Our planet’s ecosystems – from its oceans to its forests – are teeming with microbes. DNA analysis of environmental samples shows that many of these microbes belong to a group known as protists. This group consists of single-celled organisms that are close relatives of fungi, plants and animals. Though protists are a widespread and diverse group, scientists know little about them. One reason for this is the lack of high-throughput ways to recognize and count protists in environmental samples.

Colin, Coelho et al. set out to tackle this blind spot in ecology and cell biology by developing an automated imaging system. The system needed to image many kinds of protist cells in enough detail to see the features inside. The end-result was a 3D-imaging technique called e-HCFM – which is short for “environmental high content fluorescence microscopy”. Colin, Coelho et al. went on to use the technique on 72 samples collected on an expedition across the world’s oceans. This allowed them to automatically image, recognize and classify over 330,000 organisms.

This approach and new dataset will benefit researchers working in many fields, from cell biology to ecology, computational biology and beyond. In the future, this imaging method might integrate with techniques that can analyze the DNA in individual cells. This would allow scientists to link protists’ visible features to their genetic information, in a way that will scale from single cells up to entire ecosystems.

Studies of organismal diversity have traditionally focused on plants and animals, and more recently on prokaryotes and viruses, while much less attention has been given to microbial eukaryotes (Pawlowski et al., 2012; del Campo et al., 2014) (mostly unicellular protists, plus small multicellular organisms between 0.5 µm and 1 mm in size). On the other hand, genetics together with molecular, cell, and developmental biology have used a limited number of model organisms to reveal over the past 40 years key principles underpinning the organization of living matter (Alberts, 2014). However, recent holistic meta-omics surveys of biodiversity across the full spectrum of life (Bork et al., 2015) are revealing nowadays a massive amount of unknown taxa and genes in microbial eukaryotes (de Vargas et al., 2015) implying that much more eukaryotic diversity and functions need to be explored. Phylogenomics (Burki et al., 2016), geological (Knoll, 2014; Falkowski and Knoll, 2007), and ecological (de Vargas et al., 2015; Mahe et al., 2016) records also show how intense symbiogenesis and diversification in protists have shaped the evolution of eukaryotes together with that of biogeochemical cycles. In order to better understand how such processes have led to the complexification of life and the Earth system, it is necessary to bridge contemporary diversity studies to environmental eukaryotic cell biology in an ecological context. As a first step, we need new automated high-content 3D-imaging systems that can cope with the broad size, abundance, and complexity ranges characterizing environmental microbial eukaryotes.

Automated imaging techniques to tackle the diversity of uncultivated aquatic organisms include in-flow systems that couple high-throughput low-resolution imaging with feature extraction from single micro-organisms (Sieracki et al., 1998; Sosik and Olson, 2007), and widefield microscope-based methods applied to phytoplankton recognition and quantification (Embleton et al., 2003; Rodenacker et al., 2006; Schulze et al., 2013). These approaches can characterize eukaryotic organisms from low contrast bright field images acquired at a single focal plane, sometimes in association with auto-fluorescence measurements of photosynthetic pigments (Schulze et al., 2013; Hense et al., 2008). The identification and classification of eukaryotes is then based on semi-automated machine learning approaches for a few to tens of taxonomic classes (Benfield et al., 2007) with a peak performance at 70–90% accuracy, which is comparable to what trained annotators can achieve (MacLeod et al., 2010; Culverhouse, 2007). None of these tools provide images of sufficient quality and resolution to assess the structural complexity of eukaryotic cells whose diversity peaks in the 5 to 50 μm size-range (de Vargas et al., 2015).

To improve both information content and the automated capture of morphological complexity of microbial eukaryotes over a broad size-range, we developed a 3D multichannel imaging workflow denoted e-HCFM (environmental High Content Fluorescence Microscopy) (Figure 1a). Here, we introduce this new method, which adapts recent protocols developed in model cell biology (Pepperkok and Ellenberg, 2006; Eliceiri et al., 2012) to the large diversity of scattered objects that characterizes environmental samples (organisms, debris, aggregates, abiotic objects, etc.). We then apply e-HCFM to 72 plankton samples collected during the Tara Oceans expeditions (Karsenti et al., 2011) (Figure 1—source data 1), containing communities of planktonic organisms in the 5–20 µm size range, which are typically analyzed automatically by flow cytometry with a very low morpho-taxonomic resolution (Marie et al., 2014; Zubkov et al., 2007). We show that e-HCFM allows its users (i) to obtain accurate 3D images of micro-eukaryotes over a broad taxonomic range; (ii) to enrich significantly the information content of each imaged organism; (iii) to segregate automatically the high diversity of planktonic micro-particles, and (iv) to archive annotated digital images of perishable samples. These advances allow linking identification and quantification of uncultured eukaryotes with ecological and functional traits.

Figure 1 with 9 supplements see all

Download asset Open asset

Figure 1—source data 1

This image acquisition registry details the e-HCFM imaging runs, their metadata, their samples of origin, and associated metadata from the Tara Oceans expedition.

https://doi.org/10.7554/eLife.26066.011

Download elife-26066-fig1-data1-v1.xlsx

Figure 1—source data 2

List of descriptors computed for each object imaged through e-HCFM.

