Technical Comments

Response to Comment on “Extended-resolution structured illumination imaging of endocytic and cytoskeletal dynamics”

See allHide authors and affiliations

Science  29 Apr 2016:
Vol. 352, Issue 6285, pp. 527
DOI: 10.1126/science.aad8396

Abstract

Sahl et al. in their Comment raise criticisms of our work that fall into three classes: image artifacts, resolution criteria, and comparative performance on live cells. We explore each of these in turn.

Sahl et al. (1) highlight a subregion of figure 3A from Li et al. (2) (reproduced here in Fig. 1A) that exhibits periodic image reconstruction artifacts (Fig. 1B and Sahl figure 1A, B, G, and I), and argue that these hamper “the interpretation of details.” Although one always needs to be vigilant to minimize such artifacts during structured illumination microscopy (SIM) reconstruction and identify them in the final image when they occur, their selected subregion represents only 1.7% of the original image area, the remainder over which such artifacts are not notable (e.g., Fig. 1C). Furthermore, these are not “difficult or impossible to disentangle from real features.” Indeed, Sahl et al. correctly note that they arise at the discrete spatial frequencies corresponding to the harmonics of the applied activation/excitation pattern. They are incorrect, however, that their appearance is intrinsic in SIM due to “the peaked shape of the (S)PA NL-SIM OTF.” Instead, they are due to strong frequency-shifted signals arising from greater out-of-focus background in the affected region than elsewhere (compare insets in Fig. 1, B and C). As such, they can be eliminated by reducing such background. They are also easily suppressed to better reveal the underlying structure (Fig. 1D) by applying notch filters confined to just these frequencies (3), rather than by the unnecessarily aggressive approach of removing all information content beyond the diffraction limit (Sahl et al. figure 1, H to K). Sahl et al. also argue that such artifacts preclude the use of our methods to study periodic structures. However, truly periodic structures are rare in cell biology and can be distinguished from reconstruction artifacts in SIM, in that only the latter will rotate if the excitation pattern is itself rotated. In short, reconstruction artifacts affect only a small fraction of the image area in more than 30 data sets of live cell dynamics in Li et al. and its supplementary materials (SM), can be suppressed when necessary, and rarely stand in the way of revealing new biological phenomena.

Fig. 1 Issues relevant to live-cell SR microscopy.

(A) PA NL-SIM image of actin filaments in a COS-7 cell, from a movie of 30 such images. (B) Subregion highlighted by a red box in (A) and used by Sahl et al. to show periodic reconstruction artifacts. (C) Another subregion denoted by a green box in (A), without such artifacts. (D) Suppression of artifacts in (B) without loss of resolution by selective notch filtering of spurious spatial frequencies. Widefield insets in (B) and (C) show higher out-of-focus background in (B). (E) A 54-nm-wide linecut through the theoretical PSF of a PS-RESOLFT microscope. (F) Simulated resolvability of a test pattern of alternating large and small squares, separated 80 nm edge-to-edge, by PA NL-SIM and PS-RESOLFT at SNR = 20. (G) Simulated PA NL-SIM and WF-RESOLFT images at SNR = 20 of a 50-nm-wide line flanked by 250-nm-wide lines for three different line gaps. (H) Spurious linewidth variability in WF-RESOLFT for different degrees of background subtraction, 200-nm line gap case. (I) Frequency spectra from images of microtubules in macrophages obtained by various SR methods. [Reprinted with permission from (5).] (J) PA NL-SIM image of actin filaments in another cell, from a movie of 30 image frames. (K) Oblique illumination images of the same cell at times relative to the PA NL-SIM imaging. Each image represents one frame from a series of 1000 taken at 0.1-s intervals at each inspection time to assess any changes to cell morphology or dynamics.

Concerning resolution, the assertion by Sahl et al. that our claims are “lacking any demonstration in real space” is erroneous. For example, Li et al. figure 3C shows a clear and quantifiable progression of increased resolution in a field of caveolae successively imaged by diffraction-limited total internal reflection fluorescence (TIRF), linear SIM, patterned-activated non-linear SIM (PA NL-SIM), and saturated PA NL-SIM. Furthermore, Li et al.’s SM figure S37 contains linecuts through caveolar rings of various sizes that confirm experimental resolution at the theoretical limit of each method. Nevertheless, we are loath to use linewidths of isolated linear structures, as is common in the super-resolution (SR) literature, as a metric of resolution. Indeed, we argue that this metric is incomplete and often overstates the practical resolution. To illustrate, the predicted point spread function (PSF) of a reversible saturable/switchable optical (fluorescence) transitions (RESOLFT) microscope under reported conditions (Li ref. 10) of live-cell imaging has a full width at half maximum of 54 nm (Fig. 1E), yet in simulations (Fig. 1F) is unable, unlike PA NL-SIM, to resolve 25-nm-wide squares separated 80 nm edge-to-edge from adjacent 125-nm-wide squares. This is because the weak signal generated by the RESOLFT maximum when over a small square is masked by the strong background from adjacent large squares excited by tails of the PSF (red in Fig. 1E) that contain 76% of the total energy.

