This post grew out of a conversation I had with Garry Nolan and Sonali Jindal.
Spatial transcriptomics has changed what is possible in tumor immunology. The ability to resolve gene expression within intact tissue architecture — to know not just what cells are present but where they sit relative to each other — has produced a genuinely new category of biological insight. That is not in dispute.
What is worth examining is what happens when transcriptomics becomes the only modality. When a gene expression profile is treated not as a starting point but as an answer.
The core limitation is this: transcriptomics tells you who is present and in what transcriptional state. It builds the atlas. But the atlas is not the mechanism.
Proteins are the functional effectors of biology. What a cell is doing — not what it is programmed to do, but what it is actually doing, right now, in that niche — is encoded in protein expression. Surface receptor density, secreted ligand gradients, checkpoint molecule display, post-translational modifications: none of these are reliably read off from transcript abundance. The interaction between two adjacent cells in a tumor is a protein-level event. The suppressive character of a given microenvironmental niche is defined by the proteins being exchanged within it, not only by the transcriptional identities of the cells involved.
The productive question is not which modality is better. It is how to use both. Transcriptomics for atlas construction — defining cell types, states, and spatial organization at scale. Proteomics for mechanistic interrogation — resolving the functional interactions, niche architecture, and protein-level signaling that the atlas points toward but cannot resolve. Integration methods now exist that allow new proteomics data to be anchored to existing transcriptomic atlases without requiring matched parallel data generation. The two modalities compound each other in ways that neither achieves alone.
There is also a clinical argument that rarely gets stated plainly: pathology is protein-based and always will be. No spatial transcriptomic signature has become a clinical diagnostic tool. The translation pathway from gene expression to clinical assay is long. Protein-based findings map directly onto the diagnostic infrastructure that already exists and onto the decades of prior pathology data that transcriptomic findings cannot reach without a protein bridge.
None of this is an argument against spatial transcriptomics. It is an argument for not stopping there. One modality is not the totality of what is happening in a tumor. Designing research programs as if it were has real costs — mechanistic, translational, and scientific.
The atlas is essential. But the mechanism lives in the proteins.