Single cell classification of macrophage subtypes by label-free cell signatures and machine learning

Pro-inflammatory (M1) and anti-inflammatory (M2) macrophage phenotypes play a fundamental role in the immune response. The interplay and consequently the classification between these two functional subtypes is significant for many therapeutic applications. Albeit, a fast classification of macrophage phenotypes is challenging. For instance, image-based classification systems need cell staining and coloration, which is usually time- and cost-consuming, such as multiple cell surface markers, transcription factors and cytokine profiles are needed. A simple alternative would be to identify such cell types by using single-cell, label-free and high throughput light scattering pattern analyses combined with a straightforward machine learning-based classification. Here, we compared different machine learning algorithms to classify distinct macrophage phenotypes based on their optical signature obtained from an ad hoc developed wide-angle static light scattering apparatus. As the main result, we were able to identify unpolarized macrophages from M1- and M2-polarized phenotypes and distinguished them from naive monocytes with an average accuracy above 85%. Therefore, we suggest that optical single-cell signatures within a lab-on-a-chip approach along with machine learning could be used as a fast, affordable, non-invasive macrophage phenotyping tool to supersede resource-intensive cell labelling.

FC, 0000-0002-5436-3857 Pro-inflammatory (M1) and anti-inflammatory (M2) macrophage phenotypes play a fundamental role in the immune response. The interplay and consequently the classification between these two functional subtypes is significant for many therapeutic applications. Albeit, a fast classification of macrophage phenotypes is challenging. For instance, image-based classification systems need cell staining and coloration, which is usually time-and costconsuming, such as multiple cell surface markers, transcription factors and cytokine profiles are needed. A simple alternative would be to identify such cell types by using single-cell, label-free and high throughput light scattering pattern analyses combined with a straightforward machine learning-based classification. Here, we compared different machine learning algorithms to classify distinct macrophage phenotypes based on their optical signature obtained from an ad hoc developed wide-angle static light scattering apparatus. As the main result, we were able to identify unpolarized macrophages from M1-and M2-polarized phenotypes and distinguished them from naive monocytes with an average accuracy above 85%. Therefore, we suggest that optical single-cell signatures within a lab-on-a-chip approach along with machine of view. It is known that M2 present a greater mitochondrial density compared with M1, which can lead to more pronounced side scattering profiles of single cells [37]. Therefore, optical cell signatures obtained from our wide static light scattering approach can significantly improve macrophage phenotype investigations by providing a morphological characterization of cells in suspension. Such fingerprint of a given phenotype is the basis of the present approach to distinguish among monocytes and different macrophage phenotypes.

