FILM: Mapping organellar metabolism by mid-infrared photothermal modulated fluorescence

Abstract

Metabolism unfolds within specific organelles in eukaryotic cells. Lysosomes are highly metabolically active organelles, and their metabolic states dynamically influence signal transduction, cellular homeostasis, and organismal physiopathology. Despite the significance of lysosomal metabolism, a method for its in vivo measurement is currently lacking. Here, we report optical boxcar-enhanced, Fluorescence-detected mid-Infrared photothermaL microscopy (FILM), together with AI-assisted data denoising and spectral deconvolution, to map metabolic activity and composition of individual lysosomes in living cells and organisms. Using this method, we uncovered lipolysis and proteolysis heterogeneity across lysosomes within the same cell, as well as early-onset lysosomal dysfunction during organismal aging. Additionally, we discovered organelle-level metabolic changes associated with diverse lysosomal storage diseases. This method holds the broad potential to profile metabolic fingerprints of individual organelles within their native context and quantitatively assess their dynamic changes under different physiological and pathological conditions, providing a high-resolution chemical cellular atlas.

Metabolism is essential for biological systems to sustain their physiological activities. Dysfunction in metabolism marks a transition into a pathological state, contributing to a variety of diseases^{1, 2}. In multicellular organisms, metabolism is compartmentalized at multiple levels, from organelles to tissues, which ensures specificity and efficiency. Thus, understanding the spatial organization of metabolism is crucial for both biological and biomedical research. However, this remains technically challenging, particularly at the scale of organelles. Organelles are fundamental structural and functional units within eukaryotic cells, each specializing in distinct metabolic processes. Recent advances in organelle-specific immunoprecipitation have facilitated the enrichment of specific organelles from different tissues for mass-spectrometry based metabolic profiling ^3–6. However, spatial information within the organelles’ native cellular context is lost during this process. On the other hand, microscopic imaging of fluorescence-labelled organelles has revealed their structural organization, dynamics and heterogeneity in vivo ^7–9; however, even with specific metabolite sensors, it provides limited insight into their metabolic complexity ^10–12.

Infrared (IR) absorption spectroscopy, with the ability to simultaneously fingerprint a wide range of molecules based on their vibrational signatures, has been used in drug response studies¹³ and metabolic tracking analyses¹⁴. Mid-infrared photothermal (MIP) microscopy, which integrates IR absorption with photothermal effects, has further advanced IR spectroscopic imaging by achieving submicron resolution^15–17 and enabling chemical imaging in living cells^18–20. Leveraging thermal sensitivity of fluorescent reporters^{21, 22}, Fluorescence-detected mid-Infrared photothermaL Microscopy (FILM, also called F-MIP) enables organelle-level imaging of certain molecules^23–25. However, the reported point-scan FILM^{23, 24} suffers from significant photobleaching and requires an exposure time three orders of magnitude longer than conventional fluorescence microscopy to capture a full fingerprint spectrum. This limitation makes it nearly impossible to comprehensively map metabolic activity in vivo across a broad spectral range and significantly hinders in-depth mapping of organellar complexity in relation to their metabolic states.

Here, we developed an optical boxcar demodulation scheme, together with a synchronized IR-visible laser scanner and an artificial intelligence (AI)-assisted self-supervised hyperspectral denoiser, to simultaneously enhance the detected photothermal signal, reduce fluorescence exposure time, and mitigate photobleaching. The FILM signal is followed by an unmixing algorithm to quantify biomolecular contents. This upgraded FILM system enables hyperspectral imaging of organelles in the entire fingerprint window (1000 to 1800 cm⁻¹). We have applied this system for in vivo metabolic profiling of lysosomes—organelles that play vital roles in nutrient sensing, macromolecular recycling, signaling transduction, diseases, and aging^26–29. We discovered that individual lysosomes exhibit distinctive metabolic fingerprints even within the same cell, and their functions undergo significant changes during physiological aging and under various disease conditions. Our approach offers a quantitative framework for profiling organelle-level metabolic heterogeneity and capturing their dynamic changes under different physiological and pathological conditions. This method is also compatible for multi-modal imaging with point-scan fluorescence microscopy, to investigate both structural and functional coordination of organelles within their native cellular environment.

Results

Optical-boxcar enhanced fluorescence-detected mid-infrared photothermal microscope (FILM)

FILM bridges fluorescence emission with IR absorption by probing photothermal changes across a broad range of IR wavelengths, which correlate with chemical information of biomolecules. As illustrated in Fig. 1a, mid-infrared photons excite the vibrational modes of target molecules, which is subsequently relaxed into heat. The resulting local temperature increase enhances the nonradiative relaxation process of nearby fluorescent probe molecules, reducing their quantum yield for fluorescence emission via dynamic quench. This modulation in fluorescence intensity serves as an indicator of IR absorption. In the previous point-scan system ^{23, 24}, a continuous-wave (CW) visible laser is used to excite fluorescent molecules (Fig. 1b). The fluorescence intensity changes are then demodulated using a lock-in amplifier (LIA) at the IR repetition rate. However, due to the low duty-cycle nature of the pulsed photothermal process, most fluorescence photons contribute to shot noise rather than photothermal signal. More importantly, prolonged CW exposure leads to significant photobleaching.

Figure 1. Development of FILM.

a, Principle of fluorescence detected mid-infrared photothermal microscopy depicted by energy diagram. Mid-IR absorption induces vibrational energy transitions in biochemical molecules, resulting in localized photothermal (PT) effects and a temperature increase. This rise in temperature accelerates the dynamic quenching of fluorophores, leading to a reduction in fluorescence emission. b, The previous continuous-wave (CW) fluorescence excitation schematic recorded the entire IR-induced PT dynamics. c, The optical boxcar schematic selectively recorded the ‘hot’ and ‘cold’ states to remove non-contributing photons, thereby mitigating photobleaching. d, Schematic of the experimental setup for the FILM microscope. M: reflection mirrors; GM: galvo mirrors; CM: concave mirrors; SL: scan lens; TL: tube lens; DM: dichroic mirror; Obj: objective; R obj: reflective objective; F: filter; SiPM: silicon photomultiplier. e, Photobleaching curves of standard fluorescence beads (n=3) under different excitation duty cycles. Shaded area indicates the standard deviation (s.d.) of photobleaching measurements. f, FILM signal of Shigella flexneri expressing GFP measured with different duty cycles visible light (n=5). Statistical data are presented as mean ± s.d. g, FILM images of S. aureus at 1650 cm⁻¹ and 1780 cm⁻¹. h, FILM spectrum of single S. aureus. a.u., arbitrary units. Scale bar: 10 μm.

To overcome these limitations, we implemented frequency-demodulated optical boxcar detection to eliminate photons that do not contribute to the photothermal signal, thereby reducing photobleaching (Fig. 1c). In this approach, in addition to the demodulation at the IR pump frequency, a pair of pulsed visible probes, synchronized with IR excitation, was used to selectively gate emission events at the peak of the temperature rise, referred to as the “hot” state, and the subsequent cooled “cold” state. Thus, the fluorescent excitation dose was reduced by chopping the light into pulses, mitigating photobleaching. The system is depicted in Fig. 1d (detailed in Methods). To achieve high-speed imaging and further minimize photobleaching, laser scanning was implemented, operating at 30 microseconds per pixel. A multichannel pulse generator synchronized the triggering of the IR laser at a repetition rate f_IR of 200 kHz and the visible laser f_vis at 400 kHz. Fluorescence intensity was detected using a silicon photomultiplier (SiPM), and the output signal was directly demodulated at f_IR using LIA.

