Time-domain single photon-excited autofluorescence lifetime for label-free detection of T cell activation

Abstract

Fluorescence lifetime imaging microscopy (FLIM) is a powerful technique, capable of label-free assessment of metabolic state and function within single cells. FLIM measurements of autofluorescence were recently shown to be sensitive to functional state and subtype of T cells. Therefore, autofluorescence FLIM could improve cell manufacturing technologies for adoptive immunotherapy, which currently require a time-intensive process of cell labeling with fluorescent antibodies. However, current autofluorescence FLIM implementations are typically too slow, bulky, and prohibitively expensive for use in cell manufacturing pipelines. Here, we report a single photon-excited confocal whole-cell autofluorescence system that uses fast field-programmable gate array (FPGA)-based time tagging electronics to achieve time-correlated single photon counting (TCSPC) of single cell autofluorescence. The system includes simultaneous near-infrared bright-field imaging and is sensitive to variations in the fluorescence decay profile of the metabolic coenzyme NAD(P)H in human T cells due to activation state. Classification of activated and quiescent T cells achieved high accuracy and precision (ROC AUC = 0.92). The lower cost, higher acquisition speed, and resistance to pile up effects at high photon flux compared to traditional multiphoton-excited FLIM and TCSPC implementations with similar SNR makes this system attractive for integration into flow cytometry, sorting, and quality control in cell manufacturing.

T cells are a critical component of the adaptive immune response with subtypes that exhibit diverse functions in pro- and anti-inflammatory response. T cells from a patient can be collected, engineered in a manufacturing facility, and introduced back into the same patient for FDA-approved adoptive immunotherapies that target blood cancers, and for other immunotherapies that are in development [1].

Currently, T cell function and subtypes are determined on a single cell level from fluorescent antibody labels of surface protein expression using flow cytometry, live cell imaging, or immunohistochemistry. Label-free and nondestructive tools for assessment of T cell type and function during cell manufacturing could remove time-intensive labeling steps, improve assay consistency by removing reagent variability, and enable in-line monitoring of cell manufacturing. Such label-free monitoring technologies could ultimately improve the efficacy of adoptive immunotherapies through enhanced cell viability and enrichment of target cells for further modification and expansion. We have recently shown that the autofluorescence decay of NAD(P)H is sensitive to T cell subtype and activation state [2], which provides an attractive label-free method to improve cell manufacturing processes.

Among the current methods to measure fluorescence lifetime, time-correlated single photon counting (TCSPC) has the higher timing resolution and signal to noise ratio (SNR) needed to resolve small changes in complex multi-exponential autofluorescence decays that arise due to activation of T cells. TCSPC implementations of photon arrival timing include time to amplitude conversion (TAC) and time to digital conversion (TDC) schemes [3]. The analog TAC approach provides better accuracy, but its long dead time limits the maximum count rate. The digital TDC approach provides flexible integration into electronic chips and its short dead time allows high count rates. Both time-domain TAC [4] and TDC [5] as well as frequency-domain [6,7] approaches have been used for fluorescence lifetime cytometry. Time-resolved sampling of the detector signal using an ultrafast digitizer has been used for video-rate FLIM of NAD(P)H [8]. Here, we employ an FPGA-based digitizer time tagging device, the Time Tagger Ultra 8 (Swabian Instruments GmbH, Stuttgart, Germany) and characterize its performance for FLIM TCSPC by imaging a standard fluorescent sample on a commercial multiphoton microscope and comparing its timing performance to the industry standard SPC-150 card (Becker & Hickl GmbH, Berlin, Germany).

