Metabolic Reprogramming in Myeloid Cells: AI Therapies

In the rapidly evolving landscape of cancer research, the metabolic reprogramming of myeloid cells has emerged as a pivotal area of investigation, with artificial intelligence (AI) technologies poised to revolutionize our understanding and treatment strategies. Myeloid cells, key players in the immune system, undergo complex metabolic shifts within the tumor microenvironment (TME), influencing cancer progression and therapeutic resistance. Recent advances in AI-driven multi-omics integration are unraveling these intricate metabolic networks, offering unprecedented insights into myeloid cell heterogeneity and functional states.

At the forefront of this scientific revolution lies the utilization of high-dimensional single-cell data analysis platforms such as Scanpy and Seurat. These tools facilitate the precise dissection of myeloid cell diversity by preprocessing, clustering, and visualizing single-cell RNA sequencing (scRNA-seq) datasets. However, transforming transcriptional data into actionable metabolic phenotypes demands sophisticated AI-powered methods. Packages like scMetabolism and UCell extend these analyses by scoring metabolic pathway activities—such as glycolysis and the tricarboxylic acid (TCA) cycle—at single-cell resolution, effectively bridging gene expression and metabolic function.

To augment this, machine learning frameworks such as Compass and graph neural networks (GNNs) like scFEA have been developed to predict metabolic flux. Compass integrates scRNA-seq profiles with genome-scale metabolic models encapsulating biochemical constraints, providing dynamic flux estimations reflective of actual cell metabolism. Conversely, scFEA leverages GNN architectures to model metabolic pathways as interconnected graphs, enabling the inference of cell-wise metabolic fluxes directly from gene expression. These methods collectively push the boundaries of what is possible in deciphering the metabolic underpinnings of myeloid cells within tumors.

Beyond sequencing data, AI-driven metabolomic tools contribute crucial layers of knowledge. For instance, SIRIUS employs machine learning algorithms to decode metabolite structures from complex tandem mass spectrometry data, while MetaboAnalyst harnesses statistical and machine learning approaches for identifying metabolic signatures and performing pathway enrichment analyses. When woven together, these heterogeneous datasets construct a comprehensive tapestry of myeloid immunometabolism, revealing subtle nuances previously obscured.

Understanding the functional implications of metabolic states requires elucidating the interplay between cells and their environment. Cutting-edge tools such as CellChat enable the inference of cell–cell communication networks, linking metabolic profiles to immune signaling pathways, and thereby clarifying how metabolic states influence intercellular dynamics. In parallel, SCENIC reconstructs gene regulatory networks that orchestrate these metabolic phenotypes. With increasingly sophisticated AI applications, including GNNs, researchers can model dynamic cellular networks capturing both metabolic rewiring and immune crosstalk within the tumor milieu, forecasting cell state transitions and therapeutic responses.

The integration of multi-omics data represents a critical frontier in myeloid metabolism research. Frameworks like MOFA+ and DIABLO deliver statistical power to unify multi-modal datasets, harmonizing transcriptomic, proteomic, and metabolomic layers. These integrative platforms expose coordinated molecular alterations, generating holistic profiles of metabolic states and their regulatory circuits. Such comprehensive insights are instrumental in mapping the complex metabolic landscape of myeloid cells, elucidating potential vulnerabilities exploitable for therapeutic intervention.

Translating these intricacies into clinical applications motivates the development of AI frameworks for predicting therapeutic outcomes. Deep learning architectures, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), are employed to anticipate metabolic responses of myeloid cells to diverse stimuli, notably immunotherapies. Platforms like TME-NET harness these models to profile tumor-infiltrating myeloid metabolic states, correlating them with patient-specific therapy outcomes. Moreover, emerging tools such as RENAISSANCE advance kinetic modeling of intracellular metabolism, supported by AI algorithms estimating enzymatic parameters on a genome-wide scale, thereby refining drug target identification prospects.

