[📄 Paper] [🌐 Project Page] [💻 Code]

MoLF is a generative framework for scalable pan-cancer histogenomics. It utilizes a Mixture-of-Experts (MoE) velocity field within a conditional Flow Matching framework to predict spatial gene expression directly from standard H&E histology.

• State-of-the-Art: Outperforms existing methods on the HEST-1k pan-cancer benchmark.
• Scalable Architecture: Decouples diverse tissue patterns by routing morphological features through specialized experts.
• Zero-Shot Generalization: Demonstrates robust cross-species transfer (Human $\rightarrow$ Mouse) by learning invariant biological motifs.
• Optimal Transport: Efficiently maps random noise to a structured latent gene manifold in a single ODE step.

💡 Need low-compute, high-throughput precision for routine diagnostics? Check out our sister project, HistoPrism(ICLR 2026). While MoLF is built for exploratory robustness across diverse datasets, HistoPrism is highly optimized to deliver peak in-distribution accuracy for standardized human tissue screening and biomarker discovery.

1. Latent Compression: A Transformer-VAE compresses high-dimensional gene expression into a structured manifold.
2. Generative Flow: A Conditional Flow Matching model, parameterized by an MoE velocity field, generates expressions conditioned on H&E patches.

Project top-level layout:

.
├─ gene_autoencoder_latent_visulization.py
├─ gene_autoencoder_train.py
├─ index.html
├─ LICENSE
├─ post_training_experts_analysis.py
├─ post_training_trajectory_analysis.py
├─ README.MD
├─ sampling.py
├─ train.py
├─ configs/
│  ├─ autoencoder_config_local.yml
│  ├─ autoencoder_config.yml
│  ├─ training_config_latent_moe_local.yml
│  └─ training_config_latent_moe.yml
├─ data/
│  ├─ Hallmark/
│  ├─ Hest_Bench/
│  └─ sample_processed_HEST1K/
├─ figures/
├─ gene_vae/
│  └─ ...
├─ models/
├─ training_log/
└─utils/

• configs/: Configuration YAML files for VAE and flow training.
• data/: Preprocessed datasets (HEST1k, Hest_Bench, Hallmark gene sets) and split lists.
• figures/: Generated figures, overview diagrams, and assets used in the README/paper.
• gene_vae/: Outputs from gene VAE training (configs, checkpoints, UMAP embeddings and reducers).
• models/: Model definitions, baseline implementations, and architecture utilities.
• training_log/: Training run outputs and checkpoints, sampling/trajectory subfolders, and run configs.
• utils/: Helper modules for data loading, preprocessing, training utilities, and analysis tools.
• Top-level scripts: train.py (flow training), gene_autoencoder_train.py (VAE training), sampling.py (generate samples), post_training_* scripts (analysis & visualization).

• Download HEST1k from huggingface https://huggingface.co/datasets/MahmoodLab/hest. to data/HEST1K or [PATH/TO/YOUR/HEST1K]
• Please follow respective pathology foundation model (PFM) repositories for preprocessing. For example, preprocessed data with PFM uni_v2 is in data/sample_processed_HEST1K/univ2 with names [sample_id]_embeddings.pt which contains barcodes and correspongding uni features. This path is referred in utils/custom_dataset.py for dataloading.
• Also download hest_bench https://huggingface.co/datasets/MahmoodLab/hest-bench and put it in data/Hest_Bench [PATH/TO/YOUR/hest_bench] for fixed held-out test set and HVG-50 genes list
• Data split is merged from hest-bench and provided in data/Hest_Bench/hest_pathway_clean_split_0 and data/Hest_Bench/hest_pathway_clean_split_1.

Following code could be used to preprocecss HEST1k dataset.

python3 utils/pfm_preprocesing.py

• Stage 1 gene VAE

python3 gene_autoencoder_train.py -f configs/training_config_latent_moe.yml

• stage 2 latent flow matching

python3 train.py -f configs/training_config_latent_moe.yml

Pretrained weights can also be downloaded from huggingface.

# test set sampling
python3 sampling.py -l training_log/run_20260506_162107_molf -g 2.0

# external mouse validation sampling
python3 sampling.py -l training_log/run_20260506_162107_molf -g 2.0 -e

python3 gene_autoencoder_latent_visulization.py -l gene_vae/gene_vae_w_hvg_clean_split_0

add flag with chosen cfg scale from previous step

python3 post_training_trajectory_analysis.py -l training_log/MoLF_split_0 -g 3.0

Given sampling dir, to plot experts utilization.

python3  post_training_experts_analysis.py -s training_log/MoLF_split_0/sampling_2euler_cfg_guidance_3.0_epoch158 -t training_log/MoLF_split_0/trajectory_2euler_cfg_guidance_3.0_epoch158

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). For full details, see the LICENSE file.

If you find our work useful, please consider cite us.

@inproceedings{hu2026molf,
  author    = {Hu, Susu and Speidel, Stefanie},
  title     = {MoLF: Mixture-of-Latent-Flow for Pan-Cancer Spatial Gene Expression Prediction from Histology},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year      = {2026},
  url       = {https://icml.cc/virtual/2026/poster/63940}
}