Autofluorescence is a biomarker of neural stem cell activation state.

Neural stem cells (NSCs) must exit quiescence to produce neurons; however, our understanding of this process remains constrained by the technical limitations of current technologies. Fluorescence lifetime imaging (FLIM) of autofluorescent metabolic cofactors has been used in other cell types to study shifts in cell states driven by metabolic remodeling that change the optical properties of these endogenous fluorophores. Using this non-destructive, live-cell, and label-free strategy, we found that quiescent NSCs (qNSCs) and activated NSCs (aNSCs) have unique autofluorescence profiles. Specifically, qNSCs display an enrichment of autofluorescence localizing to a subset of lysosomes, which can be used as a graded marker of NSC quiescence to predict cell behavior at single-cell resolution. Coupling autofluorescence imaging with single-cell RNA sequencing, we provide resources revealing transcriptional features linked to deep quiescence and rapid NSC activation. Together, we describe an approach for tracking mouse NSC activation state and expand our understanding of adult neurogenesis.

3D convolutional neural networks predict cellular metabolic pathway use from fluorescence lifetime decay data.

Fluorescence lifetime imaging of the co-enzyme reduced nicotinamide adenine dinucleotide (NADH) offers a label-free approach for detecting cellular metabolic perturbations. However, the relationships between variations in NADH lifetime and metabolic pathway changes are complex, preventing robust interpretation of NADH lifetime data relative to metabolic phenotypes. Here, a three-dimensional convolutional neural network (3D CNN) trained at the cell level with 3D NAD(P)H lifetime decay images (two spatial dimensions and one time dimension) was developed to identify metabolic pathway usage by cancer cells. NADH fluorescence lifetime images of MCF7 breast cancer cells with three isolated metabolic pathways, glycolysis, oxidative phosphorylation, and glutaminolysis were obtained by a multiphoton fluorescence lifetime microscope and then segmented into individual cells as the input data for the classification models. The 3D CNN models achieved over 90% accuracy in identifying cancer cells reliant on glycolysis, oxidative phosphorylation, or glutaminolysis. Furthermore, the model trained with human breast cancer cell data successfully predicted the differences in metabolic phenotypes of macrophages from control and POLG-mutated mice. These results suggest that the integration of autofluorescence lifetime imaging with 3D CNNs enables intracellular spatial patterns of NADH intensity and temporal dynamics of the lifetime decay to discriminate multiple metabolic phenotypes. Furthermore, the use of 3D CNNs to identify metabolic phenotypes from NADH fluorescence lifetime decay images eliminates the need for time- and expertise-demanding exponential decay fitting procedures. In summary, metabolic-prediction CNNs will enable live-cell and in vivo metabolic measurements with single-cell resolution, filling a current gap in metabolic measurement technologies.

Label-free spatially maintained measurements of metabolic phenotypes in cells.

Metabolic reprogramming at a cellular level contributes to many diseases including cancer, yet few assays are capable of measuring metabolic pathway usage by individual cells within living samples. Here, autofluorescence lifetime imaging is combined with single-cell segmentation and machine-learning models to predict the metabolic pathway usage of cancer cells. The metabolic activities of MCF7 breast cancer cells and HepG2 liver cancer cells were controlled by growing the cells in culture media with specific substrates and metabolic inhibitors. Fluorescence lifetime images of two endogenous metabolic coenzymes, reduced nicotinamide adenine dinucleotide (NADH) and oxidized flavin adenine dinucleotide (FAD), were acquired by a multi-photon fluorescence lifetime microscope and analyzed at the cellular level. Quantitative changes of NADH and FAD lifetime components were observed for cells using glycolysis, oxidative phosphorylation, and glutaminolysis. Conventional machine learning models trained with the autofluorescence features classified cells as dependent on glycolytic or oxidative metabolism with 90%-92% accuracy. Furthermore, adapting convolutional neural networks to predict cancer cell metabolic perturbations from the autofluorescence lifetime images provided improved performance, 95% accuracy, over traditional models trained via extracted features. Additionally, the model trained with the lifetime features of cancer cells could be transferred to autofluorescence lifetime images of T cells, with a prediction that 80% of activated T cells were glycolytic, and 97% of quiescent T cells were oxidative. In summary, autofluorescence lifetime imaging combined with machine learning models can detect metabolic perturbations between glycolysis and oxidative metabolism of living samples at a cellular level, providing a label-free technology to study cellular metabolism and metabolic heterogeneity.

