In viv o spectrally unmixed multi-photon imaging of longitudinal axon-glia changes in injured spinal white matter

Supplementary


Introduction
Two-Photon Laser Scanning Microscopy (2P-LSM) is an ideal tool for imaging neural tissue in vivo [1][2][3][4][5][6][7].The advantages of 2P-LSM lie in recording relatively deep stacks with high z-resolution within a time course of a few minutes, minimal cell phototoxicity, and negligible bleaching effects on simultaneously excited fluorescent proteins (FPs) in labeled neural tissue, providing optimal conditions for in vivo time-lapse studies.Using this technique, reactions of a single cell type to environmental changes, e. g. of microglia to nerve damage [2,5,8,9], as well as neuron/axon-glia interactions in intact and pathologically altered environment [8][9][10][11][12][13][14] can be studied.A crucial role of pathophysiological changes in the white matter of the central nervous system (CNS) is evident not only in classic demyelinating diseases such as multiple sclerosis, but also in other diseases of the nervous system [15,16].The non-cell-autonomous pathophysiology of most neurological diseases emphasizes the critical contribution of pathological axon-glia interactions to the progression of often length-dependent axonal degeneration and resulting neuronal death [16,17].Degenerative processes of axonal structures in acute or chronic diseases occur in accordance with Abbreviations: 2P-LSM, two-photon laser-scanning microscopy; ACSF, artificial cerebro-spinal fluid; AF, autofluorescence; CNS, central nervous system; ECFP/ EGFP/EYFP, enhanced cyan/green/yellow fluorescent protein; FP(s), fluorescent protein(s); MIP, maximum intensity projection; SCI, spinal cord injury.(e) First, second and fourth channels were assigned to blue, green and red color table, respectively.Although some bleed-through is evident -ECFP positive astrocytes appear cyan (mixed blue and green) and EGFP positive microglia appear yellow (mixed green and red) -this combination allows the three different cell types to be recognized.For better visualization, EYFP fluorescence is shown with a red color table in all images.(f) By picking out bright pixels from a few somata, the spectra of ECFP, EGFP and EYFP were measured; the averages give the reference spectra.Note the broad spectrum of ECFP present in all channels and the absence of EGFP and EYFP in the first channel as well as the absence of EYFP in the second channel.(g) This optical partial separation enabled reliable unmixing of the three fluorescent proteins, as shown by the clear repartition of their abundance coefficients projected in the plane (α ECFP +α EGFP +α EYFP =1).(h-k) Spectral images of the region shown in e.In the first channel (h), where only ECFP is expected (see spectra in f), there is a superficial layer on the left side (within the dashed ellipse) and few fibers inside the spinal tissue (one of them is marked by an arrow).In contrast to ECFP, which is present in all channels (an example cell is marked by a square in h-k), both appearances are not visible in the other three channels.The extraction of this autofluorescence (AF) from the unmixed images is shown in Fig. 2. (l-o) Reconstituted images of the abundance of AF, ECFP, EGFP and EYFP by linear unmixing.As predicted in g, each image contains only one cell type.Furthermore, the AF was successfully isolated from ECFP (ellipse and arrow in l, see Fig. 2).Scale bars in d, 400 µm and in e and h-o, 50 µm.(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)pathological acute or chronic axon-glia interactions.
In contrast to the cortical grey matter, white matter accessibility to in vivo 2P-LSM in the brain is limited.Spinal white matter, with its superficial location, offers suitable access for performing longitudinal studies and repetitive optical imaging in living mice to study axon tracts and surrounding glia, for example after spinal cord injury (SCI).Chronic in vivo studies of axon tracts after SCI have been demonstrated over several days in zebrafish [18] and up to two to four months with several imaging sessions in mice [19][20][21].In addition, chronic in vivo studies of blood vessels and microglia have been performed in intact tissue [1,22] and after axonal lesion [6,20,21].However, in vivo imaging of white matter regions in the lateral spinal cord as well as motor neuron axons in ventral roots, both of which are affected by the motor neuron disease amyotrophic lateral sclerosis (ALS), have only been demonstrated by studies from our group [1,9,11,13,23].To overcome spatial and temporal limitations, we have developed techniques for in vivo imaging of dorsal and lateral aspects of the spinal cord, with different schedules, an acute one-time and a long-term recurrent schedule (up to 200 days with up to ten separate imaging sessions) [1,24].Additionally, we imaged spinal tissue labeled with up to four FPs in axons, astrocytes, microglia, and blood vessels.To clearly separate the simultaneously excited FPs, acquired images using a custom-made 4-channel 2P-LSM were subjected to spectral unmixing and response of different cell types were analyzed.

