Dissecting metabolic landscape of alveolar macrophage

Abstract

The highly plastic nature of Alveolar Macrophage (AM) plays a crucial role in the defense against inhaled particulates and pathogens in the lungs. Depending on the signal, AM acquires either the classically activated M1 phenotype or the alternatively activated M2 phenotype. In this study, we investigate the metabolic shift in the activated phases of AM (M1 and M2 phases) by reconstructing context specific Genome-Scale Metabolic (GSM) models. Metabolic pathways such as pyruvate metabolism, arachidonic acid metabolism, chondroitin/heparan sulfate biosynthesis, and heparan sulfate degradation are found to be important driving forces in the development of the M1/M2 phenotypes. Additionally, we formulated a bilevel optimization framework named MetaShiftOptimizer to identify minimal modifications that shift one activated state (M1/M2) to the other. The identified reactions involve metabolites such as glycogenin, L-carnitine, 5-hydroperoxy eicosatetraenoic acid, and leukotriene B4, which show potential to be further investigated as significant factors for developing efficient therapy targets for severe respiratory disorders in the future. Overall, our study contributes to the understanding of the metabolic capabilities of the M1 and M2 phenotype of AM and identifies pathways and reactions that can be potential targets for polarization shift and also be used as therapeutic strategies against respiratory diseases.

Introduction

Alveolar macrophages (AMs) serve as the first line of defense against respiratory pathogens and exhibit a high degree of plasticity¹. Depending on their interactions with pathogens, AMs can differentiate into distinct subsets, with the two main phenotypes being classically activated (M1) and alternatively activated (M2) macrophages^2,3. M1 macrophages, typically induced by microbial components such as lipopolysaccharide (LPS) and Th1 pro-inflammatory cytokines, are essential for the bacterial clearance and the recruitment of additional immune cells^4,5,6,7. In contrast, M2 macrophages are promoted by interleukin-4 (IL-4), interleukin −13(IL-13), and immune complexes (ICs) and are associated with anti-inflammatory functions, including the resolution of inflammation and tissue repair⁸. However, depending on the different stimuli, M2 macrophages can be further divided into four different subtypes: M2a, M2b, M2c and M2d⁹. These subtypes are highly diverse, and their functions range from recruiting of other immune cells to promoting tissue repair⁹. Given these insights, it is crucial to further explore the role of these metabolic pathways within the specific context of the lung microenvironment to better understand their functional significance in AM polarization¹⁰.

During infection, alveolar macrophage polarization can shift dynamically based on the type of pathogen and stage of immune response⁴. Alveolar macrophages polarize towards the M1 phenotype in response to microbial stimuli, such as bacterial lipopolysaccharides (LPS) or viral components. On the other hand, the macrophage polarization may shift towards the M2 phenotype⁵when the infection is controlled, and the inflammatory response needs to be regulated. The shift in phenotypes of AM plays an important role in regulating the body’s immune response and metabolism^11,12. Yet, the regulatory mechanisms governing the polarization are not completely understood. Due to the challenges associated with experimental studies, systems biology approaches can be utilized for reconstructing context specific (i.e., M1 or M2 phase) Genome-Scale Metabolic (GSM) models using available omics data. Systems biology has been proven to be very useful for pragmatic modelling and theoretical exploration of complex biological systems¹³. By the integration of high-throughput omics data (metabolomics, proteomics, or transcriptomics), the reconstruction of human GSM models of pancreatic cancer, tuberculosis infected AMs, obesity and diabetes, and neurodegenerative diseases have led to discovery of novel therapeutic targets and better understanding of the metabolic shifts^14,15,16,17.

Overall, GSM models can be considered as computational and context specific (species, cells, tissue, etc.) knowledgebases capable of dissecting systemic metabolic phenomena¹⁸. A GSM model contains all annotated metabolic reactions and pathways within a specific biological system¹⁹. Using Flux Balance Analysis (FBA) and Flux Variability Analysis (FVA) the specific values and the ranges of reaction fluxes can be predicted for a given condition/timepoint. A recent computational modeling study by Gelbach et al⁹on M1 and M2 subtypes of human colorectal cancer cells via generation of GSM models marks a significant step towards understanding and manipulating the polarization mechanism from M2 to M1 of macrophages, related to this specific type of cancer. Other computational modeling of macrophages includes, the GSM of AM curated by Bordbar et al., from the global human model, Recon1, that was further used to model tuberculosis infected AM^14,16. However, the activated states (i.e. M1 and M2 phenotypes) were not further explored. In addition, Wang et al. developed a method of creating draft metabolic models (mCADRE) of 126 human tissues including macrophages²⁰. Overall, there is a lack of understanding of the metabolic reprogramming of AM (i.e., in healthy and activated states). While comparative analysis with other types of macrophages is highly beneficial, understanding the behavior of AMs exclusively to the lung microenvironment is critical to develop effective therapeutic targets for the lung diseases.

In this study, three context specific GSM models were developed by integrating transcriptomics data of healthy AM, and its activated phases: M1 and M2. Figure 1shows the overview of steps involved in generation, curation, and analysis of the three GSM models. Transcriptomics data integration techniques such as mCADRE20, INIT21, iMAT²¹and E-flux²²were used for GSM model reconstruction from the global human metabolic model, Human1²³. The metabolic models were further curated by using a tool called OptRecon (inhouse tool, currently unpublished) that can be used to identify and resolve Thermodynamically Infeasible Cycles (TICs). These models were utilized to comprehensively analyze the metabolic landscape of activated alveolar macrophages, both in comparison to their healthy counterparts and to each other. In doing so, we highlight interesting activity in pathways such as pyruvate metabolism, glycolysis, carnitine shuttle (mitochondria) pathway, bile acid synthesis (BAS), chondroitin/heparan biosynthesis and heparan sulphate degradation as possible significant drivers of M1/M2 polarization in lungs. Additionally, we identified two sets of 30 reactions for M1 to M2 switch and vice versa, which include glycogenin formation and transport, the formation of leukotriene B4, potassium ion exchange, and movement of L-carnitine from the cytoplasm to mitochondria by developing a bilevel optimization framework (MetaShiftOptimizer). These reactions when altered (upregulated/downregulated or deleted) could more successfully shift M2 phenotype to M1; while for the reverse, it becomes more challenging. Going forward, the context specific activated AM GSM models will be used to study interaction with respiratory pathogens and identify effective therapeutic measures. In addition, a metabolic modeling framework of the relevant immune cells such as AM, Neutrophils, and Mast cells, could be developed to analyze the innate metabolic immune responses during pathogenesis in the lungs.