Perfusability and immunogenicity of implantable pre-vascularized tissues recapitulating features of native capillary network.

Vascularization is a key pre-requisite to engineered anatomical scale three dimensional (3-D) constructs to ensure their nutrient and oxygen supply upon implantation. Presently, engineered pre-vascularized 3-D tissues are limited to only micro-scale hydrogels, which meet neither the anatomical scale needs nor the complexity of natural extracellular matrix (ECM) environments. Anatomical scale perfusable constructs are critically needed for translational applications. To overcome this challenge, we previously developed pre-vascularized ECM sheets with long and oriented dense microvascular networks. The present study further evaluated the patency, perfusability and innate immune response toward these pre-vascularized constructs. Macrophage-co-cultured pre-vascularized constructs were evaluated in vitro to confirm micro-vessel patency and perturbations in macrophage metabolism. Subcutaneously implanted pre-vascularized constructs remained viable and formed a functional anastomosis with host vasculature within 3 days of implantation. This completely biological pre-vascularized construct holds great potential as a building block to engineer perfusable anatomical scale tissues.

Comparison of phasor analysis and biexponential decay curve fitting of autofluorescence lifetime imaging data for machine learning prediction of cellular phenotypes.

Introduction: Autofluorescence imaging of the coenzymes reduced nicotinamide (phosphate) dinucleotide (NAD(P)H) and oxidized flavin adenine dinucleotide (FAD) provides a label-free method to detect cellular metabolism and phenotypes. Time-domain fluorescence lifetime data can be analyzed by exponential decay fitting to extract fluorescence lifetimes or by a fit-free phasor transformation to compute phasor coordinates. Methods: Here, fluorescence lifetime data analysis by biexponential decay curve fitting is compared with phasor coordinate analysis as input data to machine learning models to predict cell phenotypes. Glycolysis and oxidative phosphorylation of MCF7 breast cancer cells were chemically inhibited with 2-deoxy-d-glucose and sodium cyanide, respectively; and fluorescence lifetime images of NAD(P)H and FAD were obtained using a multiphoton microscope. Results: Machine learning algorithms built from either the extracted lifetime values or phasor coordinates predict MCF7 metabolism with a high accuracy (∼88%). Similarly, fluorescence lifetime images of M0, M1, and M2 macrophages were acquired and analyzed by decay fitting and phasor analysis. Machine learning models trained with features from curve fitting discriminate different macrophage phenotypes with improved performance over models trained using only phasor coordinates. Discussion: Altogether, the results demonstrate that both curve fitting and phasor analysis of autofluorescence lifetime images can be used in machine learning models for classification of cell phenotype from the lifetime data.

Evolution of cisplatin resistance through coordinated metabolic reprogramming of the cellular reductive state.

BACKGROUND: Cisplatin (CDDP) is a mainstay treatment for advanced head and neck squamous cell carcinomas (HNSCC) despite a high frequency of innate and acquired resistance. We hypothesised that tumours acquire CDDP resistance through an enhanced reductive state dependent on metabolic rewiring. METHODS: To validate this model and understand how an adaptive metabolic programme might be imprinted, we performed an integrated analysis of CDDP-resistant HNSCC clones from multiple genomic backgrounds by whole-exome sequencing, RNA-seq, mass spectrometry, steady state and flux metabolomics. RESULTS: Inactivating KEAP1 mutations or reductions in KEAP1 RNA correlated with Nrf2 activation in CDDP-resistant cells, which functionally contributed to resistance. Proteomics identified elevation of downstream Nrf2 targets and the enrichment of enzymes involved in generation of biomass and reducing equivalents, metabolism of glucose, glutathione, NAD(P), and oxoacids. This was accompanied by biochemical and metabolic evidence of an enhanced reductive state dependent on coordinated glucose and glutamine catabolism, associated with reduced energy production and proliferation, despite normal mitochondrial structure and function. CONCLUSIONS: Our analysis identified coordinated metabolic changes associated with CDDP resistance that may provide new therapeutic avenues through targeting of these convergent pathways.

POSEA: A novel algorithm to evaluate the performance of multi-object instance image segmentation.

