MetaboFM

Mass spectrometry imaging (MSI) measures the spatial distribution of metabolites across tissue, producing rich but highly heterogeneous datasets that vary in instrumentation, mass range, and spatial resolution. This heterogeneity has made it difficult to build shared representations that transfer across experiments, leaving most analyses tied to individual studies. MetaboFM addresses this by consolidating thousands of public MSI datasets into a standardized format and learning a general-purpose foundation model over them.

Developed by Efe Ozturk, Felix G. Rivera Moctezuma, and Ahmet F. Coskun at the Georgia Institute of Technology and released as a bioRxiv preprint in October 2025, MetaboFM curates roughly 4,000 publicly available MSI datasets from the METASPACE repository and reshapes them into standardized spatial-spectral tensors. A Vision Transformer is then trained over these tensors to produce embeddings that capture spatial and spectral structure, which can be applied to downstream tasks with simple linear probes rather than task-specific retraining.

#Key Features

• Large curated MSI corpus: Around 4,000 public datasets from METASPACE are consolidated into standardized spatial-spectral tensors, unifying heterogeneous imaging experiments into a common representation.
• Vision Transformer backbone: A ViT architecture models the spatial-spectral tensors, learning embeddings that transfer across MSI datasets and prediction tasks.
• Zero-shot linear probing: Frozen embeddings support downstream tasks through lightweight linear probes, avoiding full model retraining for each new application.
• Multimodal VQA extension: A visual question answering component links learned embeddings to natural-language queries, enabling text-driven interrogation of metabolomic images.
• Interactive interface: A Gradio-based interface supports visualization and metadata queries over datasets and embeddings.

#Technical Details

MetaboFM standardizes MSI data from METASPACE into spatial-spectral tensors and trains a Vision Transformer to produce embeddings from them. Model quality was assessed across six prediction tasks using linear probes on the frozen embeddings, where MetaboFM achieved a mean macro-F1 of 0.74 and accuracy of 0.80, substantially outperforming classical baseline methods. Beyond representation learning, the framework adds a multimodal visual question answering extension that connects embeddings to natural-language queries, and it ships an interactive Gradio interface for exploring images, embeddings, and metadata within a single system.

#Applications

MetaboFM is aimed at researchers analyzing spatial metabolomics data who need representations that generalize across diverse instruments and tissue types. Its embeddings support classification and metadata-prediction tasks with minimal task-specific training, making it useful for annotating new MSI experiments, screening large repositories, and organizing metabolomic imaging collections. The VQA extension and Gradio interface lower the barrier to exploratory analysis, allowing biologists to query metabolomic images through natural language rather than bespoke analysis scripts.

#Impact

By assembling one of the larger curated collections of MSI data and learning a transferable Vision Transformer over it, MetaboFM brings the foundation-model paradigm to spatial metabolomics, a field that has lacked shared pretrained representations. The demonstration that frozen embeddings support strong linear-probe performance across multiple tasks points toward reusable metabolomic representations, while the VQA and interactive components broaden accessibility. As a bioRxiv preprint awaiting peer review, its results rest on the six evaluation tasks reported by the authors, and wider validation across laboratories and tissue types will clarify how far the learned representations generalize.