Moreover, measured linecuts often inaccurately report the true PSF. In parallelized, widefield (WF) RESOLFT (Li et al. ref. 11), for example, “background” is reduced by ad hoc subtraction, from each signal maximum, of 80% of the values at the neighboring points of greatest depletion, and enforcing a nonnegative result. Neither the spatial shift of information nor the nonnegativity constraint is a linear operation. Consequently, simulated WF-RESOLFT images show that the apparent width of a 50-nm-wide line flanked by 250-nm-wide lines varies along its length, as well as with the distance to the flanking lines (Fig. 1G) and the magnitude of “background” subtracted (Fig. 1H). Any “resolution” can be claimed by tuning these parameters. In addition, defining resolution by the narrowest observed linewidths, as is common in the SR field, risks self-selecting for artifacts, as well as for features artificially narrowed by noise. As a result, WF-RESOLFT and point scanning (PS) RESOLFT images from the literature of keratin (Li et al. ref. 11) and caveolin (Li et al. ref. 33), having linewidth-based resolution estimates of 80 nm and <50 nm respectively, appear more similar in resolution across their fields of view to diffraction-limited, deconvolved TIRF images of similar structures (Li et al. SM figures S6 and S34) than to PA NL-SIM or even linear SIM images of such features.

Finally, defining the resolution of a microscope by a single number is both incomplete and naïve. Every microscope acts as a filter that increasingly attenuates information associated with increasingly fine features (Li et al. SM note 1). Therefore, every raw image is a distorted map of the specimen that emphasizes large structures while blurring small ones. Deconvolution (4) corrects for this distortion but requires knowledge of the entire spectrum of attenuation in Fourier space [known as the optical transfer function (OTF)]. A practical resolution limit is reached at the spatial frequency where the OTF falls below the experimental noise floor (e.g., Li et al. SM figure S52). Therefore, in live-cell time-lapse SR imaging, where the potential for photobleaching and phototoxicity demands low signal-to-noise ratio (SNR) operation, how rapidly the OTF rolls off toward zero is even more important than where it reaches zero (i.e., the theoretical resolution limit). Experimental measurements of microtubules in macrophages (5) (Fig. 1I) show that the frequency spectrum of linear SIM data is larger than for stimulated emission depletion (STED) in the regime between 200 and 100 nm, and our calculated OTFs (Li et al. SM note 2 and figure S24) indicate that the same holds true for PA NL-SIM and saturated PA NL-SIM below 100 nm. Thus, the SIM-based methods are capable of achieving higher practical resolution than STED and RESOLFT for a given SNR or require far fewer emitted photons to achieve the same practical resolution (Li et al. SM note 5 and figures S51 and S54).

Clearly, the issue of resolution is a complex one. A full description requires the measured OTF, the SNR across each image, and the imaging conditions (as in Li et al. SM table S1) giving rise to this SNR. Even then, the noise, and hence the practical resolution, depends on the brightness, background level, and sparsity of the specimen—the latter because the image of a sparse specimen is weighted more strongly to high spatial frequencies that can rise above the noise floor (Li et al. SM note 5 and figure S55). This may help explain why many examples of live-cell STED and RESOLFT are of sparse, exceptionally bright filamentous or punctate objects and may not achieve the same resolution, speed, imaging duration, and noninvasiveness that we demonstrate on dense networks of dim, individual actin filaments (Li et al. figures 3 and 4), clathrin or caveoli clusters (Li et al. figures 2F and 3F), striated focal contacts (Li et al. figure 1, C to F), and membrane ruffles (Li et al. figure 4, D to F).

An alternative to such a complete description is simply to compare images of similar features as imaged by different methods. Sahl et al. argue that such comparisons are subjective and “inadequate,” yet exactly this approach is commonly used to compare STED or RESOLFT to diffraction-limited images (e.g., Sahl et al. refs. 8–11 and 16 and Li et al. ref. 11). We see no problem in such comparisons, provided that (i) the performance of each method is fairly presented; (ii) OTF, SNR, and imaging condition data are given; (iii) all image manipulations are described fully; and (iv) all previous biological knowledge of the specimen is used to identify possible artifacts. Supreme Court Justice Potter Stewart famously stated in a case on the limits of free speech that he couldn’t define hardcore pornography, but “I know it when I see it.” We feel similarly about resolution. We invite readers to review our comparisons of high NA SIM and PA NL-SIM to other SR methods (Li et al. SM figures. S4, S6, S13, S31, S34, and S44) and draw their own conclusions.

