Reproduce “MoLF: Mixture-of-Latent-Flow for Pan-Cancer Spatial Gene Expression Prediction from Histology” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Susu Hu, Stefanie Speidel
A generative framework named MoLF was developed to predict pan-cancer spatial gene expression from H&E histology, outperforming existing baselines in accuracy across 10 cancer types and demonstrating robust zero-shot generalization to cross-species data. It integrates a Transformer VAE for gene manifold learning with a Mixture-of-Experts conditional flow matching architecture to manage biological heterogeneity.
Key insights
- The MoE architecture allows MoLF to effectively model the extreme heterogeneity of pan-cancer data by decomposing the global transport map into specialized, local sub-problems, avoiding averaging artifacts seen in monolithic models.
- Experts within the MoE exhibit a distributed encoding strategy, learning feature-level invariances across the latent space rather than purely specializing by cancer type, which supports robust pan-cancer generalization.
- Training on diverse pan-cancer data enables the learning of broader, species-invariant morphological motifs, leading to superior zero-shot generalization capabilities across species.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.