MoLF

MoLF (Mixture-of-Latent-Flow) is a generative model that predicts spatial gene expression directly from hematoxylin and eosin (H&E) histology images, developed by Susu Hu and Stefanie Speidel at the National Center for Tumor Diseases (NCT) Dresden and posted to arXiv in February 2026. Spatial transcriptomics reveals where genes are expressed within a tissue, but the assays remain costly and are not routinely run, whereas H&E-stained slides are ubiquitous in pathology. Models that infer spatial expression from histology promise to bring transcriptomic context to the vast existing archives of stained tissue, and MoLF approaches this in a pan-cancer setting that must accommodate heterogeneity across tumor types.

The model frames histology-to-expression prediction as a generative problem solved with conditional flow matching. Rather than directly regressing expression values, MoLF learns a flow that maps noise to the latent manifold of gene expression, conditioned on image features. The velocity field that drives this flow is parameterized as a Mixture-of-Experts (MoE): inputs are routed to specialized sub-networks, allowing the model to handle the substantial tissue heterogeneity encountered across different cancers without forcing a single network to cover all cases.

MoLF is positioned against both single-tissue baselines and foundation-model approaches to spatial expression prediction. The authors report that it outperforms these baselines on pan-cancer benchmarks and, notably, demonstrates zero-shot generalization to cross-species data—suggesting that the learned mapping captures biological structure conserved across organisms rather than dataset-specific shortcuts.

#Key Features

• Conditional flow matching to a gene latent manifold: MoLF uses a flow-matching objective to map noise to the latent manifold of gene expression conditioned on histology, a generative alternative to direct regression.
• Mixture-of-Experts velocity field: The flow's velocity field is parameterized by an MoE that routes inputs to specialized experts, explicitly accommodating heterogeneity across tumor types.
• Pan-cancer scope: Rather than fitting one model per tissue, MoLF is trained and benchmarked across multiple cancer types, aiming for a single generalizable predictor.
• Zero-shot cross-species generalization: The model transfers to cross-species data without retraining, indicating it captures conserved biological signal linking morphology to expression.

#Technical Details

MoLF couples a histology image encoder with a conditional flow-matching generative head. The flow-matching objective learns to transport samples from a noise distribution to the latent manifold of gene expression, conditioned on features extracted from H&E images. The transport's velocity field is implemented as a Mixture-of-Experts, so different sub-networks specialize on different tissue contexts and a routing mechanism selects among them per input. The authors evaluate MoLF on pan-cancer spatial transcriptomics benchmarks, reporting improvements over single-tissue methods and existing foundation-model baselines, and additionally show zero-shot generalization to cross-species data. As a recent arXiv preprint, detailed parameter counts and full training configurations are not summarized here; the authors do not report publicly released code or weights at the time of posting.

#Applications

MoLF is aimed at computational pathology and spatial-omics researchers who want to enrich routine histology with predicted molecular context. Because H&E slides are abundant while spatial transcriptomics assays are expensive, a model that infers spatial expression from morphology could provide molecular hypotheses for archival tissue, support tumor microenvironment characterization, and help prioritize which samples warrant true spatial profiling. Its pan-cancer training and cross-species generalization make it particularly relevant where reference spatial data are scarce, such as less-studied tumor types or model organisms.

#Impact

MoLF advances histology-to-expression prediction by bringing modern generative modeling—flow matching with a Mixture-of-Experts—to a problem usually framed as regression, and by demonstrating that a single pan-cancer model can generalize zero-shot across species. These are notable claims that, if they hold up under independent evaluation, would strengthen the case for using ubiquitous H&E imagery as a proxy for spatial transcriptomics. As a February 2026 preprint without yet-reported released code or weights, MoLF's reproducibility and downstream adoption are still to be established, and broader benchmarking against the growing set of histology-to-expression methods will clarify where it stands.