Volume 34, Issue 8 p. 10357-10372
Open Access

Multiplex fatty acid imaging inside cells by Raman microscopy

Masaaki Uematsu

Corresponding Author

Masaaki Uematsu

Department of Lipid Signaling, National Center for Global Health and Medicine, Tokyo, Japan

Department of Lipidomics, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan


Masaaki Uematsu and Hideo Shindou, Department of Lipid Signaling, National Center for Global Health and Medicine, 1-21-1 Toyama, Shinjuku-ku, Tokyo 162-8655, Japan.

Email: [email protected] (M. U.) and [email protected] (H. S.)

Search for more papers by this author
Yoshihiro Kita

Yoshihiro Kita

Department of Lipid Signaling, National Center for Global Health and Medicine, Tokyo, Japan

Department of Lipidomics, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

Life Sciences Core Facility, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

Search for more papers by this author
Takao Shimizu

Takao Shimizu

Department of Lipid Signaling, National Center for Global Health and Medicine, Tokyo, Japan

Department of Lipidomics, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

Institute of Microbial Chemistry, Tokyo, Japan

Search for more papers by this author
Hideo Shindou

Corresponding Author

Hideo Shindou

Department of Lipid Signaling, National Center for Global Health and Medicine, Tokyo, Japan

Department of Lipid Science, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan


Masaaki Uematsu and Hideo Shindou, Department of Lipid Signaling, National Center for Global Health and Medicine, 1-21-1 Toyama, Shinjuku-ku, Tokyo 162-8655, Japan.

Email: [email protected] (M. U.) and [email protected] (H. S.)

Search for more papers by this author
First published: 27 June 2020
Citations: 6


Visualizing intracellular fatty acids (including free and esterified form) is very useful for understanding how and where such molecules are incorporated, stored, and metabolized within cells. However, techniques of imaging multiple intracellular fatty acids have been limited by their small size, making it difficult to label and track without changing their biological and biophysical characteristics. Here, we present a new method for simultaneously visualizing up to five atomically labeled intracellular fatty acid species. For this, we utilized the distinctive Raman spectra depending on the labeling patterns and created a new, extensible opensource software to perform by-pixel analysis of extracting original spectra from mixed ones. Our multiplex imaging method revealed that fatty acids with more double bonds tend to concentrate more efficiently at lipid droplets. This novel approach contributes to reveal not only the spatial dynamics of fatty acids, but also of any other metabolites inside cells.


  • AA
  • arachidonic acid
  • aLA
  • α-linolenic acid
  • Br
  • bromine
  • CRR
  • cosmic ray removal
  • DAG
  • diacylglycerol
  • DGAT
  • diacylglycerol O-acyltransferase
  • DHA
  • docosahexaenoic acid
  • ER
  • endoplasmic reticulum
  • FAME
  • fatty acid methyl ester
  • GC-MS
  • gas chromatography-mass spectrometry
  • LA
  • linoleic acid
  • LD
  • lipid droplet
  • OA
  • oleic acid
  • PA
  • Palmitic acid
  • PCA
  • principal component analysis
  • PLIN3
  • perilipin-3
  • PUFA
  • polyunsaturated fatty acid
  • TAG
  • triacylglycerol

    The visualization of multiple molecules within cells at the same time provides crucial information to understand their behavioral differences and biological roles. Thanks to recent progress in protein engineering, genetically encoded fluorescent proteins are now widely used for multicolor biological imagings. While the protein imaging has become common, there are relatively few methods for visualizing intracellular metabolites, primarily because they are too small to label with fluorescent tags such as nitrobenzoxadiazole without changing their original characteristics.1 The visualization of fatty acids (including free and esterified form), which play important roles as cellular membrane components, energy sources, and signaling mediators, is particularly difficult because most of them are usually buried inside lipid bilayers or lipid droplets in the form of fatty acyl groups in phospholipids, triacylglycerols (TAG), cholesterol esters, etc.2-4 As a result, even probes targeting fatty acyl chains inside those lipids cannot reach them, and little is known about their intracellular distribution and stimulus-dependent mobilization.

    Several methods have been developed to visualize metabolites. For example, while imaging mass spectrometry or positron emission tomography have been be used to directly visualize metabolites or radioactive tracers, the spatial resolutions of these tools (~5 μm and ~1 mm, respectively) are inadequate to visualize intracellular distributions.5, 6 We previously reported a method to visualize a bromine (Br)-labeled fatty acid by scanning X-ray fluorescence microscopy at the subcellular level.7 Another option is to use a Raman microscopy, which can detect the vibrational frequencies of chemical bonds in living specimens, in some cases even without labeling.8 For example, deuterium-labeled fatty acids have been observed by Raman microscopy.9-11 As for multiplex imaging, methods to simultaneously visualize up to 24 different chemical tags were developed using a stimulated Raman microscopy and stable isotopes such as 13C or 15N.12, 13 However, none of these methods have been capable of performing multiplex imaging of small molecules without the tagging of large molecules, hindering them from revealing the behavioral differences among metabolites.

    In this paper, we report a method for visualizing five atomically labeled fatty acids using a combination of Raman microscopy and a spectral unmixing software newly developed for this study. Based on the differences in the Raman spectra depending on labeling patterns, a combination of labeled free fatty acids were carefully chosen and how they are incorporated into HeLa cells were observed. To obtain a final image output, we created an open-source software, ImageCUBE, that can automatically separate overlapping spectra from the mixed measured spectrum within individual image pixels. ImageCUBE can also perform as a platform for sharing and carrying out other complicated analyses of hyperspectral images (data containing two spatial dimensions and spectral dimensions).

    Application of the proposed method revealed that fatty acids with larger degrees of unsaturation were more concentrated in lipid droplet areas than in nondroplet areas. We also developed a method for capturing fluorescence-Raman hybrid images of cells using only a conventional Raman microscope system. Our approach should be useful to reveal the intracellular spatial dynamics of any metabolites.