Fig. 1

Schematic of the workflow for the generation of healthy and activated Alveolar macrophages and the steps involved in analysis of the metabolic shift during polarization. (a) The schematic of the steps involved in the generation of context specific GSM models by integrating omics data into global human reconstruction, Human1. (b) Optimization techniques such as Flux Balance Analysis and Flux Variability Analysis were used to obtain flux ranges to analyze the metabolic behavior of each of the phases. (c) A t-SNE plot that shows the flux distribution of M1 and M2 phase and Modified M2 (M2 after incorporation of changes as predicted by MetaShiftOptimizer) to obtain the shift from M2 to M1 phase. The same process is applied to obtain the shift from M1 to M2 phase as describe in Results and Methods. .

Results

Metabolic model reconstruction of alveolar macrophage (AM) metabolism

To develop the context specific metabolic models of healthy AM, M1 phase, and M2 phase, we needed to integrate gene expression values of metabolic genes^24,25,26onto the Human1²³ model. These transcriptomic profiles were acquired from GEO databases (GSE8823, GSE40885, and GSE41649 for AM, M1 phase and M2 phase, respectively). These datasets effectively capture the biology of various states of alveolar macrophages, as demonstrated by Becker et al.²⁷, where the GSE40885 and GSE41649 datasets were utilized to identify macrophage polarization signatures in the lungs. Consequently, we utilized these datasets to perform gene set enrichment analysis using GSEA desktop software²⁸ to gain an initial understanding of the differences in pathway activities based solely on transcriptomic profile. The analysis captured the activity of glycolysis, bile acid synthesis, and peroxisomal pathways to be enriched for M1, while cholesterol metabolism and protein secretion were enriched for M2. In addition, OXPHOS and Fatty acid metabolism were found to be enriched in both the phenotypes (the details on the GSEA results are available in the GitHub repository). While gene set enrichment analysis provided an initial insight into pathways that could be important in M1 and M2 phase, we used GSM models to further explore how the pathway level transcriptomic difference can be manifested into metabolic differences and capabilities of each AM phenotype.

Among various methods available to develop context specific models via the integration of omics data, INIT (Integrative Network Inference for Tissues)²⁹, mCADRE (Context-specificity Assessed by Deterministic Reaction Evaluation)²⁰, iMAT (integrative Metabolic Analysis Tool)²¹, and E-flux²²were used in this study. While iMAT, INIT, and mCADRE are switch approaches in which reactions are turned on/off depending on gene expression levels, E-flux is a valve approach in which each reaction is constrained based on the gene expression values. iMAT madidates the presence/absence of a specific reaction depending on the relevant gene(s) having higher expression levels at a specific condition²¹. INIT considers evidence from various sources, such as protein levels and gene expression, assigning weights to reactions in a way that maximizes network consistency with available data while allowing for flexibility in metabolite secretion or accumulation²⁹. mCADRE algorithm creates a model by removing the non-core metabolic reactions while ensuring the consistency of high confidence set of core reactions with the ability to produce essential metabolite from simple precursors such as glucose²⁰. On the other hand, a valve approach such as E-flux, uses gene expression levels to regulate the flux of the corresponding reactions²².

The healthy AM model obtained upon the implementation of iMAT consists of 4,554 reactions (governed by 2,173 metabolic genes) and 3,967 metabolites (2,003 unique) distributed across eight intracellular compartments (Extracellular, Peroxisome, Mitochondria, Cytosol, Lysosome, Endoplasmic reticulum, Golgi apparatus, Inner mitochondria, and Nucleus); while the model generated by E-flux consists of 8,073 reactions and 5,380 metabolites (2,823 unique) across these eight compartments with the same number of metabolic genes. Similar comparisons were performed between the models obtained from INIT and mCADRE and are shown in Fig. 2(a, b, and c). INIT and mCADRE algorithms have the risk of either underestimating or overestimating the metabolic capabilities of any cellular systems which could potentially be problematic for human cells and can lead to divergence from their true physiological behavior^20,29. In addition, the sensitivity of iMAT approach to user defined threshold typically leads to higher number of reactions to be omitted or sometimes leads to exclusion of important reactions from the pruned model. In our case, these switch approaches resulted a pruned model capable of producing biomass but failed to include important pathways such as NO production, glycerolipid metabolism, heme synthesis, and porphyrin metabolism or consisted of several incomplete pathways such as fatty acid oxidation (FAO), fatty acid synthesis (FAS), and amino sugar and nucleotide metabolism. The E-flux-generated model, on the other hand, consists of comparatively more comprehensive list of reactions, metabolites, and ensured the expected activity of all the important pathways mentioned earlier. Additional information on the models built by each of these methods can be found in Supplementary Table 1, 2 and 3 along with the pathway distribution figure in Supplementary Data 1.

Fig. 2

Four different integration methods, namely iMAT, E-flux, INIT, and mCADRE are implemented for the context specific GSM model reconstruction from Human1. (a) The number of reactions and metabolites incorporated in each situation (healthy, M1 phase and M2 phase) by different methods is displayed through spider plots. In all three plots orange represents the number of metabolites and maroon represents the number of reactions. Each number of reactions and metabolites is scaled by 1000 (for example 3,967 is plotted as 3.697 and so on) to have better visualization. Methods like iMAT and mCADRE generate significantly smaller GSM models in comparison to INIT and E-flux. (b) A bar graph displaying the number of reactions carrying fluxes for each state of alveolar macrophage shows that E-flux GSM models has the highest number of reactions carrying fluxes in all of the cases. Hence, for the purpose of this study, we used E-flux generated GSM models.

Similar observations were obtained while implementing the above-mentioned algorithms with the expression values of 2,951 and 2,390 metabolic genes to reconstruct GSM models for M1 and M2 phase respectively (additional information on the gene expression values, distribution of pathways, reactions, and metabolites for all the models are available in Supplementary Data 1). Hence, the models developed with the implementation of E-flux allowed us to exhaustively investigate the metabolic shift occurring during polarization. The details on the GSM model comparison for all the generated models is available in Supplementary Data 1.

Validation of the metabolic capabilities of phase specific AM GSM models

The E-flux models were next validated to ensure their ability to simulate biologically significant processes by reproducing important metabolite production rates and characteristic behavior as reported in literature. The biomass function/reaction of these models was developed based on the AM cellular maintenance requirements such as proteins, lipids, DNA repair, ATP maintenance, and RNA turnover (adapted from Bordbar et al. 2010)¹⁶. AMs are tissue resident macrophages that populate the lung environment during birth and typically last for the lifespan of the individuals¹⁶. Flux Balance Analysis (FBA) is a mathematical approach to predict the flow of metabolites through a GSM model with a specific goal/objective, typically maximizing of growth/biomass. Since AM population remains stable under steady state condition, the choice of growth for AM models does not make sense. Hence, for all the simulations performed by our AM models, we optimized the production rates of key metabolites, while maintaining the maximal level of biomass (i.e., 0.03 h⁻¹). The healthy AM model was optimized for ATP production and NO production that yielded the flux of 0.6 mol/g cell DW/h and 0.03 mol/g cell DW/h respectively. These in silico values were very close to the values of 0.71 mol/g cell DW/h and 0.037 mol/g cell DW/h, respectively, as obtained from in vitro experiment^16,30. Hence, the healthy AM model obtained via E-flux algorithm is capable of accurately recapitulating experimentally reported production rates of important metabolites.

