Science January 2023

RESEARCH | RESEARCH ARTICLES CSF from the subarachnoid space into the was detected from the exposed cortex, includ- ent in the vascular compartment (fig. S7, A venous sinus system (7–10). The arachnoid ing from the sinus venous wall (fig. S7D). In to C), we propose that the apposition of the villi and granulations are present in the brains another set of control experiments, GeNL was venous endothelia and SLYM represents ro- of humans, primates, and larger animals such injected into the soft ear tissue, while FFz dent arachnoid villus–like structures, com- as dogs, but not in the brains of rodents was delivered intravenously. Consistent with parable to those in human brain. (28, 29). We critically reexamined this issue to the notion that peripheral blood vessels are evaluate the distribution of SLYM in relation leaky (11), light emission was clearly observed The mesothelium surrounding peripheral to the superior sagittal and transverse sinus. in the region of the ear injected with Nano- organs acts as an immune barrier (26). Does Sections obtained from decalcified heads of Luc but not in surrounding noninjected re- SLYM also impede the entry of exogenous Prox1-EGFP+ mice showed that Prox1-EGFP+ gions of the same ear. No signal was observed particles into CSF? In vivo two-photon imag- SLYM cells often were in direct contact with in the venous compartment, likely reflecting ing of Prox1-EGFP+ mice injected intraven- the venous sinus endothelial cells (Fig. 4A). that blood flow rapidly diluted the biolumines- ously with rhodamine 6G (Rhod6G) to label Thus, the arachnoid barrier cell layer (CLDN- cence signal (fig. S7E). Together, this analysis leukocytes (35) showed that a large number 11+/E-Cad+), which normally separates dura shows that a small molecule, FFz, can enter of Rhod6G+ myeloid cells are embedded in from the subarachnoid space, was lacking the central nervous system (CNS) from the SLYM (Fig. 5A). The number of Rhod6G+ in discrete areas allowing SLYM to directly blood and activate an enzyme, NanoLuc, pres- leukocytes in dura and SLYM was directly contact the venous sinus wall (Fig. 4B). Prox1- ent in CSF, resulting in the generation of comparable, suggesting a prominent role of EGFP+ SLYM cells were not positive for CLDN- photons along the wall of the venous sinus. SLYM in CNS immune responses, which sup- 11 or E-Cad, which distinguish the arachnoid On the basis of the juxtaposition of SLYM ports the finding that leptomeninges are den- barrier cell layer (fig. S6). and the venous endothelium in histological sely populated with immune cells (36) (Fig. examination (Fig. 4A), the selective generation 5A). How do systemic inflammation and aging Are the close appositions of SLYM and the of photons when luciferase was injected into affect the immune cell populations residing in venous endothelial cells permeable, allowing CSF, and the fact that the substrate was pres- SLYM? Ex vivo analysis of brain sections the exchange of small molecules between obtained from Prox1-EGFP+ mice showed that, blood and CSF? To test this, we used the prin- ciples of bioluminescence, wherein the con- Fig. 5. SLYM hosts a large number of myeloid cells. (A) (Left) In vivo two-photon microscopy of Prox1- vergence, in the same compartment, of an EGFP+ mice injected with Rhod6G (red) shows that SLYM (EGFP+, green) is permeated by myeloid cells enzyme with its substrate is needed to trigger similar to dura (collagen fibers, gray). Middle panels show orthogonal sections depicting Rhod6G + cells in light emission. First, we delivered the lucifer- dura and SLYM, respectively. (Right) In vivo quantification of the number of Rhod6G+ cells present in dura ase enzyme from Oplophorus gracilirostris (NanoLuc) fused to the fluorescence tag and SLYM. Values are expressed as mean ± SEM, two-tailed unpaired t test with Welch’s correction, P = mNeongreen (GeNL, 44 kDa) (30) into CSF 0.5748, n = 7 mice. (B) Representative image showing the accumulation of CD45+ cells along the pial via the cisterna magna of wild-type (C57bl/6) vessels. (C) The percentage of area covered by CD45+ cells was significantly increased both in aged (12- to mice, and allowed it to circulate for 30 min to ensure thorough distribution by the glym- 13-month-old) mice and in response to inflammation (LPS 4 mg/kg, ip, 24 hours). Values are expressed as phatic system. The distribution of GeNL was verified by mNeongreen fluorescence. Then, mean ± SEM, two-tailed unpaired t test with Welch’s correction, P < 0.005, n = 3 mice. (D) LYVE1 macrophages the blood-brain barrier (BBB)–impermeable substrate fluorofurimazine (FFz, 433 Da) (31) were also found in the SLYM layer, being more prominent in aged and LPS-treated animals than in healthy was administered intravenously (fig. S7, A to C) (32). After intravenous injection of FFz, young Prox1-EGFP mice. (E) The mannose receptor CD206 was detected at similar levels in the young, a bright bioluminescence signal catalyzed by GeNL was detected specifically near the large aged, and LPS-treated groups, suggesting that SLYM may act as a niche for border-associated mouse venous sinus wall (fig. S7, A and B). The bio- luminescence signal was particularly strong macrophages. Significance shown as *P < 0.05. Ctr, control. around the confluence of sinuses (fig. S7B). The distribution of the bioluminescence sig- nal was quantified by plotting the mean sig- nal intensity profiles perpendicular to the venous wall of the transverse sinus and supe- rior sagittal sinus. The mean bioluminescence signal profiles intersected with the fluores- cence signal profiles of the intravascular tracer (TMR-dextran, 70 kDa) or with shadow im- aging of the inverted GeNL signal outlining the vascular wall (fig. S7C). Thus, the biolu- minescence signal was restricted to the venous wall of the two major sinuses lacking a BBB (33, 34), consistent with the notion that FFz is BBB-impermeable and requires the cata- lyzation enzyme NanoLuc to generate photons (fig. S7, A to C). In control experiments, FFz was delivered intravenously, while the GeNL injection into CSF was omitted. In these con- trol experiments, no bioluminescence signal SCIENCE science.org 6 JANUARY 2023 • VOL 379 ISSUE 6627 87

RESEARCH | RESEARCH ARTICLES in the control group, CD45+ cells were abun- association with the brain surfaces, is likely to 30. K. Suzuki et al., Nat. Commun. 7, 13718 (2016). dant, located mostly along pial vessels in the play a prominent role in this surveillance. 31. Y. Su et al., Nat. Methods 17, 852–860 (2020). surface of the brain (Fig. 5B). This observa- Herein, we showed a large increase in the num- 32. M. P. Hall et al., ACS Chem. Biol. 7, 1848–1857 tion, together with the significant increase ber and diversity of immune cells residing in in CD45+ in inflammation-prone conditions SLYM in response to acute inflammation and (2012). [aging and lipopolysaccharide (LPS)–treated natural aging. Physical rupture of SLYM could, 33. J. Rustenhoven et al., Cell 184, 1000–1016.e27 mice, 4 mg/kg of body weight, intraperito- by altering CSF flow patterns, explain the pro- neally (ip), 24 hours] (Fig. 5C), suggests that longed suppression of glymphatic flow after (2021). SLYM can act as a CD45+ recruiting and/or traumatic brain injury as well as the height- 34. P. Mastorakos, D. McGavern, Sci. Immunol. 4, eaav0492 proliferating site in pathological conditions. ened posttraumatic risk of developing Alz- Of note, the dose of LPS used (4 mg/kg) did heimer’s disease (41, 42). Rupture of SLYM (2019). not affect the BBB (fig. S8). Additional im- will also permit the direct passage of immune 35. H. Baatz, M. Steinbauer, A. G. Harris, F. Krombach, Int. J. mune markers showed that LYVE1+ (Fig. 5D), cells from the skull bone marrow (33, 43) into CD206+ (Fig. 5E), and CD68+ (fig. S9) macro- the inner subarachnoid space, with direct ac- Microcirc. Clin. Exp. 15, 85–91 (1995). phages can be found in SLYM, together with cess to the brain surfaces, possibly explaining 36. A. Merlini et al., Nat. Neurosci. 25, 887–899 dendritic cells (CD11c+) (fig. S9). Despite the the prolonged neuroinflammation after trau- absence of CD3+ and CD19+ lymphocytes (fig. matic brain injury (44). SLYM may also be (2022). S9), our results indicate that SLYM functions directly involved in CNS immunity, in addi- 37. A. Louveau et al., J. Clin. Invest. 127, 3210–3219 as a niche for immunological surveillance. Thus, tion to being host to many immune cells. in young, healthy mice, SLYM hosts CD45+ Lymphatic-like tissues can transform quickly (2017). cells, but the number and diversity of innate in the setting of inflammation, which in the 38. S. Da Mesquita, Z. Fu, J. Kipnis, Neuron 100, 375–388 immune cells rapidly expands in LPS-induced brain may be of notable relevance for dis- inflammation and was also significantly al- eases such as multiple sclerosis (45). (2018). tered in aged mice. We conclude that SLYM 39. Z. Xu et al., Mol. Neurodegener. 10, 58 (2015). fulfills the characteristics of a mesothelium REFERENCES AND NOTES 40. L. Wang et al., Brain Pathol. 29, 176–192 (2019). by acting as an immune barrier that prevents 41. A. Z. Mohamed, P. Cumming, J. Götz, exchange of small solutes between the outer 1. J. J. Iliff et al., Sci. Transl. Med. 4, 147ra111 (2012). and inner subarachnoid space compartments 2. H. Mestre et al., Nat. Commun. 9, 4878 (2018). F. Nasrallah; Department of Defense Alzheimer’s Disease and by covering blood vessels in the sub- 3. N. E. Fultz et al., Science 366, 628–631 (2019). Neuroimaging Initiative, Eur. J. Nucl. Med. Mol. Imaging 46, arachnoid space. 4. H. Mestre et al., eLife 7, e40070 (2018). 1139–1151 (2019). 5. A. Louveau et al., Nature 523, 337–341 (2015). 42. J. D. Flatt, P. Gilsanz, C. P. Quesenberry Jr., K. B. Albers, Discussion 6. A. Aspelund et al., J. Exp. Med. 212, 991–999 R. A. Whitmer, Alzheimers Dement. 14, 28–34 (2018). 43. S. Brioschi et al., Science 373, eabf9277 (2021). The critical roles of the meningeal membranes (2015). 44. J. J. Iliff et al., J. Neurosci. 34, 16180–16193 lining the brain have only recently been 7. L. H. Weed, J. Anat. 72, 181–215 (1938). (2014). acknowledged (5, 37). It is now known that 8. A. Key, G. Retzius, Studien in der Anatomie des 45. N. B. Pikor, A. Prat, A. Bar-Or, J. L. Gommerman, CSF is drained by a network of lymphatic ves- Front. Immunol. 6, 657 (2016). sels in the meninges and that suppression of Nervensystems und des Bindegewebes (Samson & Wallin, this drainage accelerates protein aggregation 1876). ACKNOWLEDGMENTS and cognitive decline in animal models of 9. W. E. le Gros Clark, J. Anat. 55, 40–48 (1920). neurodegeneration (38–40). SLYM is Prox1+/ 10. K. Ohta et al., Kurume Med. J. 49, 177–183 We acknowledge P. S. Froh and H. Nguyen (Department of PDPN+/LYVE1−/CRABP2+/VEGFR3−/CLDN-11−/ (2002). Cellular and Molecular Medicine, Faculty of Health and Medical E-Cad− and thereby distinct from the tradi- 11. M. K. Rasmussen, H. Mestre, M. Nedergaard, Sciences, University of Copenhagen, Denmark) for their excellent tional meningeal membranes, including dura, Physiol. Rev. 102, 1025–1151 (2022). technical assistance for the histology and immunohistochemistry arachnoid, and pia, as well as the meningeal 12. A. Drieu et al., Nature 611, 585–593 (2022). of the decalcified samples. We also thank D. Xue for expert lymphatic vessels and the arachnoid trabecula. 13. I. Choi et al., Blood 117, 362–365 (2011). graphical support, B. Sigurdsson for analysis, and H. Hirase, SLYM subdivides the subarachnoid space 14. J. T. Wigle et al., EMBO J. 21, 1505–1513 N. Cankar, and N. C. Petersen for critical reading of the manuscript. into two compartments, suggesting that CSF (2002). Funding: Funding was provided by Lundbeck Foundation grant transport is more organized than currently 15. K. Masamoto et al., Neuroscience 212, 190–200 R386-2021-165 (M.N.), Novo Nordisk Foundation grant acknowledged. For example, SLYM covering (2012). NNF20OC0066419 (M.N.), the Vera & Carl Johan Michaelsen’s the vasculature in the inner subarachnoid 16. A. Nimmerjahn, F. Kirchhoff, J. N. Kerr, F. Helmchen, Legat Foundation (K.M.), National Institutes of Health grant space will guide CSF influx along the penetrat- Nat. Methods 1, 31–37 (2004). R01AT011439 (M.N.), National Institutes of Health grant ing arterioles into the brain parenchyma with- 17. K. Kothur, L. Wienholt, F. Brilot, R. C. Dale, Cytokine 77, U19NS128613 (M.N.), US Army Research Office grant MURI out circulating solutes present in the outer 227–237 (2016). W911NF1910280 (M.N.), Human Frontier Science Program grant subarachnoid space compartment. Yet the 18. H. Mestre, Y. Mori, M. Nedergaard, Trends Neurosci. 43, RGP0036 (M.N.), the Dr. Miriam and Sheldon G. Adelson Medical discovery of a fourth meningeal layer, SLYM, 458–466 (2020). Research Foundation (M.N.), and Simons Foundation grant 811237 has several implications beyond fluid trans- 19. M. Tomooka, C. Kaji, H. Kojima, Y. Sawa, Acta Histochem. (M.N.). The views and conclusions contained in this article are port. The observation that SLYM is a barrier Cytochem. 46, 171–177 (2013). solely those of the authors and should not be interpreted as for CSF solutes that have a molecular weight 20. S. Banerji et al., J. Cell Biol. 144, 789–801 (1999). representing the official policies, either expressed or implied, of larger than 3 kDa will require more detailed 21. M. A. Asson-Batres, O. Ahmad, W. B. Smith, Cell Tissue Res. the National Institutes of Health, the Army Research Office, or the US studies but indicates a need to redefine the 312, 9–19 (2003). Government. The US Government is authorized to reproduce and concept of CNS barriers to include SLYM. The 22. C. B. Brøchner, C. B. Holst, K. Møllgård, Front. Neurosci. 9, 75 distribute reprints for Government purposes notwithstanding any meningeal membranes are hosts to myeloid (2015). copyright notation herein. The funding agencies have taken no part on cells responsible for immune surveillance of 23. J. DeSisto et al., Dev. Cell 54, 43–59.e4 (2020). the design of the study, data collection, analysis, interpretation, or in the CNS (5, 37), and SLYM, owing to its close 24. J. Derk, H. E. Jones, C. Como, B. Pawlikowski, writing of the manuscript. Author contributions: K.M. and M.N. J. A. Siegenthaler, Front. Cell. Neurosci. 15, 703944 designed the study. F.R.M.B., P.K., L.M.M., C.D., V.P., M.K.R., R.S.G., (2021). N.L.H., T.E., and Y.M. performed the experiments, collected the data, 25. M. M. Mortazavi et al., World Neurosurg. 111, 279–290 and performed the analysis. K.M. and M.N. wrote the manuscript. (2018). All authors read and approved the final version of the manuscript. 26. S. E. Mutsaers, F. J. Pixley, C. M. Prêle, G. F. Hoyne, Curr. Opin. Competing interests: The authors declare that they have no Immunol. 