We first evaluated the level of photobleaching reduction upon modulating the duty cycle of the visible fluorescence excitation beam. To this end, we maintained a constant peak power for the visible laser while adjusting the pulse duration across different duty cycle configurations (Fig. 1e). We found that after 100 frames of scanning, the pulsed mode at 40% and 20% duty cycle reduced photobleaching to 29.3% and 16.6% of that observed in the CW mode, respectively (Fig. 1e). Next, we assessed the signal intensity level at different duty cycles. Strikingly, when the duty cycle was reduced from 100% (i.e., CW) to 30%, the signal remained within a “stable zone” highlighted by the yellow shadow (Fig. 1f). We ultimately selected a 30% duty cycle, corresponding to a 750 ns pulse duration, as it effectively recovered the same signal amplitude as the CW condition. Overall, optical-boxcar enhanced FILM reduces fluorescence excitation time by over 100 times compared to the point-scan system reported previously^{23, 24} (Supplementary Fig. 1).

Additionally, the pulsed excitation light served as a 2f carrier, shifting high odd-order harmonic signals into the demodulation frequency, thereby enhancing the signal amplitude compared to CW conditions (Extended Data Fig. 1, Supplementary Fig. 2 and detailed in Supplementary note 1). Moreover, we found that the two gating windows of the excitation light used in optical boxcar inherently function as a time-resolved measurement, making it less sensitive to slow heat diffusion background that universally exists in water environment (Extended Data Fig. 2 and detailed in Supplementary note 2).

Using the developed system, we performed hyperspectral FILM on Rhodamine 6G-labelled S. aureus (Fig. 1g). The hyperspectral dataset spans the mid-infrared range from 980 cm⁻¹ to 1780 cm⁻¹ across 160 frames. Minimal photobleaching was observed during acquisition, enabling high-quality spectral resolution at the single-bacterium level. In the resolved spectrum, distinct IR peaks corresponding to nucleic acids, Amide II, and Amide I from proteins were clearly identified (Fig. 1h). To validate the spectral fidelity of FILM, we compared bacterial spectra with those obtained using scattering-based MIP, as well as the optical-boxcar enhanced FILM spectrum of a DMSO solution stained with LysoSensor DND-189, benchmarked against ATR-FTIR measurements (Extended Data Fig. 3). Both comparisons confirmed that optical-boxcar enhanced FILM reliably fingerprints fluorescently labelled objects.

Hyperspectral FILM imaging of lysosomes and AI assisted data analysis

To image lysosomes with optical-boxcar enhanced FILM, we identified a lysosome-specific thermo-sensitive dye (LysoSensor DND-189), the fluorescence intensity of which decreased by 12% when the temperature increased by 10 Kelvin (Supplementary Fig. 3). After labeling lysosomes in intestinal cell of wide type C. elegans with this thermo-sensitive dye, we applied optical-boxcar enhanced FILM to capture lysosome-specific hyperspectral images (Supplementary Video 1, top-left). As the pixel integration time was reduced to 30 microseconds to avoid photobleaching, the photothermal signal exhibited a relatively low signal-to-noise ratio (SNR).

To restore the SNR without using long integration time, we harnessed a self-supervised deep learning denoising algorithm, Self-permutation Noise2Noise Denoising (SPEND)³⁰. This algorithm uses a single low-SNR image stack to produce two independent stacks of the same objects. A 3D U-Net then takes the pair as input and target to learn noise statistics and object priors directly, which can effectively eliminate the learning data bias introduced by photobleaching and information leakage between adjacent pixels due to point-scan imaging (Supplementary Fig. 4 and detailed in Supplementary note 3). To generate the image pair, SPEND (Fig. 2a, left panel) employs a stack permutation strategy, splitting raw data into odd and even slices along the ω dimension, which are alternately concatenated to form input-target pairs. This approach ensures independent measurements of the same field of view (FOV), enabling full utilization of signal and noise information for unbiased noise estimation. Once trained, the model can directly batch process the entire hyperspectral datasets with the same target objects and experimental conditions for downstream analysis.

Figure 2. AI-assisted FILM hyperspectral imaging and analysis.

a, Workflow of AI-assisted hyperspectral data analysis. The left panel represented a deep learning based self-supervised denoising algorithm, called Self-permutation Noise2Noise Denoising (SPEND). The raw noisy hyperspectral data were first rearranged into two different sequences with permutation process. Next, the two sets of noisy data were served as the input and target for a U-net training. The trained network was then applied to denoise raw hyperspectral data. In the schematic, concat denotes concatenate, conv(+BN)+ReLU indicates convolution followed by batch normalization and rectified linear unit activation, max pool represents max pooling, up-conv refers to up-convolution (transposed convolution), and conv denotes convolution. The right panel represented the ratiometric analysis and MCR-LASSO spectral unmixing process. Reference spectrum of pure chemicals, acquired with the same instrument, were modified with augmented MCR based on the lysosomal data and then fed to LASSO for spectral unmixing and quantification. b, The comparison of FILM images of lysosomes acquired with IR at 1711 cm⁻¹ and 1797 cm⁻¹ before and after SPEND denoising. c, Intensity profiles along the red dotted lines marked in b. d, The comparison of raw FILM spectrum without calibration before and after SPEND processing. e, Quantification of image SNR and spectral SNR before and after SPEND denoising (n=13). f, LASSO unmixing with unmodified references and comparison of original calibrated input and reconstructed spectrum (n=13). g, LASSO unmixing with MCR modified references and comparison of original calibrated input and reconstructed spectrum (n=13). h, The comparison of cosine similarity and Euclidean distance with and without augmented MCR modification (n=13). Scale bar: 10 μm. In e, f, g, and h, the boxes show the interquartile range (IQR), the centerlines indicate medians and the lines outside the boxes extend to 1.5 times the IQR.

As we verified in Fig. 2b, when the IR was set to 1711 cm⁻¹, lysosomes were visible in the raw FILM image, whereas they disappeared at 1797 cm⁻¹, indicating an off-resonance condition. The chemical selectivity of FILM imaging is further corroborated by the complete set of wavenumber frames available in the Supplementary Fig. 5. For comparison, both images without SPEND denoising exhibited much higher noise level than those after SPEND processing. The intensity profile along the red-dotted line marked in Fig. 2b shows that lysosomes with weaker signals, indicated by orange arrowheads, were nearly obscured by noise fluctuations (Fig. 2c). After SPEND processing, noise in non-lysosomal regions was almost eliminated (Fig. 2b and Supplementary Video 1, top-right). The intensity profile demonstrates substantial noise suppression, allowing the weak signal marked by the orange arrow to be clearly resolved (Fig. 2c), which were further validated by hyperspectral stack projection and DC fluorescence images (Supplementary Fig. 6). Beyond the spatial (x-y) domain, we also assessed noise reduction in the spectral (ω) dimension. In the heatmap of 13 extracted single-lysosomal hyperspectral, SPEND processing resulted in a noticeably smoother spectral profile (Fig. 2d). A comparison of the fifth spectrum (marked with red) further highlights the reduction in frame-to-frame noise (Fig. 2d). Quantitative analysis demonstrates that SPEND enhances the image SNR by 26.9× and the spectral SNR by 5.3× (Fig. 2e and detailed in Supplementary note 4). Compared to other denoising methods, SPEND exhibited the best image and spectral SNR for analyzing FILM data (Extended Data Fig. 4).

Following baseline correction, power normalization, and spectral internal normalization (Methods), we successfully extracted the fingerprint spectrum of single lysosomes. The spectral data can be used for direct ratio-metric analysis of chemical groups with defined IR peaks. However, to quantitatively analyze multiple biomolecular contents in a systematic manner, spectral unmixing is required for decomposition (Fig. 2a, right panel). To this end, we first extracted calibrated lysosomal spectra to construct a new matrix where each row represented a lysosomal spectrum. We then carried out least absolute shrinkage and selection operator (LASSO) ^{31, 32} based spectral decomposition using eight reference spectra acquired on the same system from corresponding pure standards (inset of Fig. 2a; enlarged in Supplementary Fig. 7). The rationale for the selection of these standards and their chemical information were detailed in the Methods and Supplementary Table 1. We evaluated the accuracy of unmixing by comparing the input spectrum with ones reconstructed from the quantitative results of eight compositions as shown in Fig. 2f. We noticed that the reconstructed spectrum shows several mismatches indicated by arrows, when compared with the original input (Fig. 2f). This is due to the chemical complexity of lysosomes that contain over 600 components³³.