    2.1 Raman spectroscopy

    The sources of the labeled free fatty acids analyzed in this study are listed in Table S1. Each labeled free fatty acids was dissolved in ethanol at a concentration of 10 mM. Aliquots (0.5 μL) of the free fatty acid solution was dropped onto CaF2 glass substrates (OPCFU-25C02-P; Sigma Koki, Tokyo, Japan), the ethanol was evaporated and Raman spectra were collected using WiRE 5.1 software (Renishaw, Wotton-under-Edge, United Kingdom) via confocal Raman spectroscopy (inVia Reflex; Renishaw, Wotton-under-Edge, United Kingdom) at room temperature. The samples were excited by a 532-nm diode laser through a 50× air objective (566072, NA = 0.75; Leica, Wetzlar, Germany). Scattered light was collected and separated by a 600-line/mm grating.

    2.2 Sample preparations for Raman and fluorescence imaging

    HeLa cells were cultured with Dulbecco's modified Eagle's medium (08459; Nacalai Tesque, Kyoto, Japan) supplemented with 10% of fetal bovine serum (12676-029; Thermo Fisher Scientific, Waltham, MA, USA) in a 5% of CO2 incubator. To conduct imaging experiments using Br-labeled free fatty acids, HeLa cells were seeded onto a CaF2 glass substrates (OPCFU-25C02-P; Sigma Koki, Tokyo, Japan) dipped in a 35-mm dish. In other experiments, a 35-mm quartz bottom dishes (SF-S-D12; Techno Alpha, Tokyo, Japan) were used. After 24 hours of incubation, the media were replaced with the desired combination and concentration of labeled and/or nonlabeled free fatty acid solutions, and then, incubated for a further 24 hours. The cells were then washed with phosphate-buffered saline and fixed with 4% of paraformaldehyde (160-16061; Wako, Osaka, Japan) for 20 minutes at room temperature, followed by five washings with phosphate-buffered saline. To conduct the DGAT1 and DGAT2 inhibition experiment, HeLa cells were treated with 20 μM of T863 (SML0539; Sigma Aldrich, St. Louis, MO, USA) and 10 μM of PF-06424439 (PZ0233; Sigma Aldrich, St. Louis, MO, USA) prior to the 24 hours free fatty acid treatment.

    2.3 Raman imaging

    To obtain Raman hyperspectral images, the samples were excited by a 532-nm diode laser through a 63× water immersion objective (506148, NA = 0.90; Leica, Wetzlar, Germany). The exposure time and the spatial resolution were set to 0.1 seconds/pixel and 0.3-μm/pixel, respectively, by using the WiRE 5.1 software.

    2.4 Fluorescence acquisition by Raman microscope

    Prior to fixing, HeLa cells were treated with 1 μM of Lipi-Red (LD03; Dojindo, Kumamoto, Japan), 1 μM of ER-Tracker Red (E34250; Molecular Probes, Eugene, OR, USA), or 50 nM of Mito Tracker Red CMXRos (M7512; Molecular Probes, Eugene, OR, USA) for 30 minutes. The fluorescence images were taken just prior to taking the corresponding Raman images. Samples were excited using a 532-nm diode laser with a 0.05% of neutral-density filter through a 63× water immersion objective (506148; Leica, NA = 0.90; Leica, Wetzlar, Germany) at an exposure time of 0.001 seconds/pixel. The area under the curve ranges from 2000 to 4000, 2000 to 3500, and 1500 to 3000 cm−1 were used to reconstruct fluorescence images for Lipi-Red, ER-Tracker Red, and MitoTracker Red CMXRos, respectively. For photobleaching, the neutral-density filter was removed and the same settings used for obtaining fluorescence images were applied for other conditions. When the photobleaching was insufficient, the process was repeated or a longer exposure time (0.1 seconds/pixel) was applied.

    2.5 Development of imageCUBE software

    ImageCUBE is available in macOS Mojave (10.14 or later). It is written in Python, and the source code, the installer, and the demonstration files are available at https://github.com/MasaakiU/ImageCUBE and http://www.ncgmlipidsp.jp/english/research.html.

    2.6 Cosmic ray removal

    All of hyperspectral Raman data produced by the WiRE 5.1 software were first converted to SPC files using the WiRE Batch Converter (Renishaw, Wotton-under-Edge, United Kindgom) and the subsequent data processing steps were performed using ImageCUBE software. The cosmic ray removal (CRR) process can be divided into three steps: detection of candidates for cosmic ray positions from each spectrum, exclusion of false positive cosmic ray positions from the candidates, and the replacement of detected cosmic rays. These steps are detailed as follows:

    2.6.1 Detection of candidate list for cosmic ray positions

    Each spectrum data are represented by the vector a, from which a differential spectrum b and the vector of its absolute values c were calculated as follows:
    b n = a n + 1 - a n , (1)
    c n = b n , (2)
    where a n, b n, and c n represent the components of the nth data point of vectors a, b, and c, respectively. From this, a candidate list of cosmic ray positions was determined as follows:

    where the value 10 were used for both window and threshold.

    2.6.2 Exclusion of false positive cosmic ray positions

    After executing the first step for all spectra of the pixels in a data set, the candidate list of cosmic ray positions within a range of ±window data points were clustered as follows:
    Each vector b containing candidate cosmic ray clusters was then compared to the differential vector of the spatially nearby (top, bottom, left, and right) spectra b′, which was calculated using Equation (1), to exclude false positive cosmic rays such as spectra with sharp peaks. Each b′ was first transformed to fit b according to the following equation:
    Corrected b n = b n - b l ¯ × std b k std b k + b l ¯ , (3)
    where x ¯ and std(x) are the average and the standard deviation of a set of elements in vector x, respectively, and x(condition) is a sub-vector of x extracted by collecting elements whose positions meet the condition. The conditions k and l in Equation (3) are defined as follows using ps and pe, which represent the start and end positions (or minimum and maximum positions) of the candidate cosmic ray clusters, respectively:
    k : p s - w i n d o w position p e + w i n d o w , l : p s - w i n d o w position < p s or p e < position p e + w i n d o w , (4)
    Finally, the candidate cosmic ray clusters that met the following conditions for all of the nearby pixels were retained while the others were excluded to obtain a final list of cosmic ray positions:
    var b k - c o r r e c t e d b k > t h r e s h o l d × var b k - b k , (5)

    where var(x) represents the variance of a set of elements in vector x.

