread for self-seeded mode is about 5 keV higher than for SASE. 32 To suppress pedestal effects due to microbunching instability substantially, a calm longitudinal-phase-space with energy modulation further suppressed is required (Fig. 1e). However, the total sum of the spectrum (blue dotted line in Fig. 2c) remains almost constant until at LH = 27 keV; this result supports the hypothesis that unsuppressed MBIs channel the radiation power into the sidebands.
Single-shot spectra maps (Fig. 3a) were obtained using a 0.26-eV resolution Si(333) curved-crystal single-shot spectrometer 33 ( Supplementary   Fig. 3) for 9.7-keV X-rays in SASE and self-seeded modes. The measured bandwidths of X-rays were 13.0 ± 0.1 eV in SASE and 0.35 ± 0.01 eV in self-seeded mode (Fig. 3b); the latter dropped to 0.24 eV after deconvolution from the spectrometer resolution. More-accurate measurements than these were obtained using a 0.09-eV-resolution Si(333) at-crystal scanning spectrometer; they reveal a time-averaged bandwidth of 0.21 ± 0.01 eV in self-seeded mode, which drops to 0.19 eV after deconvolution (Fig. 3c). The FBD seed bandwidth in the 33-3 Bragg re ection from the diamond crystal is 0.06 eV (Supplementary Table 1), but the resultant bandwidth of the self-seeded XFEL increased to 0.19 eV because of the energy chirp of the e-bunch. Assuming the same pulse duration in the self-seeded mode as the SASE FEL radiation pulse (20 fs), and Gaussian pulse shape, the Fourier-transform-limited HXRSS FEL radiation bandwidth should be ~ 0.1 eV. The single-shot pulse bandwidth is de nitely smaller than the 0.19-eV averaged bandwidth, so the PAL-XFEL HXRSS pulses are less than a factor of two larger than the Fourier-transform limit.
Self-seeded mode had 12 times higher average peak spectral intensity than SASE mode (Fig. 3b), but this number is limited by the spectral resolution of the single-shot spectrometer. The average pulse energy of the HXRSS at PAL-XFEL is 0.85 mJ, or 57% of the 1.5-mJ average pulse energy in SASE mode, as measured by the electron energy loss scan 34 . Appropriate undulator tapering was applied for 20 undulators ( Supplementary Fig. 4). The ratio of the integrated spectral area of the single-shot spectrometer for SASE to self-seeded mode is 1.64; this ratio is consistent with the XFEL intensity ratio of 1.76 (1.5 mJ to 0.85 mJ) that was measured by the electron energy loss scan. Overall, the peak brightness of the PAL-XFEL HXRSS FEL is calculated to be 3.2×10 35 photons/(s·mm 2 ·mrad 2 ·0.1%BW), which is 40 times higher than that of SASE, the highest achieved to date.
The radiation pulse energy of the PAL-XFEL HXRSS is both high and very stable. The self-seeded mode has a consistently higher intensity than SASE mode (Fig. 3d). In self-seeded mode, > 94% of the shots have an intensity higher than the average SASE intensity (i.e., > 1 a.u.). Such seeding stability is mainly due to the stability of the PAL-XFEL, which has a very small shot-to-shot electron-energy jitter of 0.012% (r.m.s.) 24,31 . The resultant shot-to-shot uctuation of the central radiation wavelength of SASE FEL was measured to be 0.025% (r.m.s.), which is one-half the Pierce parameter ρ = 5 x 10 -4 (Relative SASE bandwidth, ~5.6 x 10 -4 ), so self-seeded pulses are almost always ampli ed.

Serial Femtosecond Crystallography With A Self-seeded Xfel
To solve a structure for the SFX, the necessary number of indexed snapshot patterns of crystals depends on the SNR of the individual patterns, the symmetry of the crystal, and the variability of parameters on which the diffraction depends from shot to shot (such as the chaotic spectrum of FEL pulses) 6 . These factors in uence the nal accuracy of the merged data. To determine de novo the structure of a protein in which no homologous structures exist, the experimental phasing of SFX data, the data must have high resolution and a very high multiplicity of data sets for phase determination 7,10,35-38 . The large shot-by-shot variations in X-ray intensity and photon energy may make experimental phasing of XFEL data very challenging.