With regard to live-cell imaging , there exist inevitable trade-offs between spatial resolution, imaging speed, and phototoxicity. Yet, in cases where direct experimental comparisons are available, high NA SIM and PA NL-SIM outperform RESOLFT and single-molecule localization microscopy (SMLM) by nearly every metric relevant to live cell imaging simultaneously, often by one or two orders of magnitude (Table 1). These gains are consistent with our theoretical estimates (Li et al. SM notes 2–7 and figures S28, S45, S51, and S54) and underscore again how a strong OTF (Li et al. SM figure S24) in the SR regime is essential for optimal live-cell performance. They are also indicative of the penalties imposed by saturated depletion, which is at the heart of STED and RESOLFT, where fewer and fewer fluorescence state transitions produce a useful signal as higher and higher resolution is demanded. Thus, although Sahl et al. argue that “STED nanoscopy…has been used to record various structures quickly and repeatedly in living cells and…neurons,” their cited example (Sahl et al. ref. 15) (i) does not report acquisition times; (ii) uses a peak intensity (~600 MW/cm2) 20-fold higher than that shown to induce photodamage in spermatozoa (6) and Escherichia coli (7) in optical trapping experiments at similar wavelengths; (iii) resolves a structure barely below the diffraction limit; and (iv) provides no examples of time-lapse imaging of subcellular dynamics—arguably the main point of live-cell imaging.

Table 1 Comparative live-cell imaging performance of high NA TIRF SIM and PA NL-SIM to other SR imaging modalities.

Colored groups represent comparisons for similar subcellular features. Parameters of competing SR methods are normalized to results of Li et al. and are shown in green or red if better or worse, respectively. Total dose across all time-lapse image frames (bottom section) is normalized to wavelength-dependent phototoxic levels estimated from (8) as follows: 1.0 kJ/cm2 at 405 nm; 5.0 kJ/cm2 at 488 nm; and 100 kJ/cm2 at 561 nm.

View this table:

Even when multiple frames are shown, as in the example of “dendritic spines…followed over hours by RESOLFT” (Sahl et al. ref. 16), the small field of view (7.7 × 16.3 μm), prolonged acquisition time (348 s), long imaging interval (900 s), and limited number of frames (9) preclude study of dynamical processes at speeds and durations common in living cells. In fact, time-lapse images by STED and RESOLFT are often notable for how little motion is observed, given the long acquisition times. This may be a harbinger of phototoxicity or even light-induced fixation that produces “frozen” cells (8). It is also notable that independent metrics of cell health, such as differential interference contrast (Li et al. ref. 6) or oblique illumination imaging (Fig. 1K) before, during, and after SR imaging (Fig. 1J) to check for changes in cell morphology or dynamics, are largely absent from the STED and RESOLFT literature. This is perhaps not surprising, given that these methods often exceed expected levels of toxic exposure (Table 1 and Li et al. SM movie S1).

Viability metrics are essential to have any confidence in the validity in live SR findings. Even so, more sophisticated measures are needed. For example, microtubule growth rates have been shown to decline significantly at a dose of 640-nm irradiation, only 3% of that at which nearly all cells still divide (8). Thus, the development of noninvasive, broadly applicable assays to measure the immediate and local effects of light on cellular physiology should be a major thrust of live-cell imaging research, for diffraction-limited as well as SR methods. Until then, every effort should be made to minimize both the light dose and peak intensity and to restrict the illumination in time and space. We view the low-intensity, short-exposure, TIRF-confined methods of high NA SIM and PA NL-SIM that we have developed to be important steps in this regard.

Space restrictions preclude us from addressing all our points of disagreement with Sahl et al. However, researchers considering building or purchasing an SR microscope for their own work are well advised to insist on demonstration of the practical resolution, field of view, and imaging speed on their own specimens, and to evaluate the extent to which photobleaching and phototoxicity are debilitating, before making a decision.

In the final analysis, the only metric that matters for a new imaging technology is whether it can address important questions to the satisfaction of the biological community that could not be answered by previous methods. In our case, we studied: the relationship between the sizes and lifetimes of clathrin-coated pits (CCPs), the role of actin in clathrin-mediated endocytosis, the dynamics at clathrin patches and hot spots of CCP formation, the formation of nanoscale actin rings, the diversity of caveolin-rich structures, and the nanoscale recruitment of α-actinin during the formation of actin-rich structures. Elsewhere, in the past year or two live-cell SIM has been used to investigate the dynamics of MreB filaments in bacteria (9) and cortical microtubules (10), microtubule-dependent transport of vimentin filaments (11), vesicle formation at the inner nuclear membrane (12), centrosome assembly in Drosophila (13), a dorsal actomyosin network that drives cell motility (14), and the assembly of isoforms of nonmuscle myosin II (15) and combinations of myosin II and myosin 18 (16) into bipolar filaments. We look forward to revisiting our debate with Sahl et al. a decade from now to see which SR methods have had the broadest effect in biology.

References

View Abstract

Navigate This Article