Many techniques and software packages have been developed to segment individual cells within microscopy images, necessitating a robust method to evaluate images segmented into a large number of unique objects. Currently, segmented images are often compared with ground-truth images at a pixel level; however, this standard pixel-level approach fails to compute errors due to pixels incorrectly assigned to adjacent objects. Here, we define a per-object segmentation evaluation algorithm (POSEA) that calculates segmentation accuracy metrics for each segmented object relative to a ground truth segmented image. To demonstrate the performance of POSEA, precision, recall, and f-measure metrics are computed and compared with the standard pixel-level evaluation for simulated images and segmented fluorescence microscopy images of three different cell samples. POSEA yields lower accuracy metrics than the standard pixel-level evaluation due to correct accounting of misclassified pixels of adjacent objects. Therefore, POSEA provides accurate evaluation metrics for objects with pixels incorrectly assigned to adjacent objects and is robust for use across a variety of applications that require evaluation of the segmentation of unique adjacent objects.

Volumetric imaging of human mesenchymal stem cells (hMSCs) for non-destructive quantification of 3D cell culture growth.

The adoption of cell-based therapies into the clinic will require tremendous large-scale expansion to satisfy future demand, and bioreactor-microcarrier cultures are best suited to meet this challenge. The use of spherical microcarriers, however, precludes in-process visualization and monitoring of cell number, morphology, and culture health. The development of novel expansion methods also motivates the advancement of analytical methods used to characterize these microcarrier cultures. A robust optical imaging and image-analysis assay to non-destructively quantify cell number and cell volume was developed. This method preserves 3D cell morphology and does not require membrane lysing, cellular detachment, or exogenous labeling. Complex cellular networks formed in microcarrier aggregates were imaged and analyzed in toto. Direct cell enumeration of large aggregates was performed in toto for the first time. This assay was successfully applied to monitor cellular growth of mesenchymal stem cells attached to spherical hydrogel microcarriers over time. Elastic scattering and fluorescence lightsheet microscopy were used to quantify cell volume and cell number at varying spatial scales. The presented study motivates the development of on-line optical imaging and image analysis systems for robust, automated, and non-destructive monitoring of bioreactor-microcarrier cell cultures.

Autofluorescence lifetime imaging classifies human lymphocyte activation and subtype.

New non-destructive tools are needed to reliably assess lymphocyte function for immune profiling and adoptive cell therapy. Optical metabolic imaging (OMI) is a label-free method that measures the autofluorescence intensity and lifetime of metabolic cofactors NAD(P)H and FAD to quantify metabolism at a single-cell level. Here, we investigate whether OMI can resolve metabolic changes between human quiescent versus IL4/CD40 activated B cells and IL12/IL15/IL18 activated memory-like NK cells. We found that quiescent B and NK cells were more oxidized compared to activated cells. Additionally, the NAD(P)H mean fluorescence lifetime decreased and the fraction of unbound NAD(P)H increased in the activated B and NK cells compared to quiescent cells. Machine learning classified B cells and NK cells according to activation state (CD69+) based on OMI parameters with up to 93.4% and 92.6% accuracy, respectively. Leveraging our previously published OMI data from activated and quiescent T cells, we found that the NAD(P)H mean fluorescence lifetime increased in NK cells compared to T cells, and further increased in B cells compared to NK cells. Random forest models based on OMI classified lymphocytes according to subtype (B, NK, T cell) with 97.8% accuracy, and according to activation state (quiescent or activated) and subtype (B, NK, T cell) with 90.0% accuracy. Our results show that autofluorescence lifetime imaging can accurately assess lymphocyte activation and subtype in a label-free, non-destructive manner.

Label-free optical imaging of cell function and collagen structure for cell-based therapies.

Cell-based therapies harness functional cells or tissues to mediate healing and treat disease. Assessment of cellular therapeutics requires methods that are non-destructive to ensure therapies remain viable and uncontaminated for use in patients. Optical imaging of endogenous collagen, by second-harmonic generation, and the metabolic coenzymes NADH and FAD, by autofluorescence microscopy, provides tissue structure and cellular information. Here, we review applications of label-free nonlinear optical imaging of cellular metabolism and collagen second-harmonic generation for assessing cell-based therapies. Additionally, we discuss the potential of label-free imaging for quality control of cell-based therapies, as well as the current limitations and potential future directions of label-free imaging technologies.

Decellularized organ biomatrices facilitate quantifiable in vitro 3D cancer metastasis models.