Ethics statement
The experiments were carried out in accordance with the ethical guidelines of the national animal protection law and were approved by the ethics committee of the State of Lower Saxony.

Mice
Adult double-transgenic TgH(CX3CR1-EGFP)xTgN(THY1-EYFP) and triple-transgenic TgN(GFAP-ECFP)xTgH(CX3CR1-EGFP)xTgN(THY1-EYFP) reporter mice (2-5 months old at the start of experiments, male and female) were used for in vivo imaging experiments [1,9,11,13,25].EYFP is expressed in various types of neurons, but particularly in motor and sensory neurons such as motor neurons and dorsal root ganglion cells.Axon-glia interactions in the spinal white matter of the mice were examined in vivo by simultaneously detecting cell-type specific expression of fluorescent proteins.

Anesthesia and surgery
Anesthesia and surgery for acute one-time and chronic recurrent imaging experiments have been described in detail in previous publications from our group [1,13].The tail vein served as access for injecting fluorescent dyes.Most commonly, we used Texas Red Dextran as a fluorescent dye to visualize the blood vessels in vivo.

2-Photon laser-scanning microscopy and image acquisition
We built a custom-made microscope capable of simultaneously recording four channels in vivo [1,11,13,24].The ScanImage software [26] we modified was used to drive the microscope.To simultaneously excite all fluorescent proteins (FPs) in the tissue, the laser wavelength was set to 925 nm for optimal excitation of all three FPs (ECFP, EGFP, EYFP) and Texas Red Dextran [27,28].As shown in Fig. 1a,b, emitted light was split by a 505 nm longpass (dic1), a 440/520 nm bandpass (dic2) and a 462/523 nm bandpass (dic3) dichroic mirror (Semrock, Rochester, USA, except the 505 nm longpass which was from Chroma Technology Corp., Rockingham, Vermont, USA) and collected by photomultiplier tubes (Hamamatsu, Japan) through four bandpass filters (Semrock): a 554 ± 105 nm (em1), a 460 ± 40 nm (em2), a 510 ± 41 nm (em3) and a 583 ± 60 nm (em4).Parallel, evenly spaced planes were recorded, digitized, and processed to obtain z-stacks of images.The total acquisition time for a stack of 20 to 30 images was approximately 2 min.Stacks were obtained with a maximum depth of 60 µm.In the experiments, multiple stacks were recorded continuously to obtain time-lapse series.Axonal injuries were applied either by the titanium-sapphire laser focused at a plane of one axon diameter for several seconds until the fluorescence signal began to increase, or mechanically by a fine needle.

Spectral unmixing
Almost all visible dyes used in biology today exhibit spectral overlap.By using two channels to separate the contribution of two labels, only a fraction of the available fluorescence is utilized.However, a complete separation is often not possible and false misinterpretations easily occur [29].Spectral unmixing solves the problem by calculating the abundance coefficients of each dye based on knowledge or estimation of their spectra [30][31][32][33][34][35][36][37].Four channels are sufficient to unmix at least four dyes.In a system based on dichroic mirrors in cascade, more detection channels increase the distance between the objective lens and the PMTs, so we chose the compromise of four detection channels to unmix two or three dyes, resulting in a figure of merit (FoM) not too far away from the optimum [38].The filters were selected semi-quantitatively by applying the general rules known to optimize the FoM [38].Spectral unmixing including iterative unmixing procedures is described in the Supplementary Methods section.

Statistical analysis
Origin software (Northampton, USA) was used for statistical analysis.Mean values were reported ± standard of the mean (SEM).Statistical significance (p < 0.05) was determined using ANOVA followed by Tukey test.

