Macrophages are involved in the body’s immune responses and tissue homeostasis. They play a critical role in infectious diseases (Atri et al., 2018), cancer progression (Chanmee et al., 2014), autoimmunity (Funes et al., 2018), wound healing (Krzyszczyk et al., 2018), and many other diseases (Murray et al., 2014; Wynn and Vannella, 2016). Two main subsets of activated macrophages with different functional phenotypes, M1 (classically activated, pro-inflammatory) and M2 (alternatively activated, anti-inflammatory), have been identified (Martinez and Gordon, 2014; Murray et al., 2014). In tumor progression, M1 and M2 macrophages assume distinct roles. Specifically, the M2 subtype, M2d, demonstrates pro-neoplastic characteristics, while M1-like macrophages exert anti-tumor effects (Sica et al., 2008). Recent studies have shown that besides M1 and M2 types, a continuum of macrophage subtypes exists (Mosser and Edwards, 2008; Martinez and Gordon, 2014).

The study of macrophage subtypes plays a crucial role in identifying strategies for disease control (Kotwal and Chien, 2017; Chatterjee et al., 2021; Du et al., 2021; Kuntzel and Bagnard, 2022). Consequently, developing effective methods for detecting macrophage subtypes in vitro is essential.

The conventional method to identify M1 and M2 subtypes involves analyzing multiple cell surface markers, transcription factors, and cytokine profiles, which can be time-consuming and resource-intensive. Furthermore, uncertainty remains about how to identify macrophage subtypes confidently. This is in part due to the existing continuum of states. Recent studies of macrophage cultures led researchers to hypothesize that cell morphology could indicate macrophage activation status (Rostam et al., 2017).

Previous research on the classification of macrophage subtypes using machine learning has been based on fluorescent dyes and cell shape parameters (Mcwhorter et al., 2013; Rostam et al., 2017; Pavillon et al., 2018). More recently, it was suggested that motility parameters like cell speed could be used to classify macrophage subtypes (Kesapragada et al., 2024). These publications show that, although a continuum of phenotypes exists, there indeed are three primary shape modes associated with three distinct phenotypes, respectively (inclusive of the so-called “naïve” macrophages). Furthermore, it was shown that these shape modes are closely linked with predetermined cell motility metrics.

This research aims to gain a deeper understanding of macrophages’ migratory patterns and explore the potential of classifying macrophage subtypes based on raw trajectory patterns without relying on cell shape analysis. This could provide more robust methods to analyze blurry images.

We suggest that a holistic use of cell motility information, i.e., a time series of cell coordinates, could enhance the differentiation of macrophage subtypes. We develop a deep learning model that uses cell position over time as input and demonstrate that our model effectively distinguishes between M1 and M2 macrophage phenotypes. This classification method could potentially be used to understand the continuum of states further.

In this study, we leverage labelled data published in our previous work (Kesapragada et al., 2024)). In brief, bone marrow-derived macrophages (BMDMs) were isolated and cultured, resulting in an M0 macrophage culture. M0 macrophages were further activated into either M1 (with LPS) or M2 (with IL-4) (see Methods) in correspondence with existing protocols (Zhang et al., 2008; Ying et al., 2013). Time-lapse recording of cell images, segmentation, and tracking of cell trajectories was performed in correspondence with Magnusson et al. (2015). This global track-linking algorithm links cell outlines generated by a segmentation algorithm into tracks. Tracks are incrementally added to the image sequence using information from the complete image sequence in every linking decision.

Overall, we obtained three videos of single cells and three videos of cell cultures.

Single-cell video sequences were captured for one M0 cell, one M1 cell, and one M2 cell, respectively. Each single-cell video comprises 240 images, with a frame interval of 1 min, resulting in a total duration of 240 min.

Cell-culture video sequences were obtained from a non-activated M0 cell culture, an M1-activated cell culture, and an M2-activated cell culture, respectively. Each cell culture video comprises 37 phase contrast frames, each frame captured at a 5-minute interval, spanning a total duration of 180 min.