    2.6.3 Replacement of detected cosmic rays

    Cosmic rays detected in the previous steps were replaced with a list of values calculated based on the spectra in nearby pixels. To do this, the nearby spectra were transformed to fit the central spectrum by using a region without cosmic rays based on a linear least squares regression, and the coordinates (α and β) for fitting were calculated as follows:
    α β = A T A - 1 A T a m , m : ¬ p s position p e for all p s , p e pairs in a spectrum , (6)
    where the matrix A can be described as follows by using the number of data points M that satisfy the condition m, and all-ones vector of size M, JM (ie, a vector containing only numeral ones as elements):
    A = a m J M T . (7)
    The data points in the cosmic ray cluster region were then replaced with the estimated spectrum, which was calculated by averaging the following equation for all of the nearby spectra:
    Estimated spectrum = a m ~ J M ~ T α β , (8)
    m ~ : p s position p e for any p s , p e pairs in a spectrum , M ~ : the number of data points that satisfy the condition m ~ . (9)

    2.7 Principal component analysis-based noise reduction

    Following the CRR process, principal component analysis (PCA)-based noise reduction was performed. First, the matrix X was prepared as follows:
    X = a 1 a 2 a N T , (10)
    where N represents the number of pixels in a data set. Then, PCA was performed against X - X ¯ to calculate the eigenvalues (λ1, λ2, … λS) and corresponding eigenvectors (e1, e2, … eS), where
    λ 1 > λ 2 > λ S , S : number of data points in a spectrum , X ¯ = 1 N i a i T J N (11)
    The noise-reduced data set Y was then calculated using the top x eigenvalues and their corresponding eigenvectors as follows:
    Y = X - X ¯ T e 1 e 2 e x T e 1 e 2 e x + X ¯ . (12)

    2.8 Spectral unmixing using non-negative linear least squares method

    To perform spectral unmixing, the measured spectra were assumed to consists of the reference spectra from labeled fatty acid (represented by p or p1, p2, …, pL, where L equals the number of labeled free fatty acids added to the culture medium), a background spectrum q, and a baseline drift (represented by r, r1, or r2) (Figure 1). To obtain reference spectra from the labeled fatty acids p, the spectra of the labeled free fatty acid standards, p′, were measured on a CaF2 glass plate, and then, the spectrum of CaF2 glass plate s was subtracted from p′ as follows:
    p = p - t × s . (13)

    In this case, the coefficient t was calculated to make the regional spectrum of p from 250 to 450 cm−1, where the sharp peak derived from CaF2 exists, as linear as possible (Figure 2B). As for the reference spectrum of palmitic acid-d2, it was generated by subtracting the spectrum of a cell-free-position (corresponds to the vector s in Equation [13]) from the spectrum of lipid droplet region (corresponds to the vector p′ in Equation [13]]) in HeLa cells treated with 30 μM palmitic acid-d2 for 24 hours. The coefficient t was calculated to make the regional spectrum of p from 3300 to 3800 cm−1, where the large peak derived from water exists, as linear as possible (Figure 2A).

    As for the other components consisting of the measured spectrum, a spectrum from a cell-free position in the same region of interest was used as a background spectrum q, and the baseline drift was divided into two components—the slope and the intercept—represented by r1 and r2, respectively. The coefficients for each spectrum (P1, P2, …, PL, Q, R1, and R2) were then calculated as follows:
    arg min p 1 o p 2 o p L o q o r 1 o r 2 o T P 1 P 2 P L Q R 1 R 2 - a o , subject to P 1 , P 2 , , P L , Q , R 2 0 , (14)

    where the condition o indicates a user-defined region for unmixing. For all of the experiments undertaken in this study, spectral ranges from 450 to 680 and from 1900 to 2400 cm−1 were used for the unmixing of Br-labeled fatty acids and the unmixing of fatty acids with other labelings (deuterium and alkyne), respectively. Of note, even when the spectra of the labeled fatty acids did not overlap with each other or the cells were treated with only one fatty acid, spectral unmixing has to be performed in all cases because signal from the background or the baseline drift should always be considered.

    2.9 Calculation of P values

    P values of the spectral unmixing were calculated based on the variance as follows:
    Var = p 1 o p 2 o p L o q o r 1 o r 2 o p 1 o p 2 o p L o q o r 1 o r 2 o T - 1 · P 1 P 2 P L Q R 1 R 2 T P 1 P 2 P L Q R 1 R 2 · 1 freedom , (15)

    where Var is a diagonal matrix in which the diagonal components represent the variance of each coefficient (P1, P2, …, PL, Q, R1, and R2), and freedom equals the number of data points minus (L + 3). Based on variances and t-distributions, the P values were then calculated for each coefficient.

    2.10 Reconstruction of nucleus images

    The nucleus images were reconstructed by calculating the signal-to-baseline area ranges between 1560 and 1590 cm−1 (Figure 3A). In this process, area values lower than zero were regarded as zero.

    2.11 Separation of lipid droplets

    Cells were collected from four 10-cm dishes, resuspended in a 1-mL ice-cold homogenization buffer (50 mM Tris HCl, pH 7.4, 150 mM NaCl, 5 mM EDTA, 1 mM phenylmethanesulfonyl fluoride [10837091001; Roche, Basel, Switzerland], and complete EDTA-free protease inhibitor [05056489001; Roche, Basel, Switzerland]), and disrupted by N2 cavitation (4639; Parr Instrument, St, Moline, IL, USA) with 800 psi under 4°C for 20 minutes. The homogenized cells were centrifuged twice at 3000 g for 10 minutes to remove cell debris and undisrupted cells. An aliquot (500 μL) of the supernatant was diluted with an equal volume of 80% of sucrose solution (195-07925; Wako, Osaka, Japan), resulting in a 40% of sucrose sample solution. A discontinuous sucrose gradient was then generated in a centrifuge tube (347357; Beckman Coulter, Brea, CA, USA) by overlaying 1 mL of 40% of sucrose sample solution with 200 μL each of 30%, 25%, 20%, 15%, 10%, 5%, and 0% of sucrose solutions in a homogenization buffer. The gradient was then centrifuged at 44 000 g in a swing rotor (TLS-55; Beckman Coulter, Brea, CA, USA) using an Optima MAX-E ultracentrifuge (Beckman Coulter, Brea, CA, USA) for 3 hours and, finally, nine fractions (280 μL each for fractions [Fr] 1-8 and 160 μL for Fr 9) were collected from the top. For the following immunoblotting and GC-MS experiments, aliquots of 40 and 220 μL were used, respectively.