64, 88–109 (2020). competing interests. Data and materials availability: All data are 27. B. A. Hills, J. R. Burke, K. Thomas, Perit. Dial. Int. 18, 157–165 available in the main text or the supplementary materials. (1998). License information: Copyright © 2023 the authors, some rights 28. A. Jayatilaka, Ceylon J. Med. Sci. 18, 25–30 (1969). reserved; exclusive licensee American Association for the 29. D. G. Potts, V. Deonarine, J. Neurosurg. 38, 722–728 Advancement of Science. No claim to original US government (1973). works. https://www.science.org/about/science-licenses-journal- article-reuse SUPPLEMENTARY MATERIALS science.org/doi/10.1126/science.adc8810 Materials and Methods Figs. S1 to S9 Table S1 References (46–54) MDAR Reproducibility Checklist Movie S1 Submitted 6 May 2022; resubmitted 13 September 2022 Accepted 7 December 2022 10.1126/science.adc8810 88 6 JANUARY 2023 • VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES M E TA L L U R GY roots, beyond the apparent cause of process- ing parameter drifting. During a laser scan Machine learningÐaided real-time detection of with an unstable keyhole condition, the exact keyhole pore generation in laser powder bed fusion locations where gas bubbles form are random, and which bubbles will become pore defects Zhongshu Ren1, Lin Gao1, Samuel J. Clark2, Kamel Fezzaa2, Pavel Shevchenko2, Ann Choi3,4, eventually and which will be recaptured by the Wes Everhart3, Anthony D. Rollett4, Lianyi Chen5, Tao Sun1* keyhole and vanish is not deterministic. Porosity defects are currently a major factor that hinders the widespread adoption of laser-based We discovered two modes of keyhole os- metal additive manufacturing technologies. One common porosity occurs when an unstable vapor cillation in Ti-6Al-4V under unstable keyhole depression zone (keyhole) forms because of excess laser energy input. With simultaneous high-speed conditions using simultaneous high-speed synchrotron x-ray imaging and thermal imaging, coupled with multiphysics simulations, we discovered synchrotron x-ray and thermal imaging. We two types of keyhole oscillation in laser powder bed fusion of Ti-6Al-4V. Amplifying this understanding then developed a methodology for detecting with machine learning, we developed an approach for detecting the stochastic keyhole porosity keyhole pore generation by integrating exper- generation events with submillisecond temporal resolution and near-perfect prediction rate. The highly imental data, multiphysics simulation, and ma- accurate data labeling enabled by operando x-ray imaging allowed us to demonstrate a facile and chine learning. We used the thermal signals practical way to adopt our approach in commercial systems. emitted from the keyhole region for predict- ing the pore-generation events. X-ray images A fter more than three decades of intense absorption by the metal and that benefits the of the LPBF process provide a data-rich ground research and development, laser powder manufacturing process by improving the en- truth for calibrating and validating the theo- bed fusion (LPBF) has advanced from ergy efficiency and increasing the build rate, retical model and for training the machine- a convenient rapid prototyping tool for the nonuniform laser absorption on the keyhole learning algorithm. We achieved near-perfect shortening the design cycle toward a wall generates local hotspots and causes im- detection accuracy with submillisecond tem- manufacturing technology for producing end- balance between the recoil pressure, vapor dy- poral resolution for both powder bed and bare use metallic components (1–3). Although some namic pressure, capillary force, and Marangoni substrate samples. industries fully embrace LPBF now, others are force. Under unstable keyhole conditions, gas more cautious about quality control as they bubbles pinch off the keyhole tip, and some Operando experiment and data analytics integrate it into their product lines. As a pri- eventually become pore defects when they mary metal additive manufacturing (AM) tech- are trapped by the advancing solidification We conducted our operando experiment at nique, LPBF is capable of fabricating parts with front (4, 5, 12). For LPBF of a given material, the the 32-ID-B beamline of the Advanced Photon complex geometries and fine features. How- unstable keyhole zone in the power-velocity Source of Argonne National Laboratory (Fig. 1A). ever, some technical barriers still need to be (P-V) process map can be defined with small High-energy x-rays passed through the single- overcome before LPBF can reach its full poten- uncertainty (4, 5, 12). Setting the initial laser layer powder bed or bare substrate samples tial as a disruptive manufacturing technique. parameters outside the unstable keyhole zone to reveal the subsurface structure dynamics, In a typical LPBF process, a high-power laser helps mitigate the generation of keyhole po- and we set up a thermal camera to collect an beam is used to locally melt and consolidate rosity. However, multiple factors involved in angled top view of the melt pool. A continuous- metal powder to form three-dimensional (3D) LPBF can still offset the laser melting mode wave–mode fiber laser with a Gaussian pro- objects layer by layer. The extreme thermal and create keyhole-porosity–prone conditions, file and wavelength of 1070 nm scanned the conditions involved in the printing process such as the drift of laser spot size, power, and sample along a single straight line at various trigger transient phenomena and complex scan speed, as well as the scan strategies that powers and speeds. We collected full-field x-ray structural dynamics. Their interplay often result in local overheating. Therefore, keyhole images (Fig. 1C) from single-track melting events leads to structural defects, such as porosity. porosity may still exist in a part even if the ini- with a spatial resolution of 2 to 3 mm/pixel, tem- One common porosity is caused by the momen- tial machine setting is optimized for printing poral resolution of 0.1 ns to 7.5 ms, and frame tary collapse of the vapor depression zone, certain material. rates of 50 kHz to 1.08 MHz. Simultaneously, known as keyhole porosity (4–8). we collected thermal images (Fig. 1B) of the Real-time detection of keyhole pore gener- melt track in the spectra range of visible light Under the condition of excess laser energy ation in LPBF is critical not only for facilitating to infrared, view angle of 38° to 58°, spatial res- input (high power and slow scan velocity), post-build part interrogation and qualifica- olution of 5 to 30 mm/pixel, temporal resolu- metal vaporization exerts a recoil pressure that tion but also for developing closed-loop con- tion of 0.3 to 5 ms, and frame rates of 50 to pushes down the melt pool surface, forming a trol systems that can anticipate the need for 200 kHz [(19), sections 1.2 to 1.4]. narrow and deep keyhole in which multiple local variations during the build process. Op- laser reflection and absorption events occur tical and acoustic sensors are commonly used as By setting a threshold, we extracted the (9–11). Although this increases the overall laser process monitors, and data analysis approaches average light emission intensity from the key- have been developed to correlate the process hole region, which condensed the 2D optical 1Department of Materials Science and Engineering, University signatures with porosity (13–15). Although some images into 1D plots (Fig. 1D). Such data com- of Virginia, Charlottesville, VA 22904, USA. 2X-ray Science successes were achieved by the community, pression can alleviate the burden of computer Division, Advanced Photon Source, Argonne National particularly with application of machine learn- storage substantially in real-time monitor- Laboratory, Lemont, IL 60439, USA. 3Kansas City National ing (16–18), for differentiating pore-prone con- ing because the high-resolution images can be Security Campus Managed by Honeywell Federal ditions from normal conditions, detection of immediately processed with on-camera-chip. Manufacturing and Technologies, US Department of Energy, the generation of a keyhole pore locally and Meanwhile, 1D time-series signals facilitate Kansas City, MO 64147, USA. 4Department of Materials instantaneously remains challenging. This identification of frequency-related features Science and Engineering, Carnegie Mellon University, problem is because the stochastic nature of associated with keyhole dynamics. We per- Pittsburgh, PA 15213, USA. 5Department of Mechanical keyhole pore generation has other physics formed wavelet analyses over the segmented Engineering, University of WisconsinÐMadison, Madison, WI 1D datasets to create scalograms (Fig. 1E), 53706, USA. which reveal the characteristic oscillations *Corresponding author. Email: [email protected] localized in the time domain. We fed short- window scalograms, labeled by operando SCIENCE science.org 6 JANUARY 2023 • VOL 379 ISSUE 6627 89

RESEARCH | RESEARCH ARTICLES A we define the two modes as intrinsic and per- turbative oscillations. The intrinsic oscillation BD occurs for both stable and unstable keyholes CE (Fig. 2, A and C to F; figs. S6 to S12; and movies F S1 to S9) and is a consequence of varying bal- ance between Marangoni force, surface tension, Fig. 1. Real-time keyhole porosity detection in LPBF. (A) Schematic of the simultaneous synchrotron and recoil pressure. In one oscillation cycle (Fig. x-ray and thermal imaging experiment on scanning laser melting of Ti-6Al-4V. (B) A representative angle 2, C to F; figs. S6 to S12; and movies S1 to S9), top-view thermal image. (C) A representative side-view x-ray image. (D) Typical time-series signal of when the keyhole has the smallest opening at the average emission intensity from the keyhole region [(B), dashed oval] extracted from the thermal image the rim, the rear keyhole wall is directly ex- sequence. (E) Wavelet analysis performed over the time-series signals in (D). The scalogram is sectioned posed to the incident laser beam and heated to into a few windows, which are then labeled as either “Non-pore” or “Pore” on the basis of the operando x-ray a high temperature (Fig. 2C, 15 ms, and figs. S8 imaging result. (F) Machine-learning approach with sectioned scalograms as input data. A CNN was used, to S12, 0 ms). The Marangoni force then con- which is composed of a series of alternating convolution and pooling layers and a final layer. Each convolution vects the hot liquid (the hump structure near layer extracts features from its previous layer, using filters learned from the trained model, to form a the rear keyhole rim shown in Fig. 2D, 40 ms, feature map. The feature map is then down-sampled by a pooling layer to reduce the number of parameters to and figs. S9, 5 ms; S10, 20 and 60 ms; and S12, learn. The final layer of the CNN classifies the input scalogram as either “Non-pore” or “Pore.” 22 ms) back toward the rear of melt pool, where the temperature is lower. The liquid flow at x-ray imaging results as cases of “Pore” and Intrinsic keyhole oscillation the melt pool surface expands the opening “Non-pore,” into a convolutional neural net- Our approach for detecting keyhole pore gen- of the keyhole and cools it. When most of the work (CNN) for predicting the keyhole pore- eration by using thermal imaging was developed hot liquid flows backward [the hump struc- generation events (Fig. 1F and fig. S3) [(19), on the basis of our observation and under- ture is now at the rear location (Fig. 2D, 60 ms, section 1.6]. standing of keyhole oscillations. In this work, and fig. S12, 54 ms)], the keyhole temperature drops to its minimum (Fig. 2C, 90 ms, and fig. S8, 14 ms), and the opening reaches its max- imum. After this point, the Marangoni force becomes negligible owing to the much reduced temperature gradient around the keyhole. In- stead, the surface tension starts to dominate the keyhole dynamics and drives the liquid in the rear keyhole wall to flow forward and close up the keyhole. As the keyhole opening shrinks, more liquid is exposed to the incident laser beam, and the thermal gradient in the keyhole region increases. The Marangoni force starts to drive the liquid flow again, initiating a new cycle of intrinsic oscillation [(19), section 2.4]. The type of keyhole oscillation is consistent with those observed in the previous studies (13, 20–23). We used the thermal information of the keyhole collected by our near-infrared (NIR) camera as an indicator of the intrinsic oscilla- tion (representative intrinsic oscillation is shown in in Fig. 2, E and F, 16 kHz). The mech- anism for the intrinsic oscillation of an unstable keyhole conditions is the same as that of a stable keyhole (figs. S9 to S12 and movies S5 to S9). The intrinsic oscillation is not respon- sible for the keyhole pore generation under unstable keyhole conditions. This is because even though there may be temporary keyhole collapse and bubble generation, the potent forward liquid flow involved in the intrinsic oscillation can push the bubble toward the advancing laser and allow it to be captured by the rapidly drilling keyhole (Fig. 2A, figs. S11 and S12, and movies S8 and S9). Perturbative keyhole oscillation The perturbative oscillation only occurs under unstable keyhole conditions (Fig. 2, B and G to J; figs. S13 to S16; and movies S10 to S14). The “perturbative” term is because the associated 90 6 JANUARY 2023 • VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES A B CE DF GI HJ Fig. 2. Intrinsic and perturbative keyhole oscillations in Ti-6Al-4V. in (E). (G) Thermal images of perturbative keyhole oscillation. (H) High-speed (A) Megahertz x-ray images of intrinsic keyhole oscillation with no keyhole pore x-ray images and corresponding multiphysics simulation of perturbative generated. (B) Megahertz x-ray images of perturbative keyhole oscillation keyhole oscillation. (I) Time-series signal of the average light emission intensity with a keyhole pore generated. (C) Thermal images of intrinsic keyhole around the keyhole, extracted from the thermal image sequence in (G). oscillation. (D) High-speed x-ray images and corresponding multiphysics (J) Scalogram corresponding to the time-series signals in (I). The laser power simulation of intrinsic keyhole oscillation. (E) Time-series signal of the average and speed are 200 W and 400 mm/s, respectively. The color markers in (E) light emission intensity around the keyhole, extracted from the thermal and (I) indicate data points with corresponding frames shown in (C), (D), image sequence in (C). (F) Scalogram corresponding to the time-series signals (G), and (H), which have borders with the same colors. keyhole behavior disrupts the intrinsic oscil- develops, creating complex flow patterns. crease of recoil pressure at the keyhole bottom. lation. An unstable keyhole features highly Toward the keyhole opening, the Marangoni Both of these effects increase the chance of dynamic protrusion structures on the keyhole force drives the hot liquid moving upward wall (4, 22, 24). A protrusion on the front key- and backward, which opens up the keyhole keyhole wall collapsing (Fig. 2B, 342 and hole wall has two effects. First, the protrusion rim (Fig. 2H, simulation at 20, 40, and 60 ms; 344 ms, and figs. S13, 0 and 2 ms, and S16, 2 and shadows the liquid right underneath it from and fig. S15, 0 ms) (5). Second, the protrusion 4 ms). Once the collapsing occurs, the new being heated by the incident laser directly. reflects the incident laser ray toward the mid- keyhole (upper part of the previous keyhole) Thus, a large temperature difference between dle and upper regions of the rear keyhole wall, the liquid above and below the protrusion causing local hotspots and consequently a de- is directly exposed to the laser, and this key- hole’s temperature increases (Fig. 2G, 40 ms, and fig. S15, 22.5 ms). The high recoil pressure SCIENCE science.org 6 JANUARY 2023 • VOL 379 ISSUE 6627 91