Results and discussion
The different macrophage phenotypes (M0-unpolarized, M1-and M2-polarized) investigated in this work were first examined via molecular analysis using reverse transcription polymerase chain reaction (RT-PCR) [38] to confirm their polarization state. Data showed a relevant upregulation of CD68 gene expression in M1 samples ( Figure 1. Experimental set-up of label-free approach to classify polarized macrophage subtypes based on in-flow optical signature investigation. Fluid forces three-dimensionally align cells from a cell sample to the centreline of a microfluidic device (from the left to the right), where a collimated laser beam (incident light) interacts with passing individual cells. The light interaction reveals significantly different scattering patterns (optical signature) for each macrophage phenotype as well as monocytes, which a camera-based readout system record. The obtained data are processed and classified with machine learning to obtain a labelfree macrophage phenotype classification. The inset indicates the classified morphological differences between the different macrophage phenotypes correlated to the distinct illustrative optical signatures shown above.  royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 9: 220270 was found in M0 and M2 samples. Instead, M2 exhibited a significant upregulation of IL-10 (79.8 ± 2.45 arb. units) compared with M0 and M1, demonstrating the acquisition of M2-polarized macrophage phenotypes. A slight non-significant IL-10 expression (11.02 ± 3.9) was revealed in M1 samples (figure 2). To classify suspended macrophage phenotypes according to their morphological properties, we investigated their main cellular structures after fluorescence staining using a standard confocal microscope. In more detail, we used a green nucleic acid stain to highlight nuclear and chromatin content combined with a plasma membrane staining to indicate membrane structures. Our observations revealed an evident structural difference of cell cytoplasm contents between M1 versus M0 or M2 macrophage phenotypes, while M0 versus M2 macrophage phenotypes show similar cell staining results (figure 3a). Before each scattering experiment cells were observed at quiescent bright-field condition to investigate possible structural alterations (figure 3b) [38]. It is well known that macrophages present different sizes and shapes in tissue compared with suspension. Moreover, the cell detaching method can alter macrophages shape and properties. For instance, trypsin can down-modulate the surface CD163 level for M2 macrophages [39]. Therefore, we decided to use a cell scraper to minimize possible cell recovery issues after the detachment procedure. However, we observed in suspension a general round shape for all macrophage phenotypes (figure 3c), while characteristic cytosolic granules were observed. Cell observations result in a cell circularity ≥92% for all investigated macrophage phenotypes (figure 3c), which confirms a physiological cell shape after the detaching method and before in-flow scattering experiments. Furthermore, we observed a median monocyte diameter [38]  After quiescent cell structure observations, we performed separate in-flow optical signatures analysis-using the microfluidic-based experimental set-up indicated in figure 1-to investigate in more detail morphological cell feature differences between monocytes, M0, M1 and M2 macrophage phenotypes. Figure 4a summarizes in illustrative three-dimensional scatter plots the biophysical properties [38] obtained from optical cell signatures ( pooled data) and below detailed information of statistical relevant property differences between the different cell types. Not surprisingly, substantial property changes are visible in the overall cell dimension (figure 4b), from monocytes with 9.82 ± 1.59 µm to macrophage phenotypes, which range from approximately 11 to 16 µm. In more detail, M0 are the biggest cells with 15.66 ± 2.84 µm in dimension compared with 11.76 ± 1.72 µm and 14.25 ± Our outcome confirms morphometric cell property changes during cell polarization. In fact, it is well known in literature that the macrophage cytoplasm consists of lysosomes, mitochondria and granules, which composition can change for different phenotypes. From a scattering point of view, the contribution of inner cell structures depends on their dimension, number, position and composition (e.g. cytokine type and concentration). In fact, the optical signature of M2 macrophages illustrates significant scattering differences compared with other phenotypes (figures 1 and 4). For instance, a significantly higher RI N value is detected compared with M0 and M1 phenotypes, which could simply imply a more active cell state or other cell content which is recognized as nucleus. Note that our single-cell scattering approach simplifies a cell as core-shell construct. However, it has been demonstrated that a change in optical density of nucleus can be related to an active state of chromatin (active state of gene transcription). Regarding this, Rostam et al. observed that the fluorescence intensity of nuclear staining is significantly low for M2 cells compared with M1, clearly showing a more intense activity of M2, which may be related to the different action of transcriptional factors on M1 and M2 polarized macrophages [40,41]. Interestingly, the work of Halaney et al., reports a detailed analysis of light scattering distribution of M1 and M2 macrophage phenotypes [37]. The authors found that M1 and M2 present a significantly different amount of scattering intensity at side angles between 2 and 3°, which is in good agreement with our findings (see optical signature distributions in figure 1) for M2 compared with M0 or M1 phenotypes. Compared with Halaney et al., a significantly wider scattering angle range (2-30°) can be observed with our measurement technique, at single cell level resolution [37]. In fact, such additional information allows a more detailed macrophage subtype measurement, with a higher cell throughput rate, thanks to our microfluidic-based measurement concept. From a structural point of view, such behaviour seems to be related to small inner structure differences such as nuclei, lysosomes or/and mitochondria. In fact, mitochondria are known to add a significant contribution at side scattering angles, due to their dimension, optical density and structural location in the cell cytoplasm [37,42].
*** *** *** *** *** *** *** *** *** *** *** *** **   royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 9: 220270 Monocytes (n = 107) [38] show significantly different optical cell signatures compared with M0 macrophage phenotype (n = 149) [38] as reported in the three-dimensional scatter plots ( figure 4). As expected, a ML prediction accuracy greater than or equal to 99.2% was reached for e.g. linear, quadratic or cubic support vector machine (SVM) classifier using the following parameters: crossvalidation = 5, misclassification cost = 1, box constraint level = 1, multi-class method = one-versus-one) without principal component analysis resulting in a prediction speed approximately 19 000 observation s −1 and training time approximately 0.24 s.   royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 9: 220270 Next, we trained a ML algorithm with the biophysical cell properties of (un-)polarized macrophage cell subclasses. First, we trained M1 (n = 167) [38] versus M2 (n = 215) [38] macrophage phenotypes, where we found that the quadratic SVM classifier resulted as the most suitable algorithm according to calculation speed and prediction accuracy (training time approximately 0.51 s with approximately 31 000 observations s −1 and a total misclassification cost of 60) using the following ML parameters: one neighbour, Euclidean metric distance, equal distance weights, box constraint level of 1 and a kernel scale of 1. In fact, such ML parameters are used as a standard penalty for margin-violating observations, to prevent significant overfitting of the experimental data. We repeated the classification process five times resulting in an average prediction accuracy of approximately 85. Our results indicate that the mentioned ML classifier has better sensitivity in classifying M2 macrophage phenotype cells compared with other cells, due to the significantly different combination of biophysical properties (figure 6b). This could be ascribable to the different chromatin condensation and mitochondria presence in M2 cells, compared with other investigated cells.