    2.12 Antibodies

    Mouse monoclonal anti-Calnexin (610523; BD, Franklin Lakes, NJ, USA) and rabbit polyclonal anti-Perilipin-3 (10694-1-AP; proteintech, Rosemont, IL, USA) were used as the primary antibodies. HRP-conjugated anti-mouse IgG (NA931; GE Healthcare, Chicago, IL, USA) and HRP-conjugated anti-rabbit IgG (NA9340; GE Healthcare, Chicago, IL, USA) were used as the secondary antibodies.

    2.13 Immunoblotting

    Aliquots of 40 μL sample were mixed with SDS-PAGE sample buffer (46.7 mM Tris HCl, pH 6.8, 5% of glycerol, 1.67% of SDS [31607-65; Nacalai Tesque, Kyoto, Japan], 2.5% of 2-mercaptoethanol [M7522; Sigma-Aldrich, St. Louis, MO, USA], 0.003% of bromophenol blue) and heated (95°C) for 5 minutes. Cell lysates were resolved using SDS-PAGE, transferred to an Immobilon-P PVDF membrane (IPVH00010; Millipore, Burlington, MA, USA), and blotted with primary and secondary antibodies. The protein bands were visualized using Clarity Western ECL substrate (1705060; Biorad, Hercules, CA, USA). Signals were captured using a ImageQuant LAS 500 CCD camera (GE Healthcare, Chicago, IL, USA), and the resulting images were processed using ImageJ 2.0.0 software.14

    2.14 Measurement of fatty acids using GC-MS

    For fatty acid analysis by GC-MS, lipid samples were extracted using the Bligh and Dyer method.15 After adding free fatty acid C23:0 (R420230; Sigma-Aldrich, St. Louis, MO, USA) to the extracted samples as an initial standard, the free fatty acids and fatty acyl groups were methylated with the Fatty Acid Methylation Kit (06482-04; Nacalai Tesque, Kyoto, Japan) and purified using the Fatty Acid Methyl Ester Purification Kit (06483-94; Nacalai Tesque, Kyoto, Japan) following the manufacturer's instructions. Fatty acid methyl ester (FAME) samples were then measured using a gas chromatograph quadrupole mass detector (GCMS-QP2010 Ultra; Shimadzu, Kyoto, Japan) and a FAMEWAX fused silica capillary column (30 m × 0.25 mm internal diameter × 0.25 μm) (12497; Restek, Bellefonte, PA, USA). An aliquot of 1 μL-sample was injected in splitless mode, with helium used as the carrier gas at a flow rate of 45 cm/s (linear velocity) and the injection port temperature set at 250°C. In this process, the oven temperature was first set to 40°C and maintained for 2 minutes, raised to 140°C at a rate of 20°C/min, to 200°C at a rate of 11°C/min, to 240°C at a rate of 3°C/min, and finally held at 240°C for 10 minutes. The interface and ion source temperatures were maintained at 250 and 200°C, respectively, and the mass spectrometer was operated in electron impact ionization mode with 70 eV ionization energy. Mass spectra were obtained in scan mode from m/z 50 to 600 with the cycle time set to 0.3 seconds.

    2.15 Analysis of GC-MS data

    Quantification of oleic acid, docosahexaenoic acid, oleic acid-d17, α-linolenic acid-d14, docosahexaenoic acid-d5, and C23:0 methyl ester was performed by calculating the area under the curve in the mass chromatograms of their respective characteristic fragments (74, 79, 281, 79, 105, and 74, respectively). Quantification of palmitic acid and palmitic acid-d2, linoleic acid and linoleic acid-d4, arachidonic acid, and arachidonic acid-d8 was performed by mass spectral unmixing, as these pairs of compounds could not be separated completely either by retention time or m/z. The coefficients of the nonlabeled and labeled fatty acids (represented by V and V′, respectively) were calculated as follows:
    arg min v u v u T V V - w u , (16)
    subject to
    V , V 0 , u : position m / z list for unmixing , (17)

    where v and v are the reference mass spectra of nonlabeled and labeled fatty acid standards, respectively, and w is the measured mass spectrum of the sample. As a m/z list for unmixing, m/z = 270, 271, 272, 273, 274, m/z = 294, 295, 296, 297, 298, 299, and m/z = 91, 83 were used for palmitic acid and palmitic acid-d2, linoleic acid and linoleic acid-d4, and arachidonic acid and arachidonic acid-d8, respectively. Calibration curves for each compound were prepared by using either known concentrations of FAME (for nonlabeled fatty acids) (CRM47885; Sigma-Aldrich, St. Louis, MO, USA) or that of fatty acids methyl esterified as described in “Measurement of fatty acids using GC-MS” section of Materials and methods (for labeled fatty acids and C23:0 internal standard).

    2.16 Statistical analysis

    All statistical calculations were performed using Python 3.6 (https://www.python.org).


    3.1 Raman spectroscopy of atomically labeled free fatty acids

    To observe labeled fatty acids incorporated into cells by Raman microscopy, they should contain characteristic spectral patterns that are located in the silent region, in which the Raman spectral intensity of cellular biomolecules is relatively low. In addition, to enable multiplex imaging, each spectroscopic pattern of chosen fatty acids must be distinctive from others in terms of shape or wavenumber. Thus, to find suitable fatty acid combinations for multiplex imaging, we examined the spectra of available atomically labeled free fatty acids. We prepared free fatty acids labeled with either deuterium, alkyne, or Br and classified them based on their respective Raman spectra (Table S1 and Figure S1). Deuterium atoms located in methylene groups (–CD2–) within free fatty acids have a peak at ~2105 cm−1, while those in methyl groups (–CD3) have peaks at ~2069 and ~2132 cm−1 and those in alkenyl groups (–CD=CD–) have a peak at ~2248 cm−1 (red, blue, and orange dotted lines,respectively, in Figure S1A). Alkyne-labeled free fatty acids have a peak at around ~2118 cm−1 (green dotted lines in Figure S1A) distinct from any of the deuterium-labeled free fatty acids. Br-labeled free fatty acids have characteristic peaks at around 530 and 613 cm−1 (magenta lines in Figure S1B), where spectra of nonlabeled counterparts and a cellular background are nearly flat. Of note, the spectral patterns did not always follow this classification. For example, palmitic acid-d2 exhibited vibrational patterns (peaks at 2148 and 2203 cm−1) that differed from those of other free fatty acids labeled in the methylene groups (Figure S1A), as patterns of the Raman spectra from chemical bonds are also affected by the atoms surrounding the bonds.