https://doi.org/10.7554/eLife.26066.012

Download elife-26066-fig1-data2-v1.xlsx

Size-fractionated plankton communities collected across the world oceans during the Tara Oceans expeditions (Pesant et al., 2015) (Figure 1a) and fixed onboard were used as starting material. Prior to imaging, sample aliquots are loaded into an optical chamber coverglass slide, which is centrifuged to concentrate the organisms onto the glass-bottom of the well, coated for promoting cell adherence (Figure 1a, Figure 1—figure supplement 3). Subsequent preparation steps (rinsing/staining/washing) are minimized to avoid cell loss and manipulations, and thus preserve fragile bodies for reliable quantification. Two environmental samples are prepared manually in 1 hr. Key cellular features were targeted with a combination of fluorescent dyes and auto-fluorescence (Figure 1b, Figure 1—video 1): (i) DNA content; (ii) (intra)cellular membranes and hydrophobic bodies; (iii) chloroplasts and potential photosymbionts, and (iv) the overall contour of eukaryotic cells including cell walls and external matrices. For this last purpose, we specifically designed the PLL-AF546 dye (conjugation between α-poly-L-lysine and Alexa Fluor 546, see method section below) that labels a wide range of exoskeletons made of polysaccharides, proteins, carbonate, silicate, etc., abundant in marine microbial eukaryotes (Figure 1b and Figure 1—figure supplements 1 and 2). The fluorochrome moiety can be easily adapted for other multicolor labeling protocols, and the staining strategy can be applied to either live (Figure 1—figure supplement 2) or fixed cells in suspension (Figure 1—figure supplement 1). Overall, our protocol is highly effective at revealing internal and external cellular structures in organisms representing the known phylogenetic diversity of eukaryotes (Figure 1—figure supplement 1), including their diverse biotic interactions (Figure 2, Figure 2—video 1, and Figure 2—figure supplement 1).

Figure 2 with 2 supplements see all

Download asset Open asset

We used a confocal microscope (Leica Microsystem SP8) equipped for automated high-content imaging. The magnification and resolution are adjusted by selecting a suitable lens for the size-range of the organisms and structures of interest. While the XY size range (1 to 1500 µm) is limited by the fields of view (fov) of the lenses (Figure 1c), the range in the Z-axis (thickness) is constrained by the transparency of the object. Objects are optically sectioned to spatially resolve the fluorescence signal, resulting in sharp images of the entire cells. For each imaged position, five channels are recorded (bright field and four fluorescent signals) and the raw imaged Z-stacks are then segmented to extract single planktonic particles (Figure 1c, Figure 1—figure supplement 4). Particles are further processed individually to sub-segment regions of biological interest (e.g. nuclei or chloroplasts) for each fluorescent channel. A collection of 480 numeric 2D/3D features (Figure 1—source data 2) is computed for each captured particle, including biovolumes, intensity distribution, shape descriptors and texture features. The features were used in combination with a visually curated learning set (see below) to build a machine learning model (Boland et al., 1998), which associates each particle to a category (taxonomic or morphological) of the training set (Figure 3a; Figure 3—source data 1).

Figure 3 with 4 supplements see all

Download asset Open asset

Figure 3—source data 1

Organization of the hierarchical classification scheme for the automated classification, the training set categories abundance and the recall value for each category of the four levels (four tables).

https://doi.org/10.7554/eLife.26066.028

Download elife-26066-fig3-data1-v1.xlsx

Figure 3—source data 2

Confusion matrix generated by the classifier at the classification level 4.

https://doi.org/10.7554/eLife.26066.029

Download elife-26066-fig3-data2-v1.xlsx

Figure 3—source data 3

Relative abundance of each taxon in each sample.

The relationship between sample label and sampling location is provided in ‘Figure 1—source data 1.’.

https://doi.org/10.7554/eLife.26066.030

Download elife-26066-fig3-data3-v1.xlsx

Figure 3—source data 4

Assignment of stations to oceanic provinces.

https://doi.org/10.7554/eLife.26066.031

Download elife-26066-fig3-data4-v1.xlsx

Figure 3—source data 5

Object counts (normalized to seawater volume) per taxonomic group (panel d).

https://doi.org/10.7554/eLife.26066.032

Download elife-26066-fig3-data5-v1.xlsx

Figure 3—source data 6

Measured PO₄ concentrations (panel d).

https://doi.org/10.7554/eLife.26066.033

Download elife-26066-fig3-data6-v1.xlsx

Figure 3—source data 7

Values of N, Spearman correlation (rho), and number of samples (N) for each sub-group (panel d).

https://doi.org/10.7554/eLife.26066.034

Download elife-26066-fig3-data7-v1.xlsx

Imaging and processing of each sample with a 40x lens lasted 8 hr, leading to 396 fovs (Figure 1a,c), over a 22 µm Z-range. Overall, e-HCFM imaging of the 72 Tara Oceans plankton samples (Figure 1—source data 1) produced ~2.5 Tb of raw data, corresponding to 336,655 planktonic particles, each of them associated with a series of derived images for visual inspection including summary thumbnails (Figure 1c, Figure 1—figure supplements 4 and 5), Z-stack animations, 3D segmentation masks, and 3D reconstructions. The segmentation performance of each particle is enhanced by the sharp contrast generated by the fluorescent signals. In particular, the thin transparent biomineral and/or organic structures shaping many groups of eukaryotic taxa can be precisely detected (Figure 1—figure supplement 5), generating more accurate computed features.

We then generated a reference training set to build an automated classifier. A total of 18,103 objects (5.4% of the whole dataset) were manually curated and classified into a 4-level hierarchical framework including 155 categories (taxonomic lineage for organisms and morphological types for other particles, Figure 3a; Figure 3—source data 1). A random forest classifier was empirically determined to perform the best, combining high accuracy and fast computation time. We estimate the overall accuracy of the classifier at 82.2% at the finest taxo-morphological level (corresponding often to genus level), rising to 93.8% at the phylum/class level (Figure 3a, Figure 3—source datas 1 and 2 and Figure 3—figure supplement 1). This shows that our cell descriptors provide efficient segregation of the main plankton taxonomic groups, despite considerable intra-group morphological diversity. The same method applied on image descriptors obtained only from the bright field channel (which are a subset of the 480 features used in this study) resulted in only 56.8% accuracy. This further demonstrates the added value of fluorescence labeling for revealing common taxonomic traits. The method also provides a confidence score that can be used to filter only high-confidence predictions for in-depth analysis, or to focus on the weakest predictions to generate additional curation efforts (Figure 3—figure supplement 2). To further evaluate the value of the features, we computed the classification accuracy using a reduced number of features. We observed that at least 400 features are needed to push accuracy above 82% (see Figure 3—figure supplement 4).