To improve the unmixing performance, we introduced Multivariate Curve Resolution (MCR)³⁴ prior to LASSO unmixing to refine the reference spectra based on the actual lysosomal data. MCR utilizes an alternating least squares strategy to iteratively update both the reference spectra and concentration maps. To avoid generating physically unexplainable reference, we implemented an augmented MCR strategy to stabilize the spectra update via incorporating reference spectra as part of the fitting dataset. The modified spectra obtained through this process then served as the input reference for LASSO. The quantitative results from unmixing the lysosomal spectrum and its reconstructed fit using MCR-LASSO are shown in Fig. 2g. Further analysis of the difference between the fitted and actual spectra using cosine similarity and Euclidean distance confirmed that the MCR-corrected reference spectra yield superior spectral fitting, ensuring accurate quantification (Fig. 2h and detailed in Supplementary note 5). Together, these AI-driven analyses enabled us to quantitatively measure biomolecular contents within individual lysosomes and compare them between conditions.

FILM reveals hydrolytic heterogeneity of lysosomes

Using the AI-assisted FILM system, we imaged lysosomes in live C. elegans (Supplementary Video 2). Based on the fluorescence image (Fig. 3a) that localizes individual lysosomes, hyperspectral FILM imaging unveiled a previously unseen chemical dimension of these organelles through photothermal IR spectra (Fig. 3b). For the first time, we observed that the lysosomes (Fig. 3a, circles) exhibit distinct spectral features compared to the surrounding regions (Fig. 3a, rectangle) visualized by autofluorescence (Fig. 3b, orange vs. blue shadow) (Supplementary Fig. 8 and Supplementary note 6). Spectral phasor analysis enabled clear segmentation of lysosomes from surrounding regions (Supplementary Fig. 9). Strikingly, the spectra varied among different lysosomes (Fig. 3b), indicating a highly heterogeneous lysosomal population even within the same cell. By comparing the photothermal IR spectrum with that of the dye itself and the evaluated dye concentration in the lysosome based on fluorescence intensity, we determined that the lysosomal features were not derived from the dye (Supplementary Fig. 10). By comparing the spectra of lysosomes and their surrounding regions, we noticed two lysosomal characteristic peaks around 1587 cm⁻¹ and 1711 cm⁻¹ and tentatively assigned them as amino acids (AA) and free fatty acids (FFA) by examining the FILM spectra of standard mixtures, including proteins, amino acids (AA), triglycerides (TAG, lipid ester) and FFA (Supplementary Fig. 8c and Supplementary note 7). We further validated the peak assignment of AA and FFA by inactivating CTNS-1 transporter that exports cysteine from the lysosome³⁵ and by overexpressing LIPL-4 lipase that hydrolyzes triglyceride into FFA within the lysosomal lumen³⁶. As shown in Extended Data Fig. 5a–b, compared to their controls, CTNS-1 inactivation (increased AA accumulation) resulted in a modest but statistically significant increase at 1587 cm⁻¹, whereas LIPL-4 upregulation (increased FFA production) yielded a significant increase at 1711 cm⁻¹. These results from modifying lysosomal activities provide independent support for the assignment of the 1587 cm⁻¹ and 1711 cm⁻¹ feature to AA and FFA, respectively. Our results revealed that the spectrum of surrounding regions shows typical protein and lipid ester characteristics, while the lysosomal spectrum exhibits a higher presence of AA and FFA, which are consistent with the active hydrolytic function of lysosomes^{37, 38} and thus support the ability of FILM to specifically profile the metabolic composition of lysosomes in vivo.

Figure 3. Hydrolytic heterogeneity of lysosomes revealed by FILM.

a, Fluorescence image of C. elegans labelled with LysoSensor DND189. b, FILM spectra of individual lysosomes and surrounding region marked in a. c, Ratio-metric mapping of intensity ratios at 1587/1649 cm⁻¹ (proteolytic activity) and 1711 /1741 cm⁻¹ (lipolytic activity), representing proteolysis activity and lipolysis activities, respectively. d, Classification of lysosomal subpopulations based on the two ratios shown in c. e, Classification of lysosomal subpopulations of mammalian lysosomes. Scale bar: 10 μm. Representative results are shown from three independent experiments.

Based on the distinctive IR features of macromolecules and their degraded small units, FILM imaging of lysosomes offers a unique opportunity to measure hydrolytic activities of lysosomes in living cells and organisms. To this end, we calculated the ratio of 1587 cm⁻¹ and 1649 cm⁻¹, which are derived from IR absorption of AA and proteins, respectively, to represent protein hydrolytic activity; we also calculated the ratio of 1711 cm⁻¹ and 1741 cm⁻¹, which are derived from IR absorption of FFA and lipid esters, respectively, to indicate lipid hydrolytic activity. It was previously reported that transgenic worms with LIPL-4 overexpression (lipl-4 Tg) exhibit enhanced lysosomal lipolysis, as measured by lysosome-targeted lipidomics using mass spectrometry³⁶. Consistently, FILM imaging demonstrated an increased 1711/1741 cm⁻¹ ratio in lysosomes of lipl-4 Tg worms (Extended Data Fig. 5c), supporting the use of this ratio as a proxy for lysosomal lipolysis in vivo. Notably, although this ratiometric strategy does not provide an absolute quantitative measurement, it offers a qualitative way to compare metabolic states among individual lysosomes and visualize their heterogeneity within and between cells. The pixel-wise ratiometric ratio displayed in Extended Data Fig. 6a and the parallel set map visualization shown in Extended Data Fig. 6b visually demonstrate the differences between lysosomes and their surrounding regions. We found that lysosomes (orange) exhibit stronger hydrolytic activities compared to surrounding regions (light blue) (Extended Data Fig. 6). More importantly, the lysosomes with high proteolytic activity and those with high lipolytic activity do not completely overlap (Fig. 3c), suggesting metabolic heterogeneity among lysosomes. Based on the two ratios for each lysosome, we categorized them into three groups: high proteolytic activity, high lipolytic activity, and high activity in both (Fig. 3d). We found that the correlation between lysosomal hydrolytic activity and lysosomal size is relatively low, with Pearson coefficients of 0.48 for proteolytic activity and 0.44 for lipolytic activity, indicating no significant relationship between them (Extended Data Fig. 7a and b). We did not observe a significant correlation between hydrolytic activity and fluorescence intensity either (Extended Data Fig. 7c and d), indicating that the spectral features reflect lysosomal chemistry rather than dye effects. This metabolic heterogeneity was also detected in mammalian lysosomes by FILM (Fig. 3e and Extended Data Fig. 8). These results not only define the metabolic landscape of normal lysosomes within a single cell of wild-type organisms but also reveal the functional basis underlying lysosomal heterogeneity.

FILM tracks lysosomal metabolic changes during aging

Metabolic dysfunction is a key hallmark of aging². To investigate age-related metabolic changes at the organellar level in lysosomes, we first generated the ratio-metric images of 1587 and 1649 cm⁻¹ (proteolytic activity), as well as 1711 and 1741 cm⁻¹ (lipolytic activity), for lysosomes in C. elegans at adult Day 2, Day 4, Day 6, and Day 10 (Fig. 4a). The corresponding quantitative statistical results are shown in Fig. 4b. We found that the lysosomes of Day 2 worms exhibit high proteolytic and lipolytic activities, consistent with the findings in Fig. 3d. More importantly, both hydrolytic activities of lysosomes decline with increasing age (Fig. 4b), with the decrease occurring early in life, as early as Day 4 of adulthood, when animals have not yet begun to experience aging-related mortality. These results indicate an early-onset dysfunction in lysosomal metabolism during aging.