    3.2 Spectral unmixing using a newly developed ImageCUBE software

    To perform multiplex imaging, it is necessary to extract the original fatty acid spectra from the mixed spectra in samples. To this aim, we utilized the linear least squares method, which is commonly used to analyze spectral data.16, 17 We hypothesized that the measured spectrum is a mixture of labeled fatty acid spectra, a background spectrum, and a baseline drift (slope and intercept) (Figure 1). We then calculated the coefficients of each component (P1, P2, Q, R1, and R2 in Figure 1) by making the difference between the sum of the unmixed and measured spectra as small as possible (see Materials and Methods for details).

    To perform this spectral unmixing process at each pixel of measured data, we developed open-source software, which we named ImageCUBE (the installer, demonstration files, and the source code is available at https://github.com/MasaakiU/ImageCUBE and http://www.ncgmlipidsp.jp/english/research.html). Using ImageCUBE, methods for spectral unmixing can be easily created and executed depending on the combination of labeled free fatty acids used. Unlike two-dimensional data including z-slice and time-lapse images, there are few open-source software packages for hyperspectral images that can create, share, and execute complicated processes; therefore, we equipped ImageCUBE with a plugin functionality through which users can flexibly expand the use of the application. Because plugin functions can be written in Python code as simple text files, custom user-written plugins can easily be installed into ImageCUBE. Using this plugin functionality, ImageCUBE can perform as a platform to accommodate a broad range of applications for the scientific analysis of hyperspectral images, including large-scale Raman hyperspectral data sets. As a demonstration, we created several plugins including ones that were used to analyze data in this paper.

    Details are in the caption following the image
    Procedure for spectral unmixing. The solid and dashed black lines both indicate the measured spectra from the sample. The solid red lines at the top row indicate the reference spectra used for unmixing, which consists of spectra of fatty acid standards, cell free position, and two types of baseline drift. Coefficients for each component, P1, P2, Q, R1, and R2, can be calculated based on the measured and reference spectra. The unmixed spectra can be generated by multiplying coefficients of each component with their corresponding reference spectra (solid red lines at the bottom row). The red line in the leftmost panel at the bottom row indicate the sum of the unmixed spectra

    3.3 Multiplex imaging of five fatty acids incorporated into HeLa cells

    Based on the set of standard spectra (Figure S1), we estimated that a maximum of five fatty acid signals incorporated into cells could be extracted using spectral unmixing. To demonstrate this, we chose the following five biologically important fatty acids whose spectra were distinctive enough to be unmixed: palmitic acid-Br, oleic acid-alkyne, α-linolenic acid-d14, arachidonic acid-d8, and eicosapentaenoic acid-d5. HeLa cells were treated with these labeled free fatty acids (30 µM each) for 24 hours, fixed, and then, observed using Raman microscope. When performing spectral unmixing, the reference spectrum of palmitic acid-d2 was generated as follows; the spectrum of cell free area in the culture dish were subtracted from that of lipid droplets in HeLa cells treated with palmitic acid-d2 (Figure 2A). This is because the spectrum pattern of palmitic acid-d2 from cells were different from the standard spectrum measured on CaF2 (Figures 2A and S1). The reference spectra of the other free fatty acids and the corresponding spectra from cells were almost identical with each other (unpublished observations), so that the spectrum of CaF2 was subtracted from standard free fatty acid spectra measured on CaF2 to get pure reference spectra (Figure 2B). By using these reference spectra, spectral unmixing was performed on each pixel in the acquired data using ImageCUBE software (red lines in Figure 3A). The sum of the unmixed spectra closely matches with the measured spectrum (black lines in Figure 3A), suggesting a successful spectral unmixing. To further confirm the results, the p-values for each compound at each pixel were calculated using ImageCUBE plugin function (Figure 3B), with the results confirming that the signal of each compound was reliable even in pixels with low signal intensity. We also reconstructed nucleus images by calculating the signal-to-baseline area values from 1560 to 1590 cm−1 (filled with cyan in Figure 3A), which should represent the adenosine vibrational signal of DNA.18 Although the molecular structures and/or functions differ between fatty acid species, their distributions were found to be very similar in that they formed punctate structures at almost exactly the same positions in the cytosolic area and little were localized at the nucleus area.

    Details are in the caption following the image
    Calculation of reference spectra. A, Reference spectrum of palmitic acid-d2. HeLa cells were treated with 30 μM of palmitic acid-d2 for 24 hours and observed with Raman microscope. Spectrum of background in cell free position was weighed and subtracted from the measured spectrum in lipid droplet region, in a way that the region from 3300 to 3800 cm−1 of the calculated spectrum become as linear as possible. B, Reference spectrum of arachidonic acid-d8. Spectrum of CaF2 was weighed and subtracted from the measured spectrum of standard free fatty acids on CaF2, in a way that the region from 250 to 450 cm−1 of the calculated spectrum become as linear as possible
    Details are in the caption following the image
    Multiplex fatty acid imaging and fluorescence-Raman hybrid imaging. A, HeLa cells seeded on CaF2 glass substrates were treated with indicated five kinds of labeled fatty acids (30 μM each) for 24 hours, fixed and observed with Raman microscope. Spectral unmixing was performed separately for the range of 450-680 cm−1 (palmitic acid-Br) or 1900-2400 cm−1 (oleic acid-alkyne, α-linolenic acid-d14, arachidonic acid-d8, eicosapentaenoic acid-d5). The black lines indicate the representative measured spectra, and the red lines indicate each component of the unmixed spectra added to the baseline drift (first five panels from the left) or the sum of the unmixed spectra (sixth panel from the left). For the nucleus image, signal to baseline ranges from 1560 to 1590 cm−1 (indicated by cyan in the rightmost panel) were calculated. B, To validate the spectral unmixing performed in (A), P values were calculated for each component in each pixel, which were expressed in a natural logarithm. The lighter color indicates the higher P value. Pixels which meet P value <.05 for all components are shown as white pixel in the top right binary image. C, Fluorescence-Raman hybrid imaging. HeLa cells seeded on quartz bottom dishes were treated with oleic acid-d17 (30 μM) for 24 hours, and Lipi-Red was added to the medium as a lipid droplet marker 30 minutes before the fixing. Images of oleic acid-d17 and Lipi-Red were reconstructed from the unmixed spectra and the fluorescence spectra, respectively. D, Representative fluorescence spectrum derived from Lipi-Red (green) and Raman spectrum (black) in (C). The area under the curve used to reconstruct the fluorescence image (2000 to 4000 cm−1) are filled with green. E, The representative result of spectral unmixing in (C). The measured spectrum is shown in the black line, and unmixed spectra of oleic acid-d17, background, and baseline drift (left panel) or sum of them (right panel) are shown in red lines. Scale bars: 10 μm. Images of multiple cells were taken and the representative image are shown