Given the complexity of the problem of classifying different trajectory patterns and the amount of data available, our results (Figure 5A) show that the network is not overfitting and learning effectively. We observe that the training loss decreases over time while the validation loss remains low. Similarly, the training accuracy increases while the validation accuracy remains high. This indicates that the model can generalize well to new data and is likely to perform well on new unseen data. The model’s ability to accurately classify the macrophage subtypes is supported by its successful predictions on the validation data. Figure 5B represents a confusion matrix visualization where the percentage cell counts of each class are plotted on a heatmap. In this representation, the ground truth data labels “Cluster C,” “Cluster P,” and “Cluster E” are displayed in the rows, while the model predicted labels “M0,” “M1,” and “M2” are displayed in the columns. We can observe that the model has an accuracy of 91% in predicting Cluster E (elongated) cells as M2, 95% accuracy in predicting Cluster P (protruded) cells as M1, and 87% accuracy in predicting Cluster C (circular) cells as M0. We can see that the model has identified 12% of the Cluster C cells as M2 and 2% as M1. It is important to note that identifying M0 cells can be challenging, as they may exist in a continuum or may continuously transform into other subtypes. This trained model is applied to test single-cell macrophage images (Figure 1A), which accurately categorized the cells into M0, M1, and M2.

Cell-trajectory-based and cell-shape-based methods can potentially classify macrophage subtypes more accurately at the single-cell level. This differs from other methods, such as q-PCR, that give information about cell population subtypes.

As we observed in the cell morphology analysis (Kesapragada et al., 2024), each cell culture (M0, M1, or M2) is dominantly represented by a corresponding morphology cluster: M0 cells are circular, M1 cells are protruded, M2 cells are elongated, but the shape-based clusters were nonetheless diluted.

Current analysis of macrophage trajectories revealed typical cell movement patterns corresponding to cell shape clusters: circular cells spin at their initial position, protruded cells wander, and elongated cells move in a specific direction. However, trajectory-based clusters do not exactly correspond to shape-based clusters. This implies a variation in the resulting cell phenotype of macrophage culture activated into M1, M2, or non-activated (M0).

We also note that it is challenging to arrive at a set of metrics that holistically capture all the shape and movement features of the cell.

Several image-based models for macrophage type identification have been published previously. In Pavillon et al. (2018), the authors created a model for noninvasive distinguishing between cell types employing activated and non-activated macrophages for testing. The classification model was linear, allowing greater biological interpretation. However, it required multiple sources of information as input, including phase microscopy images, Raman spectra, and autofluorescence microscopy. In Rostam et al. (2017), the authors used image-based machine learning approach to classify M1, M2, naive macrophages, and monocytes. Their algorithm relied on various cell metrics, including details related to the nucleus and cytoskeleton, and requires high-quality images.

Our deep learning model relies on the raw cell trajectory data in the x-y plane, as well as the cell displacement and traveling distance. It automatically captures the features of cell motion unseen by the human eye and demonstrates good accuracy.

Our study revealed unique migratory patterns and distinct morphology in the three subtypes of macrophages. M0, M1, and M2. By analyzing their trajectories and computing various quantitative measures, such as perimeter, area of the convex hull, and pairwise distances, we observed clear differences in the migratory patterns of these macrophage cell types. However, building a classification model with predetermined features is challenging in this context.

Therefore, we take a different approach by developing a deep learning model that incorporates the trajectory path and shape of cells, which proves to be more effective in accurately classifying macrophage subtypes. The correlation between cell shape and trajectory patterns can be highly valuable in future scenarios where obtaining precise cell morphology data is challenging. Additionally, identifying cells based on their migration patterns through phase-contrast microscopy has the potential to eliminate the requirement for high-quality imaging and provide more reliable methods for analyzing blurry images.