In vivo imaging of the spinal cord
Multi-Photon Laser Scanning Microscopy is the method of choice for in vivo imaging due to its high penetration depth and low phototoxicity.However, in vivo imaging of multiple tissue levels takes time, which is why this method is often associated with movement problems of the tissue being imaged.These problems play a minor role for cortical studies because the skull can be fixed with bone cement.Consequently, it is mainly only oscillations in blood pressure that interfere with the imaging.Fixation of the vertebral column to prevent movement of the spinal cord proves to be insufficient for motionless imaging.Although the mouse is anesthetized, breathing and blood flow oscillations cause the main problems of rapid movements in the x-y-z direction.These movements can already be recognized when recording only one plane within an image stack.Scanning a plane with sufficient resolution takes about one second, while in mice we have about two breaths and about seven heartbeats during this time.Further problems arise from the objective lenses.Using water-immersed object lenses requires a pool filled with artificial cerebro-spinal fluid (ACSF).However, over time the fluid tends to drain into the surrounding tissue, and the pool may also leak.Therefore, the amount of fluid between the lens and tissue continuously decreases, resulting in change of surface tension and slow movement of the spinal cord in z-direction.This slow movement may be tolerable for individual image stacks, but is critical for time series of image stacks.To overcome these and other problems, we have developed a method for stable in vivo imaging of superficial regions within the white matter of the spinal cord [1].The necessary equipment as well as anesthesia and spinal preparation for acute one-time and chronic (recurrently in the same mouse over a time course of up to 200 days) imaging experiments are described in detail in our previous publications (caption on next page) P. Dibaj et al.Neuroscience Letters 841 (2024) 137959 [1,13].An example 2P-LSM stack of the dorsal spinal cord with labeled neurons/axons, microglia, and blood vessels is shown in Supplementary Video 1.

Spectral imaging and linear unmixing of simultaneously imaged spinal axons, astroglia and microglia
Multicolor imaging requires quantitative approaches to analyze multiple different FPs excited simultaneously in the same sample.Spectral imaging and linear unmixing offer the possibilities to clearly distinguish FPs with highly overlapping emission spectra.To test glial behavior and axon-glia interactions in the white matter of the CNS, we performed in vivo 2P-LSM imaging of the spinal cord of multipletransgenic mice expressing ECFP, EGFP and EYFP in astroglia, microglia and neurons, respectively.Fig. 1a-c shows our setup with the 4channel detection head and the transmission spectra of the optical filters for the spectral imaging of ECFP, EGFP and EYFP.Fig. 1d,e shows an overview of the exposed dorsal spinal cord.After bright-pixel measurements of each FP in all channels (Fig. 1f,g), linear unmixing was performed to achieve a clear separation of astroglia, microglia and axonal structures (Fig. 1h-o).Isolation of autofluorescence from ECFP was successfully achieved by iterative unmixing procedures (Fig. 2).Through iterative unmixing, we were able to clearly distinguish collagen fibers in the dura mater and in the nervous tissue (here especially around vessel structures) from labeled axons as well as from astroglia and microglia.Similarly, the iterative unmixing technique was used to distinguish labeled blood vessels from the three cell lines (Supplementary Fig. 1).

Glial response after axonal injury over a time course of up to 200 days
As previously shown [2,5,8,9,13], local axonal injuries in the dorsal and lateral spinal columns result in immediate attraction of microglial processes and subsequent phagocytosis and soma migration towards transected axonal structures by ameboid transformed microglia.To analyze axon-glia changes after axonal injury in spinal white matter, long-term recurrent in vivo imaging was performed in the same tripletransgenic mouse with labeled axons, astroglia and microglia over a period of 150 to 200 days (Fig. 3 and Supplementary Fig. 2).For clear differentiation, spectral imaging was followed by linear unmixing.An immediate microglial response within the first hours to days was followed by a delayed astroglial response with extensions of the processes toward the injured site beginning in the first few days and accumulation of somata at the injured site beginning at the end of the first week (Fig. 3a).Quantification of position changes in defined areas surrounding the injury highlights the contrasting behavior of glial cells, particularly in the first month after injury (Fig. 3b).In the first few days, axons degenerate similarly caudal and rostral to the injured site, i. e. proximal and distal to the respective cell somata in the dorsal root ganglia (Fig. 3c).However, distal axonal elements subsequently degenerated more rapidly, resulting in a higher proportion of remaining proximal elements (Fig. 3d).The relatively small laser-induced injuries resulted in a decrease in glial presence within and around the injured site after the first month (Fig. 3a,b,e).Comparable to laser-induced axonal injuries, mechanically induced axonal injuries of relatively similar small size resulted in similarly comparable glial responses and axonal degenerative processes (Supplementary Fig. 2).During the course of recurrent imaging, some axonal regenerative attempts were indicated by axonal sprouting of previously degenerated proximal axonal elements (Supplementary Fig. 2).