    3.4 Fluorescence-Raman hybrid imaging requiring only a conventional Raman microscope

    We speculated that the accumulated punctate structures represented the lipid droplets in the cytosol. To confirm this, it would be useful to apply fluorescent probes for lipid droplets, as their localizations have been already well studied. Typically, it is necessary to install a fluorescence microscopy system onto the Raman microscopy system to detect both Raman and fluorescence spectra19, 20; however, this is not only costly, but also requires positional alignment of two data captured by different detectors, which greatly reduce the accuracy of spatial colocalizations. Although there is another option to use anti-Stokes fluorescence emission, this technique requires the installation of a notch filter and the fluorescence background could affect the analysis of Raman spectra.21 Instead, we utilized the bleachable nature of fluorescent probes and developed a fluorescence-Raman hybrid imaging method without upgrading a conventional Raman microscope. In this approach, a narrow-bandwidth laser and a spectrometer, both are necessary components of typical Raman microscopes, were used to carry out the excitation and emission filtering functions, respectively, for acquiring fluorescence signals (Figure S2A,B). The same laser with stronger power (Figure S2C) was then used to bleach the fluorescence signals, thereby reducing the fluorescence background prior to take the Raman images.

    Using this method, we attempted to confirm that the accumulated punctate structures shown in Figure 3A were actually lipid droplets. HeLa cells treated with oleic acid-d17 were stained with Lipi-Red, a fluorescent probe used as a lipid droplet marker.22 The area under the curve of fluorescence spectra ranges from 2000 to 4000 cm−1 (from 595.3 to 675.8 nm) was used to reconstruct fluorescence images (the left panel in Figure 3C and the green line in Figure 3D). After taking the fluorescence image, bleaching was performed, the Raman spectra were obtained (the black line in Figure 3D,E), and spectral unmixing was applied (red lines in Figure 3E). The image of oleic acid-d17 was reconstructed based on the results of the spectral unmixing (the middle panel in Figure 3C). Finally, a composite image of the fluorescence and Raman components was constructed (the right panel in Figure 3C). The signal from Lipi-Red and the oleic acid-d17 were almost completely merged, suggesting that the dot-like structures observed in the Raman images were indeed lipid droplets.

    3.5 Disappearance of accumulated puncta by the inhibition of TAG biosynthesis

    Lipid droplets contain high amounts of TAG, which are biosynthesized by diacylglycerol O-acyltransferase (DGAT) 1 and DGAT2. Both enzymes localize at the endoplasmic reticulum (ER) (and, in the case of DGAT2, on lipid droplets) and catalyze the final step of TAG synthesis using diacylglycerol (DAG) and acyl-CoA as substrates.23 The two enzymes are redundant to each other, and the simultaneous inhibition of both abrogates lipid droplet biogenesis. To further confirm that the punctate structures observed in Figure 3A were actually lipid droplets, HeLa cells treated with T863 and PF-06424439, specific inhibitors for DGAT1 and DGAT2, respectively,24, 25 were cultured with four differently labeled free fatty acid species (palmitic acid-d2, α-linolenic acid-d14, arachidonic acid-d8, and eicosapentaenoic acid-d5). As a result, the punctate structures were almost completely disappeared from the images reconstructed by spectral unmixing, suggesting that the accumulated structures are actually lipid droplets (Figure 4A,C).

    Details are in the caption following the image
    Accumulated puncta disappeared by the inhibition of TAG biosynthesis. A, C, Simultaneous treatment of DGAT1 and DGAT2 inhibitors abolished lipid droplet biogenesis in HeLa cells. HeLa cells were seeded on quartz bottom dishes, and incubated with or without DGAT inhibitors for 24 hours, followed by the treatment with indicated four kinds of labeled fatty acids (30 μM each) for another 24 hours. ER-tracker (A) or mito-tracker (C) were added to the medium 30 minutes before the fixing, and then, the Fluorescence-Raman hybrid imaging were performed. Representative measured (black lines), unmixed, (red lines in the left four panels) sum of unmixed (red lines in the fifth panels from the left), and fluorescence (green lines) spectra are displayed at the bottom of each image. The area under the curve used to reconstruct the fluorescence image are filled with green. B, D, Merged images and the correlation plot of alpha-linolenic acid and organelle markers are shown. In each correlation plot, the vertical and the horizontal axes indicate the fluorescence and Raman signals, respectively. E, Quantification of correlation analysis for HeLa cells treated with DGAT1 and DGAT2 inhibitors in (B) and (D). The mean correlation coefficient values are shown. Asterisks indicate significant differences between ER-tracker and mito-tracker at P value <.001 by t test. Scale bars: 10 μm. For each condition, images of 6 cells were randomly taken and analyzed. The experiment was performed one time

    Fluorescence-Raman hybrid imaging using ER-tracker or mito-tracker suggests that under the inhibition of lipid droplet formation, at least some fatty acids (α-linolenic acid and docosahexaenoic acid) were relatively localized on ER, where remodeling of fatty acyl groups in phospholipids are most active26 (Figure 4B,D,E).