RESEARCH | RESEARCH ARTICLES then pushes down the keyhole rapidly (22, 24). After confirming that our machine-learning tially proved by our experimental result (table Meanwhile, the Marangoni force transports model is capable of identifying features with S4), yet the exact vapor plume effect on ther- the hot liquid from the bottom of the newborn predictive power from the input scalograms, mal imaging in different geometries demands keyhole toward the rear rim, facilitating the we investigated other factors that may affect more sophisticated experimental and model- drilling process. During this process, a new the keyhole porosity prediction rate. These in- ing efforts [more discussion is available in (19), protrusion appears on the front keyhole wall, clude the configuration of thermal imaging section 2.7]. and a fresh cycle of perturbative oscillation system (table S2), spatial resolution (table S3), starts (Fig. 2, B, 354 ms, and H, 60 ms; and frame rate (table S3 and fig. S24), view angle For powder bed samples, 100% prediction figs. S13, 6 ms; S15, 5 and 17.5 ms; and S16, 8 ms). (the angle between optical axis and sample rate (including “accuracy,” “precision,” and The frequency of perturbative oscillation is surface) (fig. S25 and table S4), and the win- “recall” results) of keyhole pore generation in generally above 40 kHz (figs. S14 and S15), dow length of sectioned scalograms (figs. S24 Ti-6Al-4V was achieved by performing NIR which is higher than the intrinsic oscilla- and S25). Because wavelet analysis is essentially imaging at 48° view angle, 200 kHz sampling tion frequency. Owing to the chaotic nature a form of short-window Fourier transform, in- rate, 10 mm spatial resolution, and 0.75 ms time of an unstable keyhole and the random oc- creasing the frame rate of raw data and the window (table S5). We demonstrated that the currence of protrusions on front and rear window length of each scalogram can both stochastic keyhole pore occurrence can be truly keyhole walls, the perturbative oscillation increase the information a scalogram reveals detected locally, among laser scans under var- frequency varies dramatically even during (fig. S24) and thereby improve the keyhole ious P-V conditions covering all melting modes, a single laser scan. pore prediction rate. Specifically, because the or even the same P-V condition in unstable frequency of perturbative oscillation is higher keyhole mode, in which the number of pores We discovered that the high-frequency per- than 40 kHz, the thermal imaging rate needs are smaller than the keyhole-collapsing and turbative oscillation is responsible for the key- to be above 80 kHz according to the Nyquist- bubble-forming events. hole porosity generation in Ti-6Al-4V. Once Shannon sampling theorem. Therefore, raw the keyhole collapses, the lower part of the data with 50 kHz rate, regardless of the optical Implementation without additional keyhole is separated from the upper part, ap- configuration, yielded much reduced predic- synchrotron experiments pearing as a gas bubble pinching off the key- tion rate (tables S2 to S4). hole tip. The bubble is then pushed downward A key factor for achieving high prediction rate and backward by the new keyhole (Fig. 2, B, In our approach, the off-axial thermal im- by using the machine-learning approach is the 346 to 354 ms, and H, 40 to 60 ms; and figs. S13, aging is preferred than the commonly used accurate data labeling. The larger the misla- 12 ms; S14, 20 ms; S15, 10 ms; and S16, 14 ms) (4). coaxial setup, because it (i) offers flexibility beling rate is, the worse the performance of the As the bubble moves away from the keyhole, and ease of installation to existing 3D printers, machine-learning algorithm (Fig. 3). In pre- its motion is primarily controlled by the fluid whereas the coaxial setup involves complicated vious efforts (17, 18, 31), porosity in the printed flow in the melt pool. Very few pores can es- optical systems (26–28); (ii) provides more re- samples was typically characterized by using cape by flowing to the melt pool surface or liable process signals than those of the coaxial x-ray computed tomography. Such an ex situ circulating back to the keyhole, whereas the setup, in which the thermal emission travels (post mortem) data-labeling approach iden- majority are trapped by the advancing solidifi- back through optics such as the F-Theta lens tifies the final pore locations rather than the cation front and become pore defects (4). The and dielectric mirror to the imager, influenc- moment the bubbles were initially generated. perturbative keyhole oscillation is the key to ing the signals (29, 30); and (iii) an optimum Because a bubble can move within the melt answering a question that has not been asked angle of ~50° was found in our simulation (fig. pool after it pinches off the keyhole, a sub- previously: when printing by using a single P-V S25) that yielded the highest keyhole pore pre- stantial mislabeling rate results from ex situ condition in unstable keyhole mode, why some diction rate. We speculate that the reason for labeling when the window length is small bubbles can escape the keyhole and become the lower prediction rate from coaxial detection (Fig. 3E). In our operando synchrotron x-ray pores, whereas others are recaptured by the key- is twofold. First, when looking straight down, imaging experiments, we could measure not hole and disappear [(19), sections 2.4 to 2.6]. the camera collects the thermal information of only the final location of a keyhole pore but the keyhole bottom, which remains hot most also the exact moment at which the bubble Machine learning aided keyhole pore detection of the time, so it may be less sensitive to the forms (Fig. 3, A and B). We performed statis- keyhole oscillation [more discussion is avail- tical analysis over the experiment and simula- The distinct frequencies of intrinsic and per- able in (19), section 2.11]. Second, an unstable tion data on powder bed and plate samples turbative oscillations allowed us to detect the keyhole appears at relatively slow laser scan- (Fig. 3C). The line regression shows that ex situ keyhole pore–generation events with high ning speeds. Under these conditions, the vapor labeling could induce an uncertainty of the pore fidelity and high resolution. We divided each plume ejects upward with only small inclina- generation time by ~0.4 ms, which caused con- scalogram of thermal data we obtained from tion angles. The fluctuation of the vapor plume siderable reduction in prediction rate (Fig. 3F). a single line scan into several segments, which reflects the keyhole oscillation to a great ex- correspond to frequency patterns within a tent, but at some distance above the sample With the calibrated simulation, optimized smaller time window. We used the labeled when the concentration and speed of the thermal imaging scheme and machine-learning scalogram segments to train a CNN algorithm vapor plume diminish, its dynamic motion is algorithm, we demonstrate two practical means for classifying the test data into “Pore” and also driven by the ambient gas flow. Such vapor to effectively apply our approach without the “Non-pore” cases. We used a feature visualiza- plume effect will likely weaken when introduc- need of additional synchrotron experiments. tion technique known as gradient-weighted ing shielding gas flow across the powder bed, The first approach is to calibrate the ex situ– class activation mapping (Grad CAM) (25) to and many commercial LPBF machines offer labeled time by adding 0.4 ms to better rep- highlight the important features correspond- this function now. By contrast, at ~50° view resent the moments when keyhole pores form ing to these two categories. The machine- angle, although the effective spatial resolu- (bubble separating from the keyhole tip). This learning algorithm used the low-frequency tion is slightly decreased, the camera sees only time offset appears to be independent of laser feature (intrinsic oscillation) to classify the the upper part of the keyhole wall, with lim- power and scan speed but may vary with the “Non-pore” cases and the high-frequency fea- ited influence of the secondary vapor plume laser spot size and material. The calibration of ture (perturbative oscillation) to classify the fluctuation. The effect of view angle was par- time stamps served to reduce the mislabeling “Pore” cases (fig. S21). rate (Fig. 3E) and improve the prediction ac- curacy substantially (Fig. 3F). Encouraged by 92 6 JANUARY 2023 • VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES AC B D EF Fig. 3. Effect of data mislabeling on prediction rate. (A) Illustration of in machine-learning prediction accuracy as functions of mislabeled rate and situ–labeled time t1 of a keyhole pore, the moment when a bubble pinches off scalogram window length. The variable of mislabeled rate is created by the keyhole. (B) Illustration of ex situ–labeled time t2 of the same pore as randomly assigning a certain percentage of the scalogram the inverse label of marked in (A), an artificial moment corresponding to the final pore location. the in situ labels in the training data. (E) Mislabeled rate as a function of (C) Ex situ–labeled time t2 as a function of in situ–labeled time t1 for scalogram window length for in situ, ex situ, and calibrated ex situ labeling. experimental and simulation data on powder bed and plate samples. A linear (F) The corresponding machine-learning prediction accuracy as a function of regression is indicated with the black dashed line. The error bars are not scalogram window length. The error bars in (E) and (F) are the standard displayed in this plot because the uncertainties in both t1 and t2 (~0.02 ms) are deviation of 10 repeated training, representing the randomness in partitioning much smaller than the time scale under consideration. (D) Contour map of the training data into training and validation sets. this success, we performed a proof-of-concept ticated LPBF models, in which the physics a LPBF system in two ways as just described. experiment using a commercial LPBF system governing the powder motion and vapor plume The off-axis detection scheme and the need (SLM Solutions 125) (fig. S26). The size of the dynamics should be considered. for only 1D data series could be particularly chamber and the view window location limited convenient for existing LPBF systems without the setting of thermal imaging. Even though Concluding remarks preinstalled co-axial imaging optics or high- we could not apply the optimum imaging con- performance computing hardware. The strat- dition, the prediction accuracy, precision, and The research outcome we describe highlights egy of detecting build anomalies by examining recall all improved after adjusting the ex situ the enabling character of operando synchro- the oscillation behaviors of the keyhole and labels (table S6). The second approach was to tron x-ray imaging experiments. It not only melt pool is generic and practical. We believe train the machine-learning model by use of provided key information to calibrate the mul- that the process monitoring systems built on well-calibrated simulations. When tested on tiphysics model but also allowed us to discover this core concept will promote the qualification the experiment data, prediction accuracies of the distinctive keyhole oscillation behavior and certification of metal AM parts. 85% (0.75 ms window) and 87% (1.5 ms win- associated with keyhole porosity. The quan- dow) were obtained for the plate and powder titative understanding built the foundation REFERENCES AND NOTES bed samples, respectively (table S7). It was for our machine-learning model that detects notably more difficult to predict the pore keyhole pore formation from thermal imaging. 1. T. DebRoy, T. Mukherjee, H. L. Wei, J. W. Elmer, J. O. Milewski, generation in powder bed samples by using With the ground truth acquired with operando Nat. Rev. Mater. 6, 48–68 (2021). such an approach. This is different than the x-ray imaging, our approach is capable of de- case of using experimental data for training, in tecting the generation of keyhole pores in 2. T. Sun, W. Tan, L. Chen, A. Rollett, MRS Bull. 45, 927–933 (2020). which the prediction rate is generally higher powder bed samples with 100% accuracy and 3. T. DebRoy et al., Nat. Mater. 18, 1026–1032 (2019). for powder bed samples. The discrepancy high- submillisecond temporal resolution. For those 4. C. Zhao et al., Science 370, 1080–1086 (2020). lights the need for developing more sophis- who have limited or no access to synchrotron, 5. M. Bayat et al., Addit. Manuf. 30, 100835 (2019). this approach can be readily implemented to 6. A. A. Martin et al., Nat. Commun. 10, 1987 (2019). 7. Y. Huang et al., Nat. Commun. 13, 1170 (2022). 8. Z. Gan et al., Nat. Commun. 12, 2379 (2021). 9. S. A. Khairallah et al., Science 368, 660–665 (2020). SCIENCE science.org 6 JANUARY 2023 ¥ VOL 379 ISSUE 6627 93