Fig. 1 illustrates lifetime measurements from a standard fluorescent sample, i.e., 10 mM solution of coumarin 6 in ethanol, using the two timing devices at different excitation laser power levels. The measurements were made on an Ultima two-photon imaging system (Bruker Inc., Middleton, WI, USA) built around an inverted Ti-E microscope (Nikon Instruments Inc., Melville, NY, USA) using a Chameleon Ultra II femtosecond-pulsed tunable Ti:Sapphire laser source (Coherent Inc., Santa Clara, CA, USA) and H7422PA-40 GaAsP photomultiplier tubes (Hamamatsu Corporation, Bridgewater, NJ, USA). Emission light was filtered using a 440/80 nm bandpass filter (Chroma). Images were acquired through a 40X/1.15NA Nikon objective lens with 3X optical zoom. The field of view (512x512 pixels; 100μm x 100μm) was scanned for 1 second with a 2 microsecond pixel dwell time. The laser wavelength was tuned to 750 nm and the excitation power was adjusted using a 350-80LA Pockels cell (Conoptics Inc., Danbury, CT, USA). The sample was imaged with average laser power levels ranging from 0.1 to 10 mW at the sample plane and the lifetime data were acquired separately using each timing device. Decays from all image pixels were aggregated into a single decay for each power level and displayed in Fig. 1(a–b) along with their photon count rates (Fig. 1c) and phasor [9] representations (Fig. 1d), color-coded for the excitation power level. Photon count rate and estimated lifetime for each decay is given in the legend. Instrument response function (IRF) was recorded from second harmonic generation (SHG) of urea crystals and used to correct the phasor plot. A single-exponential decay fit returns lifetimes between 2.4 and 2.5 ns, which agrees with reported values for coumarin 6 [10].

Fig. 1. Characterization of Time Tagger performance.

Two photon-excited fluorescence decay histogram of coumarin 6 at increasing excitation laser power levels (0.1 to 10 mW) using the SPC-150 and the Time Tagger Ultra 8 timing devices. (a) SPC-150 decays: photon pile-up affects the entire pulse period. (b) Time Tagger decays: only the early ~4 ns portion of the histogram skews at high photon count rates due to photon pile-up. (c) Time Tagger count rate increases quadratically with excitation power. (d) Drift in phasor representation of measured decays due to photon pile-up (color association is preserved in a, b, c, and d).

TAC modules such as those in the B&H SPC cards have a dead time of ~100 ns which translates to a theoretical maximum of 10 million photon counts per second (cps). However, to avoid photon pile-up (artifact due to missed detection of a secondary decay photon within the same excitation period), count rates should be kept well below this maximum [11]. On the other hand, the TDC-based Time Tagger Ultra 8 has a dead time of only 2 ns [12]. As such, its maximum sustained count rate is not limited by dead time, but rather by the data transfer bandwidth available between the device and the host computer where the decay is compiled. The Time Tagger Ultra 8 uses serial communication protocol, USB 3.0, which limits its sustained tag rate to ~65 million tags per second. Depending on user-defined acquisition settings, this bandwidth translates to photon count rates as high as 20-40 million cps (Mcps). Short bursts of photons at higher rates can be stored in the onboard buffer memory of the device, but long bursts at higher rates lead to data loss. The Time Tagger’s fluorescence photon count rate rises with the square of excitation power (Fig. 1c, R² = 0.99), which agrees with two-photon excitation theory. Photon pile-up has a distinct presentation on the Time Tagger decays. The histogram skews at high excitation powers, but due to the short dead time, the device can still sample the later time-points of the decay. Therefore, the pile-up effect is limited to the first ~4 ns of the decay (Fig. 1b, bracket). Another advantage of the Time Tagger is its ability to tag the decay photons at 100% duty cycle (i.e., the entire 12.5ns laser pulse period. Fig. 1b). In contrast, TAC modules have nonlinear portions at the beginning and end of their characteristic curve that must be avoided [11], which can result in truncated decay curves (Fig. 1a).

For our low-cost single-photon setup, we used a QuixX 375-70PS picosecond-pulsed diode laser (Omicron-Laserage Laserprodukte GmbH, Rodgau, Germany) that produces narrow 375 nm pulses (<50ps) with variable pulse repetition rates up to 100 MHz and adjustable output power. The schematic diagram of the system, built around an inverted Nikon Ti-S microscope, is given in Fig. 2. With a cytometry application in mind, a stationary excitation spot with roughly the same diameter as the T cells (< 10 μm) was implemented to achieve whole-cell excitation while minimizing background excitation. The laser beam was contracted to 1/3 and collimated using a Keplerian telescope lens pair with a magnifying power of 3 before being launched into the epi-illumination port of the microscope. A 100X 1.3NA oil-immersion objective lens (Nikon) was used to deliver the excitation light and collect autofluorescence. For detection of NAD(P)H fluorescence, a H10721P-210 alkali PMT (Hamamatsu) was used along with a 50μm pinhole to reject out of focus light (e.g., resuspension medium fluorescence, ambient light), and a 450/70 nm bandpass emission filter. The relay lens before the pinhole was placed such that the same cell-sized illuminated disk of the image plane would fill the face of the PMT photocathode. PMT output as well as the laser pulse sync output were routed to the Time Tagger device via shielded coaxial cables for timing of fluorescence photons. Simultaneous bright-field imaging of the sample was implemented by installing an 850 nm longpass filter in the dia-illumination light path (i.e., halogen lamp house) to minimize background noise in the NAD(P)H detection arm. Bright-field images were captured on an acA1300-200um CMOS camera (Basler Inc., Exton, PA, USA).

