Machine learning of cellular metabolic rewiring

Abstract

Metabolic rewiring allows cells to adapt their metabolism in response to evolving environmental conditions. Traditional metabolomics techniques, whether targeted or untargeted, often struggle to interpret these adaptive shifts. Here, we introduce MetaboLiteLearner, a lightweight machine learning framework that harnesses the detailed fragmentation patterns from electron ionization (EI) collected in scan mode during gas chromatography/mass spectrometry to predict changes in the metabolite composition of metabolically adapted cells. When tested on breast cancer cells with different preferences to metastasize to specific organs, MetaboLiteLearner predicted the impact of metabolic rewiring on metabolites withheld from the training dataset using only the EI spectra, without metabolite identification or pre-existing knowledge of metabolic networks. Despite its simplicity, the model learned captured shared and unique metabolomic shifts between brain- and lung-homing metastatic lineages, suggesting cellular adaptations associated with metastasis to specific organs. Integrating machine learning and metabolomics paves the way for new insights into complex cellular adaptations.

Figure 1.

Workflow and Functionality of MetaboLiteLearner. (A) Traditional targeted metabolomics uses specific ion peak areas. (B) MetaboLiteLearner uses GC/MS data acquired in scan mode in MetaboLiteLearner, which encompasses ion fragment abundances ranging from 50 to 600 m/z captured at all GC retention times. (C) Each molecule’s electron impact (EI) fragmentation spectrum is depicted as a 550-dimensional vector. These high-dimensional vectors are paired with their corresponding log2 fold change values, which serve as training labels. (D) The transformation matrices (‘w’ for spectra and ‘c’ for log2 fold changes) are the PLSR loadings used to map data into the N-dimensional latent space. This enables MetaboLiteLearner to learn the relationship between metabolite-fragment composition and metabolic rewiring.

Figure 2.

Data Acquisition and Processing for Metabolitelearner from Breast Cancer cell derivatives. (A) Cell culture and metabolite extraction: Breast cancer cell lines, including the parental MDA-MB-231 cells and its brain- and lung-homing derivatives, were cultivated. These derivatives were procured through in vivo selection using mice. Under consistent media conditions in vitro, intracellular metabolites from these cells were extracted to ensure a reliable data source for subsequent processing. (B) GC/MS processing and data aggregation: Samples underwent GC/MS analysis following the TMS derivatization protocol. The generated data matrices, unique for each sample, were amalgamated to create a virtual “bulk” sample. Peaks were identified, and their spectra were extracted from this consolidated matrix. The input (X) for MetaboLiteLearner encompasses the mass spectra of each peak. The output data (Y) indicate the comparative abundance shift of each peak in brain- and lung-homing cells relative to the parental cells.

Figure 3.

Optimization and Evaluation of MetaboLiteLearner’s Predictive Performance. (A) Model Complexity vs. Error: Through hold-out cross-validation, the training error consistently decreased as latent components (N) increased from 1 to 30. Test error reached its lowest at N = 11 before it began to rise, highlighting potential overfitting with more complex models. The chosen optimal model had N_opt = 5 components, correlating strongly with true log2 fold changes (ρ = 0.39, P-value ≪ 0.01). (B) Schematic of the model trained with five components. (B’ and B”) Variance explained by latent factors: Using the N_opt = 5 model, the transformations into the five-dimensional latent space covered 32% of the predictor variance and 68% of the response variance. The variance explained by each latent factor for the predictor and response datasets can be viewed separately. (C) Randomization test results: After shuffling the log2 fold changes, disrupting their correlations with input spectra, MetaboLiteLearner’s error with shuffled data was consistently higher than with the original dataset, confirming its ability to identify genuine relationships between metabolite spectra and abundance changes in rewired cells.

Figure 4.

Interpretation of the Model for Metastatic Breast Cancer Cells. (A) Biplot representation of the m/z ionic fragment vectors. Specific fragments, such as m/z = 104, are associated with increased levels in both cell types in contrast to fragments like m/z = 306 which are associated with decreased levels. (B) The proportional contribution of the five latent components to the variance in log2 fold changes, with components 1 and 3 indicating overlapping metabolic shifts with decreased levels in both cell types, components 2 and 5 showcasing increases in both, and component 4 underscoring the divergence between the two cell types. (C) In component 1, accounting for 27% of the response variance, most amino acids follow the trend of reduced levels in both cell types, while carbohydrates differ. (D–G) Among the latent components, we see the distinctive role of component 4 which is dominated by carbohydrates and deoxyribonucleosides and highlights potential metabolic variances between brain- and lung-homing cells.