RESEARCH | RESEARCH ARTICLES 10. R. Cunningham et al., Science 363, 849–852 (2019). PLANT SCIENCE 11. W. E. King et al., J. Mater. Process. Technol. 214, 2915–2925 (2014). 12. S. A. Khairallah, A. T. Anderson, A. Rubenchik, W. E. King, Maize resistance to witchweed through changes in strigolactone biosynthesis Acta Mater. 108, 36–45 (2016). 13. S. A. Khairallah, T. Sun, B. J. Simonds, Additive Manufacturing C. Li1, L. Dong1*, J. Durairaj2 , J.-C. Guan3, M. Yoshimura4,5,6, P. Quinodoz5, R. Horber5, K. Gaus5, J. Li7, Y. B. Setotaw7, J. Qi7, H. De Groote8, Y. Wang1, B. Thiombiano1, K. Floková1,9, A. Walmsley1, Letters 1, 100002 (2021). T. V. Charnikhova1, A. Chojnacka1, S. Correia de Lemos2,10, Y. Ding11, D. Skibbe12, K. Hermann5, 14. J. R. Tempelman et al., Addit. Manuf. 55, 102735 (2022). C. Screpanti5, A. De Mesmaeker5, E. A. Schmelz11, A. Menkir13, M. Medema2, A. D. J. Van Dijk2, J. Wu7, 15. R. Drissi-Daoudi et al., Virtual Phys. Prototyp. 17, 181–204 (2022). K. E. Koch3, H. J. Bouwmeester1* 16. S. Shevchik et al., Sci. Rep. 10, 3389 (2020). 17. Z. Smoqi et al., J. Mater. Process. Technol. 304, 117550 (2022). Maize (Zea mays) is a major staple crop in Africa, where its yield and the livelihood of millions are compromised 18. S. A. Shevchik, C. Kenel, C. Leinenbach, K. Wasmer, Addit. by the parasitic witchweed Striga. Germination of Striga is induced by strigolactones exuded from maize roots into the rhizosphere. In a maize germplasm collection, we identified two strigolactones, zealactol and Manuf. 21, 598–604 (2018). zealactonoic acid, which stimulate less Striga germination than the major maize strigolactone, zealactone. 19. Materials and methods are available as supplementary materials. We then showed that a single cytochrome P450, ZmCYP706C37, catalyzes a series of oxidative steps in the 20. L. Caprio, A. G. Demir, B. Previtali, Addit. Manuf. 36, 101470 (2020). maize-strigolactone biosynthetic pathway. Reduction in activity of this enzyme and two others involved in 21. B. J. Simonds et al., Phys. Rev. Appl. 10, 044061 (2018). the pathway, ZmMAX1b and ZmCLAMT1, can change strigolactone composition and reduce Striga germination 22. C. Zhao et al., Phys. Rev. X 9, 021052 (2019). and infection. These results offer prospects for breeding Striga-resistant maize. 23. V. V. Semak, J. A. Hopkins, M. H. McCay, T. D. McCay, J. Phys. F ood security is a growing challenge in the cular mycorrhizal fungi (AMF) and are plant D Appl. Phys. 28, 2443–2450 (1995). face of climate change and increasing hormones with developmental roles (5–9). 24. N. Kouraytem et al., Phys. Rev. Appl. 11, 064054 (2019). 25. R. R. Selvaraju et al., Grad-CAM: Visual explanations from deep food needs (1). Maize (Zea mays) is one of Thus far, more than 35 different SLs have been discovered, all containing the conserved networks via gradient-based localization, 2017 IEEE the most important staple crops in the D-ring (Fig. 1A) (10–12). The canonical SLs International Conference on Computer Vision (ICCV) (2017), include two groups, the “strigol-type” and pp. 618–626. world, especially in Africa. There, its yield “orobanchol-type,” whereas noncanonical 26. T. G. Spears, S. A. Gold, Integr. Mater. Manuf. Innov. 5, 16–40 (2016). SLs lack the A-, B-, and/or C-rings (10–12). 27. S. K. H. Everton, M. Hirsch, P. Stravroulakis, R. K. Leach, is compromised by the parasitic witchweeds Plants usually exude a blend of different SLs, A. T. Clare, Mater. Des. 95, 431–445 (2016). and the composition of the root exudate can 28. M. Grasso, B. M. Colosimo, Meas. Sci. Technol. 28, 044005 (2017). Striga hermonthica and Striga asiatica. Dam- vary greatly between and sometimes also 29. B. Brandau, T. Mai, F. Brueckner, A. F. H. Kaplan, Opt. Lasers within plant species. Many of the SLs display Eng. 155, 107050 (2022). age from these Striga species threatens the substantial differences in their biological ac- 30. M. Schürmann et al., Proc. SPIE 9351, 93510S (2015). tivity, such as the induction of AMF hyphal 31. W. G. Guo, Q. Tian, S. Guo, Y. Guo, CIRP Ann. 69, 205–208 (2020). livelihood of millions of people, particularly branching and parasitic plant germination (9, 13–15). The biological importance of SL ACKNOWLEDGMENTS in sub-Saharan regions (fig. S1) (2, 3). Striga blends is far from understood, but in sorghum (Sorghum bicolor), a change in SLs from We thank A. Deriy and V. Nikitin at the Advanced Photon Source seeds lay dormant in soil until their germina- 5-deoxystrigol to orobanchol decreased Striga for their technical support of the beamline experiments. We germination and increased field resistance (16). also thank B. Simonds at the National Institute of Standards and tion is triggered by strigolactones (SLs), sig- Technology, M. Bayat at the Technical University of Denmark, The mechanisms of SL biosynthesis have A. Mane and I. Muqica at Flow Science, and K. Jones at Carnegie naling compounds exuded by the roots of only been partially elucidated. Three enzymes— Mello University for the fruitful discussions. This research used DWARF 27 (D27) and two carotenoid cleavage resources of the Advanced Photon Source, a US Department plants, including maize. The first known SL, dioxygenases l(CCDs), CCD7 and CCD8— of Energy (DOE) Office of Science user facility operated for the catalyze the conversion of b-carotene to car- DOE Office of Science by Argonne National Laboratory under strigol, was discovered in the 1960s in the root lactone (CL) (Fig. 1A) (17, 18). In Arabidopsis, contract DE-AC02-06CH11357. All data prepared, analyzed, and CL is oxidized to form carlactonoic acid (CLA) presented have been developed in a specific context of work and exudates of cotton (4). In addition to having by a cytochrome P450 (CYP) monooxygenase, were prepared for internal evaluation and use pursuant to that CYP711A1, encoded by More Axillary Growth work authorized under the referenced contract. Reference herein been co-opted as a cue for root-parasitic plants, 1 (MAX1) homolog AtMAX1 (19). Arabidopsis to any specific commercial product, process, or service by trade has a single copy of this MAX1, whereas maize name, trademark, manufacturer, or otherwise, does not necessarily SLs serve as host signals for beneficial arbus- has three homologs, and rice has five (18, 20). constitute or imply its endorsement, recommendation, or favoring Although both the Arabidopsis AtMAX1 and by the US government, any agency thereof, or Honeywell Federal 1Plant Hormone Biology Group, Swammerdam Institute for the maize ZmMAX1b form CLA from CL, the Manufacturing & Technologies. This publication has been authored Life Sciences, University of Amsterdam, Science Park 904, rice MAX1 homologs, Os900 and Os1400, in- by Honeywell Federal Manufacturing & Technologies under 1098 XH Amsterdam, Netherlands. 2Bioinformatics Group, stead convert CL to 4-deoxyorobanchol (4DO) contract DE-NA0002839 with the DOE. A provisional application Wageningen University & Research, 6708 PB Wageningen, and orobanchol, respectively (18, 21). Dicots for a US patent (63427022) has been filed based on this work. Netherlands. 3Horticultural Sciences Department, University of also form orobanchol, but from CLA rather Funding: This work was supported by the DOE Office of Florida, Gainesville, FL 32611, USA. 4Laboratorium für than CL, and with a different cytochrome Science by Argonne National Laboratory under contract DE- Organische Chemie, Department of Chemistry and Applied P450, CYP722C. A homolog of this CYP722C AC02-06CH11357 and Honeywell Federal Manufacturing & Biosciences, ETH Zürich, 8093 Zürich, Switzerland. 5Syngenta can also produce 5-deoxystrigol from CLA Technologies under contract DE-NA0002839 with the DOE. Author Crop Protection AG, Schaffhauserstrasse 101, CH‐4332 Stein, (22, 23). contributions: Conceptualization: T.S. and Z.R. Methodology: Switzerland. 6Kyoto University, iCeMS, Yoshida Ushinomiya-cho, T.S. and Z.R. Investigation: Z.R., T.S., L.G., S.J.C., K.F., P.S., A.D.R., Sakyo-ku, Kyoto 606-8501, Japan. 7Department of Economic L.C., and A.C. Visualization: Z.R., L.G., and T.S. Funding acquisition: Plants and Biotechnology, Yunnan Key Laboratory for Wild Plant W.E., A.C., T.S., and L.C. Project administration: T.S. and A.C. Resources, Kunming Institute of Botany, Chinese Academy of Supervision: T.S. Writing – original draft: Z.R. and T.S. Writing – Sciences, Kunming 650201, China. 8International Maize and review and editing: L.G., S.J.C., K.F., P.S., A.C., W.E., A.D.R., Wheat Improvement Center (CIMMYT), PO Box 1041-00621, and L.C. Competing interests: The authors declare that they have Nairobi, Kenya. 9Laboratory of Growth Regulators, Institute of no competing interests. Data and materials availability: All data Experimental Botany, The Czech Academy of Sciences and are available in the main text or the supplementary materials. License Faculty of Science, Palacký University, Šlechtitelů 27, 783 71 information: Copyright © 2023 the authors, some rights reserved; Olomouc, Czech Republic. 10Plant genomics and transcriptomics exclusive licensee American Association for the Advancement of group, Institute of Biosciences, Sao Paulo State University, Science. No claim to original US government works. https://www. 13506-900 Rio Claro, Brazil. 11Section of Cell and Developmental science.org/about/science-licenses-journal-article-reuse Biology, University of California at San Diego; La Jolla, CA 92093, USA. 12Seeds Research, Syngenta Crop Protection, LLC, SUPPLEMENTARY MATERIALS Research Triangle Park, NC 27709, USA. 13International Institute of Tropical Agriculture, PMB 5320 Oyo Road, Ibadan, Nigeria. science.org/doi/10.1126/science.add4667 *Corresponding author. Email: [email protected] (H.J.B.), [email protected] (L.D.) Materials and Methods †Present address: Biozentrum, University of Basel, Spitalstrasse Supplementary Text 41, 4056 Basel, Switzerland. Figs. S1 to S26 Tables S1 to S7 References (32–52) Movies S1 to S22 Submitted 13 June 2022; accepted 1 December 2022 10.1126/science.add4667 94 6 JANUARY 2023 • VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES A all-trans- -carotene B 1×107 and fig. S3). Moreover, one of the genotypes, 8×106 NP2222, displayed a distinctive SL profile, lack- ZmD27 Peak area (zealactone) ing detectable levels of all but two SLs, an unknown SL and designated compound 5 (Fig. 9-cis- -carotene 6×106 1B and fig. S3). Compound 5 was previously noted in maize root exudate (24), but its low ZmCCD7 3×106 abundance and chemical instability hampered structural characterization. Therefore, on the 9-cis- -apo-10’-carotenal 2×106 basis of nuclear magnetic resonance (NMR) ZmCCD8 spectra and retrosynthetic analysis (24, 27–29), we postulated structures and subsequently syn- 16 17 19 thesized compound 5 as well as the other unknown SL (figs. S4 to S12). The synthetic 2 16 7 9 products were identical to the natural ones 3 A 10 in maize root exudate and were designated 1×106 zealactol (compound 5) and zealactonoic acid 8 (ZA) (the other unknown SL) (figs. S9 and 0 S12). Bioassay of Striga germination showed 4 5 18 O 11 O 1.6×106 that both zealactol and ZA were less inductive DO 1.2×106 than zealactone (Fig. 1C), an outcome that carlactone 12 13 14 highlights how strongly minute differences in (CL) 15 8×105 SL structure can alter their biological ac- 4×105 tivity. These findings are further supported ZmMAX1b 2×105 by work on sorghum (16). To unravel the mechanistic basis for these differences in COOH ZmCYP706C37 Peak area (zealactol) SL blends, we revealed the biosynthetic path- way of maize SLs. OO O OH 1.5×105 Three maize genes encode the carlactone carlactonoic acid O OO 1×105 biosynthetic pathway (CLA) O 5×104 Through homology, we identified the maize ZmCLAMT1 3-oxo-19-hydroxy-CL Peak area (ZA) orthologs D27, CCD7, and CCD8, which catalyze O ZmCYP706C37 0 the formation of CL from b-carotene in other COOCH3 7×105 plant species (tables S1 and S2). To confirm OH 6×105 ZmCCD8 function, we analyzed root exudate OO 5×105 of two independent zmccd8 mutants (in W22 O OO 4×105 and Mo17 backgrounds) (30). Zealactone was methyl carlactonoate O O 1.5×105 not detected, although it was the major SL in (MeCLA) wild-type exudate (fig. S13A), showing that 1×105 ZmCCD8 is a key enzyme in maize SL bio- ZmCYP706C37 synthesis (17, 31, 32). The transient expres- sion of ZmD27 (GRMZM2G158175), ZmCCD7 COOCH3 (GRMZM2G158657), and ZmCCD8 (GRMZM2- G446858) together in Nicotiana benthamiana zealactol led to accumulation of CL (Figs. 1A and 2A, fig. S14A, and table S3), which is consistent HO OO ZmMAX1b 5×104 with results from rice and tomato orthologs O (21, 33). 3-hydroxy-MeCLA COOH Identification of gene candidates for ZmCYP706C37 carlactone conversion 0 On the basis of the structures of the maize COOCH3 O OO P39 SLs identified thus far (Fig. 1A and fig. S2) O O CML69 (24–26), we postulated the involvement of M37W a methyl transferase and several CYPs in zealactonoic acid CML228 the pathway downstream of CL. Several bio- NC350 informatic approaches were combined to O OO (ZA) CML103 select candidate genes for further functional O characterization. Tzi8 CML247 Mutual Rank (MR)–based global gene coexpression analysis (34, 35) showed that of B97 the three maize MAX1 homologs, only ZmMAX1b CML333 tightly coexpressed with ZmCCD8 (fig. S15), M162W making it the strongest candidate for the next Oh7B Ky21 CML277 HP301 CML52 Tx303 Ki3 Oh43 Ki11 Ms71 Mo18W NC358 CML322 B73 Tzi9 TZEI3 TZEI14 TZSTRI101 TZSTRI107 TZSTRI108 TZSTRI110 TZSTRI112 TZSTRI114 TZSTRI119 IITASTRENT-15 IITASTRENT-70 IITASTRENT-16 IITASTRENT-67 NK NP2222 SY INFINITE SY CARIOCA SY BAMBUS SY TELLIAS SY SANDRO SY TALISMAN SY LEOPOLDO 3-oxo-MeCLA ZmCLAMT1 C *** ZmCYP706C37 *** 50 COOCH3 Striga germination 40 percentage (%) ? COOCH3 O OO 30 ns OO O OO 20 zealactone O O OH ? ? compound 4 10 O COOCH3 COOCH3 O OO O OO OO 0 ZAGR24 H2O O O zealactonzeealactol zeapyranolactone compound 3 Fig. 1. Discovery of two strigolactones with low Striga germinationÐinducing activity from maize line screening. (A) Strigolactone (SL) biosynthetic pathway of maize. The enzymes identified in this study are shown in bold. SLs detected in maize root exudate are indicated in blue. Structures in square brackets are putative. (B) Detection of three maize SLs (zealactone, mass/charge ratio (m/z) 377 > 97; zealactol, m/z 331 > 97; ZA, m/z 363 > 249) in root exudate of a collection of maize lines. Names of lines selected for further analysis are indicated in bold. Data for the other four maize SLs are shown in fig. S3. (C) Induction of germination of Striga by zealactone, zealactol, and ZA (0.347 μM). GR24 (0.335 μM) and water were used as positive and negative control, respectively. Bars indicate means ± SEM. Ns, not significant (P > 0.05), ***P < 0.001, one-way ANOVA test followed by TukeyÕs multiple comparisons test comparing the mean of each column with the mean of every other column. Maize roots exude at least six SLs, two of ing the notorious agricultural problem of which have been structurally identified as Striga infection through breeding maize for zealactone and zeapyranolactone (Fig. 1A) favorable SL composition. (24–26). However, the identities of the other four SLs remained elusive, as well as the Natural variation in strigolactone production biosynthetic differences between the six and by maize their individual roles in Striga germination. In this study, we reveal natural variation in To assess the extent of variation in the pro- the maize SL blend, identify three new maize duction of SLs by maize, we grew a collection SLs, elucidate the entire maize SL biosynthetic of maize genotypes, sampled their root exu- pathway, and show that changes in the com- date, and analyzed SLs with multiple reaction position of the SL blend correspond to differ- monitoring (MRM) liquid chromatography– ences in Striga germination and infection. tandem mass spectroscopy (LC/MS/MS) (Fig. These findings create a pathway for reduc- 1B and figs. S2 and S3) (24, 25). Quantities of exuded SLs varied among these lines (Fig. 1B SCIENCE science.org 6 JANUARY 2023 • VOL 379 ISSUE 6627 95