Metastatic cancers are chemoresistant, involving complex interplay between disseminated cancer cell aggregates and the distant organ microenvironment (extracellular matrix and stromal cells). Conventional metastasis surrogates (scratch/wound healing, Transwell migration assays) lack 3D architecture and ECM presence. Metastasis studies can therefore significantly benefit from biomimetic 3D in vitro models recapitulating the complex cascade of distant organ invasion and colonization by collective clusters of cells. We aimed to engineer reproducible and quantifiable 3D models of highly therapy-resistant cancer processes: (i) colorectal cancer liver metastasis; and (ii) breast cancer lung metastasis. Metastatic seeds are engineered using 3D tumor spheroids to recapitulate the 3D aggregation of cancer cells both in the tumor and in circulation throughout the metastatic cascade of many cancers. Metastatic soil was engineered by decellularizing porcine livers and lungs to generate biomatrix scaffolds, followed by extensive materials characterization. HCT116 colorectal and MDA-MB-231 breast cancer spheroids were generated on hanging drop arrays to initiate clustered metastatic seeding into liver and lung biomatrix scaffolds, respectively. Between days 3-7, biomatrix cellular colonization was apparent with increased metabolic activity and the presence of cellular nests evaluated via multiphoton microscopy. HCT116 and MDA-MB-231 cells colonized liver and lung biomatrices, and at least 15% of the cells invaded more than 20 µm from the surface. Engineered metastases also expressed increased signatures of genes associated with the metastatic epithelial to mesenchymal transition (EMT). Importantly, inhibition of matrix metalloproteinase-9 inhibited metastatic invasion into the biomatrix. Furthermore, metastatic nests were significantly more chemoresistant (>3 times) to the anti-cancer drug oxaliplatin, compared to 3D spheroids. Together, our data indicated that HCT116 and MDA-MB-231 spheroids invade, colonize, and proliferate in livers and lungs establishing metastatic nests in 3D settings in vitro. The metastatic nature of these cells was confirmed with functional readouts regarding EMT and chemoresistance. Modeling the dynamic metastatic cascade in vitro has potential to identify therapeutic targets to treat or prevent metastatic progression in chemoresistant metastatic cancers.

In vivo fluorescence lifetime imaging of macrophage intracellular metabolism during wound responses in zebrafish.

The function of macrophages in vitro is linked to their metabolic rewiring. However, macrophage metabolism remains poorly characterized in situ. Here, we used two-photon intensity and lifetime imaging of autofluorescent metabolic coenzymes, nicotinamide adenine dinucleotide (phosphate) (NAD(P)H) and flavin adenine dinucleotide (FAD), to assess the metabolism of macrophages in the wound microenvironment. Inhibiting glycolysis reduced NAD(P)H mean lifetime and made the intracellular redox state of macrophages more oxidized, as indicated by reduced optical redox ratio. We found that TNFα+ macrophages had lower NAD(P)H mean lifetime and were more oxidized compared to TNFα- macrophages. Both infection and thermal injury induced a macrophage population with a more oxidized redox state in wounded tissues. Kinetic analysis detected temporal changes in the optical redox ratio during tissue repair, revealing a shift toward a more reduced redox state over time. Metformin reduced TNFα+ wound macrophages, made intracellular redox state more reduced and improved tissue repair. By contrast, depletion of STAT6 increased TNFα+ wound macrophages, made redox state more oxidized and impaired regeneration. Our findings suggest that autofluorescence of NAD(P)H and FAD is sensitive to dynamic changes in intracellular metabolism in tissues and can be used to probe the temporal and spatial regulation of macrophage metabolism during tissue damage and repair.

Identification of rare cell populations in autofluorescence lifetime image data.

Drug-resistant cells and anti-inflammatory immune cells within tumor masses contribute to tumor aggression, invasion, and worse patient outcomes. These cells can be a small proportion (<10%) of the total cell population of the tumor. Due to their small quantity, the identification of rare cells is challenging with traditional assays. Single cell analysis of autofluorescence images provides a live-cell assay to quantify cellular heterogeneity. Fluorescence intensities and lifetimes of the metabolic coenzymes reduced nicotinamide adenine dinucleotide and oxidized flavin adenine dinucleotide allow quantification of cellular metabolism and provide features for classification of cells with different metabolic phenotypes. In this study, Gaussian distribution modeling and machine learning classification algorithms are used for the identification of rare cells within simulated autofluorescence lifetime image data of a large tumor comprised of tumor cells and T cells. A Random Forest machine learning algorithm achieved an overall accuracy of 95% for the identification of cell type from the simulated optical metabolic imaging data of a heterogeneous tumor of 20,000 cells consisting of 70% drug responsive breast cancer cells, 5% drug resistant breast cancer cells, 20% quiescent T cells and 5% activated T cells. High resolution imaging methods combined with single-cell quantitative analyses allows identification and quantification of rare populations of cells within heterogeneous cultures.