Beyond automated assessment of abundance and diversity of micro-eukaryotes across large geographic and taxonomic scales (Figure 3b, Figure 3—figure supplement 3, Figure 3—source datas 3–4), e-HCFM images and descriptors can be used to link fundamental cell biology features to both taxonomy and ecology. For example, whereas organismal biomasses are usually extrapolated from 2D fingerprints for investigating their relationship to nutrient uptake or to carbon flux, our method enables to use 3D biovolumes and/or sub-cellular structures which can be stratified by taxonomy (Figure 3c). We further show how concentration of a few major taxa increases differentially with the concentration of phosphate (Figure 3d, Figure 3—source datas 5–7). These results add global support to previous reports that purely photosynthetic organisms (e.g., Bacillaryophyta) are more dependent on exogenous phosphate than mixotrophic or heterotrophic (e.g., Holodinophyta) populations (Lin et al., 2016). A further potential of e-HCFM that complements DNA sequence-based inferences is the ability to directly detect and quantify biotic interactions (Figure 2), such as the finding of an unknown nanoflagellate attached to the diatom Chaeoteros simplex (Figure 2—figure supplement 1).

1. 1. Sebastien Colin
   2. Luis Pedro Coelho
   3. Shinichi Sunagawa
   4. Chris Bowler
   5. Eric Karsenti
   6. Peer Bork
   7. Rainer Pepperkok
   8. Colomban de Vargas

   (2017) idr0015-colin-taraoceans/screenA

   Publicly available at Image Data Resource platform (doi: 10.17867/10000108).

   https://idr.openmicroscopy.org/webclient/?show=screen-1201
2. 1. Colin S
   2. Coelho LP
   3. Sunagawa S
   4. Bowler C
   5. Karsenti E
   6. Bork P
   7. Pepperkok R
   8. de Vargas C

   (2016) eHCFM--TARA_OCEANS--HCS1--H5

   Publicly available at EBI-BioStudies (accession no. S-BSST51).

   https://www.ebi.ac.uk/biostudies/studies/S-BSST51

1. 1. Tara Oceans Consortium Coordinators;Tara Oceans Expedition Participants

   (2016) Registry of all samples from the Tara Oceans Expedition (2009-2013)

   Publicly available at PANGAEA (doi:10.1594/PANGAEA.859953).

   https://doi.pangaea.de/10.1594/PANGAEA.859953

1. Book

   1. Alberts B

   (2014)

   Molecular Biology of the Cell

   Garland Science.

   • Google Scholar
2. 1. Benfield M
   2. Grosjean P
   3. Culverhouse P
   4. Irigolen X
   5. Sieracki M
   6. Lopez-Urrutia A
   7. Dam H
   8. Hu Q
   9. Davis C
   10. Hanson A
   11. Pilskaln C
   12. Riseman E
   13. Schulz H
   14. Utgoff P
   15. Gorsky G

   (2007) RAPID: research on automated plankton identification

   Oceanography 20:172–187.

   https://doi.org/10.5670/oceanog.2007.63

   • Google Scholar
3. 1. Berges JA
   2. Franklin DJ
   3. Harrison PJ

   (2001) Evolution of an artificial seawater medium: improvements in enriched seawater, artificial water over the last two decades

   Journal of Phycology 37:1138–1145.

   https://doi.org/10.1046/j.1529-8817.2001.01052.x

   • Google Scholar
4. 1. Boland MV
   2. Markey MK
   3. Murphy RF

   (1998) Automated recognition of patterns characteristic of subcellular structures in fluorescence microscopy images

   Cytometry 33:366–375.

   https://doi.org/10.1002/(SICI)1097-0320(19981101)33:3<366::AID-CYTO12>3.0.CO;2-R

   • PubMed
   • Google Scholar
5. 1. Bork P
   2. Bowler C
   3. de Vargas C
   4. Gorsky G
   5. Karsenti E
   6. Wincker P

   (2015) Tara oceans. tara Oceans studies plankton at planetary scale. Introduction

   Science 348:873.

   https://doi.org/10.1126/science.aac5605

   • PubMed
   • Google Scholar
6. 1. Breiman L

   (2001) Random forests

   Machine Learning 45:5–32.

   https://doi.org/10.1023/a:1010933404324

   • Google Scholar
7. 1. Burki F
   2. Kaplan M
   3. Tikhonenkov DV
   4. Zlatogursky V
   5. Minh BQ
   6. Radaykina LV
   7. Smirnov A
   8. Mylnikov AP
   9. Keeling PJ

   (2016) Untangling the early diversification of eukaryotes: a phylogenomic study of the evolutionary origins of centrohelida, haptophyta and cryptista

   Proceedings of the Royal Society B: Biological Sciences 283:20152802.

   https://doi.org/10.1098/rspb.2015.2802

   • PubMed
   • Google Scholar
8. Conference

   1. Coelho LP
   2. Aabid S
   3. Murphy RF

   (2009)

   Nuclear segmentation in microscope cell images: a hand-segmented dataset and comparison of algorithms

   6thIEEE International Symposium on Biomedical Imaging: From Nano to Macro.