    3.6 Multiplex imaging revealed the preference for fatty acid incorporation into lipid droplets

    When HeLa cells were treated with excessive amounts of free fatty acids, they were primarily incorporated into lipid droplets regardless of their species, although weak signals were also observed in the nonlipid droplet areas (Figure 3A,B). To demonstrate that the proposed multiplex fatty acid imaging approach can visualize behavioral differences of incorporated fatty acid species, we compared the average signal ratio of lipid droplet region to nonlipid droplet region for the following six fatty acids: palmitic acid, oleic acid, linoleic acid, α-linolenic acid, arachidonic acid, and docosahexaenoic acid. Considering the maximum number of fatty acids observed at the same time and the fact that the accuracy of unmixing is higher when the number of components to unmix is smaller, we compared the ratios of two labeled fatty acids in the presence of the other four nonlabeled fatty acids over fifteen patterns (two combinations of six) of experiments (Figure 5A). Note that we calculated the mean values of linoleic acid-d4 to arachidonic acid-d11 ratio, and linoleic acid-d11 to arachidonic acid-d8 ratio, as the spectra of linoleic acid-d4 and arachidonic acid-d8 are very similar (Figure S1A and colored box in Figure 5A).

    Details are in the caption following the image
    Fatty acid with higher unsaturation showed higher value of LD/non-LD ratio. A, Combination of labeled fatty acids used for comparing LD/non-LD ratio. Ratios of two labeled fatty acids in the presence of the other four nonlabeled fatty acids (20 μM each) were measured over fifteen patterns (two combinations of six). Palmitic acid (PA)-d2, oleic acid (OA)-d17, linoleic acid (LA)-d4, α-linolenic acid (aLA)-d14, arachidonic acid (AA)-d8, and docosahexaenoic acid (DHA)-d5, and their nonlabeled forms were used to compare 14 out of 15 patterns. The comparison of the other pattern, labeled AA and LA (colored with yellow) was performed in two ways by using AA-d8 and LA-d11, or AA-d11 and LA-d11, since the spectra of AA-d8 and LA-d4 were almost exactly the same. HeLa cells were treated with these fatty acids for 24 hours. B, Procedure for the data processing. (i) Cosmic ray removal and (ii) PCA-based noise reduction procedures were performed for acquired data, and then, by using (iii) standard spectra subtracted background, (iv) spectral unmixing and (v) its validation were executed. From these results, (vi) ratio images of each fatty acid pair, (vii) image masks of LD and non-LD areas, and (v) another image mask representing validated pixels were created. Finally, LD/non-LD ratio values were calculated from those created images and masks. C, Ratio image of arachidonic acid-d8 and palmitic acid-d2. The left and the middle images represents the distribution of AA-d8 and PA-d2 inside cells, respectively, and the right ratio image is the result of the division of the AA-d8 image by the PA-d2 image. D, Representative spectra of LD area (solid crosshair) and non-LD area (dotted crosshair). In the bottom two graphs, the spectra of PA-d2 (blue lines), AA-d8 (orange lines), background (solid gray lines), and the sum of the unmixed spectra (magenta lines) added to the baseline drift (dotted gray lines) are shown. In the top two graphs, the baseline drift are flattened. E, Ratio of LD/non-LD area for each fatty-acid combination. The order is sorted by the average values. For each condition, images of 8 to 12 cells were randomly taken and analyzed. F, Quantified heatmap of (E). Colors are displayed in log2 scale. The experiments were successfully performed twice. Scale bar: 10 μm

    Workflow for comparing lipid droplet region to nonlipid droplet region are shown in Figure 5B. First, measured Raman hyperspectral data undergo (i) cosmic ray removal and (ii) PCA-based noise reduction processes to remove pulse and high-frequency noises (see Materials and Methods for details). Then, by using (iii, Figure 2) the standard spectra subtracted background, (iv, Figure 1) spectral unmixing and (v, Figure 3B) validation of unmixing were performed. From images of each fatty acids generate from unmixed results, (vi, Figure 5C) ratio images of each fatty acid pair and (vii) image masks of lipid droplet and nonlipid droplet areas were created. From the result of the validation of unmixing, another image mask representing validated pixels was created. Combining these ratio images and masks, (viii) average ratios of masked lipid droplet and nonlipid droplet areas were calculated. Finally, (ix, Figure 5E) lipid droplet/nonlipid droplet ratio values were obtained by dividing these average ratios.

    The results of the comparison between palmitic acid-d2 and arachidonic acid-d8 are shown as a representative in Figure 5C,D. First, the images of each fatty acid distribution were generated based on the result of the spectral unmixing, then, the ratio image of arachidonic acid-d8 to palmitic acid-d2 were obtained. In the ratio image, clear lipid droplet shapes can be observed, suggesting that the ratio of arachidonic acid-d8 to palmitic acid-d2 is higher in the lipid droplet regions than in the nondroplet regions. The representative unmixed spectra in the indicated pixels in each region (red crosshairs in Figure 5C) are shown in Figure 5D, which confirms that the ratio of arachidonic acid-d2 to palmitic acid-d2 is higher in the lipid droplet regions. Based on the ratio image in Figure 5C, the average ratio in lipid droplet and nonlipid droplet areas were obtained, then, lipid droplet/nonlipid droplet ratio values were calculated. The results of performing the same procedures for all of the combinations listed in Figure 5A are shown in Figure 5E. In the figure, the plots representing the division of more-unsaturated fatty acid species by less-unsaturated fatty acid species tends to show higher lipid droplet/nonlipid droplet ratio, suggesting that the polyunsaturated fatty acids (PUFA) are more concentrated in the lipid droplets than in other cellular spaces. The results were also expressed as a heatmap of average ratio values in Figure 5F. The higher values at the bottom left panel suggest that fatty acids with higher number of double bonds are, in general, more efficiently concentrated in the lipid droplets.

    We also biochemically confirmed the above results using gas chromatography-mass spectrometry (GC-MS). First, lipid droplets were extracted from HeLa cells pretreated with six types of labeled free fatty acids (palmitic acid-d2, oleic acid-d17, linoleic acid-d4, α-linolenic acid-d14, arachidonic acid-d8, and docosahexaenoic acid-d5) by density-gradient ultracentrifugation method. The extraction of lipid droplets was confirmed by Western blotting analysis (Figure 6A). In HeLa cells treated with free fatty acids, perilipin-3 (PLIN3), which localizes on lipid droplets and facilitates their synthesis,27 accumulated at the top fraction (Fr1), while the distribution patterns of the other organelle marker, calnexin (ER), did not change, suggesting that lipid droplets were enriched in Fr 1. Next, we measured the total amounts of labeled fatty acids (including free and esterified form) and their corresponding nonlabeled counterparts in the input (cell lysate prior to separation by ultracentrifugation) and those in Fr1, using GC-MS. As previously reported,28 the absolute amount of each fatty acid varies, and the ratio of Fr1 to input differs for each nonlabeled fatty acid (Figure 6B). However, the ratio of labeled fatty acids incorporated into cells at Fr1 to that in the input generally increases as the degree of unsaturation increases, supporting the results of the Raman imagings shown in Figure 5E,F.

