Antibodies and Reagents
Recombinant HIF-2α antibody and HIF-1β antibodies were purchased from Abcam (catalogue numbers ab243861 and ab2771 respectively). Monoclonal antibodies mouse anti-PD-1, rabbit anti-PD-L1 were purchased from Abcam (catalogue numbers ab52587 and ab205921 respectively). AffiniPure F(ab')₂ fragment donkey anti-mouse IgG (H + L) and peroxidase AffiniPure F(ab')₂ fragment donkey anti-rabbit IgG (H + L) were purchased from Jackson Immuno Research (catalogue numbers 715-006-150 and 711-036-152 respectively). Pierce endogenous peroxidase suppressor, TSA SuperBoost kit and Prolong Glass antifade mount were purchased from Thermofisher Scientific (catalogue numbers 35000, B40925 and P36980 respectively). ATTO488 NHS ester, bovine serum albumin and rhodamine B, were purchased from Sigma Aldrich (catalogue numbers A2153-100G and 234141-10G respectively).
Two-site assay for HIF-1β/HIF-2α and PD-1/PD-L1
The TMAs underwent antigen retrieval process using the Envision Flex retrieval solution, pH 9. The Dako PT-Link system was utilised, where the slides were heated to 95oC for 20 minutes. Using a PAP pen, an aqueous-repelling border was outlined around each tissue fragment. Pierce endogenous peroxidase suppressor was then applied to each specimen, and the slides were left to incubate for 30 minutes at 21oC room within a humid-controlled environment.
The samples underwent two washes with PBS before being incubated for an hour at room temperature with 3% BSA (10 mg/ml). For the donor-only slides, they were incubated with either anti-PD-1 (at a dilution of 1:100) or anti-HIF-1β (at a dilution of 1:100). The donor-acceptor slides were treated with the following primary antibodies: αPD-1 (1:100), or αHIF-1β (1:100) and αPD-L1 (1:500), or αHIF-2α (1:100). Primary antibodies were incubated overnight at 4oC. Samples were washed with 0.02% PBST. The samples were treated with secondary F(ab’)2 fragments. F(ab’)2- ATTO488 (at a dilution of 1:100) was introduced to the donor-only slides. As for the donor-acceptor slides, they received both F(ab’)2-ATTO488 (1:100) and F(ab’)2- HRP (1:200). These samples were incubated, for 2 hours in the dark, at room temperature in a humidified container. After the incubation period slides were washed with 0.02% PBST. The donor-only slides were mounted with 1 drop of Prolong Glass antifade mount.
The donor-acceptor slides were subjected to tyramide signal amplification (TSA). The purpose of tyramide signal amplification to amplify the acceptor labelling, thus increase the signal-to-noise ratio and enhancing resonance energy transfer. This procedure is described in detail in Veeriah et al 2014 and Magraner Sanchez et al 2020. The antibodies were labelled with species-specific F(ab’)2 fragments which were conjugated to ATTO488 (donor chromophore, used to label the receptor primary antibody) or HRP (used to label the ligand primary antibody). Tyramide signal amplification was used to conjugate the acceptor chromophore (Alexa594) to the HRP labelling the ligand.
Time-Resolved immune FRET (iFRET) determined by frequency-domain FLIM.
The quantitative molecular imaging platform utilises a custom made semi-automated frequency-domain FLIM. The first slide (donor only) was excited by a modulated (40MHz) diode 473nm laser, and the lifetime of the donor alone recorded. The second slide was excited by the diode modulated 473nm laser and lifetime of the donor in the presence of the acceptor recorded. The reduction of donor lifetime (caused by resonance energy transfer) due to the presence of the acceptor reports on distances of 1-10nm and therefore acts as a “chemical ruler” enabling to quantify receptor-ligand and HIF protein interactions.
We identified the coincidence regions where both the donor and acceptor were observed. A total of 10 to 20 regions of interest (ROIs) were selected within this coincidence regions. Subsequently, the lifetimes, along with their corresponding standard deviations, were automatically obtained and exported to an Excel spreadsheet.
Photophysical parameters for quantification of protein interactive states.
As input the algorithm takes a lifetime image of donor in the presence of acceptor (tDA) and a lifetime image of the donor (tD), followed by calculation of reduction of
tDA compared to tD, which is reflected in a metric called FRET-efficiency:
\(Ef\) = [1 − <tDA>/<tD>]x100 Eq-1
FRET-efficiency is calculated as an average for each coincident region.
\(Ef\) is directly related to the distance between the donor and acceptor fluorophores (Atto488 and Alexa 594) (Eq-2), where r is the distance between Atto488 and Alexa594 in these experiments. 'R0' the Förster radius in this case is 5.83nm and it is the distance whereby the transfer efficiency is 50%.
\(Ef=\frac{{R}_{0}^{6}}{{R}_{0}^{6}+{r}^{6}}\) Eq-2
For each region of coincidence in the two-site assay, we computed the mean distance between the donor and the acceptor.
\(r= \sqrt[6]{{R}_{0}^{6}\text{*}\frac{(1-Ef)}{Ef}}\) Eq-3
The application generates a table that summarises the quantitative characteristics for each coincident region. This includes the label, area, average FRET efficiency, and the mean distance between the donor and acceptor for each region.
Each overlapping region is colour-coded based on a colourmap that reflects a range from 0 to 50% FRET efficiency. This colour scale is used to create a heatmap, which is mapped automatically on the donor expression image.