Notably, two hallmark studies epitomize AI’s capacity to decode immunometabolic complexity. Wang et al. leveraged scFEA to unveil the metabolic landscape of myeloid subsets within hypoxic colorectal and lung cancer TMEs, highlighting lactate accumulation as a major metabolic hallmark driving myeloid cell reprogramming. Importantly, this study identified APOE+ CTSZ+ tumor-associated macrophages (TAMs) characterized by distinctive glutamate-glutamine metabolic fluxes, underscoring potential metabolic regulators of tumor immunosuppression. Complementing this, the COMPASS framework integrated transcriptomic data with detailed metabolic models to uncover critical metabolic determinants of pathogenic versus nonpathogenic Th17 cells, leading to experimental validation of the polyamine pathway as a master regulator of autoimmune T cell function. These pioneering efforts demonstrate how AI-guided inference transcends transcriptomic limitations by approximating metabolic fluxes absent direct metabolomic measurements.

A profound application of AI in this domain is the in silico repurposing of drugs targeting metabolic dependencies highly expressed by tumor-infiltrating myeloid cells. Prominent metabolic targets include the lactate axis—via enzymes like LDH-A and monocarboxylate transporters (MCTs)—and the nicotinamide adenine dinucleotide (NAD+) salvage pathway enzyme nicotinamide phosphoribosyltransferase (NAMPT). AI pipelines accelerate the identification of candidate drugs, exemplified by stiripentol, initially an antiepileptic agent repurposed to inhibit LDH-A, and KPT-9274, which targets both NAMPT and PAK4.

Central to this drug repurposing pipeline are graph-based drug–target interaction (DTI) models utilizing GNNs to encode drug molecular graphs and protein sequences through convolutional neural networks. This approach enables rapid and high-throughput prediction of binding affinities across extensive libraries of approved drugs, generating prioritized candidates against metabolic enzymes. Frameworks like DeepPurpose facilitate this integration by packaging diverse architectures—GNNs, CNNs, and transformers—within accessible software environments, further democratizing AI-assisted drug discovery.

Complementary to graph methods, knowledge graphs (KGs) represent biomedical entities as heterogeneous networks amalgamating drugs, proteins, diseases, and more, encoding nuanced biological relationships. AI algorithms exploit this structured semantic data to infer novel drug–disease links, enhancing the computational repurposing endeavor. Prominent KGs such as Bioteque and PharMeBINet consolidate multifaceted biomedical data, offering fertile ground for these sophisticated inference models.

Generative AI introduces nuanced strategies to drug discovery by creating novel molecular candidates meeting complex criteria. Molecular-based generative models, including reinforcement learning and transformer-based architectures trained on chemical languages like SMILES, optimize structures for desirable biological activity while potentially retaining prior safety profiles. This synthetic creativity enables beyond-de novo design, facilitating scaffold hopping for repurposing known drugs. Simultaneously, research-based generative models, exemplified by biomedical large language models like BioBERT, empower hypothesis generation and scoring function design, synthesizing literature-derived insights with chemical innovation.

The final critical phase involves candidate prioritization supported by explainable AI (XAI) methods such as SHAP and LIME. These tools break down AI predictions, revealing the molecular or structural features driving affinity scores or biological activity forecasts. By demystifying AI decision-making, XAI fosters confidence in predictions and guides experimental strategies, transforming opaque computations into transparent, testable scientific hypotheses.

Together, these multifaceted AI-driven methodologies forge an integrated workflow mapping myeloid metabolic states, predicting functional outcomes, and accelerating therapeutic discovery. This paradigm not only deepens our understanding of the metabolic heterogeneity underpinning cancer-associated myeloid cells but also exemplifies how AI catalyzes translational leapfrogging—transforming data deluge into actionable precision oncology.

As AI continues to evolve, its synergy with multi-omics and mechanistic biology promises to extend beyond current capabilities, unearthing previously elusive metabolic vulnerabilities and guiding personalized cancer immunotherapies. The integration of dynamic modeling, advanced machine learning, and explainable frameworks paves the way for innovative therapeutics precisely tailored to metabolic landscapes within the TME, charting a transformative course for oncology in the coming decade.