Similar to the healthy AM model, the activated phase models should also be able to recapitulate the relevant metabolic reprogramming. Based on our in silico observations, the maximal flux range for growth (i.e., biomass flux) was found to be different across all states (healthy, M1 and M2 phase) with the minimum value found (0.03 h⁻¹) for the healthy state. Hence, the biomass level was kept at this minimum value for the activated phases and was used as a maintenance function only. Despite lack of experimental evidence of production rates of any specific metabolites, the comparative enhanced/inhibited pathway and transcriptional activity were widely reported in literature for M1 and M2 phase^31,32. To this end, with the help of the flux ranges obtained through FVA with the healthy AM GSM model as the base/control, the M1/M2 phase reaction fluxes were compared. FVA is an extension of FBA that identifies the range of possible fluxes for each reaction in a metabolic model^33,34. These fluxes were categorized into five different categories, namely, complete overlap: widened flux space, complete overlap: shrunk flux space, partial overlap: increase, no overlap: definite increase in forward direction, and no overlap: definite increase in reverse direction. Complete overlap with increase (widened flux space) indicates a collective upregulation in the reaction fluxes, and similarly complete overlap with decrease (shrunk flux space) indicates a collective downregulation in the reaction fluxes. Partial overlap indicates that there are some flux values within the ranges that are shared, while other values are unique to each range. No overlap scenarios (in forward or reverse direction) represent completely unique flux ranges. Figure 3 shows the flux distribution in seven different pathways that play an important role during AM polarization including glycolysis, OXPHOS, PPP, and TCA cycle.

Fig. 3

Alveolar Macrophage acquires unique metabolic characteristics depending upon the phenotype. In the M1 phase, the reactions of glycolysis are enhanced which are highlighted by the green arrows and the PPP reactions which is a major contributor for NAPH production is also enhanced. Similarly, the pathways highlighted by yellow arrows in M2 phase are found to be enhanced. Each pie chart represents metabolic reprogramming of AM in the specific pathway in either M1 phase or M2 phase. Each component of the pie chart represents one of the four categories as color coded in the figure. The associated percentage in the pie chart represents the percentage of overall reactions of a specific pathway falling into each of the categories.

As reported in literature for M1 phase, we observe increased activity (complete overlap and partial overlap) for glycolysis and gluconeogenesis pathway^30,31. Pentose Phosphate Pathway (PPP) shows 48% of the reactions with increased fluxes which include important reactions such as formation of ribulose 5-phosphate and its conversion to ribose 5-phosphate with production of NADPH. The PPP, through the production of NAPDH, is reported to contribute to cellular functions that greatly impact cellular energy balance and overall health^30,31. We also found the increased reaction fluxes for succinate and itaconate in the TCA cycle which act as the signaling molecules for inflammatory responses during M1 phenotype^30,31. Similarly, multiple studies by Wang et al³⁵, Zhang et al³⁶, Wculek et al³⁷, and Hooftman et al³⁸provide evidence regarding the relationship between itaconate and macrophage polarization³². In line with some of the recent studies^11,12 suggesting that M1 and M2 macrophages primarily derive their energy from oxidative phosphorylation (OXPHOS) and the TCA cycle, we also noted increased flux activities in reactions that transport reducing equivalents (in the form of NADH) from the cytoplasm into the mitochondria. This includes the reversible conversions of fumarate to malate and citrate to isocitrate, as well as the oxidation of acetyl-CoA. Similarly, for the M2 phenotype, enhanced activity in OXPHOS, TCA cycle along with fatty acid oxidation (FAO) was observed. Additionally, succinate and itaconate production through TCA cycle, was found to be inhibited, contrary to what we found for M1 phenotype.

Next, we used Flux Sum Analysis (FSA)^39,40, as a proxy of metabolite pool size, to observe the differences between M1 and M2 phenotypes. FSA was conducted for metabolites such as ATP, NADPH, succinate, itaconate, and NO. Among these metabolites, the overall ATP and NADPH pool sizes were found to be significantly higher in the M2 phenotype with the major contribution from OXPHOS. However, glycolysis seems to be a significant but not the only contributor in M1 the phenotype. Similarly, in line with literature reports^31,35,38, the pool sizes for succinate and itaconate showed higher levels during the M1 phase compared to the M2 phase. The results of FSA are available in Supplementary Data 7 in detail.

Hence, the metabolic traits of each phase as predicted by our models are well supported by literature and thus establish the credibility/predictability of our context-specific GSM models of the activated phases. Supplementary Table 4 of Supplementary Data 1 shows the full list of metabolic capabilities of each GSM models. Additional information including the lists of all the metabolites and reactions of the final models are available in Supplementary Data 2.

De novo metabolic reprogramming in AM polarization mechanism

Activated phases in comparison to healthy AM

AM in the lungs act as the first line of defense against respiratory pathogens since these phagocytize pollutants and pathogens that act as a trigger to activate an innate immune response¹¹. M1 and M2 macrophages acquire distinct phenotypes which are usually driven by different stimuli. M1 macrophages are characterized by their high production of pro-inflammatory cytokines (e.g., TNF-α, IL-6, and IL-1β), reactive oxygen species (ROS), and nitric oxide (NO)⁵. In contrast, M2 macrophages, known as "alternatively activated," are associated with anti-inflammatory and tissue-repair functions⁷. They are typically induced by stimuli such as IL-4 and IL-13 and are characterized by the secretion of anti-inflammatory cytokines like IL-10 and TGF-β. M2 macrophages support wound healing, tissue remodeling, and resolution of inflammation, promoting Th2-type responses⁸. While earlier literature reported distinct upregulated glycolysis and inhibited OXPHOS/TCA cycle in M1 macrophages and upregulated OXPHOS/TCA cycle along with enhanced glutaminolysis and polyamine synthesis in M2 macrophages^9,10,11. However, recent studies, such as those by Woods et al. 2020 and Khaing et al. 2020, have suggested that M1 macrophages may not exclusively depend on glycolysis for energy metabolism in comparison to bone marrow-derived macrophages (BMDMs)^41,42. In this study, particular attention and effort were directed toward identifying the activity of pathways in activated AMs with respect to healthy AM. Additionally, the pathways beyond the central carbon metabolism such as pyruvate metabolism, arachidonic acid metabolism, chondroitin/heparan sulphate biosynthesis, and heparan sulphate (HS) degradation were exhaustively investigated. Despite the increasing interest in AM polarization and their unique contribution to the progression and suppression of diseases, there is a lot to be learned about the role of the above-mentioned pathways in lung pathogenesis. Hence, upon validating the models, we compared the metabolic shift in activated AMs to the healthy state by using the Flux Variability Analysis (FVA) as explained earlier in the model validation section.