Conclusion
Image-based machine learning is widely used in research and therapeutic applications, while the label-free investigation of scattering data is still underrated. We highlight in this work the potential of a simple and cost-effective microfluidic-based macrophage phenotype (unpolarized versus polarized) classification approach. Therefore, we analysed first in quiescent and afterwards in-flow condition the biophysical properties of polarized macrophage phenotypes as well as monocytes. Such living cell investigation resulted in a distinctive optical signature, which we used as input for a supervised MLbased cell classification. The analyses of more than 600 cells from three different donors allowed to predict macrophage phenotypes with an accuracy above 72% for M1 versus M2 macrophages and even more than 85% for M0 versus M1 and M2 phenotypes. Such outcome raises the hope for realtime-based cell analysis approaches based on scattering patterns. We believe that our measurement approach can be of significant therapeutic interest, where an identification, quantification and monitoring of both M1 to M2 phenotype is needed.

Cell collection
Human macrophages were recovered from healthy donors after obtaining informed consent, in accordance with relevant guidelines and regulations. In more detail, a standard venepuncture procedure was performed using standard K 2 EDTA tubes (Vacutainer, BD) to prevent coagulation. After sample collection, a standard density gradient separation was performed as followed: first, the whole volume of blood was diluted with an equal volume of phosphate buffered saline (PBS, Euroclone), and then gently layered on with an equal volume fraction of density gradient medium (Histopaque-1077, Sigma Aldrich) using a 50 ml centrifuge tube (Falcon). After that, a centrifugation step was performed at 300g for 30 min and disabled machine brake. After the centrifugation, the resulting peripheral blood mononuclear cells royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 9: 220270 (PBMC) were visible as a ring at the interface between the gradient medium (lower part) and the blood plasma (upper part). PBMC were collected with a disposable Pasteur pipette and washed in 10 ml of erythrocyte lysis buffer, to eliminate a possible contamination. Finally, cells were cultured in RPMI-1640 medium, supplemented with 10% fetal bovine serum, 1% L-Glu and 1% penicillin/streptomycin (Euroclone).

Macrophage phenotype polarization
PBMC were divided into three culture flasks (T-75, Corning) of equally distributed volume fractions to transform monocytes in unpolarized (M0), M1-polarized (M1) and M2-polarized (M2) macrophage phenotypes. After 24 h of incubation at 37°C and 5% CO 2 , cells in suspension (lymphocytes) were discarded, while adherent monocytes were treated for the following macrophage differentiation (day 0). First, cell medium was aspirated from the flask and substituted with RPMI-1640 and specific macrophage phenotype generation media (M0 = C-28057; M1 = C-28055; M2 = C-28056, Promocell). Complete cell medium was made of base medium with supplement mix and cytokines, following the manufacturer's instructions (Promocell). After 6 days (day 6) each flask was supplied with a volume of cell medium equal to 75% of the initial cell volume (day 0). At day 7 a new aliquot of cytokine mix (Promocell) was added to the medium. At day 9 the cell medium of each flask was aspirated to eliminate possible suspended cells, and a fresh volume of appropriate cell medium was added to each flask. At day 10, polarized and not-activated macrophages were detached from flask surfaces using a cell scraper tool and subsequently centrifuged at 200g for 10 min in 15 ml centrifuge tubes (Falcon). Finally, cells were resuspended into 200 µl of complete RPMI-1640 medium, ready to be analysed with our optical cell investigation approach.