Autofluorescence Imaging to Evaluate Cellular Metabolism.

Cellular metabolism is the process by which cells generate energy, and many diseases, including cancer, are characterized by abnormal metabolism. Reduced nicotinamide adenine (phosphate) dinucleotide (NAD(P)H) and oxidized flavin adenine dinucleotide (FAD) are coenzymes of metabolic reactions. NAD(P)H and FAD exhibit autofluorescence and can be spectrally isolated by excitation and emission wavelengths. Both coenzymes, NAD(P)H and FAD, can exist in either a free or protein-bound configuration, each of which has a distinct fluorescence lifetime-the time for which the fluorophore remains in the excited state. Fluorescence lifetime imaging (FLIM) allows quantification of the fluorescence intensity and lifetimes of NAD(P)H and FAD for label-free analysis of cellular metabolism. Fluorescence intensity and lifetime microscopes can be optimized for imaging NAD(P)H and FAD by selecting the appropriate excitation and emission wavelengths. Metabolic perturbations by cyanide verify autofluorescence imaging protocols to detect metabolic changes within cells. This article will demonstrate the technique of autofluorescence imaging of NAD(P)H and FAD for measuring cellular metabolism.

Extracellular pH affects the fluorescence lifetimes of metabolic co-factors.

SIGNIFICANCE: Autofluorescence measurements of the metabolic cofactors NADH and flavin adenine dinucleotide (FAD) provide a label-free method to quantify cellular metabolism. However, the effect of extracellular pH on flavin lifetimes is currently unknown. AIM: To quantify the relationship between extracellular pH and the fluorescence lifetimes of FAD, flavin mononucleotide (FMN), and reduced nicotinamide adenine dinucleotide (phosphate) [NAD(P)H]. APPROACH: Human breast cancer (BT474) and HeLa cells were placed in pH-adjusted media. Images of an intracellular pH indicator or endogenous fluorescence were acquired using two-photon fluorescence lifetime imaging. Fluorescence lifetimes of FAD and FMN in solutions were quantified over the same pH range. RESULTS: The relationship between intracellular and extracellular pH was linear in both cell lines. Between extracellular pH 4 to 9, FAD mean lifetimes increased with increasing pH. NAD(P)H mean lifetimes decreased with increasing pH between extracellular pH 5 to 9. The relationship between NAD(P)H lifetime and extracellular pH differed between the two cell lines. Fluorescence lifetimes of FAD, FAD-cholesterol oxidase, and FMN solutions decreased, showed no trend, and showed no trend, respectively, with increasing pH. CONCLUSIONS: Changes in endogenous fluorescence lifetimes with extracellular pH are mostly due to indirect changes within the cell rather than direct pH quenching of the endogenous molecules.

Carbomer-based adjuvant elicits CD8 T-cell immunity by inducing a distinct metabolic state in cross-presenting dendritic cells.

There is a critical need for adjuvants that can safely elicit potent and durable T cell-based immunity to intracellular pathogens. Here, we report that parenteral vaccination with a carbomer-based adjuvant, Adjuplex (ADJ), stimulated robust CD8 T-cell responses to subunit antigens and afforded effective immunity against respiratory challenge with a virus and a systemic intracellular bacterial infection. Studies to understand the metabolic and molecular basis for ADJ's effect on antigen cross-presentation by dendritic cells (DCs) revealed several unique and distinctive mechanisms. ADJ-stimulated DCs produced IL-1ß and IL-18, suggestive of inflammasome activation, but in vivo activation of CD8 T cells was unaffected in caspase 1-deficient mice. Cross-presentation induced by TLR agonists requires a critical switch to anabolic metabolism, but ADJ enhanced cross presentation without this metabolic switch in DCs. Instead, ADJ induced in DCs, an unique metabolic state, typified by dampened oxidative phosphorylation and basal levels of glycolysis. In the absence of increased glycolytic flux, ADJ modulated multiple steps in the cytosolic pathway of cross-presentation by enabling accumulation of degraded antigen, reducing endosomal acidity and promoting antigen localization to early endosomes. Further, by increasing ROS production and lipid peroxidation, ADJ promoted antigen escape from endosomes to the cytosol for degradation by proteasomes into peptides for MHC I loading by TAP-dependent pathways. Furthermore, we found that induction of lipid bodies (LBs) and alterations in LB composition mediated by ADJ were also critical for DC cross-presentation. Collectively, our model challenges the prevailing metabolic paradigm by suggesting that DCs can perform effective DC cross-presentation, independent of glycolysis to induce robust T cell-dependent protective immunity to intracellular pathogens. These findings have strong implications in the rational development of safe and effective immune adjuvants to potentiate robust T-cell based immunity.