RESEARCH | RESEARCH ARTICLES Fig. 2. Identification of gene candidates A MRM positive, m/z 303 > 97 B MRM negative, m/z 331 > 113 F MRM positive, m/z 347 > 97 for maize strigolactone biosynthesis. (A and B) Representative MRM-LC/MS/MS 100 STD CL Intensity 100 STD CLA Intensity 100 STD MeCLA chromatograms of carlactone (CL), [M+H]+m/z 303 > 97 (A), and 7.61e5 1.41e6 5.61e5 carlactonoic acid (CLA), [M-H]–m/z 331 > 113 (B), in 0 0 0 4.7 4.8 4.9 5.0 5.1 N. benthamiana leaf samples transiently expressing maize strigolactone 100 N. 5.7 5.8 5.9 100 3.7 3.8 3.9 4.0 4.1 100 N. benthamiana: Intensity (SL) precursor pathway genes. benthamiana: Intensity N. benthamiana: Intensity (C) Untargeted metabolomics to identify CLA conjugates in N. benthamiana CL pathway 7.61e5 CL pathway 4.64e3 CL pathway 5.61e5 leaf samples. m/z 539.21: CLA + hexose + formic acid – H2O; m/z 701.26: CLA + Relative intensity (%) + ZmMAX1b + ZmCLAMT1 Relative intensity (%) + ZmMAX1b + ZmCLAMT1 Relative intensity (%) + ZmMAX1b 2 hexose + formic acid – H2O (D) Venn diagram of candidate gene numbers 0 0 + ZmCLAMT1 from several analyses: module091 from 5.7 5.8 3.7 3.8 3.9 0 maizeGGM, genes differentially expressed in zmccd8 roots (compared with wild type), 100 N. benthamiana: 5.9 100 N. benthamiana: 4.0 4.1 4.7 4.8 4.9 5.0 5.1 and the top 100 genes coexpressed with CL pathway CL pathway 100 N. benthamiana: ZmCCD8 and ZmMAX1b (34, 35). (E) Putative + ZmMAX1b Intensity + ZmMAX1b Intensity CL pathway Intensity SL biosynthetic gene cluster on 7.61e5 4.64e3 chromosome 3 consisting of ZmCLAMT1, + ZmMAX1b 5.61e5 ZmMAX1b, and ZmCYP706C37, adapted from screenshot from UCSC Genome 0 0 0 4.7 4.8 4.9 5.0 5.1 Browser on Z. mays (B73 RefGen_v3) 100 N. be5n.t7hamiana5:.8 3.7 3.8 3.9 Assembly (zm3) (http://genome.ucsc.edu) 5.9 4.0 4.1 100 N. benthamiana: Intensity (39). (F) Representative chromatograms CL pathway 100 N. benthamiana: CL pathway 5.61e5 of methylcarlactonoic acid (MeCLA), Intensity CL pathway Intensity [M+H]+m/z 347 > 97, in N. benthamiana leaf 7.61e5 4.64e3 samples. STD, standard; EV, empty vector infiltrated control sample. CL 0 0 0 pathway, maize carlactone biosynthetic pathway genes, ZmD27, ZmCCD7, and 100 5.7 5.8 5.9 100 3.7 3.8 3.9 4.0 4.1 4.7 4.8 4.9 5.0 5.1 ZmCCD8. CLA pathway, CL pathway genes + N. benthamiana: Intensity N. benthamiana: 100 N. benthamiana: ZmMAX1b. MeCLA pathway, CLA pathway EV EV Intensity Intensity genes + ZmCLAMT1. Bars indicate mean ± SEM. 7.61e5 4.64e3 EV 5.61e5 0 0 0 5.7 5.8 5.9 3.7 3.8 3.9 4.0 4.1 4.7 4.8 4.9 5.0 5.1 Retention time (min) Retention time (min) Retention time (min) C 20000 EV D DEG_ccd8 vs WT CL pathway 4 (1301) 15000 CLA pathway GGM-Module091 Mass intensity MeCLA pathway (35) 1277 19 7 10000 5 13 5000 75 0 CLA-dihexose MutRank_top100 (m/z 701.26) (100) CLA-hexose (m/z 539.21) E Scale 100 kb 13,050,000 13,100,000 13,150,000 13,000,000 chr3: ZmCLAMT1 ZmMAX1b ZmCYP706C37 biosynthetic step. Analysis of root exudate from Similar conjugation has been demonstrated tative gene cluster on chromosome 3 (Fig. 2, D a zmmax1a zmmax1c double mutant (supple- for the transient production of other acidic and E, and fig. S15) (39). Genes homologous to mentary materials) showed wild-type levels compounds with N. benthamiana (36, 37). these also cluster in other Poaceae species (fig. of zealactone, thus excluding both homologs S16), but the functional importance is un- from being the biosynthetic genes we sought For selection of remaining candidate genes, known. So too is the identity of SLs produced (fig. S13B). Earlier research also demonstra- we combined three approaches: (i) MR–based by some of these species, such as switchgrass. ted that ZmMAX1b (GRMZM2G023952) con- coexpression with ZmCCD8 and ZmMAX1b as verts CL to CLA more efficiently than does baits (fig. S15), (ii) coexpression modules in ZmCLAMT1 is a carlactonoic acid ZmMAX1a (GRMZM2G018612) or ZmMAX1c MaizeGGM2016 (38), and (iii) differential gene methyltransferase (GRMZM2G070508) (18). The amounts of CL expression in a zmccd8 mutant (Fig. 2D). For in leaf extracts decreased after coinfiltration the latter, we assumed that SL pathway genes Because SLs zealactone and zeapyranolactone of ZmMAX1b with ZmD27, ZmCCD7, and downstream of CCD8 would be transcription- are methylesters, their proposed precursor has ZmCCD8 in N. benthamiana, (Fig. 2A), con- ally regulated in the zmccd8 mutant (33). The been methyl carlactonoate (MeCLA) (24). Thus, firming that ZmMAX1b uses CL as a substrate ZmCCD7, ZmCCD8, and ZmMAX1b genes clus- we sought a methyltransferase gene that (18). However, only traces of the expected tered together in MaizeGGM2016 module 091, causes the formation of MeCLA from CLA. product, CLA, were detected in this expres- suggesting that the 32 other genes in this mod- We bioinformatically identified a top candidate sion system (Fig. 2B and fig. S14B). To resolve ule were candidates for the missing pathway (GRMZM2G033126) (Fig. 2, D and E), which this enigma, N. benthamiana extracts were genes (table S5). In the roots of zmccd8 seed- successfully produced MeCLA in N. benthamiana analyzed with LC–quadrupole time-of-flight lings, 1301 genes were differentially expressed when transiently expressed together with genes (QTOF)–MS. Prominent peaks of CLA-hexose (DEGs) (less than or equal to twofold change, for the maize CLA pathway (Fig. 2F). We there- and CLA-dihexose conjugates were detected false discovery rate (FDR) < 0.05) compared fore identified GRMZM2G033126 as a carlacto- in samples expressing the maize CL pathway with the B73 wild type (tables S5 and S6). noic acid methyltransferase gene and named genes together with ZmMAX1b. These con- These three approaches shared a seven-gene the enzyme ZmCLAMT1 (Fig. 1A). The maize gene jugates were lacking in control samples and overlap (Fig. 2D and table S2) in which three is an ortholog of At4g36470, which was recently other gene combinations (Fig. 2C and table S4). [GRMZM2G033126, GRMZM2G158342, and found to encode a carlactonoic acid methyl- GRMZM2G023952 (ZmMAX1b)] formed a pu- transferase CLAMT in Arabidopsis (40, 41). 96 6 JANUARY 2023 • VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES Fig. 3. Zealactone biosynthesis. (A) Representative A MRM positive, m/z 377 > 97 B MRM positive, m/z 377 > 97 D MRM positive, m/z 377 > 97 MRM–LC/MS/MS chromatograms of zealactone, 100 STD 100 STD 100 STD zealactone [M+H]+m/z 377 > 97, in N. benthamiana leaf Intensity zealactone Intensity Intensity samples. (B and D) Representative MRM–LC/MS/MS 5.94e7 chromatograms of zealactone from in vitro assays zealactone 8.84e7 5.10e7 with yeast microsomes expressing ZmCYP706C37 or empty vector (EV) with methyl carlactonoate 0 4.0 5.0 6.0 7.0 0 (MeCLA), 3-hydroxy-MeCLA, or 3-oxo-MeCLA as 100 N. benthamiana: 5.5 6.0 6.5 7.0 7.5 substrate. (C) Proposed enzymatic conversion of MeCLA pathway 8.0 100 Yeast microsome: methyl carlactonoate (MeCLA) to zealactone. + ZmCYP706C37 ZmCYP706C37 Intensity 8.84e7 Relative intensity (%) 0 Substrate: Intensity Relative intensity (%) Relative intensity (%)5.06.07.08.00 3-oxo-MeCLA 7.07e6 Yeast 100 Intensity 5.5 6.0 6.5 7.0 7.5 1.44e7 microsome: 100 Yeast microsome: ZmCYP706C37 0 4.0 5.0 6.0 7.0 8.0 ZmCYP706C37 Intensity Substrate: 100 N. benthamiana: Substrate: 7.07e6 Intensity MeCLA 0 3-hydroxy-MeCLA 7.0 7.5 MeCLA pathway 8.84e7 5.5 6.0 6.5 Intensity 100 Yeast microsome: 0 5.0 6.0 7.0 8.0 Yeast EV 100 Intensity 0 1.44e7 Substrate: 7.07e6 100 4.0 5.0 6.0 7.0 8.0 microsome: 0 3-oxo-MeCLA N. benthamiana: Intensity EV 8.84e7 EV 5.5 6.0 6.5 7.0 7.5 Substrate: 100 Yeast microsome: Intensity EV MeCLA 7.07e6 Substrate: 0 0 0 3-hydroxy-MeCLA 4.0 5.0 6.0 7.0 8.0 5.0 6.0 7.0 8.0 5.5 6.0 6.5 7.0 7.5 Retention time (min) Retention time (min) Retention time (min) C 16 17 19 20 COOCH3 COOCH3 COOCH3 7 COOCH3 O 2 16 8 9 10 OO O 3 4 5 18 O 11 O step 1 O O step 2 O O step 3 O step 4 O HO OO O COOCH3 12 14 O 13 OO O methyl carlacton1o5 ate 3-hydroxy-MeCLA 3-oxo-MeCLA (MeCLA) COOCH3 OH COOCH3 OH step 6 step 5 O O OO OO HO O O O HO zealactone Fig. 4. Zealactol and zealactonoic acid bio- A MRM positive, m/z 331 > 97 B MRM positive, m/z 331 > 97 D MRM positive, m/z 363 > 249 synthesis. (A) Representative MRM–LC/MS/MS chromatograms of zealactol, [M+H-H2O]+m/z 331 > 100 STD Intensity 100 STD Intensity 100 STD Intensity 97, in N. benthamiana leaf samples. (B) Represent- zealactol 2.26e7 zealactol 2.39e7 ative MRM–LC/MS/MS chromatograms of zealactonoic acid 9.91e6 zealactol from in vitro assays with yeast (ZA) microsomes expressing ZmCYP706C37 or empty 0 4.1 4.3 4.5 4.7 0 vector (EV) with carlactone (CL) as substrate. 100 N. benthamiana: 4.5 5.0 5.5 6.0 6.5 7.0 7.5 Intensity (C) Reactions from CL to zealactol and ZA catalyzed CL pathway+ 5.54e5 0 4.2 4.3 4.4 4.5 4.6 100 N. benthamiana : by ZmCYP706C37 and ZmMAX1b. Structure in ZmCYP706C37 CL pathway Relative intensity (%) 100 Yeast +ZmCYP706C37 Intensity square brackets is putative. (D) Representative Relative intensity (%)microsome: 2.79e6 MRM–LC/MS/MS chromatograms of ZA, Relative intensity (%) [M+H]+m/z 363 > 249, in N. benthamiana leaf ZmCYP706C37 Intensity +ZmMAX1b samples. STD, standard; EV, empty vector control. 1.36e5 Substrate: 0 CL pathway, maize carlactone biosynthetic 0 CL 4.5 5.0 5.5 6.0 6.5 7.0 7.5 pathway genes, ZmD27, ZmCCD7, and ZmCCD8. 4.1 4.3 4.5 4.7 100 N. benthamiana : Intensity 100 N. benthamiana: Intensity CL pathway 2.79e6 CL pathway 5.54e5 0 +ZmCYP706C37 4.2 4.3 4.4 4.5 4.6 0 4.1 4.3 4.5 4.7 100 Yeast microsome: Intensity 0 100 EV 1.36e5 4.5 5.0 5.5 6.0 6.5 7.0 7.5 Intensity Substrate: 100 N. benthamiana : Intensity N. benthamiana: 5.54e5 CL EV EV 2.79e6 0 0 0 4.1 4.3 4.5 4.7 4.2 4.3 4.4 4.5 4.6 4.5 5.0 5.5 6.0 6.5 7.0 7.5 Retention time (min) Retention time (min) Retention time (min) C OH OH ZmCYP706C37 ZmCYP706C37 O OO O OO O O O OO O +ZZmmCMYAPX710b6C37 3-oxo-19-hydroxy-CL zealactol carlactone (CL) COOH O OO O O zealactonoic acid (ZA) ZmCYP706C37 catalyzes formation of several substrate (fig. S17A) and produce zealactone ring and, for zealactol, a hydroxylation at C19 maize strigolactones (Fig. 3A and fig. S2). To check for other pos- as well. To exclude the possibility of endoge- The other candidate genes were coinfiltrated sible biosynthetic pathways, we also coexpressed nous enzymes from N. benthamiana contribut- ZmCYP706C37 with genes encoding the CL ing to these complex conversions, we expressed by different combinations of precursor-pathway pathway enzymes. This combination resulted ZmCYP706C37 in yeast, isolated its micro- genes. Coinfiltration of ZmCYP706C37 (GRM- in production of zealactol (Fig. 4A and fig. S17B). somes, and analyzed product formation with ZM2G158342) (42) by those encoding the Formation of both zealactone and zealactol different substrates (Figs. 3B and 4B). This ap- MeCLA pathway decreased levels of MeCLA, involves complex rearrangement of the SL A proach confirmed that ZmCYP706C37 can indicating that this CYP can use MeCLA as a SCIENCE science.org 6 JANUARY 2023 ¥ VOL 379 ISSUE 6627 97

RESEARCH | RESEARCH ARTICLES A 70 *** D 40 For additional insight into the parallel bio- synthetic pathway of CL to zealactol, we fur- Striga germination 60 *** ***Striga germination ther analyzed samples from N. benthamiana percentage (%) *** percentage (%) and yeast microsome assays with untargeted 30 metabolomics and MRM-LC-MS/MS. This pro- 50 zmmaWx212b cess revealed another putative intermediate, 20 GR24 3-oxo-19-hydroxy-CL (compound 7) (Fig. 1A 40 H and figs. S2 and S20 and table S7). LC-QTOF- 10 2O MS analysis showed that the accurate mass 30 of compound 7 is consistent with its putative 0 structure (fig. S20). On the basis of these data, 20 we included compound 7 as an intermediate in the postulated steps required to convert CL 10 to zealactol (Fig. 4C and fig. S21). Moreover, agroinfiltration of the CL pathway genes with 0 ZmCYP706C37 and ZmMAX1b resulted in production of ZA, a result also confirmed with B CCNPMM2LL2N65292K2 LC-QTOF-MS (Fig. 4, C and D, and fig. S22). NGKCyR3225148 Last, analysis of root exudate from a zmcyp706- H c37 mutant [EMS4-045ad8, stop-codon gained 2O (fig. S23A)] showed no detectable levels of zealactol, ZA, zealactone, or three other SLs 8 NK 30 CML69 * 20 Ky21 * derived from the latter (fig. S23B) (44). Al- NC358 * though 3-oxo-MeCLA was detectable in the mu- Striga emergence NP2222 * * * * CML52 * * tant exudate, it was present at a much lower ns * 20 ** 15 level than in that of the wild type. Instead, 6 CLA and MeCLA accumulated in the mutant exudate, whereas they are absent in the wild 4 ns ns 10 type exudate (fig. S23, C and D). Together, these 2 data support our functional characterization 10 ns ns of ZmCYP706C37. ns ns 5 ns ns ns Biosynthetic control of the maize strigolactone blend ns To determine how the different maize SLs are 0 00 biosynthetically related, we applied 3-hydroxy- MeCLA, 3-oxo-MeCLA, and zealactol to seed- 30 33 36 39 42 45 48 51 25 28 31 35 38 42 45 35 38 42 45 49 52 54 lings of another commercial line, NK Falkone, which were treated with fluridone, an inhib- DAG DAG DAG itor of SL biosynthesis (45). Each of these three compounds complemented zealactone produc- NK NP2222 CML69 CML52 Ky21 NC358 tion (fig. S24A), confirming that they can serve as biosynthetic precursors for zealactone. Com- C 2×106 *** W22 ns E W22 bined transient expression of ZmMAX1b and zmmax1b zmmax1b ZmCLAMT1 in N. benthamiana leaves and 1.5×106 ** *** 40 subsequent infiltration of zealactol also showed Peak Area 30 ** ** that the latter can be converted to zealactone Striga emergence by ZmMAX1b together with ZmCLAMT1 (Fig. 1×10 6 *** 1A and fig. S25). Application of zealactone to 5×10 5 20 ** fluridone-treated plants led to the formation 4×10 5 of zeapyranolactone and two other maize SLs, * designated compounds 3 and 4, suggesting 3×10 5 10 that zealactone is their precursor (Fig. 1A and 2×10 5 fig. S24, B to D) (24). 1×10 5 ns ns * Next, we sought mechanisms underlying the 0 0 distinctive maize SL profile of NP2222 (fig. S26). 