Testing the lamella hypothesis: the next steps on the agenda

Introduction

In the accompanying Speckle Microscopy Update (Vallotton and Small, 2009), the authors present a critique of the `lamella hypothesis' – the proposal that the lamella extends to the leading edge of motile cells and overlaps with the lamellipodium (Ponti et al., 2004). In the 2004 paper, fluorescent speckle microscopy (FSM) combined with drug-induced perturbations of actin-filament organization were used to substantiate the lamella hypothesis. Vallotton and Small now question the interpretation of this FSM data. They suggest that the lamella and lamellipodium should be considered as spatially distinct. As one of the corresponding authors of the 2004 article, I have been asked to provide a response to the critique by Vallotton and Small, and this is included below. References to figures point to the data presented in Speckle Microscopy Update I (Vallotton and Small, 2009), unless stated otherwise.

EM data and the lamella hypothesis

Scepticism about the lamella hypothesis is understandable. It remains difficult to directly match up the notion of two partially colocalized actin-filament networks that have different dynamic and molecular properties (the existence of which is proposed by the lamella hypothesis) with existing electron microscopy (EM) images, which are the gold standard for ultrastructural analyses. Some of the available EM data (Delorme et al., 2007; Gupton et al., 2005) hint at the presence of two networks with distinct architectures, but I would agree that they do not provide definitive evidence for the coexistence of these structures. Obtaining such evidence by EM may prove to be a significant challenge for several reasons. First, it appears that the EM community itself is still in dispute over how best to preserve the fragile organization of actin networks for high-resolution ultrastructural analyses (Small et al., 2008). Second, the putative spatial overlap of distinct actin networks in the thin volume of the leading edge of a cell will probably be recognizable only by the application of high-resolution tomographic approaches, supported by sophisticated image segmentation and topology classification algorithms; visual inspection of projection images is insufficient to test the lamella hypothesis. Third, interactions between the lamellipodium and lamella are transient in space and time. For the epithelial-cell model considered in the present discussion, the protrusive state of the cell edge alters on a time-scale of 80-120 seconds and over a length of ∼4 μm (Machacek and Danuser, 2006), accompanied by transient changes in filament assembly and retrograde flow (Ponti et al., 2005). These variations are probably accompanied by changes in the mass balance and in the coupling of the lamellipodium and lamella networks, if they do indeed overlap. Thus, a challenge for EM analyses is the capture of well-defined functional states of cytoskeleton structures at the relevant time-scale and resolution. Moreover, the transience of cell and cytoskeleton states is expected to generate structural heterogeneity. Meaningful statistical evaluation of such structural distributions requires the combination of higher-throughput EM imaging with high-resolution live-cell light microscopy.

FSM data and the lamella hypothesis

In the accompanying paper, Vallotton and Small also contest the lamella hypothesis from an image-analysis angle, and propose that the hypothesis rests on erroneous tracking of actin speckles at the very leading edge. In the publication by Ponti et al. (Ponti et al., 2004), a group of fast-moving, short-lived speckles was shown to cluster preferentially in a band 2-3 μm wide that was juxtaposed to the cell edge. Another group of long-lived, slower-moving speckles localized throughout the entire cell front, including the first 2-3 μm. The majority of the classified speckles at the cell edge belonged to the first group. Importantly, although this is not mentioned by Vallotton and Small (Vallotton and Small, 2009), a large population of slow-moving, short-lived speckles was excluded from the analysis on the basis that they were non-classifiable. These speckles might be associated with pure tracking errors, or might consist of fluorophores that have been incorporated into a mixed population of filaments with heterogeneous turnover and motion properties, thus generating an unstable image signal (see their box 1, fig. part A). Despite the relatively low percentage of classifiable speckles, both speckle groups contained several thousand samples per cell, which was sufficient for statistical analysis of the underlying F-actin dynamics. We proposed that the differences between the two separable speckle categories indicated two different, yet partially overlapping, actin-filament populations. This conclusion was derived from the distinct spatial distribution of sites of preferential filament assembly and disassembly of the two speckle groups: The fast-moving, short-lived speckles generated a band of high F-actin assembly along the cell edge and disassembly behind the edge, which is reminiscent of the treadmilling that had been described as the characteristic lamellipodium behavior (Small et al., 2002; Pollard and Borisy, 2003), whereas the slow-moving, long-lived speckles generated random distributions of sites of assembly and disassembly. These characteristics seamlessly matched the behavior of actin filaments in the lamella region behind the lamellipodium. Whether the lamella extension into the lamellipodium region is transient or stationary could not be defined, because integration of speckle data over several protrusion-retraction cycles was required for stable separation of populations.