   • Google Scholar
9. 1. Coelho LP

   (2013) Mahotas: Open source software for scriptable computer vision

   Journal of Open Research Software 1:e3.

   https://doi.org/10.5334/jors.ac

   • Google Scholar
10. 1. Culverhouse PF

    (2007) Human and machine factors in algae monitoring performance

    Ecological Informatics 2:361–366.

    https://doi.org/10.1016/j.ecoinf.2007.07.001

    • Google Scholar
11. 1. de Vargas C
    2. Audic S
    3. Henry N
    4. Decelle J
    5. Mahé F
    6. Logares R
    7. Lara E
    8. Berney C
    9. Le Bescot N
    10. Probert I
    11. Carmichael M
    12. Poulain J
    13. Romac S
    14. Colin S
    15. Aury JM
    16. Bittner L
    17. Chaffron S
    18. Dunthorn M
    19. Engelen S
    20. Flegontova O
    21. Guidi L
    22. Horák A
    23. Jaillon O
    24. Lima-Mendez G
    25. Lukeš J
    26. Malviya S
    27. Morard R
    28. Mulot M
    29. Scalco E
    30. Siano R
    31. Vincent F
    32. Zingone A
    33. Dimier C
    34. Picheral M
    35. Searson S
    36. Kandels-Lewis S
    37. Acinas SG
    38. Bork P
    39. Bowler C
    40. Gorsky G
    41. Grimsley N
    42. Hingamp P
    43. Iudicone D
    44. Not F
    45. Ogata H
    46. Pesant S
    47. Raes J
    48. Sieracki ME
    49. Speich S
    50. Stemmann L
    51. Sunagawa S
    52. Weissenbach J
    53. Wincker P
    54. Karsenti E
    55. Tara Oceans Coordinators

    (2015) Ocean plankton. Eukaryotic plankton diversity in the sunlit ocean

    Science 348:1261605.

    https://doi.org/10.1126/science.1261605

    • PubMed
    • Google Scholar
12. 1. del Campo J
    2. Sieracki ME
    3. Molestina R
    4. Keeling P
    5. Massana R
    6. Ruiz-Trillo I

    (2014) The others: our biased perspective of eukaryotic genomes

    Trends in Ecology & Evolution 29:252–259.

    https://doi.org/10.1016/j.tree.2014.03.006

    • PubMed
    • Google Scholar
13. 1. Eliceiri KW
    2. Berthold MR
    3. Goldberg IG
    4. Ibáñez L
    5. Manjunath BS
    6. Martone ME
    7. Murphy RF
    8. Peng H
    9. Plant AL
    10. Roysam B
    11. Stuurman N
    12. Stuurmann N
    13. Swedlow JR
    14. Tomancak P
    15. Carpenter AE

    (2012) Biological imaging software tools

    Nature Methods 9:697–710.

    https://doi.org/10.1038/nmeth.2084

    • PubMed
    • Google Scholar
14. 1. Embleton KV
    2. Gibson CE
    3. Heaney SI

    (2003) Automated counting of phytoplankton by pattern recognition: a comparison with a manual counting method

    Journal of Plankton Research 25:669–681.

    https://doi.org/10.1093/plankt/25.6.669

    • Google Scholar
15. Book

    1. Falkowski P
    2. Knoll AH

    (2007)

    Evolution of Primary Producers in the Sea

    Academic Press.

    • Google Scholar
16. 1. Hense BA
    2. Gais P
    3. Jutting U
    4. Scherb H
    5. Rodenacker K

    (2008) Use of fluorescence information for automated phytoplankton investigation by image analysis

    Journal of Plankton Research 30:587–606.

    https://doi.org/10.1093/plankt/fbn024

    • Google Scholar
17. 1. Karsenti E
    2. Acinas SG
    3. Bork P
    4. Bowler C
    5. De Vargas C
    6. Raes J
    7. Sullivan M
    8. Arendt D
    9. Benzoni F
    10. Claverie JM
    11. Follows M
    12. Gorsky G
    13. Hingamp P
    14. Iudicone D
    15. Jaillon O
    16. Kandels-Lewis S
    17. Krzic U
    18. Not F
    19. Ogata H
    20. Pesant S
    21. Reynaud EG
    22. Sardet C
    23. Sieracki ME
    24. Speich S
    25. Velayoudon D
    26. Weissenbach J
    27. Wincker P
    28. Tara Oceans Consortium

    (2011) A holistic approach to marine eco-systems biology

    PLoS Biology 9:e1001177.

    https://doi.org/10.1371/journal.pbio.1001177

    • PubMed
    • Google Scholar
18. 1. Kiernan JA

    (2000)

    Formaldehyde, formalin, paraformaldehyde and glutaraldehyde: What they are and what they do

    Microscopy Today 8:8–12.

    • Google Scholar
19. 1. Knoll AH

    (2014) Paleobiological perspectives on early eukaryotic evolution

    Cold Spring Harbor Perspectives in Biology 6:a016121.

    https://doi.org/10.1101/cshperspect.a016121

    • PubMed
    • Google Scholar
20. 1. Kolisko M
    2. Boscaro V
    3. Burki F
    4. Lynn DH
    5. Keeling PJ

    (2014) Single-cell transcriptomics for microbial eukaryotes

    Current Biology 24:R1081–R1082.

    https://doi.org/10.1016/j.cub.2014.10.026

    • PubMed
    • Google Scholar
21. 1. Lin S
    2. Litaker RW
    3. Sunda WG

    (2016) Phosphorus physiological ecology and molecular mechanisms in marine phytoplankton

    Journal of Phycology 52:10–36.

    https://doi.org/10.1111/jpy.12365

    • PubMed
    • Google Scholar
22. 1. MacLeod N
    2. Benfield M
    3. Culverhouse P

    (2010) Time to automate identification

    Nature 467:154–155.