We observed that glycolysis did not show distinct inhibition in M2 phase in addition to the upregulated OXPHOS/TCA, while the OXPHOS/TCA activity was also not completely down in M1 phase. Based on the FSA predictions (as mentioned in the previous section), the energy metabolism seems to be dependent on both the glycolysis and OXPHOS/TCA, while the degree of contribution of each pathway varies based on the phenotype. One possible reason for the absence of distinct up/downregulation of these key pathways could also be the diversity of M2 subtypes. Recent findings indicate that M2 phenotypes is further divided into several subtypes, out of which M2d subtype is capable of proinflammatory responses⁴³. Hence, we postulate all M2 subtype population may not exhibit inhibited glycolysis activity based on the in silico predictions. On the other hand, upregulated activity of carnitine shuttle pathway is unique to M2 phase as we did not observe similar activity in M1 phase. In fact, for M1 phase, over 80% of the reactions from carnitine shuttle (mitochondria) metabolism showed decrease in flux ranges. Figure 4 portrays the complex regulatory pathways in AM polarization and shows the pathways that are up- or down-regulated in either state.

Fig. 4

Important AM pathways. Glycolysis, TCA cycle, and OXPHOS play major roles in energy production with the help of pathways such as the carnitine shuttle (mitochondria), which shows enhanced activity during the anti-inflammatory phase. On the contrary, Bile Acid Synthesis, and Arachidonic Acid Metabolism are heightened to induce acidic conditions to minimize pathogen survival. Pyruvate Metabolism play key roles in the immune response of the cell.

The activity of the bile acid synthesis and arachidonic acid metabolism was found to be enhanced exclusively in M1 phenotype. Specifically, the reactions involved in the formation of oxylipins such as 5,15-DiHETE and 5,6-Ep-15S-HETE from arachidonic acid metabolism were found to have increased fluxes^44,45,46. Since the M1 model is LPS induced, the presence of oxylipins and bile acids are in accordance with the immune response of pro-inflammatory AM. Similarly, the upregulated activity of chondroitin/heparan sulfate biosynthesis in the M1 model can also be linked to the proinflammatory responses. The formation of heparan sulfate is a crucial step for the recruitment, adhesion, crawling, and transmigration of leukocytes from the circulation to the site of inflammation^47,48,49. On the other hand, increased heparan sulphate (HS) degradation was observed in the M2 model. Although the role of formation and degradation of HS has been a topic of interest during lung injury and inflammation, it has not been systematically explored^47,48,49. Hence, the in silico activity of Chondroitin/heparan biosynthesis postulates that this pathway is enhanced due to inflammatory response in the M1 phase, while the HS degradation is enhanced in order to aid to the versatile function in M2^47,48,49.

In addition to the metabolite mentioned above, pyruvate is an important metabolite involved in maintenance of cellular and immune function. It is a key mediator connecting several important pathways such as glycolysis, TCA cycle, amino acid metabolism (such as arginine)^50,51. Despite being such a key modulator, the pyruvate metabolism was found to be inhibited both in the M1 and M2 phase. We observed that the reactions contributing to the direct formation of pyruvate were mostly inhibited in both of these activated macrophages. While the key reaction such as PEP to pyruvate (at the end of glycolysis) and pyruvate to oxaloacetic acid (OAA) (at the beginning of TCA) maintained high fluxes, other reactions that contribute to pyruvate production via different mechanisms (e.g., from lactate or methylglyoxal) was inhibited. Hence, the activity of pyruvate metabolism emerged as a possible polarization driver in AMs which is further explored in the next section. Additional information on the flux comparison is available in Supplementary Data 3.

Activated phases in comparison to each other

We next compared the FVA solution of activated phenotypes to each other in similar manner as explained above. Earlier (while comparing healthy AM fluxes with activated phases), we reported that despite being a key intermediatory metabolite, the metabolic shift in AM resulted in limited pyruvate production in both M1 and M2 phase with respect to healthy AM. However, while comparing the M1 fluxes with those from M2, we observed all the reactions contributing to pyruvate production in the cytoplasm were inhibited (complete overlap, shrunk flux space), whereas the mitochondrial reactions are enhanced (complete overlap, widened flux space). In Fig. 5, it is shown that in the cytoplasm, only one reaction (malate to pyruvate) has enhanced fluxes with higher fluxes toward the production of D-Lactate. On the contrary, enhanced fluxes were observed in multiple mitochondrial reactions that indicate increased pyruvate and L-lactate production. The pyruvate produced in mitochondria is directly used up for OAA production which promotes OXPHOS and TCA cycle, while lactate plays an important role in the maintenance of acid–base balance in the cell and in the maintenance and resolution of inflammation^50,51. A study by Abusalamah et al. (2020) suggests that incorporating pyruvate as sodium pyruvate in growth media for macrophages inhibited immune response of the cell and also had positive impact on the bacterial growth⁵¹. Hence, the reprogramming of the pathway as suggested by our in silico analysis shows that the cells actively inhibit multiple reactions that contribute to pyruvate production while upregulating specific reactions that channel pyruvate into mitochondria for energy production. Further experimental studies in human AM could establish if pyruvate metabolite is not only an important factor but also recognizes the regulation of pyruvate metabolism as a key step in pathogenesis.

Fig. 5

Pyruvate Metabolism activity in activated phase M2 when compared to M1 phase. The reactions indicated by green arrow are enhanced in the M2 phase and show clear redirection of metabolic activity through pyruvate production.

In addition to pyruvate production reprogramming, glycogen emerged as a significant factor in acquiring specific phenotype. We observed the category with "definite increase in forward and reverse direction" consisted of mainly reactions related to glycogen production/consumption. While this observation could also be due to glycogen being a part of the biomass function, the potential role of glycogen in modulating immune responses is also highly probable, especially due to biomass reaction merely acting as a maintenance function for all of our analysis. The upregulated reactions mainly include the production of glycogenin G8 from glycogenin in cytoplasm that results in glycogenin G4G4, which is a critical precursor in the synthesis of glycogen. To further explore the M1 and M2 metabolic models for highly impactful reactions, we constrained the models by turning off the reactions completely (also known as reaction knockout) or by severely limiting the flux of each of these reactions. We found reaction involved in conversion of UTP and alpha-D-galactose-1-phosphate into UDP-galactose within the cytoplasm as critical for sustaining vital connections in the metabolic models during activated phenotypes. When these specific reactions were completely knocked out, we observed severe negative impact in the GSM models, while constraining or relaxing the reaction fluxes resulted in more desirable output. Hence, in the next section we will discuss our bilevel optimization framework to identify reactions that could be potential drivers of AM polarization^52,53.

