CHIMERA

A hepatocyte and a Kupffer cell side by side in the same liver lobule can run entirely different lipid programmes, but no instrument reads the transcriptome and the metabolome of one cell at scale — mass spectrometry consumes what it measures. The usual workaround infers metabolism from expression through a curated pathway database, which returns fluxes or activity scores confined to already-catalogued reactions, not abundances on the scale of the measurement itself.

CHIMERA (Cell-level Hybrid Inference of Metabolome Embedded on RNA Atlas) reaches the metabolome through a measurement that does exist. Spatially co-registered 10x Visium and MALDI-MSI sections pair a transcriptome with a metabolome directly — but at spot resolution, where each spot mixes several cell types. CHIMERA treats that mixing as the learning problem rather than an obstacle: cell2location deconvolves each spot into nine cell-type proportions, a gene transformer predicts a metabolic profile per cell type, and those predictions are recombined by the same proportions and scored against the measured spectrum. Because the loss is only evaluated on the mixture, this is multiple-instance learning, and it pushes spot-level supervision onto cell types. A single cell is already pure, so inference on one is a plain forward pass.

The work comes from Huashan Hospital and the Cancer Metastasis Institute at Fudan University, with collaborators at Renji Hospital and Microsoft. The encoder borrows the gene-token idiom of scGPT but is randomly initialised and trained end to end; the name is shared with an unrelated MICCAI multimodal-oncology benchmark.

#Key Features

• Quantitative abundances, not pathway scores: Predictions land in the same log-normalised intensity space as the MALDI-MSI training data, so they compare directly against a mass-spec readout rather than a relative flux score.
• Cell-type supervision from spot-level data: Virtual pure spots — a cell type's reference expression profile plus a spot-specific residual — give the encoder a per-cell-type input while the loss stays on the observed mixture.
• Joint gene–metabolite co-embedding: The predicted metabolome attaches to cells whose transcriptome was actually measured, so both can be co-embedded for the same cells to surface differential metabolites and co-regulated modules.
• Turnkey single-cell inference: A trained HVG-256 checkpoint ships in the repository with its gene list, and the bundled example script runs it over cells exported from a Seurat object.

#Technical Details

The encoder is a six-layer pre-norm transformer — four heads, model dimension 256, feed-forward dimension 1024 — totalling roughly 5M parameters. Each of the 256 input tokens is one highly variable gene, encoded as a learned identity embedding plus a projection of its log1p expression, and a learnable CLS token carries the cell-level representation into an MLP decoder. Training minimised an equal mix of mean-squared error and a Pearson correlation penalty. The substrate was three haCCA-aligned mouse liver-tumour sections — wild-type and Padi4-knockout intrahepatic cholangiocarcinoma, 25,959 paired spots — with an 87,554-cell scRNA-seq reference annotated into nine cell types and 189 of 212 MALDI-MSI features retained. Spots were partitioned by pairwise distance into train, an excluded buffer, and test. On that internal fold the model reaches spot-level R² 0.752 and a per-metabolite median Pearson r of 0.285, against 0.177 for a PCA-plus-ridge baseline, which it beats on 158 of 189 metabolites. Cross-section figures are far smaller and reported separately: R² 0.053 on the held-out WT2 replicate, and a smoothed median r of 0.130 on WT3. The MIL objective is what keeps them positive — direct per-spot supervision fits the internal fold marginally better but falls to R² −0.043 on WT2.

#Applications

The shipped checkpoint annotates any mouse scRNA-seq object. On 33,214 CD45− hepatic cells from a Liver Cell Atlas Western-diet cohort — sharing no animals, samples, or measurement batch with training — the predicted metabolomes separate diet conditions and recover glycerophospholipid remodelling, purine metabolism, and hexosamine biosynthesis signatures consistent with published NAFLD pathophysiology. In a Rarres2 (chemerin) knock-down hepatocellular carcinoma model, macrophages that barely separate on the transcriptomic UMAP resolve into four metabolic subclusters, from a lipid-associated MC-3 state to an Spp1+ MC-2 end state.

#Impact

CHIMERA applies to single-cell metabolomics the pattern that lets histology models impute spatial expression: a paired modality becomes the training substrate for carrying an expensive measurement onto cells profiled by a cheaper one. Its envelope is narrow. The token vocabulary is a fixed 256-gene panel chosen by variance on mouse liver-tumour transcriptomes, the output a fixed 189 metabolites tied to one MALDI-MSI panel, and genes missing from a user's dataset are zero-filled; transfer to human tissue, other organs, or a different panel is not demonstrated. Cross-section agreement is modest even within the training tissue, so predicted abundances read as a comparative signal across cells rather than a calibrated measurement. The repository carries no license file, and the underlying data are available on request rather than deposited. The work awaits peer review.