    https://doi.org/10.1038/467154a

    • PubMed
    • Google Scholar
23. Preprint

    1. Mahe F
    2. de Vargas C
    3. Bass D
    4. Czech L
    5. Stamatakis A
    6. Lara E
    7. Mayor J
    8. Bunge J
    9. Sernaker S
    10. Siemensmeyer T
    11. Trautmann I
    12. Romac S
    13. Berney C
    14. Kozlov A
    15. Mitchell E
    16. Seppey C
    17. Singer D
    18. Egge E
    19. Wirth R
    20. Trueba G
    21. Dunthorn M

    (2016) Soil protists in three neotropical rainforests are hyperdiverse and dominated by parasites

    bioRxiv.

    https://doi.org/10.1101/050997

    • Google Scholar
24. Book

    1. Marie D
    2. Partensky F
    3. Vaulot D
    4. Brussaard C

    (1999)

    Current Protocols in Cytometry

    Robinson J. P, editors. John Wiley & Sons, Inc.

    • Google Scholar
25. 1. Marie D
    2. Rigaut-Jalabert F
    3. Vaulot D

    (2014) An improved protocol for flow cytometry analysis of phytoplankton cultures and natural samples

    Cytometry Part A 85:962–968.

    https://doi.org/10.1002/cyto.a.22517

    • PubMed
    • Google Scholar
26. 1. Pawlowski J
    2. Audic S
    3. Adl S
    4. Bass D
    5. Belbahri L
    6. Berney C
    7. Bowser SS
    8. Cepicka I
    9. Decelle J
    10. Dunthorn M
    11. Fiore-Donno AM
    12. Gile GH
    13. Holzmann M
    14. Jahn R
    15. Jirků M
    16. Keeling PJ
    17. Kostka M
    18. Kudryavtsev A
    19. Lara E
    20. Lukeš J
    21. Mann DG
    22. Mitchell EA
    23. Nitsche F
    24. Romeralo M
    25. Saunders GW
    26. Simpson AG
    27. Smirnov AV
    28. Spouge JL
    29. Stern RF
    30. Stoeck T
    31. Zimmermann J
    32. Schindel D
    33. de Vargas C

    (2012) CBOL protist working group: barcoding eukaryotic richness beyond the animal, plant, and fungal kingdoms

    PLoS Biology 10:e1001419.

    https://doi.org/10.1371/journal.pbio.1001419

    • PubMed
    • Google Scholar
27. 1. Pedregosa F

    (2011)

    Scikit-learn: machine learning in python

    Journal of Machine Learning Research : JMLR 12:2825–2830.

    • Google Scholar
28. 1. Pepperkok R
    2. Ellenberg J

    (2006) High-throughput fluorescence microscopy for systems biology

    Nature Reviews Molecular Cell Biology 7:690–696.

    https://doi.org/10.1038/nrm1979

    • PubMed
    • Google Scholar
29. 1. Pesant S
    2. Not F
    3. Picheral M
    4. Kandels-Lewis S
    5. Le Bescot N
    6. Gorsky G
    7. Iudicone D
    8. Karsenti E
    9. Speich S
    10. Troublé R
    11. Dimier C
    12. Searson S
    13. Tara Oceans Consortium Coordinators

    (2015) Open science resources for the discovery and analysis of Tara Oceans data

    Scientific Data 2:150023.

    https://doi.org/10.1038/sdata.2015.23

    • PubMed
    • Google Scholar
30. 1. Rodenacker K
    2. Hense B
    3. Jütting U
    4. Gais P

    (2006) Automatic analysis of aqueous specimens for phytoplankton structure recognition and population estimation

    Microscopy Research and Technique 69:708–720.

    https://doi.org/10.1002/jemt.20338

    • PubMed
    • Google Scholar
31. 1. Schulze K
    2. Tillich UM
    3. Dandekar T
    4. Frohme M

    (2013) PlanktoVision--an automated analysis system for the identification of phytoplankton

    BMC Bioinformatics 14:1.

    https://doi.org/10.1186/1471-2105-14-115

    • PubMed
    • Google Scholar
32. 1. Sieracki CK
    2. Sieracki ME
    3. Yentsch CS

    (1998) An imaging-in-flow system for automated analysis of marine microplankton

    Marine Ecology Progress Series 168:285–296.

    https://doi.org/10.3354/meps168285

    • Google Scholar
33. 1. Sosik HM
    2. Olson RJ

    (2007) Automated taxonomic classification of phytoplankton sampled with imaging-in-flow cytometry

    Limnology and Oceanography: Methods 5:204–216.

    https://doi.org/10.4319/lom.2007.5.204

    • Google Scholar
34. Data

    1. Tara Oceans Consortium, Coordinators
    2. Tara Oceans Expedition, Participants

    (authors) (2017) Registry of all samples from the tara oceans expedition (2009-2013)

    PANGAEA.

    https://doi.org/10.1594/PANGAEA.859953
35. 1. Williams E
    2. Moore J
    3. Li SW
    4. Rustici G
    5. Tarkowska A
    6. Chessel A
    7. Leo S
    8. Antal B
    9. Ferguson RK
    10. Sarkans U
    11. Brazma A
    12. Salas REC
    13. Swedlow JR

    (2017) The Image data resource: a bioimage data integration and publication platform

    Nature Methods 14:775–781.

    https://doi.org/10.1038/nmeth.4326

    • PubMed
    • Google Scholar
36. 1. Zubkov MV
    2. Burkill PH
    3. Topping JN

    (2007) Flow cytometric enumeration of DNA-stained oceanic planktonic protists

    Journal of Plankton Research 29:79–86.

    https://doi.org/10.1093/plankt/fbl059

    • Google Scholar

Thank you for submitting your article "Quantitative 3D-Imaging for Cell Biology and Ecology of Environmental Microbial Eukaryotes" for consideration by eLife. Your article has been reviewed by two peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Wendy Garrett as the Senior Editor. The following individual involved in review of your submission has agreed to reveal her identity: Heidi Sosik (Reviewer #1).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

The reviewers made a number of constructive comments, which you should carefully address. These include:

Essential Revisions:

1) At the review stage, it was extremely difficult to find, understand, and evaluate the supplementary tables and videos. In place of the links on the download page, the editorial office sent me a zip file of the tables and videos with filenames that match the designations in the main manuscript and in your supplementary information, but some of those appear to be mislabeled or to contain the wrong information. For example, the file "Figure 3D—source data 1" looks like it just contains the statistical values listed for Figure 3—figure supplement 3—source data 1; whereas that latter file contains the principal component values from the figure. These all need to be checked and their contents appropriately described.