Fig. 2. Schematic diagram of the single photon-excited whole-cell NAD(P)H fluorescence lifetime detection system.

A confocal pinhole rejects out-of-focus light. The CMOS camera provides simultaneous near-infrared bright-field imaging. LP, longpass filter or dichroic beam splitter. Representative decays from T cells and medium fluorescence are plotted and the fitted bi-exponential decay model equations are given.

To test the sensitivity of the system, peripheral blood was obtained from a healthy human donor according to IRB-approved protocol (#2018-0103, UW-Madison) and T cells were isolated using the protocol described previously [2]. Half of the T cells were activated using a tetrameric antibody against CD2/CD3/CD28 surface ligands. Both activated and control (quiescent) dishes were cultured in ImmunoCult-XF T cell Expansion Medium (StemCell Technologies Inc., Vancouver, BC, Canada) for 72 hours before suspending in low-fluorescence medium (PBS + 2% FBS) commonly used for flow cytometry. Cells were then plated on No. 1.5 coverglass-bottom Poly-D-Lysine coated dishes and allowed to settle for 15 minutes. Autofluorescence decays were acquired from n=400 activated, and n=471 quiescent T cells by translating the dish atop the motorized PZ-2000 microscope stage (Applied Scientific Instrumentation Inc., Eugene, OR, USA) to expose the excitation beam to one cell at a time. The laser was pulsed at a repetition rate of 50 MHz (i.e., 20 ns interval) with an average power of 0.75 mW at the sample plane. The decays were compiled over an integration time of 200 ms per cell, resulting in median photon count rates of 0.71 and 0.26 Mcps for activated and quiescent T cell decays, respectively, while the PBS+FBS medium alone and red blood cells (hemoglobin fluorescence) had median count rates of 0.1 and 0.45 Mcps, respectively. Note that the pinhole removed most fluorescence from the medium when a cell occupied the focal volume. Representative decays from an activated and a quiescent T cell as well as the PBS+FBS medium are shown in Fig. 2. The IRF was estimated from the decay of a red blood cell (FWHM 350ps) and validated by recovery of a standard coumarin 6 lifetime. A constant fraction (0.30) of the medium decay was subtracted from the measured T cell decays (to account for medium fluorescence not rejected by the pinhole) and the result was deconvolved from the IRF. Next, the decays were fit, via nonlinear least squares optimization, to a bi-exponential model $\left[I\left(t\right)={\alpha }_0+{\alpha }_1{e}^{-t/{\tau }_1}+{\alpha }_2{e}^{-t/{\tau }_2}\right]$ using built-in Matlab function ‘fit’ to determine lifetime fit parameters for each cell including the mean lifetime $\left[{\tau }_m=\left({\alpha }_1{\tau }_1+{\alpha }_2{\tau }_2\right)/\left({\alpha }_1+{\alpha }_2\right)\right]$.