Discussion
A major problem in in vivo imaging of the spinal cord using 2P-LSM is the drift of the imaged area due to the relatively long acquisition time and motion artifacts due to breathing and heartbeat.For a lowresolution image of 256 dpi in the x and y directions, approximately one second is required for the image of one z-slice when the x-scanner speed is set to four ms per line.Four times as long is required for highresolution images of 1024 dpi in x and y.Therefore, z-stack image recording usually takes a few minutes.In addition, time-lapse recording results in a longer image course of several hours.Unlike the skull, movement artifacts arise because of the proximity of the spine to the heart and lungs.Although blood pressure-related fluctuations cannot be avoided, they can be minimized by keeping a sufficient distance from large arteries.However, breathing-related movements could be influenced by the experimental procedure.To overcome respiratory movements, anesthesia by i. p. injection of ketamine, xylazine, and acepromazine and suspending the entire mouse in air were proposed by one group [39].In our chronic recurrent experiments with volatile anesthesia, as used by another group [20], it was not necessary to fixate the mouse with another clamp at the tail root (which would block access to the tail veins).There was also constant contact with the heated support.Our experimental set-up offers the advantage of easy temperature control and permanent access to a tail vein for infusion with Ringer's solution or with the fluorescent marker Texas Red for labeling blood vessels.In long-term experiments, dehydration could thus be minimized through tail vein access.Other groups apply anesthesia by i. p. injection and do not need to suspend the entire mouse in the air [21,40].
The method of imaging the spinal cord through the window between two adjacent vertebrae [1,40] not only significantly minimizes the traumatic impact of surgery compared to tissue irritation caused by laminectomy [20] but also the rate of artificial movements.The Unmixing of the four dyes with four channels leads to unreliable estimates.Therefore, we initially neglected AF and only unmixed ECFP, EGFP and EYFP.The triangular plot of the abundance coefficients (see Fig. 1) shows that each exogenous fluorescent protein (ECFP, cyan dots; EGFP, green dots; EYFP, red dots) is different from the others, but the pixels with AF (blue dots) generate aberrant abundance coefficients falling in the ECFP region.(c) Two-dimensional (ECFP-EGFP) reduction of three-dimensional (ECFP-EGFP-EYFP) abundance coefficients from regions containing AF, ECFP, or EGFP confirming that AF pixels are interpreted as ECFP pixels and not as EGFP pixels.Note the expected huge confidence intervals (ci) of the AF coefficients.(d) The same applies to the ECFP-EYFP pair as shown in c.(e) Diagram of the iterative AF extraction method.We initially unmixed ECFP, EGFP and EYFP from the four spectral channels ch1, ch2, ch3, and ch4 (step one).AF is only present in ch1 in a layer on the left (ellipse) and as one visible fiber (arrow).The images of the abundance coefficients illustrate the conclusion of b: EGFP and EYFP are correctly unmixed, while the ECFP image (ECFP ) contains ECFP and AF (ellipse and arrow).Then we used all four channels again, but only considering AF and ECFP (Step two).EGFP and EYFP spectra are closer to the ECFP spectrum than to the AF spectrum (see a), therefore they are interpreted as ECFP − the image ECFP ́́ therefore contains all three exogenous fluorescent proteins.On the contrary, AF is detected accurately (image AF ).The 2-dimensional plot of AF abundance coefficients versus ECFP coefficients (inset) illustrates the separation of AF from the other components.Note the small ci of the AF estimates, which shows that the pixels detected as AF are actually AF positive.The ECFP estimates have asymmetric ci: a rather small number along the AF axis, showing that these pixels really do not contain AF, and a large number along the ECFP axis if they contain a protein other than ECFP (EGFP or EYFP).This step allowed us to create the image AF ́ containing all and only AF positive pixels.Finally (step three), unmixing ECFP and AF from a pseudo two-channel image formed by mixing ECFP ́ (containing AF and ECFP) and AF ́ (containing only AF) enabled to split ECFP (image ECFP) from AF (image AF), as shown in f by the two-dimension plot of AF abundance coefficients versus ECFP abundance coefficients of AF (blue dots) and ECFP (cyan dots) pixels, respectively.The ci are too small to be visualized.In the end, we successfully unmixed all four signals: AF and ECFP (merged in g), EGFP  implantation of a spinal chamber [20] is therefore not necessarily a requirement for longitudinal studies, especially when lengthy surgical protocols should be