2) Samples were fixed in a high concentration of glutaradehyde (25%) which contributes strong autofluorescence, especially in the green region of the spectrum, that may confound the specific (dye-based) fluorescent signals. Have the authors compared the current images to images of samples fixed only with paraformaldehyde? Please comment.

3) The authors extract a large number of features from each sample. Have they performed a simple principal component analysis to determine which features are most informative? I would suspect that most of the discrimination power is contained within a relatively small subset of these parameters. Please discuss.

4) The thresholding procedure ("mean + 1.5 standard deviations") is arbitrary and depends on the density of fluorescent objects in the image as well as the level of the background. How is the fidelity of the classification impacted by the threshold level? Classification results using at least one other threshold level (e.g. mean+3 standard deviations) or an automated thresholding algorithm should be presented and discussed.

5) The authors claim to have imaged the "entire extent" of each well. However, the bright-field images will be strongly degraded near the edges of the wells due to the cone of transmitted light hitting the vertical edges of the chambers and thus missing the condenser. How was this issue addressed? If this is a simple case of over-statement, then a more accurate description should be used.

6) In terms of throughput, did the authors assess the utility of simple widefield microscopy (with deconvolution) as an alternative to confocal? This will allow for much faster data acquisition. This warrants some discussion.

7) Did the authors correct for channel bleed-through in the 4-color images? This is especially the case when using DiO and AF546. This combination may introduce some cross talk, confounding the classifier.

8) Did the authors compare the static threshold of 1.5 standard deviation above average with a more robust/sophisticated algorithm such as Otsu's method?

9) Very often, classifiers are built such that 50% of the data set is used to train, while the remaining 50% is used to test. Was there a reason the training set was such a small portion of the total data? Was this limited by manually classifying the organisms?

In summary, the reviewers generally appreciated this paper and we look forward to receiving a revised manuscript.

Essential Revisions:

1) At the review stage, it was extremely difficult to find, understand, and evaluate the supplementary tables and videos. In place of the links on the download page, the editorial office sent me a zip file of the tables and videos with filenames that match the designations in the main manuscript and in your supplementary information, but some of those appear to be mislabeled or to contain the wrong information. For example, the file "Figure 3D—source data 1" looks like it just contains the statistical values listed for Figure 3—figure supplement 3—source data 1; whereas that latter file contains the principal component values from the figure. These all need to be checked and their contents appropriately described.

We have cross-checked all caption and files. Reviewer was correct that we had incorrectly labeled the data files corresponding to “Figure 3—figure supplement 3”. We have now both corrected the labeling and added the full relative abundance matrix. We have also two additional tables with the averaged measurements depicted in the panel 3D of Figure 3. Note that the full raw data (including the assignment of each object to a taxonomic group and the raw feature measurements) is, as before, provided through EcoTaxa due to its large size.

2) Samples were fixed in a high concentration of glutaradehyde (25%) which contributes strong autofluorescence, especially in the green region of the spectrum, that may confound the specific (dye-based) fluorescent signals. Have the authors compared the current images to images of samples fixed only with paraformaldehyde? Please comment.

Reviewer is right in mentioning the green/yellow emitted fluorescence induced by the formaldehyde and glutaraldehyde fixative which occurs mainly under blue and green excitation [see Lee, K., Choi, S., Yang, C., Wu, H. C., and Yu, J. Autofluorescence generation and elimination: a lesson from glutaraldehyde. Chemical Communications, 49(29), 3028-3030 (2013)]. However, we did not use a final concentration of 25% Glutaraldehyde, but a mix of 1% formaldehyde (prepared from paraformaldehyde) and only 0.25% Glutaraldehyde, as mentioned in the first version. We have modified the sentence to avoid confusion between stock concentration and working concentration: “For e-HCFM, 45 ml were poured into a 50 ml tube pre-aliquoted with 5 ml of 10% monomeric formaldehyde (1% final concentration) buffered at pH7.5 (prepared from paraformaldehyde powder) and 500 μl of 25% EM grade glutaraldehyde (0.25% final concentration) [Maria et al., 1999].”

We did not compare our low-glutaraldehyde protocol to a protocol using paraformaldehyde only, as it is anyways very hard to disentangle signals from native autofluorescences (e.g. green autofluorescence or fluorescence from photosynthetic pigments like phycobiliproteins), which can be strong and out of prediction/control in environmental samples. However, we strongly decreased the impact of glutaraldehyde autofluorescence by minimizing its final concentration. In addition, both the DiOC6 and AlexaFluor546 staining overlap the glutaraldehyde excitation/emission spectra, but they generate much brighter fluorescence signal than the glutaraldehyde autofluorescence, further reducing its overall contribution. Note that the signal quantification of these two channels was mostly used for classification rather than function quantification.