Key differences between the contradicting analyses

What are the sources of difference between the speckle tracking presented in the accompanying Speckle Microscopy Update by Vallotton and Small and the methods published by Ponti et al. (Ponti et al., 2004)? Overall, there is qualitative agreement on the nature of the speckles that have been tracked in both papers. The disagreement, I believe, originates in the speckles that are not tracked by Vallotton and Small (Vallotton and Small, 2009).

The key issue in tracking dense speckle flow fields is that the frame-to-frame assignment of detected speckles is performed simultaneously for all speckles to generate a topologically consistent set of links that is globally optimal. This can be achieved through the graph-theoretical algorithms developed in Vallotton et al. (Vallotton et al., 2003) and mentioned again in Speckle Microscopy Update I, or through neural networks (Ponti et al., 2004; Ponti et al., 2005). Hence, in terms of speckle assignments, both methods are equivalent. Notably, in Delorme et al. (Delorme et al., 2007), which reproduced the classification of lamellipodium and lamella speckles in a different epithelial-cell model under different conditions, we employed the same graph-theoretical formalism for tracking as Vallotton et al. (Vallotton et al., 2003).

However, the first difference between the method reported in Speckle Microscopy Update I and the methods used by us in Ponti et al. (Ponti et al., 2004) and follow-up papers is in the approach to tracking large speckle displacements. As Vallotton and Small point out, at a sampling of 10 seconds per frame, speckle displacements between frames can be close to or greater than half of the inter-speckle distance. Under these conditions, it is impossible to assign speckles accurately, even when using global optimization. In Vallotton et al. (Vallotton et al., 2003), this problem was tackled by constraining the motion of individual speckles. To be tracked, speckles had to move straight and at constant speed. Speckles that fell outside the tolerated deviation from this model were not considered. Speckle Microscopy Update I mentions the introduction of an additional constraint to enforce spatial coherence in speckle motion. Although no specifics of this constraint are given, it is clear that a priori homogenization of speckle motion biases the tracking towards the behavior of the majority. Should a minority with different motion behavior co-exist, such as a few lamella speckles in a sea of lamellipodium speckles, it will be disregarded.

In Ponti et al. (Ponti et al., 2004), we refrained from making a priori assumptions about speckle behavior. To capture the fast motion of speckles, an iterative tracking approach was developed, in which the flow fields acquired in previous iterations were used to `push' speckles in the direction of their most likely movement before assignments to speckles in the next frame were made. Hence, evidence for fast and slow speckle movements was accumulated from the data itself and not enforced by the user. The requirement for predicting speckle motion was articulated in detail in Ji and Danuser (Ji and Danuser, 2005), in which we also proposed that correlation-based and particle-based tracking be combined; correlation-based tracking is a robust technique for following large speckle displacements, whereas particle-based tracking enables heterogeneous movements to be followed. Importantly, motion prediction renders the assignment of speckles less sensitive to the selection of the search radius for possible speckle assignments. In combination with a prediction scheme, the search radius becomes a measure of how much individual speckle movements can vary from the mean behavior. Thus, the assertion in Speckle Microscopy Update I that Ponti et al. (Ponti et al., 2004) missed fast movements because large speckle displacements in the lamellipodium were excluded by overly short search radii does not apply. On the contrary, the velocity of speckles classified as lamellipodium speckles in Ponti et al. (Ponti et al., 2004) ranged from 0.7 to 1.6 μm/min, which is significantly faster than the lamellipodium speckles shown in the histogram in fig. 1F. Moreover, the same family of algorithms reported speckle speeds of up to 6 μm/min when lamellipodium retrograde flow was accelerated by increased cofilin activity in movies also sampled at 10 seconds per frame (Delorme et al., 2007). These comparisons indicate that the algorithms applied in Ponti et al. (Ponti et al., 2004) and follow-up papers are perfectly suited to tracking high speckle speeds.