We conducted test of self-seeded XFEL for SFX, because the self-seeded XFEL that we achieved performs extremely well. A previous did not show any difference in the data quality metrics of the SFX compared to SASE 11 , but the peak spectral brightness of our XFEL is about ten times higher compared to the XFEL used previously and 40 times higher than SASE, with excellent stability. We expect that reduction in the relative bandwidth from ΔE/E=1.3×10 -3 (SASE) to ΔE/E=1.9×10 -5 (SS) will sharpen diffraction patterns, especially those collected at large scattering angles, which are responsible for increasing the resolution. Also, we expect an increase in ltration rate of raw data owing to the higher spectral intensity of the self-seeded XFEL compared to SASE.
We performed a demonstration experiment by mapping out the three-dimensional structure of the lysozyme from chicken eggwhite and performing a comparative analysis of the results obtained using the narrowband HXRSS FEL and the broadband SASE FEL (see Methods for the crystal preparation and experimental conditions).
We collected and processed three data sets that had different numbers of images for both self-seeded and SASE modes: SS1/SASE1 (111,467/101,443), SS2/SASE2 (38,510/38,686), and SS3/SASE3 (20,209/20,530). The indexing rates were substantial in all cases. For example, SS1; 70,656 crystal diffraction patterns (63.4%) were identi ed as crystal hits, and 33,663 of them were indexed (47.6%). The index rates of the self-seeding data sets were higher than those of the SASE data sets (Table 1). SFX data quality metrics such as SNR [or I/σ], multiplicity, R split (i.e., the consistency of merged intensity distributions between two half-datasets separated from the full dataset), and correlation coe cient [CC*] strongly depend on the number of images, as is known (Fig. 4, Supplementary Table 2). However, the self-seeding data shows superior metrics than the SASE data at high resolutions, unlike a previous report 11 . Remarkably, the self-seeding data sets had twice the multiplicity of the SASE data set at all resolutions (Fig. 4b), so the nal accuracy of the merged data is improved, even with the same number of hit images (see Methods for SFX data processing).  Bank code 1VDS) as a search model, then conducted atomic model re nement using phenix.re ne, then inspected of (mFo-DFc) omit maps. 40 (see Methods for structure determination, re nement, and analysis). To compare and analyze the structures and their electron density maps without bias or error, we performed structural determination using the same numbers of hit images for self-seeding and SASE data sets (SS1/SASE1, SS2/SASE2, and SS3/SASE3).
After re nement, when we compared the models with their structure maps (SS1 and SASE1), we found apparent improvements in 2mFo-DFc maps of the self-seeded mode (Fig. 5a), even though lysozyme is a globular protein and has some buried residues that strongly interact with other residues. To get a much better view, we obtained bias-free mFo-DFc omit maps by sorting out the residues (Fig. 5b). Comparison of the mFo-DFc omit maps at 1.75-Å resolution (Fig. 5b) clearly shows that the maps of the ten residues (Phe21/Ala28/Tyr41/Trp46/Phe52/Asn62/Tyr71/Trp81/Trp126/Trp141) are not blurred in self-seeded mode; the maps, including the side chains and the main chains (carboxyl groups, nitrogens on the peptide backbones, and α-carbons), are sharper than those obtained in SASE mode. For instance, in the Phe21 and Asn62 maps, β-carbons and side chains are revealed clearly only in self-seeded mode. Re ned models without a speci c residue were deleted from the original structure (Supplementary Table 3).
Comparative analysis of the mFo-DFc electron density maps of the ten residues reveals the superiority of the self-seeded data set over the SASE mode data sets ( Table 2, Supplementary Fig. 5). For example, even though the data-quality metrics of the SS3 data are inferior to those of the SASE1 (SS3 dataset has one-fourth as many indexed images as the SASE1), the omit maps of the ten residues from the SS3 data are better than those from the SASE data. B-factors are crystallographic parameters to explain this big difference. The average B-factors 41 of both protein and solvent waters models are relatively lower in the models from the self-seeded than in those from SASE mode, and the average Bfactors are independent of the number of indexed images (Table1: Model re nement). These traits indicate that the atomic displacement uctuations are relatively weaker when a narrowband self-seeded FEL is used, than when a broadband SASE FEL is used. The reduced uctuations might help increase the re nement of the model with sharpened electron density maps. The overall sharpening of the omit maps obtained from the self-seeding data resulted from phasing-quality data with fewer patterns. The high quality of data obtained in self-seeded mode is a result of the use of recurrent shots from a highly-stable self-seeded XFEL.