In order to quench the autofluorescence from glutaraldehyde (coming from double bond with primary amines and free aldehyde groups), the fixative could also be reduced by a NaBH4 treatment. However, this treatment would negatively impact other critical parts of our protocol: the generation of H₂ gas bubble can interfere with cell attachment to the bottom of the well; preserving free aldehyde groups contributes to cross-link poly-L-lysine with cellular proteins.

3) The authors extract a large number of features from each sample. Have they performed a simple principal component analysis to determine which features are most informative? I would suspect that most of the discrimination power is contained within a relatively small subset of these parameters. Please discuss.

We thank the reviewer for this question as it prompted us to quantitatively address it. We used the random forest method to select features that are most discriminative and to test increasing number of features (selected from most informative to least informative). The results demonstrate that almost 200 feature are necessary until accuracy similar to the full dataset is reached and that even after 400 features there is still some improvement observed.

Note too that features are computed in groups and features from the same group may have differing discriminatory power. We find overlap features (features computed from the overlap of fluorescence between channels) to be present both among the most discriminative and least discriminative features. Similarly, the zero-th Zernike moments (effectively, average fluorescence after normalization) are not very discriminative, but other Zernike moments are much more important.

We have now summarized this analysis in the main text (Results and Discussion, fourth paragraph) and added Figure 3—figure supplement 4 as a supplemental figure (with its corresponding source data, Figure 3—figure supplement 4—source data 1).

4) The thresholding procedure ("mean + 1.5 standard deviations") is arbitrary and depends on the density of fluorescent objects in the image as well as the level of the background. How is the fidelity of the classification impacted by the threshold level? Classification results using at least one other threshold level (e.g. mean+3 standard deviations) or an automated thresholding algorithm should be presented and discussed.

We chose this procedure based on previous work in fluorescence image analysis which showed that the mean could outperform methods based on automated thresholding algorithms such as Otsu [Coelho, Aabid and Murphy, 2009]. This value is computed for the whole plate and reused for the different fields since we did not expect (and did not observe) any large differences in background fluorescence between fields. Since background occupies the majority of the volume, this can provide a good separation between background and objects. Due to the high variance in object brightness, we feared that a method such as Otsu may erroneously classify dim objects as background. Indeed, quantitatively, using Otsu results in much fewer objects being detected (231,610 compared to 336,655 with the method in the text).

We followed the reviewer’s suggestion and implemented the threshold using the “mean + 3 standard deviations” and the segmentation is similar enough that the classification system achieves almost identical accuracy (82.1% compared to 82.2% for the version in the text). When using Otsu, we see a slight decrease (78.5%), which may be due to the fact that several objects in the training data are no longer detected (see above).

In the revised version, we have added the following two sentences justifying our choices 1) “This method was chosen based on previous work which identified the mean as a good threshold for fluorescence microscopy images [Coelho, Aabid and Murphy, 2009]; 2) “Tests with Otsu thresholding led to fewer objects being detected (231,610 compared to 336,655 with the chosen method).”

5) The authors claim to have imaged the "entire extent" of each well. However, the bright-field images will be strongly degraded near the edges of the wells due to the cone of transmitted light hitting the vertical edges of the chambers and thus missing the condenser. How was this issue addressed? If this is a simple case of over-statement, then a more accurate description should be used.

We confirm that the signal from transmitted light is degraded for the field of view in the vicinity of the well edge. The meniscus of the liquid in the well also impacts this signal depending on the position of the objective. Nonetheless, in terms of classification, the information extracted from the bright field channel is still valuable (perhaps because many of the features are robust to illumination changes). As we describe in the text (Results and Discussion, fourth paragraph), we obtain 56.8% accuracy using only the bright field images. Even though this is much worse than the results when combining all channels (including the fluorescence channels), it is still significantly better than random (recall that we have 155 classes). To further explore this question, we attempted classification while excluding the features from the bright field channel. The result was slightly worse accuracy (82.0% vs. 82.2%).

6) In terms of throughput, did the authors assess the utility of simple widefield microscopy (with deconvolution) as an alternative to confocal? This will allow for much faster data acquisition. This warrants some discussion.

The reviewer is right. Acquisition by widefield microscopy with deconvolution was considered when we initiated the project a few years ago. Unfortunately, the 3D deconvolution of large fields of view from fluorescence microscopy datasets required intensive computation (several seconds per stacks and per channels). The time saved at the acquisition level was lost at the computation level without having the optical sectioning performance and the excitation/emission versatility of our SP8 CLSM setup. We also explored the use of spinning disk systems, equipped with a Yokogawa CSU-X1 Confocal Scanner Unit. The time saved in scanning at similar resolution was lost in the multiplication of overlapping positions because the fields of view were much smaller than the one acquired with the SP8 CLSM setup.

However we agree that both widefield and spinning disk microscopy deserve further evaluation in the future. Recent GPU-based deconvolution algorithms [see Bruce, M. A. and Butte, M. J. Real-time GPU-based 3D Deconvolution. Opt. Express 21, 4766–4773 (2013).] and open source softwares [see Sage, D. et al. DeconvolutionLab2: An open-source software for deconvolution microscopy. Methods 115, 28–41 (2017).] are making the widefield microscopy approach more time-efficient and economically attractive; both methods can also take advantage of the higher sensitivity and dynamical range of the latest camera as compared to the photomultiplier tubes used in CSLM. Following the reviewer’s suggestion, we now mention this in the Conclusion.

Note that we are investigating another way to speed up the acquisition process by implementing a 2-steps acquisitions in e-HCFM. The first step is a fast prescan of the cell preparation aiming at spotting the object of interest based on a reduced set of features. The second step triggers dedicated imaging job at positions of interest. This method significantly improves the triangular trade off: ‘imaging time / information quality / quantity of object’, while still taking advantage of the great optical sectioning performance and the ex/em versatility of CLSM. It is specifically adapted for rare events or discrete distribution of cells without wasting time in imaging at high resolution empty area. It also significantly decreases image storage requirement and associated cost. We are not mentioning this work in progress in this paper presenting the primary version of e-HCFM.