Finally, to analyze whether the pathway activities observed so far could be corroborated from a metabolite perspective, we also conducted a metabolite-centric analysis that identifies highly significant metabolites based on the transcriptomics date of each phenotypic state³⁹. To this end, metabolites with positive normalized score were reported as significant metabolites in this study. The algorithms identified numerous metabolites from leukotriene metabolism, arachidonic acid metabolism, and chondroitin/heparan sulphate biosynthesis and degradation as significant for the M1 phase, while the major metabolites for the M2 phase were from glycolysis, OXPHOS and TCA cycle. This observation helped further substantiate the in silico predictions as mentioned earlier. The complete list of significant metabolites is available in the Supplementary Data 4.

Minimal modifications for AM polarization shift

It is extremely crucial for AM to acquire right phenotype for proper immune response. The imbalance in the M1/M2 cells can be deleterious to the lungs that can cause prolonged and unwanted inflammation in the absence of the process that shuts it down⁵⁴. In addition, without the necessary inflammation, AMs cannot effectively activate other immune cells to fight invading organisms⁵⁵. For example, the interaction between tuberculosis and AM is sometimes reported to promote M2 cells as opposed to M1, and reports on progression of cancer cells also mention the positive role of the M2 phenotype⁵⁶. Hence it is very important for AM to shift toward the phenotype which is best suited to fight the invading pathogens. To this end, we formulated a bilevel optimization framework named as MetaShiftOptimizer (as described in the Methods) that predicts a set of reactions when altered (i.e., up-/down-regulation and minimal deletion) can force the shift from pro-inflammatory (M1) state to anti-inflammatory (M2) state and vice versa. The identified reactions in both the scenarios can be linked to important immune responses and have been linked to signaling pathways involved in macrophage polarization⁵⁵. For example, one of the reactions identified by MetaShiftOptimzer is involved in the production and storage of glycogenin. Glycogen metabolism is known to play a significant role in regulation of macrophage mediated acute inflammatory responses and can be linked to several signaling networks, particularly those involved in metabolic regulation and cellular energy balance such as AMP-activated Protein Kinase (AMPK) Pathway, and mTOR pathway^55,57. Similarly, Leukotriene B4 is also known as a potent inflammatory mediator that is closely linked to NF-KB, MAPK, PI3K/AKT pathways^55,57. The possible links of different reactions and pathways to various signaling mechanism is illustrated in Fig. 6.

Fig. 6

The five signaling pathways associated with macrophage polarization which are highlighted by pink color. The connection of each signaling pathway to the polarization to either M1 phase or M2 phase is denoted by brown and blue arrows respectively. The black arrows connect the signaling pathway to its key signaling protein(s) which are then connected to each activated phase. The complex interrelationship between signaling pathway and regulatory network is shown by connecting the pathway highlighted in yellow to the signaling pathway in pink by green arrows.

The MetaShiftOptimizer result indicates that the transition from the anti-inflammatory (M2) to the pro-inflammatory (M1) phase could be achieved by altering 30 reactions. For example, the transport of glycogenin G4G4, the transport of 5-hydroperoxyeicosatetraenoic acid (5(S)-HPETE) to mitochondria followed by the conversion of 5(S)-HETE to leukotriene B4 in arachidonic metabolism, the movement of L-carnitine from cytoplasm to mitochondria in the carnitine shuttle, and reactions in cholesterol metabolism^58,59,60. Among these 30 reactions mentioned above, six reactions were to be upregulated, including the transport of glycogenin G4G4 and the transport of cofactors such as dTTP and dCTP to mitochondria. Two reactions (namely the transport of pitavastatin and exchange of potassium) were recommended for deletion, while the rest of the reactions were to be suppressed. By incorporating the suggested alterations, we achieved a modified version of M2 which shows flux modulation shifted more towards M1 state. On the other hand, the transition from the M1 to M2 state requires adjustments in a different set of 30 reactions across various pathways. These include the conversion of fructose-1,6-bisphosphate to fructose-6-phosphate and 2-phospho-D-glycerate to 3-phospho-D-glycerate—key initial reactions in glycolysis. Other necessary reactions involve the conversion of arachidonate to 5(S)-HPETE, which is subsequently metabolized to Leukotriene A4 (a precursor for LTB4 and LTC4) in arachidonic metabolism, the transformation of UDP-glucose to glycogen, and the production of adenosine⁶⁰. In this scenario, eight reactions such as transport of 3-hydroxyisobutyrate from mitochondria to cytoplasm, the movement of succinate to cytoplasm, and the conversion of cyclic-3-hydroxymelatonin to hydroxide were suggested for upregulation, while one reaction that governs the movement of gliclazide was to be deleted. The rest of the reactions were needed to be downregulated. Further details regarding the reactions identified by the algorithm are available in Supplementary Data 5.

Following the adjustments made for the shift of M1 to M2 and vice versa, quantifying the difference in the metabolic flux is crucial to assess the impact of modifications on the overall metabolic behavior of the system. We used methods such as Hausdroff distance, Jaccard Distance, and Jaccard Index to measure the similarity/dissimilarity between M1 and M2 before and after adjustments (as described in Methods)^61,62. Among the three methods mentioned, Hausdroff distance is a more reliable measure of dissimilarity in GSM models due to its interval-based representation (hence, can be used for FVA-predicted flux ranges), ability to incorporate uncertainty, and its sensitivity to the extreme values. The Hausdroff distance between the flux intervals of M1 and M2 is found to be 3.6789 e⁰⁵ which decreases to 3.5018 e⁰⁴ between normal M1 and modified M2. The distance between the normal M2 and the modified M1 did not show any significant change (both having a value of 3.6775 e⁰⁵). This prompted us to further explore if the change is more prominent for select pathways essential for each activated phase, as reported in literature and this study, such as glycolysis/gluconeogenesis, OXPHOS, PPP, TCA cycle, OXPHOS, urea cycle, bile acid synthesis, pyruvate metabolism, and cholesterol metabolism were examined^61,62. The Hausdroff distance for these select pathways before modification between M1 and M2 was found to be 4.30 e⁰⁴ which decreases to 4.507 e³ after modification in M2. This clearly signifies that the trend of reduced distance between normal M1 and modified M2 holds true for these select pathways. However, the distance measure between normal M2 and modified M1 did not show any prominent change even for these select pathways. This resistance of M1 phase to transition to M2 phase could be due to the absence of crucial M2 reactions in M1. Study by Gelbach et al²³ also mentioned that the activity of pathways such as chondroitin/heparan sulphate degradation and nucleotide sugar metabolism is unique to M2. Hence, we added 130 unique reactions present only in M2 in addition to the modification of 30 reactions as mentioned earlier and generated a new modified M1 model. We next calculated Hausdroff distance for the select pathways which was found to decrease from to 4.3037 e⁰⁴ to 4.3009 e⁰³ between modified M1 and normal M2. The details on the calculation of Hausdroff distance and other distance measures considered are available in Supplementary Data 5. The visual representation of flux distribution if the select pathways mentioned above, before and after modifications is shown using t-SNE plots in Fig. 7.