Results are shown in Fig. 3(a). Activated T cells had lifetime parameters (mean ± std) of τ_m = 724 ± 99ps, α₁% = 86% ± 3%, τ₁ = 422 ± 61ps, τ₂ = 2635 ± 268ps, while quiescent T cells presented with τ_m = 738 ± 173ps, α₁% = 83% ± 3%, τ₁ = 342 ± 87ps, τ₂ = 2767 ± 309ps. Activated and quiescent T cells have distinct distributions of the lifetime fit parameters, particularly the free NAD(P)H lifetime (τ₁) and fractional contribution (α₁%), in agreement with previously reported two-photon FLIM results with a SPC-150 card [2]. Additionally, the same separation can be directly observed in the phasor representation of the decays as seen in Fig. 3(b). Finally, logistic regression classifiers using a logistic binomial model and a logit link function were trained, via Matlab’s built-in ‘fitglm’, on half of the data, randomly selected from the activated and quiescent groups, and tested on the other half to quantify classification performance using either fit-derived lifetime parameters (τ₁, τ₂, α₁%, τ_m) or fit-free phasor parameters (G, S) as features. The resulting receiver operating characteristic (ROC) curves shown in Fig. 3(c) suggest that using phasor coordinates as classifier features can separate the two populations well (AUC=0.912), matching the performance of lifetime parameters (AUC=0.919). Dish condition was assumed to be the ground truth for training and testing purposes. To validate this assumption, both activated and quiescent dishes were stained using a PerCP-conjugated antibody (BioLegend Inc., San Diego, CA, USA) against the early activation surface marker, CD69, and imaged on a wide field epi-fluorescence microscope (Nikon Ti2). The images, shown in Fig. 4, confirm that over 90% of the T cells in the activated dish and almost none of the T cells in the quiescent dish express CD69.

Fig. 3. Single-photon excitation and Time Tagger detection resolves NAD(P)H lifetime changes due to activation of primaiy human T cells.

Measurements were acquired from T cells in low-fluorescence PBS+FBS medium that is commonly used for flow cytometry. (a) Distributions of cell NAD(P)H fluorescence lifetime parameters show separation of activated and quiescent T cells, particularly in short NAD(P)H lifetime (τ₁) and the fractional contribution of the short NAD(P)H lifetime (α₁%). ****p<0.0001, determined using two-sided unpaired t-tests. Horizontal lines show the median (solid) and the interquartile range (dashed). (b) Phasor representation of autofluorescence decays visualizes the separation of the two T cell populations, along with the PBS+FBS medium alone and red blood cells. (c) ROC curves show that logistic regression classifiers trained with fit-free phasor parameters can robustly separate activated (positive) and quiescent (negative) T cells with high sensitivity and specificity and match the performance of classifiers trained with lifetime decay parameters obtained from curve fitting.

Fig. 4. Epi-fluorescence images of immunolabeled activated and quiescent T cells.

PerCP (red, Ex: 480/40 nm, Em: 720/60 nm) staining of the activation surface marker CD69 in a majority of the T cells in the activated dish (left) and almost none of the T cells in the quiescent dish (right) along with NAD(P)H autofluorescence (blue, Ex: 378/52 nm, Em: 447/60 nm).

We presented a single photon-excited whole-cell autofluorescence lifetime system using a fast FPGA-based time tagging device for efficient photon counting. Compared to multiphoton FLIM systems, our 1P system combines significantly reduced cost (~20-fold) and footprint (~40-fold) with increased maximum count rate (~5-fold) (see specifics at [13]). The sensitivity of this system was validated by resolving NAD(P)H lifetime changes due to activation in primary human T cells that were previously only resolved using multiphoton FLIM. Here, we opted to use a relatively wide excitation beam waist and large pinhole to greatly improve fluorescence photon count rate via whole-cell excitation and detection. While we only used 0.75 mW excitation power and achieved ~1 Mcps in emission, higher count rates (10× or more) can be expected in fast flow cytometry settings where higher laser powers (20-100 mW) are used. We note that the previous multiphoton FLIM study included more cells in training (n=4131) and test (n=696) sets, from more donors (n=6) with same-cell CD69 validation in the test set to achieve a higher AUC (0.97) [2]. However, our current feasibility study indicates encouraging performance (AUC=0.92) with a simplified single-photon system. This low-cost alternative can integrate with flow cytometry and sorting equipment to create accessible label-free metabolic assessment technologies for cell manufacturers. Such direct metabolic readouts help reduce dependence on labor-intensive and variable reagent-based assays. The non-invasive nature of this autofluorescence measurement means it could be integrated in closed-loop bioreactors for continuous monitoring and quality control of cell products, in compliance with good manufacturing practice (GMP) requirements.