Figure 4. Age-related metabolic changes at lysosomal scale.

a, Ratio-metric mapping of intensity ratios at 1587/1649 cm⁻¹ and 1711/1741 cm⁻¹ across worms of different ages. b, Quantitative comparison of the two intensity ratios among four age groups (Two-sample t-test comparing with Day2 group: *, p<0.05; **, p<0.01; ***, p<0.001; ****, p<0.0001; ns, not significance). c, Heatmap of lysosomal fingerprint spectra extracted from worms in four age groups (n=69 for Day2, n=97 for Day4, n=68 for Day6 and n=80 for Day10 derived from five to seven independent experiments), highlighting spectral variations with age. Each row represented a lysosomal spectrum. d, Representative average spectra for each age group, showing age-dependent metabolic differences. Shaded area indicates the standard deviation. e, Z-score heatmap of different age groups. Red boxes highlight signal regions with the higher intensity for Day2 group. Orange boxes indicate signal regions with the higher intensity for Day4 group. Yellow boxes highlight signal regions with the higher intensity for Day6 and Day10 groups. f, t-SNE visualization of all spectra, displaying clustering patterns based on age-related spectral features. Each dot indicates a lysosomal spectrum. Shaded area indicates 85% confidence interval. g, Intra-cluster distance analysis from t-SNE, where larger distances indicate poorer clustering and greater heterogeneity within the data. h, High-content analysis of metabolic profiles across the four age groups, identifying age-related trends. All comparisons were made relative to Day 2 group, whose lysosomal profiles are representative of a normal metabolic state. (Two-sample t-test: *, p<0.05; **, p<0.01; ***, p<0.001; ****, p<0.0001; ns, not significance). Scale bar: 10 μm. In b and h, the boxes show the interquartile range (IQR), the centerlines indicate medians and the lines outside the boxes extend to 1.5 times the IQR.

To better analyze the characteristics of the entire fingerprint spectrum, we extracted dozens of lysosomal spectra from each age group and generated heatmaps (Fig. 4c–d). Comparison of the heatmaps reveals clear differences in the lysosomal spectrum at different ages (Fig. 4c), which supports the early metabolic changes observed at Day 4 of adulthood. We found that with increasing age, the strong peak around 1700 cm⁻¹, which represents C=O of FFA, weakens, while the peak around 1100 cm⁻¹, representing the phosphate group of nucleic acids, gradually increases (Fig. 4d). Additionally, the peak shape representing proteins and AA in the range of 1530-1680 cm⁻¹ also changes (Fig. 4d). Based on these spectra shown in Fig. 4c, we further performed z-score analysis to highlight spectral differences across age groups relative to the total average spectrum (Fig. 4e). As indicated by the rectangular boxes, lysosomes in Day 2 animals exhibited higher AA and FFA peak signatures, whereas those in Day 4 showed stronger intensities in the Amide I and Amide II bands, corresponding to proteins. Additionally, lysosomes in Day 6 and Day 10 animals displayed more prominent features in the lower wavenumber region (1060–1350 cm⁻¹) that can be attributed to nucleic acids and carbohydrates.

When performing t-SNE dimensionality analysis of these spectral data, we observe that the data points of Day 2 form a compact cluster separated from other data points from Day 4, 6 and 10 (Fig. 4f). Day 4 can also be clustered into a relatively independent group from Day 6 and 10 (Fig. 4f). Furthermore, by calculating the Euclidean distance of data points, we verified that each age group data point exhibits the shortest distance within its own class, confirming distinct clustering between groups. The inter-group differences between Day 6 and Day 10 were the smallest, approaching the clustering range observed within each group (Supplementary Fig. 11), suggesting that the spectral differences between these two ages are less pronounced. We also uncovered that the intra-group differences for Day 4, 6 and 10 increase compared to Day 2, indicating that spectral heterogeneity across lysosomes becomes more pronounced with aging, though differences were not significant from Day 4 to Day 10 (Fig. 4g).

Next, we interpreted the spectral data through spectrum decomposition to determine the chemical content of various components. We collected spectra for proteins, AA, FFA, TAG, ceramide, glycogen, nucleic acids (NA), and cholesterol ester (CE) and modified them with augmented MCR as detailed in Fig. 2. We were able to quantitatively compare eight chemicals’ contents of lysosomes, including protein, AA, FFA, TAG, ceramide, glycogen, NA and CE, between ages by leveraging LASSO to decompose the lysosomal spectrum (Fig. 4h). We found the average levels of macromolecules, including NA, ceramides, triglycerides, and glycogens, increase with increasing age (Fig. 4h), suggesting their age-related accumulation within lysosomes.

FILM profiles metabolic changes associated with lysosomal storage diseases

Metabolic dysfunction of lysosomes underlies lysosomal storage diseases (LSD), leading to the accumulation of undegraded macromolecules within lysosomes²⁷. To date, it remains challenging to assess metabolic changes at the lysosomal level under those pathological conditions. We hypothesized that FILM provides an avenue to address this challenge. To test this hypothesis, we knocked down several well-conserved LSD genes using RNA interference (RNAi) in C. elegans and performed FILM imaging at Day 2 of adulthood (Methods). Among these genes, nuc-1 encodes acid deoxyribonuclease, and its deficiency contributes to Autoinflammatory-Pancytopenia Syndrome; aagr-2 encodes acid alpha-glucosidase, whose loss causes Pompe Disease; asah-2 encodes acid ceramidase, and its defect is associated with Farber Disease; lipl-3 encodes lysosomal acid lipase, which is involved in Wolman Disease; and ncr-1 encodes lysosomal cholesterol transporter and its deficiency results in Niemann-Pick Type C Disease.

We first fingerprinted the lysosomes of five Day 2 RNAi groups together with their controls, whose lysosomal profiles represent the normal metabolic state. (Fig. 5a–b). From their heatmaps, the metabolic composition of lysosomes exhibits obvious changes upon RNAi inactivation of these LSD genes (Fig. 5a). Consistent with the results shown in Fig. 4d, the spectrum of the control group predominantly exhibits two peaks: one around 1587 cm⁻¹, representing AA, and another around 1711 cm⁻¹, associated with FFA. When compared to the controls, we found that the intensity around 1100 cm⁻¹ and 1294 cm⁻¹ is significantly increased in the nuc-1 RNAi condition, while showed slight increases in the other RNAi conditions as well (Fig. 5b). In the 1530-1730 cm⁻¹ range, the peak shapes for the various RNAi conditions were notably different, particularly in the ratio of the main peaks at 1587, 1649, and 1711 cm⁻¹. The differences in these spectra indicate that RNAi inactivation of different LSD genes leads to distinct changes in the metabolic composition of lysosomes.

Figure 5. Profiling of metabolic changes associated with lysosomal storage diseases.

a, Heatmap of fingerprint spectra extracted from lysosomes under different RNAi conditions (n=22 for control, n=32 for nuc-1, n=35 for aagr-2, n=21 for asah-2, n=34 for lipl-3 and n=24 for ncr-1 derived from two to four independent experiments), illustrating spectral variations across groups. Each row represented a lysosomal spectrum. b, Representative average spectrum for each RNAi condition, highlighting distinct metabolic profiles associated with specific RNAi treatments. Shaded area indicates the standard deviation. c, High-content analysis of lysosomal contents across RNAi groups, revealing differences in chemical composition and metabolic activity. All comparisons were made relative to the Day 2 control group, whose lysosomal profiles represent the normal metabolic state. (Two-sample t-test: *, p<0.05; **, p<0.01; ***, p<0.001; ****, p<0.0001; ns, not significance). d, Pearson correlation analysis of eight lysosomal contents from C. elegans samples visualized using a chord diagram. Blue curves represent negative correlations lower than −0.5, and red curves represent positive correlations higher than 0.5, with curve thickness indicating the strength of the correlation. e, Heatmap of fingerprint spectra extracted from WT and NPC1KO of HEK293T cells (n=59 for WT, n=61 for NPC1KO derived from five independent experiments). f, Representative average spectra of WT and NPC1KO cell lines. Shaded area indicates the standard deviation. g, High-content analysis with statistical comparison of lysosomal chemical contents between WT and NPC1KO groups (Two-sample t-test: *, p<0.05; **, p<0.01; ***, p<0.001; ****, p<0.0001; ns, not significance). h, Pearson correlation analysis of eight lysosomal contents from mammalian cells visualized using a chord diagram. Blue curves indicate negative correlations lower than −0.5, and red curves indicate positive correlations higher than 0.5, with curve thickness reflecting correlation strength. In c and g, the boxes show the interquartile range (IQR), the centerlines indicate medians and the lines outside the boxes extend to 1.5 times the IQR.