Classification of T-cell activation via autofluorescence lifetime imaging.

The function of a T cell depends on its subtype and activation state. Here, we show that imaging of the autofluorescence lifetime signals of quiescent and activated T cells can be used to classify the cells. T cells isolated from human peripheral blood and activated in culture using tetrameric antibodies against the surface ligands CD2, CD3 and CD28 showed specific activation-state-dependent patterns of autofluorescence lifetime. Logistic regression models and random forest models classified T cells according to activation state with 97-99% accuracy, and according to activation state (quiescent or activated) and subtype (CD3+CD8+ or CD3+CD4+) with 97% accuracy. Autofluorescence lifetime imaging can be used to non-destructively determine T-cell function.

Fluorescence intensity and lifetime redox ratios detect metabolic perturbations in T cells.

The auto-fluorescent coenzymes reduced nicotinamide dinucleotide (NADH) and oxidized flavin adenine dinucleotide (FAD) allow label-free detection of cellular metabolism. The optical redox ratio, which is traditionally computed as the ratio of NADH and FAD intensities, allows quantification of cell redox state. In addition to multiple formulations of the optical redox ratio from NADH and FAD intensity measurements, a fluorescence lifetime redox ratio (FLIRR) based on the fractions of protein-bound NADH and FAD was developed to overcome the limitations of experimental factors that influence fluorescence intensity measurements. In this paper, we compare fluorescence-intensity computations of the optical redox ratio with the fluorescence lifetime redox ratio for quiescent and activated T cells. Fluorescence lifetime images of NAD(P)H and FAD of T cells were acquired with a two-photon fluorescence lifetime microscope. Metabolic perturbation experiments, including inhibition of glycolysis, oxidative phosphorylation, glutaminolysis, and fatty acid synthesis revealed differences between the intensity and lifetime redox ratios. Statistical analysis reveals that the FLIRR has a lower standard deviation and skewness (two-tail T-test, P value = 0.05) than the intensity redox ratio. Correlation analysis revealed a weak relationship between FLIRR and intensity redox ratio for individual cells, with a stronger correlation identified for activated T cells (Linear regression, R-value = 0.450) than quiescent T cells (R-value = 0.172). Altogether, the results demonstrate that while both the fluorescence lifetime and intensity redox ratios resolve metabolic perturbations in T cells, the endpoints are influenced by different metabolic processes.

Classifying T cell activity in autofluorescence intensity images with convolutional neural networks.

The importance of T cells in immunotherapy has motivated developing technologies to improve therapeutic efficacy. One objective is assessing antigen-induced T cell activation because only functionally active T cells are capable of killing the desired targets. Autofluorescence imaging can distinguish T cell activity states in a non-destructive manner by detecting endogenous changes in metabolic co-enzymes such as NAD(P)H. However, recognizing robust activity patterns is computationally challenging in the absence of exogenous labels. We demonstrate machine learning methods that can accurately classify T cell activity across human donors from NAD(P)H intensity images. Using 8260 cropped single-cell images from six donors, we evaluate classifiers ranging from traditional models that use previously-extracted image features to convolutional neural networks (CNNs) pre-trained on general non-biological images. Adapting pre-trained CNNs for the T cell activity classification task provides substantially better performance than traditional models or a simple CNN trained with the autofluorescence images alone. Visualizing the images with dimension reduction provides intuition into why the CNNs achieve higher accuracy than other approaches. Our image processing and classifier training software is available at https://github.com/gitter-lab/t-cell-classification.