Set-up for super-registration microscopy

Super-registration refers to the ability to generate an internal registration signal from the sample, for example, each cell imaged, that can be used to register spectrally different channels relative to each other to achieve spatial precision below the optical resolution limit. Image series were acquired on a customized dual channel set-up (see Supplementary Fig. 1c) using an Olympus 150× 1.45 N.A. oil immersion objective lens. The right side port of an IX 71 (Olympus) was modified by removing the tube lens. Outside of the stand we placed a 514.5-nm notch filter (Semrock), a 300-mm focal length lens, followed by a 568-nm notch filter (Semrock) that was rotated by 17 degrees to the normal to achieve blocking of 561-nm scattered light. The effective magnification of the optical system was 250× resulting in a pixel size of 64 nm. A dichroic mirror (z543rdc, Chroma) was used to split the fluorescence onto two EMCCDs (Andor iXon, Model DU897 BI). A combination of mirrors and CCD supports (x,y,z, ϕ- and θ-angle) was used to physically pre-align both CCDs to optimize super-registration after image processing. A resolution standard (Gellermicro), focal check beads (Invitrogen) and diffraction-limited multi-colour beads (Invitrogen) were used for pre-alignment. Using two cameras it is possible to adjust their focal plane to account for small axial chromatic shifts. Super-registration is achieved by the combination of precise mechanical alignment and image processing using transformations based on the registration signal that is detected on both cameras. CCDs were synchronized by a start signal generated by one CCD that was directly delivered to the second CCD. The offset between the two CCDs was determined to be three orders of magnitude below our integration time (2.1 ± 0.2 ns per frame per ms). For excitation of fluorescent proteins an Argon laser with 514.5-nm emission (Melles Griot) and a 561-nm laser line (Cobolt) were merged into a mono mode optical fibre (Qioptiq). The output of the fibre was collimated and delivered through the back port of the IX71 stand and reflected towards the objective by a dichroic mirror (z514-561-1064rpc, Chroma). Alignment onto the optical axis of the objective was achieved with a 4-axis controlled support for the collimator. A size-adjustable iris was used to restrict the illumination to an area of approximately 25 µm in diameter. The intensity profile in this area had a flatness of about 5%. Each laser had a shutter (Uniblitz) that was controlled from the imaging software. To allow reasonably fast switching (100 ms) between high- and low-power settings with the 561-nm line, a motorized filter wheel with appropriate neutral density filters was placed behind the shutter but before the merging dichroic of the laser module. The microscope was equipped with a heated stage inset (Warner Scientific) and an objective heater (Bioptechs). During the experiment the stage was covered by a 100-mm cell culture dish wrapped with aluminium foil to exclude stray light. Heating devices were run overnight before an experiment. One hour before an experiment three small dishes with a few millilitres of water were placed on the stage inset to provide humidity. Cells were imaged in a closed dish.

Image acquisition

Simultaneous imaging of nuclear pores and mRNA enabled a relative measurement of distances (drifts are accounted for by the tracking of both entities) and hence overcomes a limitation in earlier work on imaging nucleocytoplasmic transport, namely missing information on the exact position of the nearest nuclear pore during the acquisition of the cargo signal. To achieve this goal with both sufficient spatial and temporal resolution, EMCCDs, laser shutters and the filter wheel were controlled from the camera software using customized scripts written in Andor Basic. Using sub-frames (∼2/3 of each chip, 330 × 330 pixel) on both cameras we were able to observe whole nuclei at a frame rate of 50 Hz equalling a time resolution of 20 ms for tracking single mRNAs. The effective integration time was 19.92 ms. A frame rate of 20 ms was chosen to gain sufficient tracking resolution. Test experiments at 50-ms frame rates showed blurring of mRNA signals whereas 20-ms frame rates offered adequate signal accumulation to ‘freeze’ the RNAs with a positional accuracy sufficient for tracking (see section on image processing and equation (2)). To generate the super-registration signal used for post experimental, computational fine alignment of the two detectors the following imaging protocol was implemented: potential cells of interest were selected and brought into focus (equatorial plane) at very low power settings (0.5 W cm−2) in the red channel using maximal gain on the camera, by avoiding excitation at 514.5 nm bleaching in the green channel was minimized. Next, an automated protocol was used to image NPCs only at 561-nm laser using ‘high’ power setting (180 W cm−2) for 50 frames, followed by a 100-ms break to save data, switch gain settings and filter wheel position, followed by 400 frames with both laser lines (514.5 nm used at 15 W cm−2, 561 nm used at 18 W cm−2). While the green channel CCD was used with ×1,000 gain during both imaging cycles the gain on the red channel CCD was adjusted between 450 for the first cycle and ×1,000 for the second cycle. The first imaging cycle generated a detectable signal from the NPC staining on both cameras, due to surface reflection on the dichroic mirror between the cameras. The front surface reflection was more pronounced than the back surface reflection and could be detected well enough to use an average time projection of the 50 images collected in the first imaging cycle as a reference for image alignment (Supplementary Fig. 1). Power measurements were done using an objective power meter (Carpe). Stage drifts during data acquisition were minimal and as the nuclear pores and the mRNA were imaged simultaneously no extrinsic drift control was needed.