Next, the spectra were interpreted using MCR-LASSO that allowed for the translation of the spectral data into eight specific chemical components for quantification (Fig. 5c). We found that RNAi inactivation of nuc-1 and asah-2 led to the accumulation of NA and ceramides in lysosomes, while the level of FFA decreases (Fig. 5c). With the RNAi inactivation of aagr-2, the lysosomal accumulation of NA and ceramides are also increased (Fig. 5c). RNAi inactivation of lipl-3 and ncr-1 resulted in the lysosomal accumulation of NA, TAG, while FFA levels decreased (Fig. 5c). These results suggest the lysosomal accumulation of NA and ceramides as shared metabolic dysfunction among the loss of five different LSD genes. When analyzing the correlations among the eight components, we found that in addition to the negative correlation between protein and AA (Pearson’s r = −0.74) and between TAG and FFA (Pearson’s r = −0.65), consistent with proteolysis and lipolysis, FFA and DNA also exhibit a negative correlation (Pearson’s r = −0.63), suggesting that defects in DNA degradation may impact the lipolytic activity of lysosomes (Fig. 5d).

Furthermore, we have applied FILM to investigate metabolic changes of lysosomes in mammalian cells with NPC1 knockout (NPC1KO). Niemann-Pick disease type C is a fatal hereditary neurodegenerative lysosomal storage disorder characterized by the accumulation of cholesterol and glycosphingolipids in late endosomal/lysosomal compartments³⁹. We conducted FILM imaging and extracted the lysosomal fingerprint spectra for both WT and NPC1KO cells (Fig. 5e–f). Unlike C. elegans lysosomes, which contain dominant AA and FFA signals, lysosomes in mammalian cells exhibit primary spectral features composed of Amide I and Amide II bands, characteristic of proteins. We found that compared with WT, lysosomes in the NPC1KO cells exhibited stronger signals between 1100 cm⁻¹ and 1250 cm⁻¹, as well as changes in the 1530-1730 cm⁻¹ range (Fig. 5f). Changes in those ranges were also observed in C. elegans lysosomes with the inactivation of LSD genes (Fig. 5b).

When quantifying the eight components obtained through MCR-LASSO spectral analysis, we found that compared to WT, FFA and AA levels in NPC1KO lysosomes were decreased, while the level of macromolecules, including NA, ceramide, TAG, CE, and glycogen were increased (Fig. 5g). These results support that constitutive NPC1 knockout in mammalian cells lead to defective degradation of nearly all macromolecules. No significant changes were observed in protein levels, likely due to the high protein contents in mammalian lysosomes (Fig. 5g). In the correlation diagram among the eight components, we observed a positive correlation between TAG and glycogen (Pearson’s r = 0.64), indicating that the lysosomal accumulation of these macromolecules likely occurs simultaneously in mammalian lysosomes. By contrast, FFA exhibit negative correlations with DNA (Pearson’s r = −0.67), ceramide (Pearson’s r = −0.78) and glycogen (Pearson’s r = −0.63), suggesting that lysosomal lipolysis may be modulated by defects in DNA and ceramide degradation as well as by glycogen accumulation.

Discussion

FILM equipped with AI-based denoiser and spectral decomposition, provides a technical platform to quantitatively profile metabolic composition of specific organelles in living cells and organisms, enabling the investigation of dynamic biochemical changes under various physiological and pathological conditions. By incorporating the optical boxcar strategy, FILM effectively mitigates the photobleaching issues that were prevalent in early point-scan F-MIP studies^{23, 24}. Additionally, this advancement not only enhances signal strength but also significantly reduces solvent background interference, particularly in aqueous environments, beyond wide-field geometry, which is primarily confined to dry samples²⁵. On the data analysis front, we harnessed a self-supervised hyperspectral denoising algorithm, improving both image and spectral SNR by 26.9× and 5.3×. To refine the extraction of component distributions, we optimized the MCR-LASSO spectral decomposition, enabling quantitative characterization of diverse chemical contents. These methodological advancements collectively enhance the sensitivity, specificity, and robustness of FILM, paving the way for deeper insights into organelle-specific metabolic processes in living systems.

Vibrational microspectroscopy techniques, leveraging coherent Raman scattering or optical photothermal detection of vibrational absorption, offer powerful tools for spatial metabolic profiling^{19, 40–43} with high spatial resolution but typically lack organelle specificity. The FILM system effectively bridges vibrational and fluorescence imaging modalities, which provides chemical fingerprints and organellar specificity simultaneously. Unlike stimulated Raman or IR up-conversion fluorescence techniques^44–46, which rely on the co-excitation of specific fluorophores to achieve superior sensitivity but confine chemical information to the dye itself, FILM operates as a decoupled process, where fluorescent molecules function as reporters to sense the surrounding molecules. Furthermore, fluorescence-guided MIP for co-localizing vibrational imaging^{47, 48} often suffers from spatial mismatches caused by focal plane shifts, while scattering-based MIP is prone to ring artifacts that are heightened by environmental solvent interference.

In this study, we chose lysosomes to demonstrate the application of FILM, given their involvements in diverse macromolecular processing and their highly dynamic metabolic activities ^{49, 50}. We also explored the inter-cellular heterogeneity under normal physiological conditions by analyzing the lysosomal variations among twenty-one wild type cells from twenty-one worms at day-2 young age (Extended Data Fig. 9). Our analysis revealed that lysosomal metabolism is highly heterogeneous among individual worms, even at the same age and under identical physiological conditions. The t-SNE embedding consistently separated lysosomal spectra into three functional subpopulations: C1, dominated by proteolytic activity; C2, enriched in lipolytic activity; and C3, displaying high activity in both, consistent with the classification shown in Fig. 3. These findings demonstrate that the distribution of distinct lysosomal subpopulations with different metabolic activity varies markedly from cell to cell. Future applications are expected to extend this technique to other metabolically active organelles, such as lipid droplets and mitochondria (Extended Data Fig. 10). Further advancements in instrumentation, such as enhancing hyperspectral acquisition speed through thermal deposition multiplexing, could further improve imaging efficiency and throughput⁵¹. Overall, we believe that this imaging technique holds significant promise for advancing in vivo metabolic investigation across scales. By offering a more precise and versatile imaging approach, this technique has the potential to uncover new insights into cellular mechanisms underlying aging, metabolic disorders, and disease pathogenesis.

Methods

FILM hyperspectral imaging.