avoided.Our minimally invasive surgical protocol has allowed us to conduct longitudinal studies up to 200 days with up to ten separate imaging sessions without observing any motor or other behavioral deficits.In addition, this method offers the advantage of a natural pool created by removing only the muscle and tendon layer between two adjacent spines.Two technical limitations have made high-resolution multi-cellular in vivo imaging in spinal cord difficult to perform: i) breathing-induced movements, as mentioned previously; ii) cell-type specific GFP-based fluorescent proteins (FPs) usable in vivo show overlapping emission spectra.Multi-cellular imaging has been performed primarily in vitro, whereas in vivo studies in the spinal cord have focused primarily on a single cell-type [3,18,19,41], thus precluding any assessment of dynamical cellular interactions.Imaging of microglia in addition to axonal structures has now been performed by some groups [20,22,40,42,43].Here, we present an experimental protocol to circumvent technical problems by coupling sophisticated surgery procedures with 2P-LSM spectral imaging followed by iterative linear unmixing for unambiguous differentiation of FPs.We enable simultaneous imaging of up to four dyes in the spinal cord in vivo with 2P-LSM resolution.Application of this technique in acute one-time and chronic recurring experiments allowed us to reveal subtle structural interactions between axons, astroglia and microglia in the healthy or injured white matter of the spinal cord in vivo.The technique of iterative linear unmixing also allowed us to clearly distinguish the three cell lines from intravenously labeled blood vessels and collagenous structures.Our longitudinal study of microglial and subsequent astroglial invasion after laser and mechanically induced axon transections provides important temporal and morphological aspects for the development of therapeutic strategies to positively influence glial response, scar formation and regeneration attempts.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.Fig. 3. Long-term multi-cellular imaging of axon-glia changes after a laser-induced injury over 150 days.(a) Recordings over months of the same lesion site in the dorsal column in vivo.Spectral unmixing was applied to separate the signal coming from astrocytes (CFP, cyan), microglia (GFP, green) and axons (YFP, red).Microglia density progression was measured in three regions shown in the first image (pre lesion, low magnification)."r1" is a circle that directly covers the lesion site; "r2" is defined as the area between the inner ellipse and "r1", while "r3" corresponds to the area between the outer and the inner ellipses.Note the extension of microglia processes in the first hour after injury and the accumulation of their somata at the lesion core already one day later.Although astroglia did not show any significant migration (not even in the next few days; see the marked astrocytes in circles), three days after injury they begin to send out thick processes (arrows) that envelop the lesion core.On the seventh day, they formed a diffuse scaffold that was localized to the injured site together with the microglia.Injured axons were fragmented rostrally and caudally as early as one hour after lesion without involvement of microglia that focused the lesion core.Axonal debris was removed at later time points (dashed square, rostral) or remained in the same location for weeks (arrowhead).After 3 months the reaction had almost completely subsided.The few axons that appeared to cross the lesion site were actually underneath and were not affected by the injury.Unlabeled blood vessels are visible as dark structures.Images are Maximum Intensity Projections (MIP) of 15-µm z stacks.Scale bar is 50 µm in the first image and 25 µm in the others.All images are arranged rostral side up.(b) Quantification of microglia and astrocyte responses.They differ in terms of temporal and intensity changes: microglia reacted faster and more strongly.In the first hours after injury, they even behaved oppositely: microglial processes immediately invaded the core of the lesion, while astrocytes appeared to leave the core of the lesion.(c) Quantification of axonal dye-back.Rostral and caudal parts of injured axons seem to degenerate with the same speed.In particular, the measurement was carried out on axons whose tips were still visible.(d) The proportion of the degenerated axons decreased much more rapidly on the rostral side.(e) Quantification of astrocytes response by measuring CFP intensity driven by the GFAP promoter normalized to YFP intensity.CFP expression peaks between 3 days and 2 weeks after injury and then returns to basal levels.Error bars are 90 % confidence intervals assuming Gaussian distribution.(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 1 .