Fig. 7

t-SNE plot visualizing the M1 phenotype, M2 phenotype, and the modified M2 phenotype represented by blue, red, and green, respectively. (a) The flux distribution of M1 and M2 polarized state before any modifications (green represents the M2 and yellow represents M1 flux). (b) The t-SNE plot of Modified M2 (adjustment of 30 reactions) represented by red. (c) Understanding the metabolic shift in M1 select pathways with modifications on 30 reactions. (d) The new flux distribution after adding 130 unique M2 reactions that shift M1 to M2 more effectively. Modified M1 in both (c) and (d) is marked in blue.

Discussion

Advancing respiratory medicine, improving disease management, and developing targeted therapies necessitate a deep understanding of alveolar macrophage (AM) polarization mechanisms⁵⁴. In this study, we reconstructed genome-scale metabolic (GSM) models for healthy and activated states of AM to investigate the metabolic shifts that occur when AMs adopt either a pro-inflammatory (M1) phenotype or an anti-inflammatory (M2) phenotype.

Among the integration techniques applied, E-flux effectively captured the biological intricacies of all activation states of alveolar macrophages (AM), generating more comprehensive metabolic models. For instance, genome-scale metabolic (GSM) models derived through E-flux successfully incorporated key AM-related pathways, including nitric oxide (NO) production, succinate and itaconate synthesis, and complete fatty acid oxidation (FAO) and fatty acid synthesis (FAS) pathways³⁰. Thus, E-flux preserves the entire metabolic framework, enabling a holistic exploration of metabolism and providing a more in-depth analysis of metabolic reprogramming compared to models generated using switch-based methods^{20,21,22,23,24,29}.

E-flux integrates gene expression data with gene-protein-reaction (GPR) associations to impose constraints on metabolic reactions, making it a robust tool for constructing biologically relevant metabolic networks when suitable transcriptomic datasets are available. However, due to limitations in dataset availability, the closest approximation of the M2 macrophage phenotype in the lung was derived from bronchial tissue biopsies (GSM41084)²⁶, which served as a proxy for lung-specific data. Despite multiple studies reporting that this dataset as a good representation of M2 macrophage in lungs, an ideal case will be to get iL-4 stimulated AM. To the best of our knowledge, at the time of reconstruction of these models, no such datasets were available. Recent advances such as study by Travaglini et al. (2020)⁶³and Sun et al. (2021)⁵⁶ have led to scRNA and snRNA sequencing of AM to study certain diseased states; however, such advanced data are not available for the M1 and M2 phase specifically. Therefore, the most suitable transcriptomics data available for all the conditions were selected.

One of the most important aspects of GSM models is determining the objective function that does justice to the biology of the cell. In most cases the biomass equation that defines the growth of the relevant cellular system is used as an objective function. However, an immune cell’s metabolic objective cannot be confined to a growth function (especially not for a tissue resident AM) and/or a specific metabolic function such as ATP or protein synthesis. Hence in this study, we used objective-independent techniques such as FVA, FSA and metabolite centric analysis along with a basal level of biomass as a constraint, thus ensuring the maintenance to biological activities. Next, we conducted exhaustive analysis of the metabolic pathways of AM and its phenotypes. In doing so, our GSM models highlighted the pathways and reactions that can be linked to the inflammatory/anti-inflammatory responses. These pathways, specifically bile acid synthesis, arachidonic acid metabolism, chondroitin heparan sulphate biosynthesis can be linked to the M1 phenotype acquirement and the initiation of inflammation. Chen et al. (2016) investigated three bile acids and the effect of change in their concentration to cytotoxicity in the cell⁴⁴. Additionally, oxylipins which are obtained via arachidonic acid metabolism play a very important role in the regulation of inflammation and the formation of other important leukotriene metabolites such as LTA₄^59,60.

Our results add to the preexisting potential of chondroitin/heparan sulphate biosynthesis and degradation as a significant modulator for initiation and resolution of inflammation^47,48,49. Chondroitin/heparan sulphate biosynthesis and degradation pathways are known to either contribute to the formation or degradation of an important metabolite called Heparan sulphate⁴⁷. The formation of heparan sulfate is a crucial step for the recruitment, adhesion, crawling, and transmigration of leukocytes from the circulation to the site of inflammation⁴⁸. Despite their important roles, these pathways have not been fully explored for their potential contribution in lung injury and inflammation. The enhanced HS biosynthesis in M1 and enhanced HS degradation in M2 phenotype as predicted by the models show the potential to be key regulator of initiation and resolution of inflammation in AMs.

Our bilevel optimization framework MetaShiftOptimizer is able to identify set of reactions originating from several metabolic pathways such as production and transport of glycogenin and production of Leukotriene B4, that can be linked to various important signaling network such as the PI3K/AKT, TLRs/NF-kB, JAK/STAT, Notch, and JNK signaling that are known to drive the macrophage polarization^55,57. Figure 6shows such metabolic reactions/pathways identified by MetaShiftOptimizer and the possible link to the signaling mechanisms. While several optimization frameworks such as OptForce, and MOMA ae extensively used to identify the reactions to enable a metabolic network to enhance a certain metabolite production, MetaShiftOptimizer identifies reactions, which when modified, can shift one phenotype to the other^64,65. To this end, the decrease in distance between the flux spaces of normal and modified conditions show that the modifications were able to shift the phenotype to the desired state. Even the small differences in fluxes can be statistically and biologically significant due to cascading effects on downstream pathways. For example, a slight increase in a bottleneck reaction flux could indicate metabolic shifts that are biologically meaningful, particularly in tightly regulated systems or pathways involved in complex immune cells such as AM. Thus, this framework can be used to identify targets points to shift polarization to desired phenotype to effectively fight pathogens and diseases.

In conclusion, our models are able to successfully captivate the biological essence of the healthy and activated state of AM. We have used available experimental data and phenomenon reported in literature to validate our models. Our results indicate several potential targets such as pyruvate, glycogen, chondroitin/heparan sulphate biosynthesis and degradation, bile acid synthesis, arachidonic acid metabolism as driving forces in macrophage polarization and/or significant markers of each activated phase. However, the activated AM models in this study are simplified version unlike the complex M1/M2 dynamics as typically observed in vivo. In reality, AM polarization is highly plastic, and context dependent and influenced by a complex array of signals in the lung microenvironment that results in a wide range of functional states beyond the M1/M2 dichotomy which unfortunately is very challenging to capture by any GSM model due to lack of suitable omics data. Additionally, our study is also limited to the in silico analysis and does not directly include experimental validation while substantiating the observed phenomena based on literature evidence. In future, with further advancement in study of AM and its activated phenotypes and when advanced omics data become available, these models will further be tested and refined for better prediction capability. Moving forward, experimental validation would be our focus (through establishing collaboration) so that some of the model findings can be tested.