The pulsed mid-infrared pump beam is generated by a wavelength-tunable quantum cascade laser (QCL, Daylight Solutions, MIRcat-QT-Z-2400). Fluorescence excitation light is provided by either a 488 nm fixed-wavelength diode laser module (Cobolt, 06-MLD 488 nm) or a femtosecond laser (Insight DeepSee, Spectral Physics, Insight DS DUAL), depending on the fluorophore used. The 1040 nm output of the femtosecond laser is frequency-doubled using an LBO crystal and temporally broadened with SF57 rods to generate picosecond 520 nm light. The 488 nm laser can be digitally modulated into pulsed light via an external trigger, while the 520 nm laser is modulated using an acousto-optic modulator (AOM). A function generator synchronizes the visible excitation light and the mid-IR pump beam, with their modulation frequencies set to 2f (400 kHz) and f (200 kHz), respectively. The IR pulse width is set to 200 ns, and the visible light operates with a 30% duty cycle. The fluorescence excitation light is rapidly scanned using a pair of dual-axis galvo mirrors (GVS002, Thorlabs). After passing through a scan lens (f = 100 mm; a pair of AC508-100-A, Thorlabs) and a tube lens (f = 200 mm; TTL200-A, Thorlabs), the beam is reflected by a dichroic mirror (DM) into a water-immersion objective (UPlanSApo, Olympus, 60×, NA=1.2) and focused onto the sample. The IR beam is scanned independently with another pair of X-Y galvanometer mirrors (GVS002, Thorlabs). The IR beam path employs a concave mirror as the scan lens (f = 200 mm; CM508-200-P01, Thorlabs) and a tube lens (f = 500 mm; CM508-500-P01, Thorlabs) to relay the scan to the back pupil of a reflective objective (PIKE, 40×, NA=0.78), achieving counter-propagation alignment with the visible excitation light. Before imaging, the IR beam is carefully aligned to overlap with the visible focus. During imaging, the IR and visible foci are synchronously scanned, ensuring uniform excitation and detection over the FOV. The two galvanometer pairs are synchronized with the focal lengths of the visible and IR objectives and scaled based on the beam expansion ratio of the relay system. This scaling factor is calibrated at the start of the experiment. The backward fluorescence emitted from the sample is collected by the water-immersion objective and directed through the DM. After further filtering with a bandpass or long-pass filter, the fluorescence signal is detected by a silicon photomultiplier (SiPM, Hamamatsu, C13366-3050GA). The resulting electrical signal is fed into Moku:Pro (Liquid Instrument, Multi-instrument Mode), filtered, and input into the slots of two lock-in amplifiers for demodulation at 2f and f frequencies, corresponding to the FILM and fluorescence DC signals, respectively. These demodulated signals are simultaneously acquired through two input ports of an acquisition card, enabling real-time dual-channel imaging. To perform hyperspectral imaging, the quantum cascade laser (QCL) operates in Multi-Spectral mode using a preset scanning list that that spans the entire fingerprint region. For S. aureus imaging, the hyperspectral range covers 980 to 1780 cm⁻¹ with 160 frames, while for organelle imaging, including lysosomes, lipid droplets, and mitochondria, it covers 1000 to 1800 cm⁻¹ with 126 frames.

IR spectral calibration.

Since the signal of FILM is proportional to the DC fluorescence intensity, IR light power and IR absorption cross section of the molecules as described below:

$$Signal={Fluorescence}_{DC}\ast I_{IR}\ast {\sigma }_{IR}$$

To obtain the IR absorption spectrum of the molecules, ${\sigma }_{IR}$, we need to calibrate the spectrum of the collected FILM.

As depicted in Supplementary Fig. 12, the raw FILM spectra were initially corrected for the noise-induced baseline and then divided by the fluorescent photobleaching curve. The baseline was estimated using the average intensity under the IR-off condition and approximated by the wavenumber at the IR power dip. Subsequently, the spectra were divided by the IR power spectrum to calibrate the peak resulting from the power profile. Finally, the spectra were smoothed with 3-5 pixels neighboring average and normalized with the area under the curve. Since there is a power dip around 1450 cm⁻¹ caused by the switching of laser chips, which may introduce artifacts during power calibration, the spectral band from 1380 to 1480 cm⁻¹ was excluded from the quantification analysis.

Self-supervised hyperspectral denoising.

To suppress the spatially correlated and spectrally varied noise intrinsic to hyperspectral FILM imaging, we implemented a self-supervised denoising framework termed SPEND (Self-Permutation Noise2Noise Denoising)³⁰. In contrast to conventional approaches that depend on explicit noise modeling or on high-SNR reference data for training, SPEND can be trained directly from a single noisy hyperspectral stack.

The key of SPEND is to learn the noise characteristics using two independent measurements of the same field of view (FOV). To achieve this, we proposed a permutation-based method accompanied by correlation identification. Because the scanning step size is much smaller than the features of interest both spectrally and spatially, adjacent pixels or frames can be treated as independent measurements of the same FOV. However, chemical imaging often contains correlated noise, which violates the independence assumption. In our case, noise correlation was primarily observed along the spatial domain (Supplementary Fig. 4). Comparison along the spectral axis, which is perpendicular to the correlation axis can help to avoid the noise leakage into signals. Accordingly, the raw hyperspectral stack was split into odd and even frames along the spectral axis, recombined into two complementary sequences, and used as input–target pairs for network training. These paired sequences could thus be regarded as independent noisy measurements of the same underlying signal.

A 3D U-Net architecture was adopted as the backbone network. Its encoder–decoder structure, comprising convolution, max-pooling, and upsampling layers, efficiently captures both spatial and spectral features even with limited training data. During training, the network learns to map one noisy sequence onto its paired counterpart, thereby suppressing noise while retaining consistent structural and spectral information.

In the prediction phase, the original (non-permuted) hyperspectral stack is fed into the trained model, thereby preserving both spectral continuity and spatial integrity. This approach enables effective removal of spatially correlated and spectrally heterogeneous noise, yielding denoised hyperspectral datasets with substantially improved signal-to-noise ratio. The enhanced data quality facilitates downstream analyses, including ratiometric mapping, spectral unmixing, and quantitative profiling of biochemical contents.

As a self-supervised method based on a permutation strategy, SPEND does not require external labels during training. The trained model can then be applied to new hyperspectral datasets with similar SNR, including those involving different dyes or biological samples, without the need to modify the network architecture. In our implementation, the training set consists of six fields of view, each with dimensions of 200 × 200 × 126 pixels. The training process takes ~30 minutes on an NVIDIA RTX 4090 24GB GPU. Once trained, SPEND performs inference on new hyperspectral image stacks efficiently, requiring ~2 seconds per stack.

Construction of an eight standard reference library for spectral unmixing.

Lysosomes are widely recognized as the primary catabolic centers of the cell, responsible for degrading various macromolecules. Accordingly, these macromolecules and their building blocks represent the major chemical constituents of lysosomes. To construct a comprehensive and biologically relevant reference spectral library, the standards were selected based on the four major classes of biomolecules: proteins, lipids, carbohydrates, and nucleic acids. For proteins and lipids, we included both the macromolecules and their respective monomers (protein and amino acids, triglycerides and free fatty acids). In contrast, for carbohydrates and nucleic acids, we didn’t include the monomers of glycogen and DNA. This decision was based on the fact that, unlike proteins and lipids, their monomers (glucose and nucleotides) exhibit minimal spectral differences from the polymers, lacking clear peak shifts that could serve as reliable diagnostic markers. Moreover, glycogen and DNA better represent the biological context of lysosomal dysfunction, which primarily involves defective macromolecule degradation, as demonstrated in RNAi studies targeting NUC-1 and AAGR-2. Finally, ceramide and cholesterol esters were included because they are well-known metabolites that frequently accumulate in lysosomal storage disorders, making them essential for accurately modeling pathological lysosomal metabolism.

Detailed information for the eight compounds, including their chemical identities and commercial suppliers, is provided in Table S1. To obtain FILM spectra of the non-fluorescent standards, we used 50 μM rhodamine 6G (R6G) as a fluorescence reporter by dissolving powdered compounds or mixing liquid samples with R6G, followed by measurement on air-dried gels for powdered compounds or on solutions for liquid samples. The characteristic wavenumbers for each standard, together with the corresponding vibrational mode assignments, are listed in Table S2.

MCR-LASSO spectral unmixing.

To enhance unmixing performance, we input the spectrum of the original standard as the initial estimate into the MCR. Additionally, to mitigate the over-adjustment of the reference spectrum by MCR, we adopted an augmentation MCR strategy, incorporating the reference spectrum into the lysosome spectral dataset as an additional constraint. Following this, the corrected spectrum output by the augmented MCR was used as input for the LASSO spectral unmixing. We segmented individual organelles (e.g., lysosomes) and extracted their fingerprinting spectra. Following photobleaching correction and IR power calibration, the calibrated spectra were reshaped into a 2D matrix, with rows representing individual spectra, for LASSO-based concentration decomposition³².