Conclusion
The PAL-XFEL HXRSS successfully demonstrated a forward Bragg diffraction self-seeded XFEL with unprecedented peak brightness (3.2 × 10 35 photons/(s·mm 2 ·mrad 2 ·0.1%BW)) and stability; the average pulse energy is 0.85 mJ at 9.7 keV, the bandwidth (0.19 eV) is about 1/70 as wide, the peak spectral brightness is 40 times higher, and the stability is excellent with > 94% of shots exceeding the average SASE intensity. We used high-index Bragg re ections (33 − 3 or 115) to exploit a narrow seed bandwidth < 0.1 eV and a long wake duration ~ 65 fs. A calm longitudinal-phase-space with energy modulation further suppressed is required to suppress pedestal effects due to microbunching instability substantially. We demonstrated that high-spectral-intensity and high stability self-seeded XFEL improves the data-quality metrics of SFX: it achieves outstanding quality in the SNR, multiplicity, CC*, and R split compared to the large-bandwidth SASE. The high multiplicity of the selfseeding data sets yields phasing-quality data with fewer patterns than in SASE datasets, and improve the re nement of the model with sharpened electron density maps. The self-seeded data set achives superior quality of electron-density map over the SASE mode data sets. Even with one-fourth of the indexed images of the SASE data set, the self-seeded data set shows a better or similar electron density maps for the residues. The improved structure map by the self-seeded XFEL indicates that high brightness narrowband XFEL increases the resolution of signal collection, and helps to solve three-dimensional macromolecular structures with high resolution, especially for a very small crystal.

Methods
Laser heater to suppress MBI. The laser heater adds a slice energy spread to the 150-MeV e-bunch. The IR laser beam size is comparable to that of the electron bunch, so the energy spread distribution assumes a super-Gaussian pro le that can effectively suppress the MBI. The induced slice-energy spread of the e-bunch at LH as a function of the IR laser energy was measured using a transverse de ector and an energy spectrometer located after the rst bunch compressor. The accelerating sections (L1 and XLIN) and the bunch compressor BC1 downstream of the laser heater were all turned off (Fig. 1). The longitudinal phase space of e-bunch was simulated for three cases of the laser heater condition, where no laser heater case (top), optimized cases for SASE (middle), and self-seeding (bottom) (Fig. 1e). The spectral purity of the self-seeded FEL is very sensitive to the energy modulation of the e-bunch, so the optimal condition of laser heater for self-seeding is different from that for SASE. The CSR (green dotted line in Fig. 2c) measured at the third bunch compressor using a visible CCD camera 42 is due to the MBI; this result shows that MBI should be more suppressed by the laser heater for the self-seeding than for SASE.  35 . In both selfseeded FEL and SASE modes, the X-ray pulse was focused to a beam size of 2.5 μm (horizontal) × 2.5 μm (vertical) (FWHM) using a Kirkpatrick-Baez mirror 44 . The diffraction data were collected using an MX225-HS detector with a 4×4 binning mode (pixel size: 156 μm×156 μm) (Rayonix, LLC, Evanston, IL, USA) at room temperature and monitored by OnDA 45 . The 4 × 4 binning mode was used to match the XFEL repetition rate of 30 Hz. The distance between the sample position and the detector was 111 mm and was validated by comparing the index rate of each data set. Aside from using the self-seeded FEL or SASE mode, all other conditions were identical in both experiments. We collected six data sets: three in self-seeded mode (SS1, SS2, and SS3) and three in SASE mode (SASE1, SASE2, and SASE3).
SFX data processing. After data collection, the hit images were ltered using Cheetah (version 8). 46 The parameters for peak detection were optimized for Cheetah including min-snr of 4.0. The pre-processed images were further indexed, integrated, merged, and post-re ned using CrystFEL (version 0.6.3) 47,48 . The experimental geometry was also re ned for CrystFEL. Indexing was performed using DirAx (version 1.17) 49 with peak integration parameters of int-radius = 3, 4, 5. The measured diffraction intensities were merged with process_hkl in the CrystFEL suite 47,48 . To investigate the data statistics from the two modes (self-seeded and SASE) carefully, we processed data with the same methods and the same parameters for consistency.