21 24 28 31 35 38 42 45 This line produces zealactone in fluridone- treated seedlings, as does NK Falkone, but ns DAG only from MeCLA and 3-oxo-MeCLA, not from zealactol (figs. S24A and S26A), suggesting *** *** inactivity of MAX1b and/or CLAMT1. As pre- viously noted, ZA accumulated in the root zeap3-yccrooozaxneomm-oappllooMaaeuuccttnnCooddLnnA43ee W22 zmmax1b exudate of NP2222 (Fig. 1B and fig. S26D), zealactol indicating dysfunction of CLAMT1. Zealactol ZA Fig. 5. Changes in the maize strigolactone blend result in changes in Striga resistance. (A and D) Induction of Striga germination by root exudates of selected maize lines. GR24 (0.335 μM) and water were used as positive and negative control, respectively. (B and E) Striga infection of selected maize lines. Emerged Striga numbers were recorded; representative photos highlight the differences. DAG, days after germination of maize. (C) SL levels in the root exudate of zmmax1b and its wild type, W22. Bars indicate means ± SEM, ns = not significant (P > 0.05), *P < 0.05, **P < 0.01, ***P < 0.001, two-tailed, unpaired t test. convert MeCLA to zealactone and CL to zea- detection of these compounds (fig. S2) and lactol (Fig. 1A). identified them as intermediate products in the conversion of MeCLA to zealactone (fig. To form zealactone from MeCLA, ZmCYP706- S18). Moreover, analysis of maize root exudate C37 must catalyze several consecutive oxidative revealed that 3-oxo-MeCLA is also a natural reactions with 3-hydroxy-MeCLA and 3-oxo- maize SL previously referred to as compound MeCLA as putative intermediates (Figs. 1A 6 (fig. S19 and Fig. 1A) (24). These results de- and 3C). The latter two compounds were prev- monstrate that a single enzyme, ZmCYP706- iously synthesized as intermediates in the C37, can catalyze the many oxidative steps total synthesis of heliolactone (43). We used necessary for the conversion of MeCLA to zea- them here as substrates in our ZmCYP706C37- lactone that were previously hypothesized expressing yeast-microsome assay, and both to require several enzymes (Figs. 1A and were successfully converted to zealactone 3C) (24). (Fig. 3D). We developed an MRM method for 98 6 JANUARY 2023 ¥ VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES added to either NK Falkone or NP2222 was con- the balance between zealactone and zealactol 36. L. Dong et al., Metab. Eng. 20, 198–211 (2013). verted to ZA, showing that ZmMAX1b is active 37. X. Xu et al., J. Exp. Bot. 72, 5462–5477 (2021). in NP2222 (fig. S26, B and C). Inspection of the plus ZA. Zealactol and ZA induce much less 38. S. Ma, Z. Ding, P. Li, BMC Plant Biol. 17, 131 CLAMT1 sequence in a proprietary NP2222 genome database revealed a large insertion in Striga germination, thus imparting a strong (2017). the second exon of this gene, and reverse tran- 39. W. J. Kent et al., Genome Res. 12, 996–1006 scriptase polymerase chain reaction (RT-PCR) reduction in Striga infection to genotypes that showed that regions flanking the insertion (2002). were not transcribed (fig. S26E). These col- exude more zealactol and ZA than zealactone. 40. T. Wakabayashi et al., Planta 254, 88 (2021). lective data indicate disfunction of CLAMT1 41. K. Mashiguchi et al., Proc. Natl. Acad. Sci. U.S.A. 119, in NP2222. Future research should investigate whether e2111565119 (2022). To analyze biological consequences of the these changes in the SL blend affect coloni- 42. Y. Li, K. Wei, BMC Plant Biol. 20, 93 (2020). different SL profiles, several maize lines were 43. M. Yoshimura et al., Helv. Chim. Acta 102, e1900211 selected for Striga germination and infection zation by AM fungi, which was not observed assays. The NP2222 root exudate induced much (2019). lower germination than that of NK Falkone. for lgs sorghum (16). Our results offer a per- 44. X. Lu et al., Mol. Plant 11, 496–504 (2018). Results were consistent with their respective 45. J. A. López-Ráez et al., New Phytol. 178, 863–874 SL profiles and differences in germination- spective for breeding Striga resistance through inducing activity of the individual SLs (Figs. 1C (2008). and 5A and fig. S26D). CML52 and NC358, modification of the SL blend in maize and thus both with high proportions of zealactol and ACKNOWLEDGMENTS ZA, induced significantly less Striga germi- potentially reducing the devastating effects of nation than did CML69 and Ky21, which pro- We acknowledge S. Al Babili from King Abdullah University of duced mostly zealactone despite similar total this parasitic weed in Africa. Science and Technology and D. Werck-Reichhart from the University SL peak areas (Figs. 1C and 5A, and fig. S27, A of Strasbourg for helpful discussions, as well as L. Hagmann and B). These differences were also reflected in REFERENCES AND NOTES from Syngenta for his support in NMR analyses and interpretation. a Striga infection assay with a containerized Funding: This work was funded by the China Scholarship Council system, in which Striga emergence was less for 1. T. Wheeler, J. von Braun, Science 341, 508–513 (CSC) PhD scholarship 201706300041 (C.L.), the European low-zealactone genotypes (Fig. 5B). In addi- (2013). Research Council (ERC) Advanced grant CHEMCOMRHIZO 670211 tion to their SL blend, these lines may have (H.J.B.), the Dutch Research Council (NWO/OCW) Gravitation other genetic differences that could affect these 2. B. Badu-Apraku, F. M.A.B., Advances in Genetic program Harnessing the second genome of plants (MiCRop) results. However, we also analyzed a gene- Enhancement of Early and Extra-Early Maize for 024.004.014 (H.J.B.), the Marie Curie fellowship NEMHATCH suppression mutant of ZmMAX1b (transposon Sub-Saharan Africa (Springer Cham, 793795 (L.D.). K.E.K and J.G. acknowledge funding from the US insertion in a W22 background) (fig S28, A 2017). National Science Foundation (NSF) Plant Genome Research and B). This mutant exuded significantly less Program (PGRP) (1421100 and 1748105). Author contributions: izealactone and zealactone-derived SLs, whereas 3. I. Dörr, Ann. Bot. (Lond.) 79, 463–472 (1997). C.L., L.D., and H.J.B. conceived and designed the project. C.L. the level of zealactol was higher than in the 4. C. E. Cook, L. P. Whichard, B. Turner, M. E. Wall, G. H. Egley, discovered and characterized the candidate genes, grew the W22 control (Fig. 5C). The zmmax1b mutant plants, collected and analyzed the root exudate, cloned the genes, also induced less Striga germination and em- Science 154, 1189–1190 (1966). performed agroinfiltration, yeast microsome assay and plant ergence (Fig. 5E). Results confirm that a change 5. A. Besserer et al., PLOS Biol. 4, e226 (2006). compound treatment assays, and coordinated the project; K.F., in activity of specific SL biosynthetic enzymes 6. V. Gomez-Roldan et al., Nature 455, 189–194 T.V.C. and A.C. developed LC-MS methods and helped with in maize can change the SL composition and SL analysis; T.V.C., J.D., and A.D.J.V.D. helped to establish the confer Striga resistance. Although the under- (2008). biosynthesis mechanisms; J.G. and K.E.K. developed and provided lying mechanisms are completely different, 7. S. Al-Babili, H. J. Bouwmeester, Annu. Rev. Plant Biol. 66, maize seeds (NAM, zmccd8, zmmax1azmmax1c, and zmmax1b) these findings resemble those of lgs sorghum and analyzed RNA-seq and related data. B.T. and L.D. supported (16) and present a promising prospect for Striga 161–186 (2015). the metabolomics analysis; M.Y., K.G., A.D.M. synthesized zealactol resistance breeding in maize. The zmmax1b 8. M. Umehara et al., Nature 455, 195–200 (2008). and provided zealactone, 3-hydroxy-MeCLA, and 3-oxo-MeCLA; mutant did not exhibit a branching pheno- 9. K. Akiyama, K. Matsuzaki, H. Hayashi, Nature 435, 824–827 P.Q., R.H., and A.D.M. synthesized zealactonoic acid; J.L., Y.B.S., type, in contrast to zmccd8 (fig. S28C). Also, J.Q., and J.W. grew the zmcyp706c37 EMS mutants, performed zmcyp706c37, which is located parallel to or (2005). genotyping, selfing, and root exudate collection; H.D.G. collected downstream of ZmMAX1b, did not display 10. H. Bouwmeester, C. Li, B. Thiombiano, M. Rahimi, L. Dong, and prepared the maps of maize and Striga occurrence. Y.W. an obvious branching phenotype either. This helped with the agroinfiltration and yeast microsome assays; all suggests that the downstream SLs are not Plant Physiol. 185, 1292–1308 (2021). C.L., A.W., and B.T. performed the Striga germination and infection nor precursors of the branching inhibiting 11. K. Yoneyama et al., J. Exp. Bot. 69, 2231–2239 bioassays; S.M.C.d.L. and M.H.M. carried out the gene cluster hormone and are therefore safe breeding analysis; Y.D. and E.A.S. provided support on coexpression targets that will not result in unwanted pleio- (2018). analysis; D.K., K.H. and C.S. provided all commercial maize seeds tropic effects. 12. K. Mashiguchi, Y. Seto, S. Yamaguchi, Plant J. 105, 335–350 from Syngenta and coordinated the collaboration with Syngenta. A.M. provided African inbred maize lines. C.L., L.D., and H.J.B. Conclusions (2021). wrote the manuscript, with contributions from other authors. 13. K. Akiyama, S. Ogasawara, S. Ito, H. Hayashi, Plant Cell Physiol. Competing interests: M.H.M. is a consultant to Corteva We have shown that two parallel SL biosyn- Agriscience, but that company was not involved in this work. All thetic pathways operate in maize and that 51, 1104–1117 (2010). the other authors declare that they have no competing interests. both pathways produce the major maize SL, 14. N. Mori, K. Nishiuma, T. Sugiyama, H. Hayashi, K. Akiyama, Data and materials availability: The maize mutants zmccd8 zealactone. Changes in flux through these path- and zmmax1azmmax1c were obtained via a material transfer ways can alter the maize SL profile by shifting Phytochemistry 130, 90–98 (2016). agreement (MTA) with the University of Florida Board of Trustees. 15. H. I. Kim et al., J. Pestic. Sci. 35, 344–347 (2010). The RNA-seq data of zmccd8 and B73 root tissues are available 16. D. Gobena et al., Proc. Natl. Acad. Sci. U.S.A. 114, 4471–4476 in the NCBI database (BioProject PRJNA757767) under accession numbers of SRR15613590, SRR15613591, SRR15613599, (2017). SRR15613593, SRR15613594, and SRR15613595. All the other 17. A. Alder et al., Science 335, 1348–1351 (2012). data are presented in the main text and in the Supplementary 18. K. Yoneyama et al., New Phytol. 218, 1522–1533 Materials. License information: Copyright © 2023 the authors, some rights reserved; exclusive licensee American Association (2018). for the Advancement of Science. No claim to original US 19. S. Abe et al., Proc. Natl. Acad. Sci. U.S.A. 111, 18084–18089 government works. https://www.science.org/about/science- licenses-journal-article-reuse (2014). 20. C. Cardoso et al., Proc. Natl. Acad. Sci. U.S.A. 111, 2379–2384 SUPPLEMENTARY MATERIALS (2014). science.org/doi/10.1126/science.abq4775 21. Y. Zhang et al., Nat. Chem. Biol. 10, 1028–1033 (2014). Materials and Methods 22. T. Wakabayashi et al., Sci. Adv. 5, eaax9067 (2019). Figs. S1 to S28 23. T. Wakabayashi et al., Planta 251, 97 (2020). Tables S1 to S8 24. T. V. Charnikhova et al., Phytochemistry 137, 123–131 References (46–70) (2017). Submitted 12 April 2022; accepted 30 November 2022 25. T. V. Charnikhova et al., Phytochem. Lett. 24, 172–178 10.1126/science.abq4775 (2018). 26. X. Xie et al., J. Pestic. Sci. 42, 58–61 (2017). 27. M. Yoshimura et al., Helv. Chim. Acta 103, e2000017 (2020). 28. M. C. Dieckmann, P.-Y. Dakas, A. De Mesmaeker, J. Org. Chem. 83, 125–135 (2018). 29. T. Kumagai et al., Heterocycles 36, 1729–1734 (1993). 30. J. C. Guan et al., Plant Physiol. 160, 1303–1317 (2012). 31. K. C. Snowden et al., Plant Cell 17, 746–759 (2005). 32. W. Kohlen et al., New Phytol. 196, 535–547 (2012). 33. Y. Zhang et al., New Phytol. 219, 297–309 (2018). 34. E. Poretsky, A. Huffaker, PeerJ 8, e10264 (2020). 35. S. Stelpflug et al, The Plant Genome 9, plantgenome2015.04.0025 (2016). SCIENCE science.org 6 JANUARY 2023 ¥ VOL 379 ISSUE 6627 99

RESEARCH | RESEARCH ARTICLES MPOX for VACV could also be applied to MPXV. How- ever, we are still awaiting a reliable high- Structure of monkeypox virus DNA resolution structure of the replicating state of polymerase holoenzyme the Orthopoxvirus polymerase holoenzyme, and the mode of operation of the processiv- Qi Peng1†, Yufeng Xie2†, Lu Kuai1†, Han Wang1,3, Jianxun Qi1, George F. Gao1,2,4,5,6*, Yi Shi1,5,6* ity factor needs to be elucidated. The World Health Organization declared mpox (or monkeypox) a public health emergency of Results international concern in July 2022, and prophylactic and therapeutic measures are in urgent need. Biochemical characterization of the purified The monkeypox virus (MPXV) has its own DNA polymerase F8, together with the processive cofactors polymerase proteins A22 and E4, constituting the polymerase holoenzyme for genome replication. Here, we determined the holoenzyme structure in complex with DNA using cryo–electron microscopy at the global resolution We coexpressed MPXV F8 polymerase and the of ~2.8 angstroms. The holoenzyme possesses an architecture that suggests a “forward sliding clamp” A22-E4 heterodimer using the baculovirus ex- processivity mechanism for viral DNA replication. MPXV polymerase has a DNA binding mode similar pression system and purified the homogeneous to that of other B-family DNA polymerases from different species. These findings reveal the mechanism F8-A22-E4 heterotrimer protein for enzymatic of the MPXV genome replication and may guide the development of anti-poxvirus drugs. and structural studies (fig. S1). When a 38- nucleotide (nt) template DNA was used with a A s of 2 December 2022, over 82,000 hu- MPXV is a large double-stranded DNA virus 24-nt primer DNA, the wild-type holoenzyme man mpox (or monkeypox) cases have that replicates exclusively in the cytoplasm of heterotrimer displayed weak primer extension been laboratory confirmed in 110 coun- the infected cells. It belongs to the Orthopox- activity in the reaction buffer with deoxynu- virus genus of the Poxviridae family, which cleotide triphosphate (dNTP) (fig. S1). More- tries worldwide (https://www.cdc.gov/). also includes the variola virus that causes over, exonuclease activity was confirmed using Most infection cases have been reported smallpox and has killed millions of humans an enzymatic assay without dNTP substrates. in recorded history. Similar to the vaccinia The holoenzyme could completely degrade in Europe and other non-endemic countries, virus (VACV), the prototype of poxviruses, the primer DNA in an adenosine 5´-triphosphate including China (1), and these cases were MPXV may enter host cells by either fusion (ATP)–independent manner (fig. S1). We also mostly found in homosexual young men (2). with the plasma membrane or endocytosis, prepared an exonuclease-deficient F8 mutant and at least 16 proteins in the virion mem- protein and found that the F8-mutant-A22-E4 Human-to-human transmission usually occurs brane are involved in the entry process (13). holoenzyme did not cleave the primer-template through close contact with lesions, respiratory After entry, the virus initiates early gene tran- DNA, thereby demonstrating a much stronger scription events, and viral DNA synthesis oc- product band than the wild-type holoenzyme droplets, body fluids, and contaminated mate- curs at perinuclear sites called viral factories protein, and the polymerization product could rials, such as bedding (3). (14, 15). The MPXV replicative holoenzyme be efficiently inhibited by heparin (fig. S1). consists of catalytic polymerase F8 (equivalent Although the monkeypox virus (MPXV) was to E9 in VACV), a heterodimeric processivity Overall architecture of F8-A22-E4 first isolated from a monkey in Denmark in factor consisting of A22 (equivalent to A20 in polymerase holoenzyme VACV) and uracil-DNA glycosylase E4 (equiv- 1958, its natural host was thought to be rodent