Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow
Authors: Mengbo Wang, Shourya Verma, Aditya Malusare, Luopin Wang, Yiyang Lu, Vaneet Aggarwal, Mario Sola, Ananth Grama, Nadia Lanman
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Using spatial transcriptomic gene expression and corresponding histology data, we construct a novel framework, Gene Flow... we generate high-resolution images... Our rectified flow based method outperforms diffusion methods and baselines in all experiments. |
| Researcher Affiliation | Academia | Purdue University *Equal Contribution (EMAIL, EMAIL), Corresponding Authors 1 Computer Science, 2 Industrial Engineering, 3 Comparative Pathobiology, 4 Institute For Cancer Research. |
| Pseudocode | Yes | D.1 RNA Encoder: The RNA Encoder is a deep neural encoder designed to transform gene expression profiles from multiple single cells into a compact, biologically enriched embedding. This encoder supports enhanced cell representation through three mechanisms... D.2 Conditioned U-Net Architecture with RNA and Timestep Embeddings: The conditioned U-Net integrates residual blocks with both timestep and RNA conditioning throughout its encoder, bottleneck, and decoder. |
| Open Source Code | Yes | https://github.com/wangmengbo/Gene Flow. |
| Open Datasets | Yes | To curate our training data with high resolution H&E stained image and real single-cell level resolved spatial transcriptomics data, we used three large publicly available spatial transcriptomics datasets prepared with 10x Genomics Xenium platform [36]... We also extended experiments to 59 human Xenium samples from 12 organs in the HEST-1k dataset [37]... |
| Dataset Splits | Yes | For each dataset, we performed 3-fold cross-validation to ensure robustness and generalizability of our results across tissue variations. |
| Hardware Specification | Yes | All experiments were conducted on a single NVIDIA H100 GPU, with training times ranging around 12 hours per experiment (on full sample) requiring up to 78 GB of VRAM. |
| Software Dependencies | No | The paper mentions optimizers and learning rate schedules but does not explicitly list software dependencies with specific version numbers (e.g., Python 3.8, PyTorch 1.9). |
| Experiment Setup | Yes | We trained our models for 100 epochs using the Adam W optimizer [34] with a batch size of 96. The learning rate followed a cosine annealing schedule [35] with a minimum learning rate set to 1% of the initial value, helping the model converge more smoothly during later stages of training. Additionally, Appendix E.4 specifies: 'k-nearest neighbors: k = 5 Spatial loss weight: λs = 0.1 Warmup epochs: 5 epochs with linear ramp-up Activation threshold: Spatial loss begins at 70% of total epochs or when validation loss drops below a predetermined threshold'. |