Vallotton and Small allegedly measured speckles that moved as fast as ∼4.0 μm/min, and note that `jumps of up to 10 pixels are not exceptional'. Unfortunately, we do not learn how many such speckles were measured. The histogram in fig. 1F only presents velocities up to ∼0.8 μm/min. However, a significant population of such speckles seems unlikely, because of the average lamellipodial flows of 0.4 and 0.6 μm/min reported in fig. 1C. To achieve this average value, each speckle of ∼4.0 μm/min would require >8 speckles in the neighborhood moving at 0.2 μm/min (the lowest value shown for speckles in the lamellipodium). No such slow speckle population can be found in the histogram. Also of note, the histogram of lamellipodium velocities contains 85 speckles, acquired over 10 frames, i.e. 100 seconds (Pascal Vallotton, personal communication). In combination with the homogenization of speckle motion imposed by the tracking algorithm (see above), this low sampling over not even a full protrusion cycle might be insufficient to characterize the transient interactions of lamellipodium and lamella.

The speckle-tracking methods in the Speckle Microscopy Update and Ponti et al. (Ponti et al., 2004) also differ in the speckle-detection step. It is not made clear in the Speckle Microscopy Update how speckles were detected, but key features of the speckle detection in Ponti et al. (Ponti et al., 2004) are a comparison of the signal of each individual speckle against a calibrated model of background and shot noise, and an iterative search for partially overlapping speckles (Ponti et al., 2005). The latter is crucial for recovering heterogeneous and intermingled speckle movements, in which speckles can slide past one another and transiently cancel out their signals. Along the same lines, the tracking algorithm in Speckle Microscopy Update I probably does not account for temporary speckle occlusion, whereas the methods in Ponti et al. (Ponti et al., 2004) and follow-up papers explicitly deal with short-term interruptions of speckle trajectories. Overall, the density of speckles extracted and tracked in Ponti et al. (Ponti et al., 2004) and follow-up papers is much greater than in Speckle Microscopy Update I (according to fig. 1F). False detection negatives in Vallotton and Small (Vallotton and Small, 2009), combined with the assumption of motion coherence, are a further source of tracking bias towards the behavior of the majority. Thus, the results in Ponti et al. (Ponti et al., 2004) and the accompanying Speckle Microscopy Update differ in that, through the choice of algorithms, the latter may have systematically excluded from the tracking a putative minority of lamella speckles at the cell front.

At this point, I wish respectfully to object to the statement in Vallotton and Small (Vallotton and Small, 2009) that our paper (Ponti et al., 2004) did not offer a validation of the tracking code used on simulated data or manually tracked scenes. After a summary of the methods, the supplement of Ponti et al. (Ponti et al., 2004) refers to a second article `in revision for Biophysical Journal', which was later published (Ponti et al., 2005). The Biophysical Journal paper contains extensive illustrations of the importance of iterative speckle detection and tracking for the measurement of F-actin dynamics. It also contains validation of the algorithms by simulations and refers to more comprehensive analyses of algorithm performance in the publicly accessible PhD thesis of Aaron Ponti (http://e-collection.ethbib.ethz.ch/view/eth:26913).

Why are the slow-moving speckles not seen in the kymographs and manual-tracking analyses in the accompanying Speckle Microscopy Update? There are at least three possible answers to this question. First, kymographs average image dynamics in space and time. Therefore, they are `majority voting schemes'. It takes single-particle analysis to dissect a potentially heterogeneous speckle population.

Second, the dissection of heterogeneous speckle movements by hand tracking is very challenging. In the past, our laboratory has made several attempts to generate manual ground truth data sets in these images. Our approach differed from the approach of Vallotton and Small (in which the operator selects a few speckles), in that we tried to avoid tracking bias as a result of the limited detection sensitivity of the human eye (our eyes tend to focus on the bright and coherently moving spots) by presenting to the operator computer-detected speckles for subsequent manual tracking. Our program presented the same speckle to one or different operators, allowing us to calculate the intra- and inter-user reproducibility. Values of 70-80% and 50-70%, respectively, indicated to us the difficulties of manual tracking. Indeed, in previous work Vallotton et al. also made the observation that manual tracking and kymographs might be of limited use in validating speckle-tracking algorithms (Vallotton et al., 2003). In that paper, the authors excluded manual tracking as a flawed means of validation, and cautioned that kymograph analysis is limited in regions with significant heterogeneity in speckle flow.

