Methodological considerations for global analysis of cellular FLIM/FRET measurements

Abstract

Global algorithms can improve the analysis of fluorescence energy transfer (FRET) measurement based on fluorescence lifetime microscopy. However, global analysis of FRET data is also susceptible to experimental artifacts. This work examines several common artifacts and suggests remedial experimental protocols. Specifically, we examined the accuracy of different methods for instrument response extraction and propose an adaptive method based on the mean lifetime of fluorescent proteins. We further examined the effects of image segmentation and a priori constraints on the accuracy of lifetime extraction. Methods to test the applicability of global analysis on cellular data are proposed and demonstrated. The accuracy of global fitting degrades with lower photon count. By systematically tracking the effect of the minimum photon count on lifetime and FRET prefactors when carrying out global analysis, we demonstrate a correction procedure to recover the correct FRET parameters, allowing us to obtain protein interaction information even in dim cellular regions with photon counts as low as 100 per decay curve.

Fig. 1

(a) Fluorescence intensity images of GPax (top) and GPax/FATmCh (bottom cells). (b) Images of the same cells in (a) segmented by intensity and arbitrarily colored to distinguish different intensity segments. Only pixels with LL1000 were selected for intensity segmentation. Scale bar is 10 μm.

Fig. 2

(a) Four representative variations instrument response (IR) functions were taken over a period of 2 min. A width variation of over 100 ps was seen. (b) Scaled IRs for two data sets. For the solid line plot of data set 1, the blue line is IRmsr(GaAs), the green line is IRext(GPax), and the red line is IRext(GPax / FATmCh). For the dashed line plot of data set 2, the blue line is IRext(flsn), the green line is IRext(GPax ), and the red line is IRext(GPax/FATmCh). The plot for data set 2 has been shifted to the right by 1.5 ns for clarity.

Fig. 3

(a) GPax cell segmented by intensity at LL1000. (b) Histogram of FR distribution of cell in (a), color-coded by cell intensity segment. A double-exponential fit was carried out to obtain the FR. Scale bar is 10 μm.

Fig. 4

Double-exponential fits were carried out with ${\tau }_{\mathit{D}}$ fixed at 2.6 ns for both cell and simulated data. (a) Lifetime decay data and the associated curve fits for five representative pixels of the LL100 data set (containing 1458 pixels in total) are shown. The normalized ${\chi }^2$ for fitting the whole data set is 0.26 without reweighting and is 0.67 with reweighting (see also Table 5). (b) Variation in fitted ${\tau }_F$ with varying LL. Left axis is fitted ${\tau }_F$ (circle) right axis is percentage difference in fitted ${\tau }_F$ compared to fitted ${\tau }_F$ at LL2000 (triangle). Open markers: two independent cell measurements; closed markers: simulated data. (c) Master curve for scaling ${\tau }_{\mathit{F}}$ fit to true ${\tau }_{\mathit{F}}$. Squares are percentage deviations from ${\tau }_{\mathit{F}}$ (LL2000) (500%dev2000 and 1000%dev2000), circles are percentage deviations from ${\tau }_{\mathit{F}}$ (LL1000) (500%dev1000), and triangles are percentage deviations from true ${\tau }_{\mathit{F}}$ (500%dev${\tau }_{\mathit{F}}$, 1000%dev${\tau }_{\mathit{F}}$, and 2000%dev${\tau }_{\mathit{F}}$). Dotted lines are fits of LL500 %dev points, dashed lines are fits of LL1000 %dev points, and solid line is a fit of LL2000 %dev points.

Fig. 5

(a) Scaled histogram of fit-truFR at a ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}$ ratio of 0.8 and varying LL values. Color scheme is as follows: LL100 (dark blue), LL500 (green), LL750 (red), LL1000 (yellow), LL1500 (magenta), and LL2000 (light blue). (b) Plot of true FR versus fit FR at a ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}$ ratio of 0.8 and LL1000. Red points: raw data. Plot is pseudocolored by data density; yellow: high density of data; green: low density of data, with a gradation for intermediate values. Solid line: line of best fit through data, dashed line: line with 0 $\mathit{y}$-intercept and slope of 1. (c) Slope (triangle) and intercept (circle) of true versus fitted FR plots at a ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}$ ratio of 0.8 and with varying LL. A best-fit line is drawn through the data points based on a third-degree spline fit (using polyfit function of MATLAB) to generate a master curve.

Fig. 6

(a) Scaled histogram of fit-truFR at LL1000 and varying ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}$ ratio. Color scheme is as follows: ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}=0.25$ (magenta), ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}=0.5$ (red), ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}=0.7$ (green), ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}=0.8$ (blue). (b) Slope and intercept of true versus fit FR plots at LL1000 and varying ${\tau }_{\mathit{F}}/{\tau }_{\mathit{D}}$ ratio. A best-fit spline is drawn through the data points.

Fig. 7

(a) Intensity image of GPax/FATmCh cell. (b) Cell image pseudocolored by fit FRET ratio. (c) Cell image pseudocolored by scaled FRET ratio. (d) Histogram of FRET ratio. Dotted line: fit FR, solid line: scaled FR. Scale bar is 10 μm.