This dissertation develops multiscale computational and AI-driven frameworks for spatial metabolomics, spanning single-cell-resolved models to large-scale representation learning, and proposes to integrate them into a virtual-cell framework that defines metabolic states and reveals biologically meaningful cellular heterogeneity not captured by existing microscopy-defined labels alone. Mass spectrometry imaging (MSI) resolves the spatial distribution of metabolites and lipids directly from tissue, but its native resolution is coarse relative to individual cells, and the resulting object-, tissue-, and cell-scale gaps motivate the three completed studies and the proposed research that this dissertation combines. The first completed study, Metabobarcoding, introduces a graph neural network framework that relates native-resolution MSI signal to microscopy-defined spatial objects, generating interpretable, saliency-derived molecular barcodes of protein-marker states across four independent mouse and human tissue studies, and packages the workflow into an accessible executable software tool, MetaBar. The second, MetaboFM, is a two-stage self-supervised representation-learning framework trained on one of the largest public MSI corpora assembled to date; it shows that representations learned directly from ion-image spatial content, rather than from mass values, molecular structure, or acquisition metadata, are chemically and biologically organized and generalize across acquisitions and instruments unseen during training. The third, guided super-resolution (GSR), is a self-supervised pipeline that reconstructs single-cell-scale molecular detail from native low-resolution MALDI/MALDI-IHC data by leveraging paired high-resolution structural imaging, validated quantitatively against published baselines and negative controls across independent tissue contexts. Building directly on GSR, the proposed research represents each cell not as a fixed, microscopy-derived segmentation mask but as a continuous, learned geometry, a Gaussian-splat-style field whose spatial extent and per-channel molecular intensity are fit directly against GSR-enhanced signal, extended toward a volumetric three-dimensional representation via registered serial-section data. Combining this geometry with transfer-learned MetaboFM embeddings, and optionally an agentic AI interpretation layer that retrieves and analyzes the resulting states and queries public metabolite-protein/pathway databases, the proposed work asks whether the resulting metabolite-defined virtual-cell states reveal cellular heterogeneity that existing microscopy-defined labels and marker-based clustering do not capture, using association with available protein-marker ground truth as one concrete, testable check of biological validity.