Fig. 1.In vivo 2P-LSM unmixing of ECFP, EGFP and EYFP expressed in astrocytes (blue), microglia (green) and axons (red) of the dorsal spinal cord, respectively.(a) Photo of the setup with the four-channel detection head and an anesthetized mouse under the objective lens (lower left).The position of the Photo-Multiplier Tubes (PMT 1 to 4) close to the objective lens and the arrangement of the dichroic filters in cascade (dic0, dic1, dic2, and dic3) insure high collection efficiency.(b) Outline of the setup highlighting the relative position of the dichroics and emission filters (em1, em2, em3, and em4) used to record four spectral channels simultaneously.Ellipses show the position of lenses.(c) Transmission spectra of optical filters (modified data from Semrock).The filters dic2 and dic3 are multi-edges dichroics, but only one band is included in the appropriate spectral window defined by dic1: dic2 left to dic1and dic3 right to dic1.(d) Bright-field image of the dorsal spinal cord prepared by removing the connective tissue between the two vertebral arches L1 and L2 (arrows) down to the intact dura matter.All images are arranged rostral side up.The central vein is marked by an arrowhead.The square box marks the region imaged with 2P-LSM, shown in e as an RGB Maximum Intensity Projection (MIP) image.(e)First, second and fourth channels were assigned to blue, green and red color table, respectively.Although some bleed-through is evident -ECFP positive astrocytes appear cyan (mixed blue and green) and EGFP positive microglia appear yellow (mixed green and red) -this combination allows the three different cell types to be recognized.For better visualization, EYFP fluorescence is shown with a red color table in all images.(f) By picking out bright pixels from a few somata, the spectra of ECFP, EGFP and EYFP were measured; the averages give the reference spectra.Note the broad spectrum of ECFP present in all channels and the absence of EGFP and EYFP in the first channel as well as the absence of EYFP in the second channel.(g) This optical partial separation enabled reliable unmixing of the three fluorescent proteins, as shown by the clear repartition of their abundance coefficients projected in the plane (α ECFP +α EGFP +α EYFP =1).(h-k) Spectral images of the region shown in e.In the first channel (h), where only ECFP is expected (see spectra in f), there is a superficial layer on the left side (within the dashed ellipse) and few fibers inside the spinal tissue (one of them is marked by an arrow).In contrast to ECFP, which is present in all channels (an example cell is marked by a square in h-k), both appearances are not visible in the other three channels.The extraction of this autofluorescence (AF) from the unmixed images is shown in Fig.2.(l-o) Reconstituted images of the abundance of AF, ECFP, EGFP and EYFP by linear unmixing.As predicted in g, each image contains only one cell type.Furthermore, the AF was successfully isolated from ECFP (ellipse and arrow in l, see Fig.2).Scale bars in d, 400 µm and in e and h-o, 50 µm.(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 2 .
Fig. 2. Extraction of the autofluorescence by iterative unmixing.(a) Emission spectra of autofluorescence (AF), ECFP, EGFP and EYFP measured in vivo.(b)Unmixing of the four dyes with four channels leads to unreliable estimates.Therefore, we initially neglected AF and only unmixed ECFP, EGFP and EYFP.The triangular plot of the abundance coefficients (see Fig.1) shows that each exogenous fluorescent protein (ECFP, cyan dots; EGFP, green dots; EYFP, red dots) is different from the others, but the pixels with AF (blue dots) generate aberrant abundance coefficients falling in the ECFP region.(c) Two-dimensional (ECFP-EGFP) reduction of three-dimensional (ECFP-EGFP-EYFP) abundance coefficients from regions containing AF, ECFP, or EGFP confirming that AF pixels are interpreted as ECFP pixels and not as EGFP pixels.Note the expected huge confidence intervals (ci) of the AF coefficients.(d) The same applies to the ECFP-EYFP pair as shown in c.