Structure determination, re nement, and analysis. The structure of lysozyme from chicken eggwhite was determined (Table 1) by the molecular replacement method using the Phaser-MR in PHENIX (version 1.14-3260), 40 using a model of lysozyme (Protein Data Bank code 1VDS) as a search model. During the calculation of molecular replacement, we excluded water molecules from the template model to avoid model bias. Water molecules were inspected and added manually using Coot (version 0.8.9), 50 by reference to mFo−DFc maps. Water molecules were placed in correct positions depending on the density map where positive peaks higher than 1.5σ and 3.0σ occurred in the 2mFo-DFc map and mFo-DFc map, respectively. The molecular replacement model was rst re ned with a rigid-body protocol and Cartesian simulated annealing (starting at 5,000 K) using phenix.re ne to reduce model bias. After ve cycles of restrained re nement, the model was evaluated by MolProbity (version 4.4) 51 . The data of lysozyme crystals from both self-seeded and SASE modes belonged to the tetragonal space group P4 3 2 1 2, with unit cell parameters of a = b = 77.56~77.88 Å, c = 37.32 Å, α = β = γ = 90°. To inspect effects on map quality using the self-seeded mode, we made all of the omit maps on residues of the lysozyme model excluding glycine, which cannot present meaningful maps. Therefore, we manually deleted each residue from the lysozyme model and performed phenix.re ne in PHENIX 40 to generate an mFo-DFc map on each residue (Supplementary Table 2). Omit maps were generated to reduce the possible effect of model bias. Six models of the self-seeded and SASE modes were calculated in the same manner for a fair comparison.

Data availability
The coordinates and structural factors have been deposited in the Research Collaboratory for Structural Bioinformatics (RCSB) under the accession codes 7BYO/7D01/7D04 (for lysozyme from self-seeded mode) and 7BYP/7D02/7D05 (for lysozyme from SASE mode).
Declarations Figure 2 Suppression of microbunching instability by a laser heater (LH) centered at 9.7 keV. a, Spectra of self-seeded XFEL as a function of the induced-energy spread by LH. b, Fraction of enclosed within the bandwidth for four different induced-energy spreads by LH. c, Peak intensity of self-seeded FEL (solid red line), total sum of spectrum in a (dotted blue line), and fraction of FEL intensity enclosed within ±0.5 eV (magenta dotted line) as a function of energy spread induced by LH. Green dotted line: coherent synchrotron radiation (CSR) due to the MBI, as measured at the third bunch compressor using a CCD visible-light camera. The optimized LH condition for self-seeding and SASE are different, where the induced energy spread is about 5 keV is higher for self-seeding than than for SASE.

Figure 3
Spectral intensity of self-seeded vs. SASE XFEL. a, Color maps of 1,000 SASE and self-seeded FEL spectra measured by a single-shot spectrometer. b, SASE and self-seeded FEL spectra averaged over 1,000 shots, with the peak value of the SASE spectrum set to 1, and Ec = 9.7 keV. c, Crystal angle-scanning spectrum measurement with Si-333 at-crystal scanning spectrometer. Each data point is an average of 150 shots. Vertical axis represents the photo-diode current normalized by a quadratic beam position monitor (QBPM) for the FEL intensity measurement. The step of the crystal angle scan is 0.0001°, which corresponds to 0.022 eV. d, Histogram of radiation intensity (for 1-eV bandwidth around peak) expressed relative to average SASE intensity (1 a.u.) for SASE and self-seeded (SS) modes. Histogram data represent the 1000 single-shot spectra measurements in Fig. 3a. The 33-3 Bragg re ection is used in the diamond FBD monochromator. The ebunch delay in the magnetic chicane is 30 fs.

Figure 4
Data quality indicators as a function of resolution. a, signal-to-noise ratio (SNR or I/σ), b, multiplicity, c, Rsplit, and d, correlation coe cient (CC*) derived from three HXRSS and three SASE data sets. The sets SS1/SASE1, SS2/SASE2, and SS3/SASE3 are calculated from 70,656, 27,926, and 12,377 total hit images, respectively. The resolution scale (x-axis) of each gure ranges from 1.75 Å to 3.0 Å to show differences between the self-seeded and SASE modes clearly (speci c values, Supplementary Table 3). CC* represents a direct comparison of crystallographic model quality and data quality on the same scale, especially for multiplying measured data.