Methods

Transcriptomics data processing

An exhaustive literature search was conducted to identify the appropriate set of transcriptomics data which included the transcriptomic profiles of healthy non-smokers²⁴AM induced by Lipopolysaccharides (LPS) and bronchial tissues induced by interleukin-4 (iL-4) resembling M1 phase²⁵and M2 phase²⁶, respectively. The data obtained were used as input for Gene Set Enrichment Analysis (GSEA) tool²⁸. GSEA is a tool that is used for pathway analysis based on the transcriptomic state of the cells. It was used to compare the pathway activity of healthy AM with the M1 phase, healthy AM with the M2 phase, and the M1 with the M2 phase. In this process, genes are ranked based on the correlation between their expression and the class distinction using any suitable metric. GSEA calculates the enrichment score (ES) and its significance level using p-values²⁸. The default metrices such as meandiv was used in the desktop GSEA for this study. The output from the GSEA run generated lists of enriched pathways for the M1 and M2 phase that mainly focused on signaling pathways and are available in the GitHub repository. The obtained results are impar with the evidence reported in literature for M1/M2 and healthy phenotype of alveolar macrophages.

Using the raw data set, we deduced a list of genes that were also present in the Human1 metabolic model. Quantile normalization technique was used for all the datasets to obtain the gene expression values at all conditions. After normalization, the average was taken for each condition from all the samples and then was scaled to −1000to 1000. The list of genes for healthy AM, M1 phase, and M2 phase were 2,173, 2,951, and 2,390 respectively. The expression values of these genes were integrated into Human1²³model by using different integration techniques (iMAT²¹, mCADRE²⁰, INIT²⁹and E-flux²²) as described below, to reconstruct models of healthy AM and its activated phases. The details on the datasets and normalization are available in Supplementary Data 6.

GSM model reconstruction

The transcriptomics data obtained for each of the phenotypes of AM was integrated into Human1, a global human metabolic reconstruction consisting of 13,417 reactions, 10,138 metabolites (4,164 unique), and 3625 genes²³. Three context-specific AM metabolic reconstructions were obtained by implementing both switch and valve approaches of omics integration. Among various methods available in both the categories of switch and valve approach, iMAT²¹, INIT²⁹, mCADRE²⁰, and E-flux²²were used in our study. iMAT (integrative metabolic analysis tool) is an optimization-based program that can be used to integrate the available omics data with GSM models for the prediction of metabolic fluxes²⁰. The modified version of iMAT was used in which instead of classifying the overall reactions into three categories (highly expressed, lowly expressed, and moderately expressed), the reactions were divided as either highly expressed or lowly expressed with the biomass precursors always included in the highly expressed set. The cutoff point of a p-value less than 0.05 was used to include reactions in the highly expressed set (the same threshold was used for all conditions). The biomass reaction for the healthy and the activated state was adopted from Bordbar et al¹⁶. The formulation was constructed in such a way that all the reactions from the highly expressed set were always made active and the minimum number from the lowly expressed reaction set was added to obtain the specified objective. This resulted in a pruned model which is significantly smaller than the original human model with the reactions, and metabolites specific to AM and its activated stages. On the other hand, E-flux only requires the change of the upper bound and lower bound on each reaction depending on the gene expression level²². The forward reactions consisted of a lower bound of 0 and a unique upper bound according to the gene expression levels. The backward reactions ranged from a unique lower limit to 0 as an upper bound. And for the reversible reactions, the range spans from the negative value specific to each reaction to its corresponding positive value based on the gene expression levels. Additionally, GSM models via INIT (Integrative Network Inference for Tissues) and mCADRE (Context-specificity Assessed by Deterministic Reaction Evaluation) methods were reconstructed^20,21,22,29. Both the approaches integrate the expression data into the global Human metabolic model while ensuring the production of certain important metabolites. However, INIT, mCADRE, and iMAT prune the human model specific to a context by turning off a number of reactions. This could lead to the omission of certain reactions that are crucial for the specific phase (healthy or M1 or M2), while E-flux does not deactivate any reaction. By not pruning reactions based on gene expression, E-flux provides a more inclusive representation of the metabolic models, allowing for a broader exploration of pathway activities. This approach is particularly useful to capture the overall functionality of the metabolic models under different conditions without biasing these towards specific gene expression patterns.

Hence, the GSM models for healthy AM, M1, and M2 phases were obtained by implication of the mentioned approaches. We compared the distribution of metabolites, reactions, and pathways in all the GSM models and the details are discussed in the Results and Discussion section. To ensure the biological relevance of these GSM models, we used techniques such as Flux Balance Analysis (FBA)³³and Flux Variability Analysis (FVA)³⁴ to analyze and improve model connectivity.

Flux balance, flux sum, flux variability, metabolite centric analysis

Flux Balance Analysis (FBA) is used in this study to analyze the flow of metabolites in different conditions. FBA is a widely used approach to study biochemical networks, namely, genome-scale metabolic models that contain the known metabolic reactions in a biological system and the associated genes and enzymes⁶⁶. The GSM model is represented by a stoichiometric matrix which contains the metabolites as columns and the reactions as rows. The upper and lower bound act as a constraint on each of the reactions based on nutrient availability and other (microenvironment or genetic) conditions. FBA generates a flux value for each reaction. Flux Variability Analysis (FVA) is an extension of FBA which calculates the maximum and minimum possible flux for each of the reactions in the model at a specific condition³³. Additionally, Flux Sum Analysis³⁹ is used to obtain the metabolite pool size in different conditions (M1 and M2). FSA is conducted for important metabolites such as ATP, NADPH, NO, succinate, and itaconate in normal (without turning off any pathways), glycolysis off and OXPHOS/TCA off condition. The metabolite centric analysis was conducted by using RPA^m function in COBRA toolbox in MATLAB (the details are available in the Supplementary Data 4).

Model curation

The three metabolic models were curated by using the classic design-build-test-refine cycle to be able to ensure proper network connectivity and accurate reflection of the metabolic capabilities of the AM cell. To ensure all the important AM pathways such as NO cycle, FAS, FAO, were activated, thermodynamically infeasible cycles (TICs) need to be properly handled. TICs are cycles created by reactions that carry fluxes even in the absence of nutrients essential for cellular growth and functionality. The TICs can cause the metabolic model to produce metabolites higher/lower than expected, by activating reactions that would be off in a biological scenario. However, if essential reactions are eliminated or the directionality of these reactions are changed without proper review, the behavior of the metabolic model might shift away from the known biological phenomena of the cell. Hence, it is extremely important to refine metabolic models by using efficient and effective methods.