alent to D4 in VACV). To capture the replicating conformation of the (3, 4). Since the first human mpox case was MPXV F8-A22-E4 polymerase holoenzyme, we Previous genetic, biochemical, and structural incubated the 3′-H modified primer-template identified in the Democratic Republic of the studies on the VACV E9-A20-D4 core repli- DNA and the exonuclease-deficient polymer- Congo (5), it has been endemic to several cen- cation machinery have advanced our under- ase holoenzyme in the reaction buffer with tral and western African countries (6, 7). standing of poxvirus DNA replication. VACV deoxythymidine triphosphate (dTTP) substrate. Sporadic infection cases have been reported E9 was recognized as a member of the We then prepared cryo–electron microscopy B-family DNA polymerase, and structural (cryo-EM) samples using a graphene grid to outside Africa, including England, the United analysis has revealed the canonical features avoid preferential orientation observed with States, Singapore, and Israel, and are mainly of DNA polymerases and five poxvirus-specific ordinary grids. The holoenzyme–DNA com- insertions (16). The E9 polymerase alone does plex was resolved to ~2.8 Å (figs. S2 and S3). associated with travelers from endemic coun- not have processive DNA synthesis activity The EM map shows the key structural features tries, nosocomial infections, or direct contact unless it is bound to its heterodimeric cofactor of all proteins and DNA elements (fig. S4). A20/D4 (17–20). Although poxvirus DNA poly- Although the density of the 5′-end template with imported rodents infected with MPXV merase shares many features with other B-family was weak, we traced the main chains using the (4, 8, 9). Phylogenetic analysis has revealed polymerases, the processivity factor is dis- unsharpened EM density map to demonstrate tinctive. In VACV, A20 serves as an essential the template entry channel (see below). that MPXV can be classified into two genetic bridge to link E9 and D4 together and shares clades: the West Africa clade and the more no homology with viral proteins beyond pox- The structure of the holoenzyme–DNA com- pathogenic Congo Basin clade (10, 11). The 2022 virus. The N-terminal domain of A20 binds to plex contains one F8, one A22, one E4, and the outbreak of MPXV belongs to the West Africa D4 (21–23), and its C-terminal domain binds primer-template DNA, as well as an incoming to one insertion in the palm domain of E9 dTTP substrate (Fig. 1). F8, A22, and E4 form clade and most likely has a single origin that (24). Given that the DNA replication machin- pairwise interactions with each other (Fig. 1). has not been identified (12). ery is extremely conserved for orthopoxviruses, with a sequence identity of more than 97% The F8 structure can be traced for 1004 res- 1CAS Key Laboratory of Pathogen Microbiology and between VACV and MPXV, the results obtained idues, except for the last two residues, and Immunology, Institute of Microbiology, Chinese Academy of the classical N-terminal domain (NTD), exo- Sciences, Beijing 100101, China. 2Department of Basic nuclease domain (Exo), palm domain, fingers Medical Sciences, School of Medicine, Tsinghua University, domain, and thumb domain were observed in Beijing 100084, China. 3College of Future Technology, Peking a closed conformation (Fig. 1). Five “poxvirus- University, Beijing 100871, China. 4Savaid Medical School, specific” insertion regions in MPXV F8 can also University of Chinese Academy of Sciences, Beijing 100049, China. 5Center for Influenza Research and Early-warning (CASCIRE), CAS-TWAS Center of Excellence for Emerging Infectious Disease (CEEID), Chinese Academy of Sciences, Beijing 100101, China. 6Research Unit of Adaptive Evolution and Control of Emerging Viruses, Chinese Academy of Medical Sciences, Beijing 100052, China. *Corresponding author. Email: [email protected] (G.F.G.); shiyi@ im.ac.cn (Y.S.) These authors contributed equally to this work. 100 6 JANUARY 2023 • VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES be observed, like those inserts seen in the VACV Fig. 1. Overall structure of the replicating MPXV DNA polymerase holoenzyme. (A) Schematic diagrams of E9 structure (16) (Fig. 1 and fig. S5, A and B). the domain architecture of MPXV DNA polymerase F8 and processivity factors (A22 and E4). The F8 can be divided into five domains: NTD, blue; Exo, magenta; palm, green; fingers, yellow; thumb, cyan. Compared to The 218-residue MPXV E4 resembles the other B-family polymerases, F8 contains five inserted elements in which the largest one was named as insert2 VACV D4 structure (25) (Fig. 1A and fig. S5, C (purple), and the other four small inserts are indicated as rectangles. A22 is colored by domains: NTD, deep and D). The 426-residue MPXV A22 structure pink; Mid, pink; CTD, salmon. E4, orange; template strand, gray; primer strand, red. (B and C) Atomic model and can be divided into three domains: the A22 cryo-EM density map of the replicating MPXV DNA polymerase holoenzyme. The structures were colored by NTD, middle domain (Mid), and C-terminal domains, as depicted in (A). domain (CTD) (Fig. 1A and fig. S5E). When we aligned the Mid of A22 on the online Dali Exo of F8 and the E4 subunit in an orientation ribose backbones (C6 to T11) because of the server, we found that it shows high structural perpendicular to the DNA duplex (Fig. 2, C weak EM density (Fig. 2, C and D). similarities with the African swine fever virus and D, and fig. S9). The template DNA has 12 (ASFV) DNA ligase and the bacteriophage unpaired nucleotides at the 5′ end, but only Upon primer-template DNA binding, F8 poly- T4 DNA ligase (26, 27). The Mid can be fur- two of them are well-ordered with a defined merase undergoes conformational changes, a ther divided into two subdomains: an adeny- base structure. For the remaining 10 unpaired common feature of the B-family DNA poly- lation domain (OD), which mostly resembles bases, we can only trace partial phosphate- merases. Comparison of the MPXV F8 in this the OD of ASFV DNA ligase (fig. S5F), and an holoenzyme–DNA complex structure with the OB-fold domain (OB), which mostly resem- bles the OB of Thermus filiformis DNA ligase (fig. S5G). Further enzymatic assay showed that the MPXV polymerase holoenzyme did not possess ordinary ligase activity similar to that of T4 ligase (fig. S6, A and B). Compared with the structures of adenylation domains from the T4 and ASFV ligases, the putative ligase active site of the A22 Mid is replaced by hydrophobic and negatively charged resi- dues, which may prevent the binding of ATP (fig. S6C). As the Mid of A22 lacks the essen- tial DNA binding domain, A22 may comprise a degenerative ligase domain acting simply as a flexible linker. The MPXV DNA polymerase holoenzyme is stabilized by pairwise interactions between the F8, A22, and E4 subunits (fig. S7). A22 acts as a bridge to bind E4 and F8 via the A22 NTD and CTD, respectively. The interactions are almost identical to those of their VACV counterparts (fig. S8) (21–24). A previous study proposed a VACV polymerase holoenzyme model with an elongated shape of the A20-D4 cofactor, lead- ing to a ~150-Å distance between the E9 poly- merase active site and the D4 DNA binding site (28). However, in our replicating MPXV holo- enzyme structure, the A22-E4 cofactor folds back, and E4 directly interacts with the Exo domain of F8 at two sites, one where Trp36 and Arg39 of E4 form hydrogen bonds and hydrophobic inter- actions with Phe179 and Leu278 from F8 Exo (fig. S7F), and another where Asn165 of E4 forms a hydrogen bond with Asn303 from F8 Exo (fig. S7G). Primer-template DNA recognition by the polymerase complex The structure of MPXV polymerase holoenzyme– DNA complex contains 22-nt DNA in the tem- plate strand, 14-nt DNA in the primer strand, and the incoming dTTP, as well as a magne- sium ion that may serve as catalytic ion near the active site (Fig. 2, A and B). The double- stranded primer-template DNA binds in a groove formed between the palm and thumb domains of F8, and the single-stranded 5′ ex- tension of the template strand probably passes through a channel formed by the NTD and SCIENCE science.org 6 JANUARY 2023 • VOL 379 ISSUE 6627 101

RESEARCH | RESEARCH ARTICLES Fig. 2. The interactions between DNA and MPXV polymerase. (A) Sharpened EM densities and atomic models of dsDNA. (B) Enlarged view of the sharpened EM densities and atomic models of dsDNA in the active site of polymerase. (C) The cut-off view of unsharpened EM density map for MPXV polymerase holoenzyme in complex with the DNA, revealing the consecutive density of template DNA. (D) Enlarged view of the unsharpened densities and supposed atomic models of the template DNA in the template entry channel. The region (C6 to T11) was built using this unsharp- ened map. The remaining 5′-terminal five bases of template DNA were invisible, which reflects an inherent flexible conformation of the 5′-terminal unpaired region of the template strand. (E and F) The primary interfaces between F8 and DNA. The F8 mainly interacted with the minor groove of primer-template DNA, with only a few contacts to the major groove contributed by the residues of exonuclease domain. The primer-template DNA is shown in surface repre- sentation calculated from the atomic model, and F8 is shown in cartoon representation. VACV apo E9 structure, which has high se- of incoming dNTP. The rotated fingers do- plex is accommodated in a positively charged quence identity, shows that the fingers domain main interacts with the Exo, and this interac- groove of the thumb domain, as observed in rotates toward the palm domain by ~17° in the tion further stabilizes the closed conformation other B-family DNA polymerases (fig. S10). replicating state (fig. S9). This rotation drags of the fingers domain. Moreover, the thumb the positively charged Arg634 and Lys661 of the domain also makes a distinct rotation to wrap The modeled double-stranded DNA helix fingers domain closer to the active site, where around the primer-template DNA duplex on is formed by 14 base pairs from the primer- they can interact with the triphosphate group its minor groove side (fig. S9). The DNA du- template DNA and maintains a B-form con- formation (Fig. 2A and fig. S11). Extensive 102 6 JANUARY 2023 ¥ VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES protein-DNA interactions are observed be- Fig. 3. Recognition of the incoming dTTP. (A) Cut-off view of the F8 protein, which is shown in surface tween the primer-template DNA and F8, with representation to reveal the inner active site. The F8 is colored by domains as in Fig. 1; template strand, a total of 47 residues from F8 directly partici- gray; primer strand, red. (B) The binding pocket of the incoming dTTP. It is formed by the fingers domain, pating in DNA binding (within 4.0 Å distance, palm domain, and upper base pair. The incoming dTTP is shown as a ball-and-stick model, and the residues 29 residues to the template strand, and 18 res- of F8 and the upper base pair are shown in surface representation. (C) Interaction details between F8 idues to the primer strand) (Fig. 2, fig. S11, and the incoming dTTP. The key residues are shown as sticks and colored as corresponding domains. The and table S2). Most of the key residues in- magnesium ion is depicted as a black sphere. Single-letter abbreviations for the amino acid residues are as volved in primer-template DNA binding are follows: A, Ala; C, Cys; D, Asp; E, Glu; F, Phe; G, Gly; H, His; I, Ile; K, Lys; L, Leu; N, Asn; P, Pro; R, Arg; S, Ser; highly conserved among different B-family T, Thr; V, Val; W, Trp; and Y, Tyr. DNA polymerases. Protein-DNA interactions mainly involve the phosphodiester backbone to select dNTP as the substrate. This similar ratio of A22-E4 to F8 was 1:1, corresponding to of the DNA, with many interactions directly “steric gating” effect has also been described the stoichiometry of the polymerase holoenzyme, with the phosphate groups (fig. S11A). Inter- in other DNA polymerases and was first pro- the amount of the full-length product was al- actions with the template-strand phosphates posed in HIV-1 reverse transcriptase (31–33). most the same as that of the product generated are largely hydrogen bonds to the main or side by the preassembled F8-A22-E4 polymerase chains of 13 residues from the thumb, palm, Operation mode of processivity cofactor holoenzyme (Fig. 4A). This indicates that the fingers domains, Exo, and NTD of F8 (fig. S11, isolated A22-E4 and F8 can be efficiently as- A to C), whereas the primer strand is bound by For B-family DNA polymerases, proliferating sembled into functional holoenzymes to per- both electrostatic interactions and hydrogen cell nuclear antigen (PCNA) or PCNA-like pro- form processive DNA synthesis. Moreover, we bonds with the nine residues from the thumb, teins are required for high processivity. How- performed alanine scanning of critical resi- palm domains, and Exo of F8 (fig. S11, A, D, ever, for poxviruses, including MPXV and dues responsible for the interaction between and E). There is little contact of F8 with the VACV, no homologous PCNA-like proteins E4 and F8 and found that R39A and N165A base pairs of primer-template DNA, except have been identified in the viral genome. In- substitutions of E4 showed minor effect, W36A for one hydrogen bond interaction between stead, the poxvirus-specific A22-E4 hetero- reduced the synthesis of full-length products, R832 of the thumb domain and the base of dimer is responsible for the high processivity whereas the W36A/R39A and W36A/R39A/ T22 from the primer strand, which may be of DNA replication. N165A substitutions abolished the synthesis important for stabilizing the B-form confor- of full-length products (Fig. 4B). These results mation of the DNA duplex (fig. S11, A and D). A primer-extension assay using a 60-nt tem- further confirmed the important function of the This is consistent with the fact that the enzy- plate DNA in the presence of heparin, which A22-E4 heterodimer in DNA replication pro- matic activity of F8 does not rely on a specific can trap the dissociated DNA polymerase from cessivity in a pure enzymatic reaction system. sequence during the elongation step. In addi- the primer-template DNA to guarantee a single- tion, residue N675 of the fingers domain forms turnover reaction, showed that the F8 polymer- As described above, the E4 cofactor interacts a hydrogen bond with the base of unpaired ase alone dissociated from the primer-template with the Exo of F8 polymerase, and together A12 from the template strand, and this inter- DNA after incorporating less than 14 nt, whereas with the NTD of F8, they form a closed-ring action may be responsible for the kinking of the F8-A22-E4 holoenzyme was able to gener- channel to encircle the single-stranded template the single-stranded 5′ extension of the tem- ate full-length 60-nt products with few abort- DNA (Fig. 4C). By contrast, in the yeast DNA plate strand near the active site. ive ones (Fig. 4A). We then demonstrated that polymerase complex (29), a representative of the addition of the A22-E4 heterodimer con- the other B-family DNA polymerases (fig. S12), Interactions between the polymerase and the ferred processivity to F8 in a concentration- the Exo and NTD form an open semicircular incoming nucleotide dependent manner (Fig. 4A). When the molar Next to the 3′ terminus of the primer strand is the incoming dTTP, which binds to the active site of the polymerase in a manner analogous to that observed in the structures of other DNA polymerase complexes (Fig. 3A) (29). The incom- ing dTTP is accommodated in a groove formed by residues from the palm and fingers domains (Fig. 3B). The two highly conserved aspartate residues, D549 (in motif A) and D753 (in mo- tif C), together with the triphosphate tail of dTTP, coordinate one divalent metal ion (as- sumed as magnesium, which has been added to the reaction buffer) (Fig. 3C). The triphosphate tail also interacts with the main chains of Y550, S552, and L553 from motif A, and the side chains of two positively charged R634 and K661 from the fingers domain (Fig. 3C). The ribose of dTTP stacks on top of the phenyl ring of Y554 from motif A, in a manner similar to that previously observed with Y416 in the ternary complex structure of RB69 poly- merase (30) (Fig. 3C). There would be a steric clash between the 2’OH of ribonucleotides and Y554, hence providing a “steric gating” effect SCIENCE science.org 6 JANUARY 2023 • VOL 379 ISSUE 6627 103