7) Did the authors correct for channel bleed-through in the 4-color images? This is especially the case when using DiO and AF546. This combination may introduce some cross talk, confounding the classifier.

In our protocol, the recording of DiOC(6) and AlexaFluor546 fluorescence were separated sequentially through 2 acquisition steps. The first step combines an excitation at 488 nm and 638 nm which allows the measurement of DiO emission signal between 505-520 nm and chlorophyll autofluorescence between 680-720 nm. We now mention: “AF546 is poorly excited by 488 nm and 638 nm beamlines and its emission was not detected between 505-520 nm and between 680-720 nm. The chlorophyll fluorescence was recorded in the spectral range 680-720 nm where AF546 and Hoechst do not emit with 488 nm and 638 nm illumination.”

However a low autofluorescent noise can be revealed by maximum projection in this Chlorophyll channel (see Figure 1—figure supplement 5) which seems related to core cell structures. The origin of this signal is not clear but its detection was caused by the selection of high sensitivity settings to enable the detection of dim chlorophyll signal. Similarly, the second step excites simultaneously Hoechst at 405 nm and AF546 at 552 nm. We now mention: “Chlorophyll is also excited at 405 nm but it does not emit in both the hoechst (415-475 nm) and AF456 (570-590 nm) channels. […] The potential signal from Hoechst was finally not subtracted because the AF546 was mostly used to delineate cell edges.”.

We would also note that as long as any cross-talk or auto-fluorescence is consistent within the same taxonomic group, the classification system will be robust.

8) Did the authors compare the static threshold of 1.5 standard deviation above average with a more robust/sophisticated algorithm such as Otsu's method?

See response to point 4.

9) Very often, classifiers are built such that 50% of the data set is used to train, while the remaining 50% is used to test. Was there a reason the training set was such a small portion of the total data? Was this limited by manually classifying the organisms?

While it is true that the training set represents a small fraction of the whole dataset, we consider that this reflects rather the large size of the dataset. In absolute terms, though, we consider 18,103 manually labeled objects as a large training set.

In addition, we carefully screened the whole object collection for optimizing the diversity of classes in the training set, and created classes each time a group of similar objects reached at least 30 (subsection “(3) Object classification”). On the other hand we did not overweight classes of very abundant object and considered that few hundreds of representative objects were enough to define a class statistically which is now mentioned in the aforementioned subsection.

Author details

1. Sebastien Colin

   1. UMR 7144, team EPEP, Station Biologique de Roscoff, Centre Nationnal de la Recherche Scientifique, Roscoff, France
   2. Université Pierre et Marie Curie, Sorbonne Universités, Roscoff, France
   3. Advanced Light Microscopy Facility, European Molecular Biology Laboratory, Heidelberg, Germany

   Contribution

   Conceptualization, Methodology, Investigation, Data curation, Formal analysis, Software, Visualization, Funding acquisition, Writing—original draft, Writing—review and editing

   Contributed equally with

   Luis Pedro Coelho

   For correspondence

   colin@sb-roscoff.fr

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-4440-9396

2. Luis Pedro Coelho

   Structural and Computational Biology, European Molecular Biology Laboratory, Heidelberg, Germany

   Contribution

   Methodology, Investigation, Data curation, Formal analysis, Software, Visualization, Writing—original draft, Writing—review and editing

   Contributed equally with

   Sebastien Colin

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-9280-7885

3. Shinichi Sunagawa

   Structural and Computational Biology, European Molecular Biology Laboratory, Heidelberg, Germany

   Present address

   Department of Biology, Institute of Microbiology, ETH Zurich, Zurich, Switzerland

   Contribution

   Formal analysis, Visualization, Supervision, Funding acquisition, Writing—review and editing

   Competing interests

   No competing interests declared

4. Chris Bowler

   Institut de Biologie de l’École Normale Supérieure, École Normale Supérieure, Paris Sciences et Lettres Research University, Paris, France

   Contribution

   Funding acquisition, Writing—review and editing

   Competing interests

   No competing interests declared

5. Eric Karsenti

   1. Institut de Biologie de l’École Normale Supérieure, École Normale Supérieure, Paris Sciences et Lettres Research University, Paris, France
   2. Directors’ Research, European Molecular Biology Laboratory, Heidelberg, Germany

   Contribution

   Conceptualization, Methodology, Resource, Supervision, Funding acquisition, Writing—review and editing, Project administration

   Competing interests

   No competing interests declared

6. Peer Bork

   Structural and Computational Biology, European Molecular Biology Laboratory, Heidelberg, Germany

   Contribution

   Resource, Supervision, Funding acquisition, Writing—review and editing

   Competing interests

   No competing interests declared

7. Rainer Pepperkok

   1. Advanced Light Microscopy Facility, European Molecular Biology Laboratory, Heidelberg, Germany
   2. Cell Biology and Biophysics Unit, European Molecular Biology Laboratory, Heidelberg, Germany

   Contribution

   Methodology, Resource, Supervision, Writing—review and editing

   Competing interests

   No competing interests declared

8. Colomban de Vargas

   1. UMR 7144, team EPEP, Station Biologique de Roscoff, Centre Nationnal de la Recherche Scientifique, Roscoff, France
   2. Université Pierre et Marie Curie, Sorbonne Universités, Roscoff, France

   Contribution

   Conceptualization, Methodology, Formal analysis, Resource, Supervision, Funding acquisition, Writing—review and editing, Project administration

   For correspondence

   c2vargas@gmail.com

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-6476-6019

• 5,542

  Page views

• 566

  Downloads

• 26

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.