We used OptRecon (an inhouse tool, currently unpublished), that has been developed as an expansion of OptFill​​, a tool previously developed by our group with different functionalities⁶⁷. The initial function of OptRecon is to refine GSMs by removing TICs; however, the process of removing TICs from GSMs was found to be much more difficult than the process of incorporating reactions without creating TICs, and thus the method was upgraded to be able to expand a minimal model (i.e., minimum number of biochemical reactions required to satisfy the objective, in our models the number was found to be 143) by adding reactions from a database (the database consisted of all but these 143 reactions from Human1). OptRecon generated three possible solutions to avoid formation of any TICs and ensure optimal connectivity. These solutions consisted of either blocking a reaction completely or changing the direction of the reaction. Before incorporating any changes, an exhaustive literature search was conducted to ensure that none of the biologically relevant pathways were fully omitted or partially affected due to these changes. Special attention was given to novel AM pathways such as production of NO from arginine in healthy AM, production of succinate, itaconate and citrate in TCA cycle during M1 phase and citrulline and urea production in NO cycle during M2 phase. FBA and FVA techniques were used to check the fluxes of the metabolic models ensuring proper network connectivity. All the fluxes from FBA and FVA in the absence of nutrients were found to be zero as expected in healthy and activated AM GSM models, while in other conditions the fluxes were found to be in accordance with the biological nature of AM.

Identification of minimal modification for polarization shift

The GSM model for each activated phase comprises of numerous reactions. When comparing the flux range distribution across the M1 and M2 phases, we examined whether any reactions exhibited a unique distribution. However, manually observing and comparing each reaction, which may have different flux ranges and potentially contribute to polarization shifting when limited, was a highly time-consuming approach to address the problem.Hence, in this work we propose a bilevel optimization formulation which identifies a set of reactions capable to shift the one phenotype to another. Previous works such as OptForce ensures the overproduction of desired metabolite while considering the flux values of wildtype strain by identifying a force set^65,67. Here, our focus lies in identifying the specific reactions that, when manipulated through upregulation, downregulation, or deletion, can induce a shift in the metabolic network from one activated state (M1 or M2) to the other. To achieve this, the minimization of metabolic interventions was optimized while ensuring the minimal discrepancy in reaction fluxes between the M1 and M2 states.

Initially, we began by creating a collection of reactions that fall into three categories: those that must be upregulated (MUSTU), those that must be downregulated (MUSTL), and those that must be deleted (MUSTdel). In order to distinguish these three sets of reactions, we conducted a comparison of the flux ranges acquired via Flux Variability Analysis (FVA). These three sets have unique reactions i.e. a reaction included in a MUSTU set only belongs to MUSTU and is not present in other sets. The minimal alterations required for the transition of both M1 and M2 to the alternate phase encompassed 30 reactions, originating from diverse pathways, and metabolizing distinct compounds.

The formulation is following:

$$min\sum_{j}{y}_{u}+{y}_{l}+{y}_{x}$$

$$\text{Subject to}:\text{ min}\sum_{j}|M1\left(j\right)-M2\left(j\right)|$$

$$\text{Subject to}: \sum_{j}{S}_{ij}. {v}_{j}=0\quad \forall i \in I$$

(1)

$${v}_{j}\ge {v}_{j}max.{y}_{u}+L{B}_{j}\left(1-{y}_{u}\right)\quad \forall j \in MUSTU$$

(2)

$${v}_{j}\le {v}_{j}min+U{B}_{j}\left(1-{y}_{L}\right)\quad \forall j \in MUSTL$$

(3)

$${LB}_{j}\left(1-{y}_{x}\right)\le {v}_{j}\le +U{B}_{j}\left(1-{y}_{x}\right)\quad \forall j \in MUSTdel$$

(4)

$$L{B}_{j}\le {v}_{j}\le U{B}_{j}\quad \forall j \in J$$

(5)

$${v}_{biomass}\ge 0.03$$

(6)

In the above formulation, the y_u, y_L and y_x are the binary variables associated with the reactions in the MUSTU, MUSTL and MUSTdel sets. The binary variable is 1 if the reaction is to be modified (For example y_u is 1 when the reaction is to be upregulated and 0 otherwise). M1(j) and M2(j) are the sets of all the reaction fluxes in the M1 GSM model and the M2 GSM model, respectively, under similar nutrient conditions (obtained from FBA). S_ij is the stoichiometric coefficient for the metabolite i in the reaction j and v_j is the flux of each reaction. Sets I and J include all metabolites and reactions known to occur within AM and its activated phenotypes, respectively. Each reaction is constrained by a upper bound (UB_j) and a lower bound (LBj) that specify the direction based on thermodynamic constraints. The pseudo-steady state assumed in Eq. 1 allows the assumption that all internal metabolites are consumed and produced at an equal rate. The Eqs. 2, 3, and 4 re-adjust the upper bound and lower bound of each of the reactions that is to be upregulated, downregulated and deleted accordingly. Finally, as shown in Eq. 6, in order to ensure a minimum level of biomass can be produced, a flux of 0.03 mol/g cell DW/h is forced, which corresponds to the biomass flux under healthy condition. From this algorithm, a set of reactions is identified which when modified as suggested shift the polarized state of the activated AM. The reactions identified by the modified algorithm can be linked to the signaling pathways that are known to be directly responsible for macrophage polarization.

Measure of similarity/dissimilarity

In order to assess the overall impact and extent of change resulting from the introduction of constraints in the GSM models, we employed techniques such as Hausdorff distance, Jaccard distance, and Jaccard index⁶¹. The Jaccard index is computed by considering the intersection and union of a single data point. Flux Balance Analysis (FBA) is employed to derive a permissible distribution of metabolic fluxes in a steady-state system within a Genome-Scale Metabolic (GSM) model. However, it is important to note that the resulting fluxes are not exclusive solutions. The GSM models are typically characterized by being underdetermined, context-specific, and physiologically relevant flux solutions⁶⁶. These solutions can be further refined to a single solution by imposing additional restrictions⁶². Eflux2 is an enhanced version of FBA that deduces a metabolic flux distribution from transcriptomics data and addresses the limitation of E-flux by offering a singular solution⁶². The Jaccard similarity index is a quantitative measure that assesses the degree of similarity between two sets of data. This index was computed using this distinct solution. It is expressed as a percentage range from 0 to 100%, with a larger percentage indicating a greater level of similarity^61,68. However, the solution of E-flux2 is not constant and varies depending on the specified value of the goal function. For instance, the solution set derived from FBA’s maximum biomass differs from the solution set derived from FVA’s maximum flux set as biomass. To the best of our knowledge, there is no specific growth rate documented for alveolar macrophages. Therefore, the flux ranges obtained via Flux Variability Analysis (FVA) were utilized to produce 50,000 samples using the Flux Sampling method³⁹ for both the regular activated state and the modified activated states. These samples were then used to calculate the Jaccard Index, Jaccard distance and Hausdorff distance, and to construct t-SNE plots. Supplementary Data 5 contains detailed information on the concept and calculation of the Jaccard index, Jaccard distance, and Hausdorff distance.