(e) Diagram of the iterative AF extraction method.We initially unmixed ECFP, EGFP and EYFP from the four spectral channels ch1, ch2, ch3, and ch4 (step one).AF is only present in ch1 in a layer on the left (ellipse) and as one visible fiber (arrow).The images of the abundance coefficients illustrate the conclusion of b: EGFP and EYFP are correctly unmixed, while the ECFP image (ECFP ) contains ECFP and AF (ellipse and arrow).Then we used all four channels again, but only considering AF and ECFP (Step two).EGFP and EYFP spectra are closer to the ECFP spectrum than to the AF spectrum (see a), therefore they are interpreted as ECFP − the image ECFP ́́ therefore contains all three exogenous fluorescent proteins.On the contrary, AF is detected accurately (image AF ).The 2-dimensional plot of AF abundance coefficients versus ECFP coefficients (inset) illustrates the separation of AF from the other components.Note the small ci of the AF estimates, which shows that the pixels detected as AF are actually AF positive.The ECFP estimates have asymmetric ci: a rather small number along the AF axis, showing that these pixels really do not contain AF, and a large number along the ECFP axis if they contain a protein other than ECFP (EGFP or EYFP).This step allowed us to create the image AF ́ containing all and only AF positive pixels.Finally (step three), unmixing ECFP and AF from a pseudo two-channel image formed by mixing ECFP ́ (containing AF and ECFP) and AF ́ (containing only AF) enabled to split ECFP (image ECFP) from AF (image AF), as shown in f by the two-dimension plot of AF abundance coefficients versus ECFP abundance coefficients of AF (blue dots) and ECFP (cyan dots) pixels, respectively.The ci are too small to be visualized.In the end, we successfully unmixed all four signals: AF and ECFP (merged in g), EGFP and EYFP (merged in h).(i) Merge of all four signals.Scale bars in e and g-i, 50 µm.(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) Fig. 2. Extraction of the autofluorescence by iterative unmixing.(a) Emission spectra of autofluorescence (AF), ECFP, EGFP and EYFP measured in vivo.(b)Unmixing of the four dyes with four channels leads to unreliable estimates.Therefore, we initially neglected AF and only unmixed ECFP, EGFP and EYFP.The triangular plot of the abundance coefficients (see Fig.1) shows that each exogenous fluorescent protein (ECFP, cyan dots; EGFP, green dots; EYFP, red dots) is different from the others, but the pixels with AF (blue dots) generate aberrant abundance coefficients falling in the ECFP region.(c) Two-dimensional (ECFP-EGFP) reduction of three-dimensional (ECFP-EGFP-EYFP) abundance coefficients from regions containing AF, ECFP, or EGFP confirming that AF pixels are interpreted as ECFP pixels and not as EGFP pixels.Note the expected huge confidence intervals (ci) of the AF coefficients.(d) The same applies to the ECFP-EYFP pair as shown in c.(e) Diagram of the iterative AF extraction method.We initially unmixed ECFP, EGFP and EYFP from the four spectral channels ch1, ch2, ch3, and ch4 (step one).AF is only present in ch1 in a layer on the left (ellipse) and as one visible fiber (arrow).The images of the abundance coefficients illustrate the conclusion of b: EGFP and EYFP are correctly unmixed, while the ECFP image (ECFP ) contains ECFP and AF (ellipse and arrow).Then we used all four channels again, but only considering AF and ECFP (Step two).EGFP and EYFP spectra are closer to the ECFP spectrum than to the AF spectrum (see a), therefore they are interpreted as ECFP − the image ECFP ́́ therefore contains all three exogenous fluorescent proteins.On the contrary, AF is detected accurately (image AF ).The 2-dimensional plot of AF abundance coefficients versus ECFP coefficients (inset) illustrates the separation of AF from the other components.Note the small ci of the AF estimates, which shows that the pixels detected as AF are actually AF positive.The ECFP estimates have asymmetric ci: a rather small number along the AF axis, showing that these pixels really do not contain AF, and a large number along the ECFP axis if they contain a protein other than ECFP (EGFP or EYFP).This step allowed us to create the image AF ́ containing all and only AF positive pixels.Finally (step three), unmixing ECFP and AF from a pseudo two-channel image formed by mixing ECFP ́ (containing AF and ECFP) and AF ́ (containing only AF) enabled to split ECFP (image ECFP) from AF (image AF), as shown in f by the two-dimension plot of AF abundance coefficients versus ECFP abundance coefficients of AF (blue dots) and ECFP (cyan dots) pixels, respectively.The ci are too small to be visualized.In the end, we successfully unmixed all four signals: AF and ECFP (merged in g), EGFP and EYFP (merged in h).(i) Merge of all four signals.Scale bars in e and g-i, 50 µm.(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)