C. elegans strains.

C. elegans N2 strain was obtained from Caenorhabditis Genetics Center (CGC). C. elegans strains were maintained at 20°C on standard NGM agar plates seeded with OP50 E.coli (HT115 E. coli for RNAi experiments) using standard protocols⁵².

C. elegans RNAi treatments.

RNAi clones used in this study were sourced from the RNAi library generated by Dr. Marc Vidal’s lab⁵³, including aagr-2, asah-2, ncr-1, lipl-3, and nuc-1. All RNAi colonies were selected for resistance to both 50 μg ml−1 carbenicillin and 50 μg ml−1 tetracycline, and verified by Sanger sequencing. RNAi bacteria were cultured for 14 hours in LB with 25 μg/ml carbenicillin, then seeded onto RNAi agar plates containing 1 mM IPTG and 50 μg/ml carbenicillin. Each RNAi bacteria clone was allowed to dry on the plates before overnight incubation at room temperature to induce dsRNA expression. RNAi-based experiments were conducted using E. coli HT115 bacteria, with L4440 empty vector bacteria used as controls. For RNAi plates containing LysoSensor, LysoSensor was added to RNAi plates at 0.5 μM final concentration.

Synchronized L1 N2 worms were added onto 6 cm RNAi plates and raised at 20 °C for two days till the L4 stage. Around 50 L4 worms were transferred to the RNAi plates containing LysoSensor. At Day 2, worms were anesthetized with 1% NaN₃ and imaged using FILM, mounted between a glass coverslip and a CaF₂ substrate.

C. elegans aging experiments.

Synchronized L1 N2 worms were added onto 6 cm plates and raised at 20 °C. Around 50 worms at L4, day 2, day 4, day 8 were transferred LysoSensor containing plates and imaged at day 2, 4, 6 and 10, respectively.

Physiological validation of the 1587 and 1711 cm⁻¹ spectral features.

A functional perturbation experiment by knocking down the lysosomal cystine transporter cystinosin (CTNS-1)³⁵, whose loss is known to cause cystine accumulation in the lysosome was performed to further validate the 1587 cm⁻¹ assignment of amino acids. As shown in Extended Data Fig. 5a, FILM imaging revealed a modest but statistically significant increase in the lysosomal amino acid signal following CTNS-1 knockdown (P = 0.013). This effect likely reflects the selective impact of CTNS-1 loss on cystine rather than a global change in all amino acids. This physiological validation provides independent evidence supporting our interpretation of this spectral feature as a readout of lysosomal amino acid levels.

A transgenic C. elegans strain with upregulated lysosomal lipase LIPL-4 (lipl-4 Tg), which is known to enhance lysosomal lipolysis, based on lysosome-targeted lipidomics using mass spectrometry (MS)³⁶, was used to validate the ratiometric measurement of lipolytic activity and the 1711 cm⁻¹ assignment of fatty acids. As shown in the Extended Data Fig. 5b–c, FILM imaging reveals increased lipolytic activity and higher levels of fatty acids in the lysosomes of lipl-4 Tg worms compared to wild type controls. The consistency with the MS-based lysosome-targeted lipidomics results supports the validity of the FILM approach for in vivo assessment of lysosomal metabolic activity.

Gene knockout cell line.

HEK293T sgNT (NPC1 +/+, control) and sgNPC1 (NPC1 −/−, knockout) cells were a gift from Dr. Roberto Zoncu (University of Berkley, PMID: 33308480⁵⁴). Mycoplasma contamination was regularly tested, and cells were confirmed to be mycoplasma-fee using MycoAlert Mycoplasma Detection Kit (Lonza, LT07-318).

Bacteria strains.

Shigella flexneri expressing GFP was grown overnight at 37 °C on a tryptic soy agar plate. Colonies with green fluorescence were picked up by sterile inoculation loops and then resuspended in PBS. The bacterial solution was diluted by optical density at 600 nm (OD600) to 0.1. The bacteria were then fixed by 10% formalin for 30 minutes at room temperature. The bacterial solution was washed twice with deionized water, dried on CaF₂, and sandwiched with a coverslip. The samples were then observed using the FILM setup with a DM (500 LP, Edmund, #69-899) and a filter (520/36, Edmund, #67-016).

Staphylococcus aureus (S. aureus) was incubated in a MHB medium for 10 h. After centrifuging and washing in phosphate-buffered saline (PBS), the bacteria were fixed by formalin solution for 30 minutes. Rhodamine 6G at 10⁻⁴ M was then added to the bacteria pellet, which was subsequently resuspended and incubated for 1 hour. Following the final washing steps with deionized water, the bacterial suspension was dried on CaF₂ and sandwiched with a coverslip for FILM imaging with a DM (550 LP, Edmund, #69-900) and filter (575/27, Edmund, #33-333).

Spectral phasor analysis.

In spectral phasor analysis, the spectrum of each pixel was interpreted through the discrete Fourier transform of first-order harmonics. By scattering the pixels of the entire image across the complex plane, we were able to identify specific clusters that represented target chemical channels. Phasor analysis was performed with the standardized phasor analysis plug-in in ImageJ (1.49v). The phasor domain segmentation was shown in Supplementary Fig. 9.

Fluorescence labeling of mammalian cells.

HeLa cells labelled by LipiRed:

HeLa cells were seeded on CaF₂ substrates at a density of 1 × 10⁵ cells/ml in 2 ml of high-glucose DMEM supplemented with 10% FBS and penicillin–streptomycin and incubated for 24 h at 37°C in a humidified atmosphere with 5% CO₂. The following day, the medium was replaced with fresh serum-free medium containing 6 μM LipiRed, and the cells were incubated at 37°C for 30 minutes. After incubation, the cells were gently washed three times with warm PBS to remove excess dye. For imaging, the cells on CaF₂ are sandwiched with coverslip, maintaining in PBS and observed using FILM setup with DM (550 LP, Edmund, #69-900) and filter (600 LP, Edmund, #62-985).

HeLa cells labelled by MitoTracker Green:

HeLa cells were seeded on CaF₂ substrates at a density of 1 × 10⁵ cells/ml in 2 ml of high-glucose DMEM supplemented with 10% FBS and penicillin–streptomycin and incubated for 24 h at 37°C in a humidified atmosphere with 5% CO₂. The following day, the medium was replaced with fresh serum-free medium containing 100 nM MitoTracker Green, and the cells were incubated at 37°C for 15 minutes. After incubation, the cells were gently washed three times with warm PBS to remove excess dye. For imaging, the cells on CaF₂ are sandwiched with coverslip, maintaining in PBS and observed using FILM setup with DM (500 LP, Edmund, #69-899) and filter (520/36, Edmund, #67-016).

HEK293T cells labelled by LysoSensor DND189:

WT and NPC1KO HEK293 cells were seeded on CaF₂ substrates at a density of 1 × 10⁵ cells/ml in 2 ml of high-glucose DMEM supplemented with 10% FBS and penicillin–streptomycin and incubated for overnight at 37°C in a humidified atmosphere with 5% CO₂. The following day, the medium was replaced with medium containing 1 μM LysoSensor DND189, and the cells were incubated at 37°C for 20 minutes. After incubation, the cells were gently washed three times with warm PBS to remove excess dye. For imaging, the cells on CaF₂ are sandwiched with coverslip, maintaining in PBS and observed using FILM setup with DM (500 LP, Edmund, #69-899) and filter (520/36, Edmund, #67-016).

ATR-FTIR spectroscopy.

The FTIR spectra of all samples were measured on an attenuated total reflection (ATR) FTIR spectrometer (Nicolet Nexus 670, Thermo Fisher Scientific). The measurements were conducted with a spectral resolution of 2 cm⁻¹, and each spectrum was measured with 32 scans. Prior to measurement, the ATR crystal was carefully cleaned with ethanol and dried to prevent contamination between samples. All spectra were automatically normalized using the built-in baseline correction feature of the spectrometer.

Thermal sensitivity measurement of fluorescence dye.