Statistics
To evaluate FRET-efficiencies in two distinct experimental configurations (PD-1/PD-L1 and HIF-1β/HIF-2α), box and whisker plots were employed to visually represent the data from each TMA independently. This approach aimed to facilitate obtaining the graphs and comparison of FRET efficiency distributions among the group of patient samples. Following this, the Mann-Whitney U test was utilised to statistically analyse and compare the FRET efficiencies between ccRCC samples and the normal tissue. The null hypothesis presumes no significant difference between the two groups, while the alternative hypothesis poses a significant difference. The implementation of the Mann-Whitney U test resulted in a p-value. This p-value signifies the likelihood of observing the differences in FRET efficiencies between the ccRCC and the adjacent normal tissue samples.
Computational Analysis of FuncOmap
The initial data are the images captured through a multiple frequency Fluorescence lifetime imaging microscopy. There are two problems to solve from a user perspective: one is how to process that data to provide actionable clinical information, the other is how to enable clinicians to explore the data so they can confirm their diagnosis. We solved the first by constructing a false-colour composite image from the donor and donor/acceptor images which allows the clinician to observe levels of protein expression and protein function on the same image. We solved the second by making the image interactive so that mousing over the image displays additional data (\(Ef\) and molecular distance (rnm), while a drop-down control permits exploration of thresholding to shrink or grow the areas of protein function according to the measured \(Ef\). From a computational perspective, the contributions are the choice and composition of the appropriate, validated library software, the construction of the image mapping between expressional and functional states and the spatial enhancement of the coincident image to highlight sites with significant levels of functional activation. The computational process, linked to the corresponding mathematics shown in the equations above, appears in Fig. 4A.
The computational approach outlined in this figure is a significant advance on the existing method because it is automated – hence reproducible - more precise and provides whole sample functional information. As noted above, the initial data capture provides a donor and a donor-acceptor image because of the optical configuration. For each pixel we record the lifetime in the control donor and the donor/acceptor image and hence calculate the FRET efficiency, which provides the data to carry out a thresholding operation across the image and hence identify all the regions where protein function is above noise level. We use the Otsu thresholding method [24] to convert the original grey level expression image into a binary one, where zeros represent the background and ones indicate regions where the proteins exist. To facilitate user exploration of the data, we provide a set of thresholds for user selection. After applying the threshold to the donor and donor/acceptor images, a logical “and” operation of the two pinpoints the coincident region where both proteins are present. Any pixels outside these overlapping regions are set to zero. Consequently, we identify the (image) coordinates of all coincident regions. Each region is assigned a label and the area it encompasses is calculated, represented as the total number of pixels within each specific coincident region. The “right” level for thresholding is an automatic function of the Otsu algorithm, which adapts to the image supplied to identify the best value to differentiate background from signal. This typically works out at a signal level 4x that of the background but results in images where contrast is too low for human visual perception. This is resolved by increasing the contrast through multiplication of the signal up to 3-fold (12). This factor only serves to assist human assessment of the images and does not affect the \(Ef\) or r (nm) information in any way.
In contrast, the previous computational approach uses human annotation of the image to identify coincidence, then generates spreadsheets for donor and donor lifetime in the presence of acceptor. From these changes in lifetime the FRET efficiencies, \(Ef\) were calculated (Eq-1) and box and whisker statistical distributions plotted. The box and whisker plots provide an average across all the regions, which hides critical spatial localisation of protein activation states. Thus, the drawbacks here are the use of time-consuming human annotation, with attendant inaccuracy (marking regions using a mouse) and incompleteness (some regions that should be marked are not, not least because they can be too small for reliable and consistent identification), the homogenisation of the regions through the analysis process and the inaccessibility (to most clinicians) of a box and whiskers plot as a interpretable diagnostic tool.
The critique above notes that small regions are likely to missed with the manual approach. Small regions remain a problem, in the new approach – not in their detection; the thresholding automatically handles that – but in their presentation to the clinician, since a single pixel, for example, is still too small for ready identification on screen or selection by mouse. For this case, we have introduced a spatial enhancement in the rendering process that aggregates the pixels around one with a high \(Ef\), in effect using homogenisation to reveal rather than hide information.
One of the main advantages of the new computational approach is that it allows for a one-step analysis, as opposed to multiple steps. In the new approach the lifetime changes, \(Ef\) and thresholding are all processed in an individual pixel. Furthermore, the computational complexity of the new approach is O(n). Each step is linear in the dimensions of the image, since lifetime, FRET efficiency and thresholding all process individual pixels, while the aggregation process for spatial enhancement examines the eight pixels surrounding an individual pixel. The entire process is coded as a Python notebook that generates a web page for the presentation of the protein expression and functional state interactive map (Fig. 4A), from where the clinician can examine the different regions of coincidence by moving a mouse pointer over the image. As illustrated in Fig. 4B, the region selected has a pop-up that shows the data associated with the region and at the same time highlights the row in the data table. The clinician can also explore the effect of choosing different thresholds through the drop-down immediately below the image.
We demonstrate the application of the framework by mapping HIF1β/HIF2α and
PD-1/PD-L1 functional states (interaction) states on expression image of HIF1β and PD-1, respectively. The dataset is partitioned into two primary directories, one for the PD-1/PD-L1 experiment and the other for the HIF1β/HIF2α experiment. Within each of these primary directories, seven separate sub-directories exist for each TMA: A1, A2, A5, B1, B5, and B7. These sub-directories contain image files, categorised into four types: Donor Expression, Donor Expression when an Acceptor is present, Donor Lifetime, and Donor Lifetime in the presence of an Acceptor.
FuncOmap allows for the visualisation of protein interaction states through a heatmap representation of average \(Ef\) within each coincidence region. By hovering over the heatmap, users can observe the \(Ef\) and distance r(nm) between proteins in each region of interest (ROI). Additionally, this information can be shown as a spreadsheet, alongside the image, in the browser. Since FuncOmap is implemented on Google Colab, there is no requirement for additional software setup. Google Colab provides a certain amount of memory at no charge and all calculations are performed using cloud-based computational resources.