Third, speckles provide a stochastic view of dynamic macromolecular assemblies. Moreover, experimental control over the speckle pattern is limited. To cope with the resulting intrinsic and extrinsic variation of the speckle signal, many events must be accumulated to draw an accurate picture of the underlying cellular processes. This is complicated further when the pattern potentially presents a convoluted stochastic view of several processes. A short sequence of a single movie is usually insufficient to capture the full spectrum of speckle behaviors. In Speckle Microscopy Update I, Vallotton and Small mention that only the very clearest speckle movies can be tracked at the single-particle level. Although it seems at first sight to be paradoxical, simulations indicate that these movies are not necessarily the most informative for revealing multiple heterogeneous processes, especially when the presence of a few bright speckles biases the detection algorithm towards rejecting weaker speckles. It is these weaker speckles that might report the dynamics of secondary structures that have different kinetics of fluorophore incorporation. Thus, much work has been done to make the particle tracking methods robust for a wide range of speckle quality (Danuser and Waterman-Storer, 2006). To detect heterogeneous polymer behaviors, complete and unbiased measurements of all speckles above noise have to be made, a task in which computational methods outperform manual analyses.

Conclusion

As with every scientific model, the lamella hypothesis should be challenged and further analyzed. I fully agree with Vallotton and Small that quantitative FSM (qFSM) should be used alongside complementary approaches to the study of cytoskeletal dynamics. The first steps towards this goal were made in Ponti et al. (Ponti et al., 2004), where we showed the differential accessibility of the speckle populations to different drugs. Later, it was suggested that manipulations of the concentration and activity of the actin-associated proteins tropomyosin and cofilin also lead to differential effects on the speckle populations (Delorme et al., 2007; Gupton et al., 2005). However, these experiments are limited in that they test the lamella hypothesis only indirectly, by testing its functional predictions. More direct evidence for diverse actin-filament dynamics at the cell edge is still needed. Whether correlative EM is equipped at this point to provide such evidence remains open (see the concerns above). However, we are in exciting times for testing the lamella hypothesis, using techniques such as single particle tracking photoactivation localization microscopy (sptPALM) of actin (Manley et al., 2008) or multicolor qFSM (Hu et al., 2007) of lamellipodium- and lamella-specific molecules (Iwasa and Mullins, 2007). sptPALM maintains many of the strengths of live-cell imaging, but can attain the resolution level of EM. Because of the requirement for repeated photoactivation, however, it may be difficult to capture transient interactions between lamellipodium and lamella. Multicolor qFSM will be complementary for this point, but will provide only diffraction-limited resolution. The spatial colocalization of lamellipodium and lamella filaments will still have to be inferred from the differential behaviors of different molecular species. Clearly, combining the strengths of all three methods in a quantitative framework will be the most powerful approach to testing the lamella hypothesis.

I do not assume that the considerations described here will defend the lamella hypothesis against its sceptics. On the contrary, these technical disputes may only strengthen their opinion that the statistical methods applied to qFSM image analysis are not robust enough to support the hypothesis, even though many of its aspects have been confirmed by functional perturbations. As described above, it is time to move on and scrutinize this issue with novel, complementary optical and molecular approaches. Should these new data provide evidence against a partial overlap of functionally distinct filament networks, new questions will follow. I will mention but two of these: First, how, if not by overlap, are the fast- and slow-moving networks mechanically integrated to sustain the forces driving cell protrusion? Second, how, if not by overlap, can it be reconciled that only ∼70-80% of the actin-polymer mass is disassembled at the lamellipodium base (Vallotton et al., 2004), whereas there is no evidence that actin filaments buckle at the lamellipodium-lamella transition, despite the different speeds of actin flow in these regions? At the light-microscopic level, actin networks can be treated as a continuum (Vallotton et al., 2004) and thus disassembly is equivalent to a reduction of material density; at the ultrastructural level, however, the velocity gradient between lamellipodium and lamella must yield a deformation of the network architecture in the absence of complete filament disassembly (or complete breaking). We have not seen this in EM of epithelial cells (Delorme et al., 2007; Gupton et al., 2005).