RESEARCH | RESEARCH ARTICLES Fig. 4. The distinctive mechanism of A22-E4 heterodimer promoting pro- own processivity factors and DNA. Polymerases and processivity factors are shown in cessivity of MPXV DNA polymerase. (A) Enzymatic assay of A22-E4 surface representation and colored by domains, as in Fig. 1, whereas the primer- heterodimer protein improving the DNA replication processivity of exonuclease- template DNA strands are shown as a cartoon (template, gray; primer, red). The deficient F8 under single-turnover conditions. The F8 alone was shown to be processivity of yeast polymerase was strengthened by trimeric PCNA to clamp the distributive and failed to generate full-length products, whereas the addition of primer-template dsDNA. While in the structure of MPXV polymerase holoenzyme, the A22-E4 complex promoted the yield of full-length DNA in a concentration- A22-E4 heterodimer does not interact with dsDNA. Instead, E4 located on the dependent manner. (B) Alanine scanning of critical residues of E4 responsible for the template entry channel combined with NTD and Exo domains of F8 to form interaction with F8 was performed to examine the effects on the DNA replication a forward clamp structure that would prevent the template strand disassociating processivity. The A22-E4 W36A mutant reduced the processivity activity, whereas the from the polymerase complex during DNA replication. (E and F) Two binding W36A/R39A and W36A/R39A/N165A mutants abolished the processivity activity. The modes of processivity factors with polymerases. The processivity factors bound F8 used in this assay was exonuclease deficient. (C and D) Comparison of the with template in poxvirus function as a “forward sliding clamp” (E) or dsDNA structures of the MPXV and yeast polymerases (PDB ID: 7KC0) in complex with their products in eukaryotes as a “backward sliding clamp” (F). channel to accommodate the single-stranded “forward sliding clamp” (Fig. 4E); whereas the to be important for high DNA replication template DNA, and the trimeric PCNA ring other B-family DNA polymerase complexes processivity. This processivity mechanism is encircles the template-product DNA duplex possess continuous DNA replication capacity different from that of other B-family DNA (Fig. 4D). This architectural difference be- by encircling the double-stranded template- polymerases that utilize PCNA or PCNA-like tween MPXV and yeast polymerase complexes product DNA helix that can be recognized as proteins to encircle the product-template DNA is responsible for the different processivity a “backward sliding clamp” (Fig. 4F). duplex (29, 34, 35). The configuration of encir- mechanisms during DNA replication events. cling the single-stranded template DNA prob- The MPXV DNA polymerase holoenzyme guar- Discussion ably allows the MPXV polymerase complex antees its high DNA replication processivity to perform continuous DNA replication by by encircling the single-stranded template The interaction between E4 and F8 could gen- preventing template DNA disassociation from DNA, and we propose that it functions as a erate a ring channel that encircles the single- the polymerase holoenzyme. The efficiency stranded template DNA, which is proposed 104 6 JANUARY 2023 • VOL 379 ISSUE 6627 science.org SCIENCE

RESEARCH | RESEARCH ARTICLES of polymerase holoenzyme assembly in the cel- studied. Moreover, the working mechanism of 33. G. Gao, M. Orlova, M. M. Georgiadis, W. A. Hendrickson, lular environment requires further study, and S. P. Goff, Proc. Natl. Acad. Sci. U.S.A. 94, 407–411 (1997). moreover, it should be investigated whether an intact replisome, including the F8-A22-E4 the F8-A22-E4 polymerase holoenzyme has 34. Y. Shamoo, T. A. Steitz, Cell 99, 155–166 (1999). other active conformations (28). holoenzyme and other replication proteins, 35. C. Lancey et al., Nat. Commun. 11, 1109 (2020). 36. D. B. Gammon, D. H. Evans, J. Virol. 83, 4236–4250 (2009). Previous studies have implicated the distinc- will be a fascinating area for future studies. 37. M. D. Hamilton, A. A. Nuara, D. B. Gammon, R. M. Buller, tive role of the polymerase holoenzyme in DNA recombination, which involves both 3′-5′ exo- REFERENCES AND NOTES D. H. Evans, Nucleic Acids Res. 35, 143–151 (2007). nuclease and DNA-joining activities (36–38). 38. M. D. Hamilton, D. H. Evans, Nucleic Acids Res. 33, 2259–2268 Although we identified DNA ligase-like do- 1. H. Zhao et al., China CDC Wkly. 4, 853–854 (2022). mains for the Mid of the poxvirus-specific 2. J. G. Rizk, G. Lippi, B. M. Henry, D. N. Forthal, Y. Rizk, Drugs (2005). cofactor A22, the F8-A22-E4 polymerase holo- 39. S. M. Kerr et al., EMBO J. 10, 4343–4350 (1991). enzyme did not have canonical ligase activity 82, 957–963 (2022). 40. O. Faye et al., Lancet Infect. Dis. 18, 246 (2018). in our assay. However, we cannot rule out the 3. J. Kaler, A. Hussain, G. Flores, S. Kheiri, D. Desrosiers, Cureus 41. N. Sklenovská, in Animal-Origin Viral Zoonoses, Y. S. Malik, possibility that A22 possesses ligase activity in a different state from the current holoenzyme 14, e26531 (2022). R. K. Singh, K. Dhama, Eds. (Springer Singapore, Singapore, structure or in an ATP-independent manner. 4. D. Focosi, F. Novazzi, A. Baj, F. Maggi, Rev. Med. Virol. 32, e2392 2020), pp. 39-68. Orthopoxviruses encode a DNA ligase that is 42. M. Esteban, J. A. Holowczak, Virology 78, 57–75 (1977). not essential for virus replication but affects (2022). virulence and sensitivity to DNA-damaging 5. D. B. Di Giulio, P. B. Eckburg, Lancet Infect. Dis. 4, 15–25 ACKNOWLEDGMENTS agents (39). Further functional dissection of the middle part of A22 and its cooperation (2004). We thank X.L., Y.M., and X.B. at the cryo-EM Center, Shanxi Academy of with viral and host ligases is needed. 6. E. M. Bunge et al., PLOS Negl. Trop. Dis. 16, e0010141 Advanced Research and Innovation, for their technical support of the cryo-EM data collection. Funding: National Key Research and Poxvirus replication processes use different (2022). Development Program of China grant 2021YFC2300700 (Y.S.); replication models, including self-priming, 7. W. Tan, G. F. Gao, China CDC Wkly. 4, 847–848 (2022). Strategic Priority Research Program of CAS grant XDB29010000 (Y.S., primer-dependent, and recombination models 8. N. Erez et al., Emerg. Infect. Dis. 25, 980–983 (2019). G.F.G.); National Natural Science Foundation of China (NSFC) grants (15). Poxviruses, including MPXV, have linear 9. A. Vaughan et al., Euro Surveill. 23, 1800509 (2018). 82241076, 81871658, 32192452, and 32100119 (Y.S., Q.P.); The double-stranded DNA genomes, and the ter- 10. N. Chen et al., Virology 340, 46–63 (2005). Youth Innovation Promotion Association of CAS grant Y201921 (Y.S.). mini of the two DNA strands are connected to 11. A. M. McCollum, I. K. Damon, Clin. Infect. Dis. 58, 260–267 Author contributions: Conceptualization: Y.S., G.F.G., Q.P. form a continuous polynucleotide chain (40). Investigation: L.K., Q.P., Y.X., J.Q., H.W. Visualization: Q.P., Y.X., L.K. A rolling cycle mechanism has been proposed (2014). Funding acquisition: Y.S., G.F.G., Q.P. Supervision: Y.S., G.F.G. Writing – to replicate DNA in the form of unbranched 12. J. Isidro et al., Nat. Med. 28, 1569–1572 (2022). original draft: Y.S., Q.P., G.F.G. Writing – review and editing: Y.S., Q.P., head-to-tail concatemers, which would be re- 13. B. Moss, Semin. Cell Dev. Biol. 60, 89–96 (2016). G.F.G. Competing interests: The authors declare that they have no solved by a Holliday junction resolvase to pro- 14. M. D. Greseth, P. Traktman, Annu. Rev. Virol. 9, 239–259 competing interests. Data and materials availability: Cryo-EM map is duce unit genomes (41). The proposed “forward available in the Electron Microscopy Data Bank with code EMD-34731. sliding clamp” mode can help interpret the (2022). Structural model is available in the Protein Data Bank (PDB) with self-priming model, as the processivity cofac- 15. B. Moss, Cold Spring Harb. Perspect. Biol. 5, a010199 accession code 8HG1. All materials are available from the authors on tor would facilitate robust continuous replica- reasonable request with materials transfer agreements (MTAs). tion along the single-stranded template DNA (2013). License information: Copyright © 2023 the authors, some rights unwound by the primase-helicase. Because 16. N. Tarbouriech et al., Nat. Commun. 8, 1455 (2017). reserved; exclusive licensee American Association for the Advancement there is also a possible presence of Okazaki 17. E. S. Stanitsa, L. Arps, P. Traktman, J. Biol. Chem. 281, of Science. No claim to original US government works. https://www. fragments for Orthopoxvirus DNA replication science.org/about/science-licenses-journal-article-reuse. In the interest (42), other replication models should also be 3439–3451 (2006). of rapid dissemination of results with immediate public health 18. A. M. Druck Shudofsky, J. E. Silverman, D. Chattopadhyay, relevance, the author will make the Author Accepted Manuscript (AAM) version available under a CC BY public copyright license. R. P. Ricciardi, J. Virol. 84, 12325–12335 (2010). 19. M. W. Czarnecki, P. Traktman, Virus Res. 234, 193–206 SUPPLEMENTARY MATERIALS (2017). science.org/doi/10.1126/science.ade6360 20. K. A. Boyle, E. S. Stanitsa, M. D. Greseth, J. K. Lindgren, Materials and Methods Figs. S1 to S12 P. Traktman, J. Biol. Chem. 286, 24702–24713 (2011). Tables S1 and S2 21. C. Contesto-Richefeu et al., Acta Crystallogr. F Struct. Biol. Commun. References (43–53) MDAR Reproducibility Checklist 72, 687–691 (2016). 22. W. P. Burmeister et al., J. Biol. Chem. 290, 17923–17934 (2015). Submitted 29 August 2022; accepted 5 December 2022 23. C. Contesto-Richefeu et al., PLOS Pathog. 10, e1003978 (2014). Published online 15 December 2022 24. B. Bersch, N. Tarbouriech, W. P. Burmeister, F. Iseni, J. Mol. Biol. 10.1126/science.ade6360 433, 167009 (2021). 25. N. Schormann et al., BMC Struct. Biol. 7, 45 (2007). 26. K. Shi et al., Nucleic Acids Res. 46, 10474–10488 (2018). 27. Y. Chen et al., Nat. Commun. 10, 387 (2019). 28. C. Sèle et al., J. Virol. 87, 1679–1689 (2013). 29. F. Zheng, R. E. Georgescu, H. Li, M. E. O’Donnell, Proc. Natl. Acad. Sci. U.S.A. 117, 30344–30353 (2020). 30. M. C. Franklin, J. Wang, T. A. Steitz, Cell 105, 657–667 (2001). 31. A. P. Silverman, S. J. Garforth, V. R. Prasad, E. T. Kool, Biochemistry 47, 4800–4807 (2008). 32. G. J. Klarmann et al., Biochemistry 46, 2118–2126 (2007). SCIENCE science.org 6 JANUARY 2023 • VOL 379 ISSUE 6627 105

LIFE SCIENCE TECHNOLOGIES Produced by the Science/AAAS Custom Publishing Office new products Photochemical Reactor System - Asynt announces the LightSyn Lighthouseñ ment challenges encountered during drug development programs a new photochemical reactor system in their LightSyn range, developed in re- generated through well-characterized kit reagents and standards, sponse to customer demand for higher re- while providing a high sensitivity and wide dynamic range to cover action yields, ease of use, high operational the full spectrum of cytokine levels seen in disease states or phar- safety, and competitive pricing. The Light- Syn Lighthouse uses new technology to - channel light through a quartz rod directly into the reaction medium, maximizing power intensity while keep- round times, and accelerate data-driven decisions. Gyros Protein Technologies to that of a more typical photoreactor, therefore maximizing the For info: +46-0-18-56-63-00 lightís interaction with the sample. The result is faster chemistry www.gyrosproteintechnologies.com - GMP-grade Human AB Serum for Cell and Gene Therapies istry, the LightSyn Lighthouse is simple to set up and to use, with serum, a cell culture supplement to enhance the development and for scaling your experiments. - Asynt For info: +44-0-1638-781-709 vitamins, nutrients, trace elements, and transport factors found in an - 3í3í-cGAMP ELISA Kit ditions, it allows many primary human cells to grow at a faster rate. and mammalian cell lysates and cell supernatants. The assay uses This is particularly important for autologous therapies that require using a personís own cells or tissues, which may divide at a slower produced using the same optimized protocols and enhanced regula- tory oversight of source material collection, production, and process- - colorimetric readout. BioIVT Cayman Chemical For info: 516-483-1196 For info: +1-800-364-9897 COVID-19 Omicron Research Products Reagent Reservoirs - metric, and chemiluminescence. These new kits are ideal for use in -- pipetting systems. genes that provide convenient readout of transduction. Also new to Porvair Sciences For info: +1-856-696-3605 - Human Cytokine Kit AMS Biotechnology For info: +1-617-945-5033 Electronically submit your new product description or product literature information! Go to www.science.org/about/new-products-section for more information. Newly offered instrumentation, apparatus, and laboratory materials of interest to researchers in all disciplines in academic, industrial, and governmental organizations are featured in this space. Emphasis is given to purpose, chief characteristics, and availability of products and materials. Endorsement by Science or AAAS of any products or materials mentioned is not implied. Additional information may be obtained from the manufacturer or supplier. 106 6 JANUARY 2023 • VOL 379 ISSUE 6627 science.org/custom-publishing SCIENCE