Image processing

The image information of the mRNA and NPC signals needed to be fine-registered post experimentally. For each cell imaged, two data sets per channel were collected as described above. The first set contained signal from the nuclear pore label, POM121 fused to tandem Tomato, which was recorded on both cameras. Time projection of the average signal yielded an image that identified single NPCs. Original image stacks were divided into two sub-stacks with only half the area but still retained the same number of images to achieve better registration because of non-monotonic distortion over the field of view. Time-projected images from both cameras were registered using ‘projective’ transformation in MatLab. The individual transformation matrixes were applied to the second movie from the red channel of each data set to overlay NPCs with the mRNA signal. The signal of the NPC label in the second movie was much lower due to bleaching during the recording of the registration data. To improve the signal-to-noise ratio a sliding average of 15–25 frames was calculated for the second movie and used to fit the NPC positions during the experiment. This averaging resulted in a reduced time resolution for the NPC signal. As nuclear pores are relatively immobile at least 6 nuclear pores per cell from 15 different cells were tracked for at least 150 frames in these averaged movies to estimate the localization precision of our nuclear pore signal. On the basis of the mean error of the localized position of these NPCs we achieved 15-nm localization precision. This value is an underestimation, as cellular movement will contribute to the error source for localization over this time range. The drift of an average NPC was 1.1 ± 0.2 nm between subsequent frames (20-ms integration time).