    Details are in the caption following the image
    Biochemical analysis of fatty-acid composition shows similar results as Raman imaging. A, HeLa cells were treated with (FA+) or without (FA−) the mixture of six kinds of labeled fatty acids (20 μM each) for 24 hours, then, the lipid droplet fraction (Fr 1) was separated using a density gradient centrifugation method. The separation was confirmed immunoblotting. B, C, The amount of endogenous nonlabeled (B) and incorporated labeled fatty acids (C) in Fr 1 (top row) and input (middle row) are shown, and the ratio of Fr1 to input (bottom row) were calculated. The experiments were successfully performed twice


    In this study, we developed a system to simultaneously visualize multiple fatty acids with higher specificity and higher spatial resolution (<1 µm) than previous work, which can provide fundamental information on the incorporation, storage, and metabolism of fatty acids. Previous localizations of intracellular fatty acids involved biochemical procedures, which have poor spatial resolution, making it difficult to uncover specific subcellular distributions. The tagging of small fluorophores has also been applied to the analysis of fatty acids at the organelle level, but these tags are still too large compared to fatty acids to preserve their native behavior and characteristics.1, 29 Our method significantly outperforms these methods by using free fatty acids labeled at the atomic level, preventing the alteration of native fatty acid characteristics such as hydrophobicity, steric hindrance, or their metabolism in cells. Although a previous and pioneering method was reportedly able to visualize two types of fatty acids in cells through the use of Raman microscope, unfortunately, the separation of signals appeared incomplete, resulting in images of mixed fatty acids.10 Our method applies spectral unmixing using the newly developed ImageCUBE software to obtain much more accurate separation of individual fatty acid signals, thereby visualization of up to five labeled fatty acids is now possible and new insights into the composition of fatty acids within lipid droplets were yielded.

    Under our current approach, the combination of metabolites must be carefully chosen by comparing similarities of their standard spectra, as more distinctive individual spectra result in more accurate spectral unmixing. The labeling of metabolites with 13C or 15N attached to deuterium atoms is a promising approach to make this process easier, since the heavier isotopes shift the vibrational spectra toward lower wavenumbers30 and should increase the variation of spectrum patterns.

    We also developed a convenient fluorescence-Raman hybrid imaging approach that is more simple to implement than previous methods,19-21 and requires no additional optical equipments. Although this approach can only be applied to fixed cells, it is useful in easily analyzing the colocalization of metabolites with fluorescent tagged proteins or organelle markers.

    Using multiplex fatty acid imaging, we detected the biological differences in the behavior of incorporated fatty acids in HeLa cells; that is, the incorporated fatty acids with higher degrees of unsaturation are more efficiently concentrated in lipid droplets. The biological mechanism of this phenomenon remains unclear, however. As lipid droplets are made from DAG and acyl-CoA by DGAT1 and DGAT2, and are degraded by various lipases (Patatin Like Phospholipase Domain Containing 2, hormone sensitive lipase, and so on),31 substrate preferences of these enzymes might be involved. Regulation of fatty acid composition in membrane phospholipid, major fatty acid storage in nonlipid droplet area, might also cause the heterogeneous fatty acid distributions, possibly by lysophospholipid acyltransferases involved in phospholipid de novo and remodeling pathways.26 The biological significance of the distribution differences of fatty acid is another interesting question. Considering the property that PUFAs are easily oxidized, lipid droplets may serve as shelters to protect them from being oxidized inside cells. Previous work also demonstrated that the abnormal enrichment of PUFAs in tumor tissues.32 Since we used HeLa cells derived from cervical cancer, the intracellular distribution differences of fatty acids may be involved in this tumor promoting effects. Further studies are required to answer these questions.

    In summary, we established three new approaches to visualize fatty acids with higher accuracy and spatial resolution within cells. First, we utilized distinctive spectra patterns of deuterium, bromine, and alkyne labeled fatty acids to enable multiplex imaging. Second, a new software named ImageCUBE was developed to perform accurate unmixing of overlapped spectra. Third, we devised a fluorescence-Raman hybrid measuring system applicable for all Raman microscopes to measure lipids and proteins at the same time. The combination of these approaches demonstrate that the proposed methods are very useful tools that can be applied to nearly any type of metabolites, including sugars, amino acids, nucleotides, etc, and chemical compounds such as inhibitors and drug candidates, to reveal their behavioral characteristics via multiplex imaging. The use of our system for the intracellular visualization of multiple fatty acids will not only pave the way to the discovery of novel biological lipid functions, but also help to broaden the analysis of a variety of intracellular metabolites.


    We thank Fumie Hamano (the University of Tokyo) for helping GC-MS measurements, Keizo Waku (Teikyo University), and Mari Shimura (National Center for Global Health and Medicine) for valuable discussion, and Saori Uematsu (the University of Tokyo) for reviewing the manuscripts. This work was supported by AMED-CREST 19gm0910011 (to HS), AMED-P-CREATE 19cm0106116 (to HS), AMED Program for Basic and Clinical Research on Hepatitis JP19fk0210041 (to HS), and Japan Society for the Promotion of Science KAKENHI Grant-in-Aid for JSPS Fellows 18J21897 (to MU). TS was supported by the Takeda Science Foundation.


      M. Uematsu performed the spectroscopy, microscopy, and GC-MS experiments, developed ImageCUBE software, and analyzed the data. M. Uematsu, H. Shindou, Y. Kita, and T. Shimizu designed the experiments, thoroughly discussed the project, and wrote the manuscript.


      The authors declare the following conflicts of interest; Department of Lipidomics, Graduate School of Medicine, The University of Tokyo, is financially supported by Shimadzu Corporation; Department of Lipid Signaling, National Center for Global Health and Medicine, is conducting joint research with Ono pharmaceutical company.