The thermal sensitivity measurement of fluorescent dye was measured with the setup shown in Supplementary Fig. 3a. A small droplet of fluorescent dye solution was placed on a silicon wafer, which served as the substrate. A coverslip was then positioned on top to ensure the sample was securely enclosed and evenly distributed. The sample is excited by light from LED (SOLIS-3C, Thorlabs), which passes through an excitation filter to select the specific wavelength required to excite the dye. The excitation light was directed toward the sample through the optical setup. The emitted fluorescence was collected and passed through an emission filter, which blocks any residual excitation light and isolates the specific fluorescence wavelength of interest. This filtered emission light was then detected by camera (Hamamatsu, C13440). Since the sample was placed on a temperature-controlled plate (Bioscience Tools, TC-1-100s), we can control the temperature and analyze the corresponding changes in fluorescence intensity.

Statistics analysis.

In the aging studies, we analyzed lysosomal function over time in live, anesthetized wild-type worms at four different ages (Day 2, Day 4, Day 6, and Day 10). This analysis revealed, for the first time, age-dependent metabolic changes at the lysosomal level. Approximately six worms were imaged per age group, and ~10 lysosomes per intestinal cell were analyzed to provide sufficient statistical capability. For disease-associated studies, all comparisons were made relative to carefully matched wild-type controls. Control data are shown in all panels, labeled as “Control” in panels a, c, and d, and as “WT” in panels e, f, and g. For each condition, three to five independent biological replicates were performed, with 5-10 lysosomes analyzed per replicate, ensuring statistical robustness of the comparisons.

All significance analyses were performed using two-sample t-tests. *, p<0.05; **, p<0.01; ***, p<0.001; ****, p<0.0001; ns, not significance.

Extended Data

Extended Data Fig. 1: Optical boxcar strategy enhances the signal by shifting high odd-order harmonics to detected frequency.

a, Principle of the higher order harmonics shifting. The fluorescence excitation light was modulated into pulses. The pulsed excitation light functioned as a 2f carrier, shifting high odd-order harmonic signals into the demodulation frequency. b, FILM signal of Shigella flexneri expressing GFP at different duty cycle (n=5). Statistical data are presented as mean ± s.d. Scale bar: 10 μm.

Extended Data Fig. 2: Optical boxcar strategy suppressed solvent background by harnessing the differential thermal dynamics between particles and water medium.

a, Photothermal dynamics of LipiRed labelled lipid droplets under different IR absorption peaks. The shaded area represents the mean derived from three independent measurements, while the solid line indicates the exponential fitting curve. b, The pulse pair serves as the gating windows to capture the time-resolved fluorescence signal, which is less sensitive to slow dynamic processes. Scale bar: 10 μm.

Extended Data Fig. 3: Spectral fidelity verification.

a, Spectral comparison of FILM and scattering-based MIP (Sc-MIP) with Rhodamine 6G labelled S. aureus. b, Spectral comparison of FILM and ATR-FTIR with LysoSensor DND189 stained DMSO.

Extended Data Fig. 4: Head-to-head comparison of SPEND with BM4D and Noise2Void (N2V).

a, FILM images of lysosomes were acquired with IR at 1711 cm⁻¹ and 1797 cm⁻¹ before and after three denoising algorithms. b, the intensity profiles along the red dotted lines indicated in a. c, compares the raw, uncalibrated FILM spectra before and after three denoising algorithms. d, the quantification of image SNR and spectral SNR before and after three denoising algorithms (n = 13). In d, the boxes show the IQR, the centerlines indicate medians and the lines outside the boxes extend to 1.5 times the IQR. Scale bar: 10 μm.

Extended Data Fig. 5: Physiological validation of the 1587 and 1711 cm⁻¹ features.

a, Quantification of the 1587 cm⁻¹ intensity for individual lysosomes. CTNS-1 knockdown strain showed a modest but statistically significant increase in the lysosomal amino acid signal. A significant increase in signal intensity at 1587 cm⁻¹ is observed in ctns-1 compared to L4440 worms (Two-sample t-test: *, p=0.013). b, Quantification of the 1711 cm⁻¹ peak intensity shows a significant increase in lipl-4 Tg worms compared to WT. c, Ratio of the 1711 cm⁻¹ (free fatty acids) to 1741 cm⁻¹ (lipid ester) peaks, indicating a relative upregulated in lipolytic activity upon LIPL-4 overexpression. ****p < 0.0001 by two-sample t-test. Each point represents a single lysosome, with box plots indicating the median, interquartile range, and whiskers representing 1.5× IQR.

Extended Data Fig. 6: Visualization of two ratios and class discrimination.

a, Pixel-wise scatter plot of two calculated intensity ratios (200×200 pixels). b, The parallel set shows the relationship between two spectral ratios (1587 cm⁻¹/1649 cm⁻¹ and 1711 cm⁻¹/1741 cm⁻¹) and the class separation of lysosomes (red) and surrounding region (blue). The curves represent different data points from the corresponding classes, illustrating the distribution and class distinction.

Extended Data Fig. 7: Correlations of lysosomal hydrolytic activity with lysosomal size and fluorescence intensity.

a, The correlation between lysosomal proteolytic activity and size is not significant with a Pearson coefficient of 0.48. b, The Pearson coefficient of lysosomal lipolytic activity and size is 0.44. c, The correlation between lysosomal proteolytic activity and fluorescence intensity is not significant with a Pearson coefficient of 0.49. d, The Pearson coefficient of lysosomal lipolytic activity and fluorescence intensity is 0.43.

Extended Data Fig. 8: Hydrolytic heterogeneity of lysosomes in mammalian cells.

a, Fluorescent and FILM images at 1587 and 1649 cm⁻¹ of LysoSensor DND189 labelled HEK293T cells. b, Ratio-metric mapping of intensity ratios at 1587/1649 cm⁻¹ and 1711/1741 cm⁻¹. Scale bar: 10 μm. Representative results are shown from three independent experiments.

Extended Data Fig. 9: t-SNE–based classification of WT lysosomal spectra and identification of functional subtypes.

a, Within-individual heterogeneity. For each spectrum, the standard deviation (STD) across wavenumbers was calculated; Mean_STD denotes the average STD per worm (y-axis ×10⁻³, x-axis = Worm ID). b, t-SNE embedding (one dot per spectrum) colored by individual worm. Three regions are drawn: C1 (red), C2 (blue), C3 (green), corresponding to three clusters of “high proteolytic activity”, “high lipolytic activity”, and “high activity in both”. c, Cluster-specific quantitative analysis of activity ratios at 1587/1649 cm⁻¹ (proteolytic activity) and 1711/1741 cm⁻¹ (lipolytic activity), confirming functional differences among clusters; colors match panel b. d, Per-worm composition across regions. For each worm, the fraction of its spectra falling into C1/C2/C3 is shown as a stacked bar; colors match panels b-c. Two-sample t-test: *, p<0.05; **, p<0.01; ***, p<0.001; ****, p<0.0001; ns, not significance. In c, the boxes show the IQR, the centerlines indicate medians and the lines outside the boxes extend to 1.5 times the IQR.

Extended Data Fig. 10: Hyperspectral FILM imaging of lipid droplet and mitochondria.

a, Fluorescence and FILM images at 1741 (ester C=O, on-resonance) and 1797 cm⁻¹ (off-resonance) of LipiRed labelled HeLa cells. b, FILM spectral of lipid droplet marked by red circle and arrowhead in a. c, Fluorescence and FILM images at 1649 (Amide I, on-resonance) and 1797 cm⁻¹ (off-resonance) of MitoTracker Green labelled HeLa cells. d, FILM spectral of mitochondria marked by red circle and arrowhead in c. Scale bar: 10 μm. Representative results are shown from three independent experiments.

Data availability.

All data related to the work are available in the article and supplementary information in this paper and are available upon reasonable request to the corresponding authors.

Code availability.

All relevant code, including SPEND denoising, and spectral unmixing, can be accessed at the Cheng Lab GitHub page (https://github.com/buchenglab).