Microfluidic device and cell alignment
Cell measurements were performed with a microfluidic device, composed of a supporting geometry fabricated with a three-dimensional printer (Objet30 pro, Stratasys) and a series of two glass channels. Briefly, a round-shaped glass channel (TSP050375, Molex)-where three-dimensional alignment of cells takes place-is inserted on one side in a square shaped readout channel (8240, Vitrocom)-where single cell investigation takes place-which permits the precise in-flow optical readout of cells. The other end of the round-shaped channel is immersed in the cell sample. By applying a certain pressure on the sample liquid, the cell medium is pushed through the channel and enters the microfluidic device. Such sample liquid consists of cells immersed in an alignment medium, consisting of a highly biocompatible viscoelastic polymer ( polyethylene oxide, PEO, molecular weight = 4 MDa, Sigma Aldrich) diluted in PBS at 0.4 wt%. Thanks to the resulting fluid properties, generated by viscoelastic fluid forces, cells are strictly aligned to the centreline of the round-shaped channel and subsequently remain aligned at the centreline of the subsequent microfluidic readout channel. Note that fluid forces have been chosen to prevent cell deformation effects, while ensuring sufficient single cell alignment. In more detail, three-dimensional cell alignment is achieved if the following relationship 3Wi b 2 ðL=2RÞ . Àln(3:5b) is satisfied. Where Wi ¼ 2lU=2R, uses l the relaxation time (0.197 ms) of the viscoelastic fluid, the average fluid velocity (1496 µm −1 ), R the channel radius (25 µm), b ¼ r 1 =R, a non-dimensional geometrical channel parameter, with r 1 being the cell radius and L the channel length (0.35 m). However, the subsequent readout channel allows precise single cell analysis due to its square shape of 400 × 400 μm and preserved three-dimensional alignment. To ensure continuity between the alignment and readout channel, the alignment section is collinearly inserted in the readout section and sealed with a soft ferrule (UP-N-123-03X, Idex). At the end of the readout channel, cells can be recovered for further cell studies.

Sample preparation and observation
Cells are diluted in alignment medium to obtain a final cell concentration of approximately 1 × 10 5 cells ml −1 . Such cell concentration ensures a throughput rate of approximately 2 cells s −1 passing through the readout laser beam. Please note that the sample concentration and fluid velocities were optimized to reduce possible cell-cell interactions and cell deformation effects, while the maximum throughput performance of the actual measurement approach is approximately 50 cells s −1 . Finally, each investigated macrophage phenotype is checked for mycoplasma infection. For off-chip cell investigations, macrophage types were observed with an inverted microscope (IX81, Olympus), to royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 9: 220270 4.6. Experimental set-up and image processing We used a small-angle light scattering technique combined with a viscoelastic microfluidic single-cell alignment approach [30][31][32] In more detail, our cell investigation approach reveals biophysical properties of living cells from individual cell scattering records generated in a continuous angular range from approximately 2°−30°and an angular resolution of approximately 0.1°. Briefly, cells flowing in the readout channel of the microfluidic device pass through a collimated laser beam (λ = 632.8 nm). The resulting scattered light is collected and mapped on a camera sensor (ORCA Flash 4.0, Hamamatsu Photonics). The recorded scattering signatures are processed by a self-written Matlab (R2020b, MathWorks) routine to directly obtain the searched-for light-scattering profile (LSP) and consequently the biophysical cell properties of each passing cell. More specifically, collected LSPs are matched with a lookup table (greater than or equal to 335 000 curves) of previously calculated theoretical LSPs to obtain biophysical cell properties (diameter, D; Nucleus/cytoplasm-ratio, N/Cratio; refractive index of the nucleus and cytoplasm, RI N and RI C , respectively) and to distinguish morphological properties within the sub-micrometric cell dimension range. More detailed information about the LSP matching is shown elsewhere [30].
royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 9: 220270 4.7. Machine learning The machine learning (ML) approach was carried out with a Matlab (R2020b, MathWorks) routine to classify circulating monocytes from macrophages, as well as the main subtypes of macrophage phenotypes (M0, M1 and M2) based on their biophysical properties retrieved from optical cell signatures. Several classification methods were set up with operational parameters chosen based on previous experiments of our working group [31,33]. For ML training, a randomly chosen subset of data (all donors) was used, followed by testing classification accuracy on the remaining data (all donors), while the classification accuracy was measured by fivefold cross-validation. The ML results are shown in a 2 × 2 matrix for M1 versus M2 and 3 × 3 matrix for M0 versus M1 versus M2, with the following values: positive predictive value (PPV) and false discovery rate (FDR).

Statistical analysis
All results are presented as the mean ± standard error. When normality assumptions were met, the statistical significance for two or more groups of data was calculated by using a one-way ANOVA with corresponding Tukey's multiple comparison. Significance is indicated by p values ( ns p > 0.05; Ã p < 0.05, ÃÃ p < 0.01, ÃÃÃ p < 0.001) combined with F-values (F ). We used Excel 365 (Microsoft Corporation) for all statistical analyses [34][35][36].
Ethics. Cells were obtained from the blood bank of the medical school of the Federico II University of Naples (Italy). At the time of blood donation, each donor signed an informed consent (model no. 5526 of Azienda Ospedaliera Universitaria 'FEDERICO II', Naples, Italy) in which it is specified that waste parts of the blood, not useful for the medical-therapeutic purposes of blood donation, could be used for scientific research purposes.
Data accessibility. Datasets for this research work are available from the Dryad Digital Repository: doi:10.5061/dryad. 1ns1rn8wh [38].