The image registration precision was tested by fitting NPC positions on the green channel registration data set and the registered red channel data set for nine cells. The resulting registration precision was better than 10 nm (see Fig. 1 and following text). Determination of the absolute co-localization precision in living cells by our method is limited by the available signal in the green channel. As photons contributing to this image are reflected off the glass surface of a dichroic that is designed to transmit light at this wavelength, the signal-to-noise ratio in the green channel is clearly worse than in the red channel. Compensation could be reached by longer imaging at high laser intensities, but at the cost of losing the capability to track nuclear pores during acquisition of export movies in the green channel. The applied transformation matrix is based on four pores that have been identified in both images. We therefore tested our co-registration precision by calculating the distance between 6, 10 and 15 nuclear pores in both images for a total of 21 registered nuclei from two of three experimental sets (a total of 33 cells; see Fig. 1). Each registered image series contained an expected number of 40 to 60 nuclear pores, depending on the size of the nucleus. On the basis of the differences in signal-to-noise ratio between the two registration images we argue that 10 nuclear pores are a fair sub-sample to estimate registration precision, leading to a registration precision of 8 ± 1 nm. Six pores might be too few as the number is almost identical with the number of pores used for super-registration, whereas 20 pores would introduce a co-registration uncertainty that would be largely determined by the signal-to-noise ratio of nuclear pores imaged in the green channel. As can be seen in Fig. 1 the resulting registration precision is 10 ± 1 nm if we apply a 15 pore criterion. NPC and mRNA signals were evaluated by Gaussian fitting. Although the localization precision for nuclear pores could be determined experimentally within our data sets to be 15 nm, the localization precision for our mRNA signal was estimated from the number of detected photons and the FWHM of the Gaussian fit by equation (1)31: The number of photons (N) was calculated from the counts detected by the camera and reported by the fitting routine using the manufacturer’s calibration data for each camera, taking into account the electron multiplying gain, electrons generated per A/D count, quantum efficiency of the CCD and the energy of a photon at the centre emission wavelength. The factor s is the standard deviation of the Gaussian approximation of the point-spread function. It is determined by fitting a steady signal repeatedly and calculating the distances between identical positions in different frames. Our mRNA is moving and hence we need to estimate this value for use in equation (1). One consequence of an inherent mobility of the signal is that it will spread and be less bright than an immobilized equally labelled sample. We used the following assumption: a signal that can be fitted has to have one brightest pixel. The brightest pixel will be a lower approximation for the true position of the mRNA. Hence s can be approximated as a. The pixel size a was 64 nm, and the background b was estimated from our data sets. The resulting localization precision for our mRNA signal was 19 nm. The co-localization precision between the NPC and mRNA signal is given by equation (2): The precision of mRNA signal is σ mRNA = 19 nm, nuclear pores are localized with σ NPC = 15 nm and the registration between the channels is σ registration = 10 nm. The overall co-localization precision that equals our achieved ‘super-registration’ is calculated to be 26 nm. All our numbers for registration precision between cameras, localization of mRNAs and nuclear pores are the average of our data. Although such an average is a reliable and well defined measure, we argue that such a number might be of limited relevance for the biological problem. In detail, the observed kinetics of transient interactions in living cells would be heavily biased if traces would be cut short because in individual frames during the total interaction time the signal of one of the observed entities drops below the threshold value for registration precision. Accordingly, selection of data points based on the localization precision, as used in single-molecule-based super resolution techniques, is not an option for tracking in living cells. The data presented here present a breakthrough in spectrally resolved super-registration microscopy as they are mostly limited by the detection precision of the mRNA signal, not the pore signal or the channel registration precision. Gaussian fits were preformed with two routines. One routine included automated particle identification and nearest neighbour tracking as described previously31. The other routine was analogous to that of ref. 32 but implemented in a semi-automated way. Upon ‘clicking’ of a signal the brightest spot in a ten-pixel environment is found and a centre of mass algorithm delivers the start point for the Gaussian fit. A number of control checks were used to validate the fit. All fit parameters are immediately reported to the user to allow direct appreciation of the fit. A graphical help was also implemented to disallow for confusion of particles. This routine was used to fit all signals within a 10–15 pixel distance of the nuclear envelope. This allowed visual identification of signals and manual tracking. As the focal thickness of our observation volume was small, owing to the high N.A. of the objective, manual tracking allowed better control of ‘blinking’ events. Both routines used raw data to perform the fitting. Localization precisions are based on fits performed according to ref. 31.

Cells

Immortalized mouse embryo fibroblast cells (MEFs) from a homogeneous transgenic knock-in mouse for β-actin-24-MBS were infected with a lentivirus coding for NLS-MCP–YFP protein. The mouse develops normally having all β-actin transcripts tagged with the 24× MBS repeats. This stable cell line was FACS sorted for low expression levels of NLS-MCP–YFP and infected with a lentivirus coding for POM121–tandem-Tomato (POM121–tdTomato). Cells were FACS sorted for double-positive signals in the green and red channels. Successive FACS analysis was used to separate cells with homogeneous NLS-MCP–YFP and POM121–tdTomato expression. Growth curves of the immortalized MEFs, MEFs derived from the β-actin 24 MBS mouse, β-actin MEFs with either NLS-MCP–YFP or MCP–GFP expression and β-actin MEFs with additional POM121–tdTomato expression were collected (Supplementary Fig. 2). Cells were seeded at 3,000 cells ml−1 density in 60-mm dishes. A total of 30 dishes for each cell line were seeded and up to four dishes a day were harvested and counted. A haemacytometer (Fisher) was used for counting, and at least four samples from each dish were counted. All five cell lines grew with the same doubling times (Supplementary Table 1), indicating that neither the MCP label for the RNA nor the POM121 label for the NPC have major effects on cellular metabolism. The labelling ratio of NPCs is discussed in more detail in Supplementary Fig. 4.