🧚🏻‍♀️ProGENIE: An interpretable AI framework for transcriptomic inference and therapy-relevant profiling from prostate cancer histology

Abstract

Molecular stratification is central to precision prostate cancer management, yet high costs and tissue requirements limit routine clinical implementation. Recent advances in AI have enabled transcriptomic inference from routine histology. However, the biological relevance of inferred transcriptomes remains incompletely characterized. Here, we present ProGENIE, an interpretable framework that integrates spatially aware transformation and multi-head attention pooling to infer genome-scale transcriptomic profiles (16,591 genes) from H&E-stained images. ProGENIE generates spatially resolved transcriptomic heatmaps that recapitulate proliferation, stromal, and immune programs within the tumor microenvironment (TME). Using an independent cohort from South Australian hospitals (SAH), the model identified 3,167 well-predicted genes and demonstrated improved performance relative to representative histology-to-transcriptomics methods. Further, we demonstrate the value of ProGENIE in stratifying recurrence risk, characterizing TME programs, and capturing therapy-relevant biological signals. By enabling genome-scale molecular profiling from routine histology, ProGENIE provides a scalable, cost-effective strategy with the potential to expand access to precision oncology.

Overview

Pre-requisites

• Linux (Tested on Red Hat Enterprise Linux 8.4)
• NVIDIA GPU (Tested on NVIDIA A100 PCIe 40GB)
• Python (Python 3.10.0), PyTorch==1.13.1+cu116, Torchvision==0.14.1+cu116, Torchaudio==0.13.1+cu116, Matplotlib==3.10.1, NumPy==1.23.5, OpenCV-Python==4.11.0.86, Openslide-Python==1.4.1, Pandas==2.2.3, Scikit-Image==0.25.2, Scikit-Learn==1.6.1, SciPy==1.15.2, Seaborn==0.13.2, Einops==0.8.1, Transformers==4.49.0, Timm==1.0.3, Tensorboard==2.19.0, TensorboardX==2.6.2.2

Installation Guide for Linux (using anaconda)

1. Clone this git repository: git clone https://github.com/ABILiLab/ProGENIE.git
2. cd ProGENIE
3. Create a conda environment: conda create -n progenie python=3.10.0
4. conda activate progenie
5. Install the required package dependencies: pip install -r requirements.txt

Preparation

1. Prepare the wsi/ directory, which contains the whole slide images
2. Prepare the reference file: example/ref_file.csv

For example:

WSI File Name	Patient ID	rna_A1BG	...	rna_ZZZ3	tcga_project
TCGA-2A-A8VL-01A	TCGA-2A-A8VL-01A	0.0658	...	2.4027	TCGA-PRAD
TCGA-2A-A8VO-01A	TCGA-2A-A8VO-01A	0.0243	...	2.5807	TCGA-PRAD
TCGA-2A-A8VT-01A	TCGA-2A-A8VT-01A	0.0195	...	3.6254	TCGA-PRAD

3. Prepare the ground truth label file: examples/true_label.csv

Preprocessing

1. Create patches from WSIs

To extract image patches from raw Whole Slide Images (WSIs), run the patch generation script provided in: pre_processing/create_patches.py

An example script to run the patch extraction: scripts/create_patches.sh

2. Extract Features from Patches

To extract features from patches using a pretrained encoder, run the feature extraction script provided in: pre_processing/extract_patch_features.py

An example script to run the feature extraction:scripts/extract_patches_features.sh

Note: Pretrained encoder weights (e.g., for UNI, CHIEF, and Prov-GigaPath) can be obtained by referring to the following repositories:

• UNI: https://github.com/mahmoodlab/UNI

• CHIEF: https://github.com/hms-dbmi/CHIEF

• Prov-GigaPath: https://github.com/prov-gigapath/prov-gigapath

3. Obtain k-Means Features

To compute k-Means features from extracted patch features, run the clustering script provided in: pre_processing/kmean_features.py

An example script to run the k-Means clustering: scripts/kmean_features.sh

Inference on independent dataset

We released the model weights for four pre-trained models on HuggingFace, please download the weights first.

1. Prepare the dataset

To combine the k-Means features with ground truth gene expression profiles for model training, run the dataset preparation script provided in: pre_processing/prepare_dataset.py

An example script to run the dataset preparation: scripts/prepare_dataset.sh

2. Inference and evaluation

To perform model inference on the test set and evaluate performance:

Run the main inference and evaluation script:inference.py

You can find an example script here:scripts/inference.sh

The output will be saved in: examples/results

This includes:

• test_pred_labels.csv: predicted gene expression values

• test_true_labels.csv: ground truth labels

• test_gene_metrics.csv: PCC, RMSE, and R²

Visualization on spatial

Generate spatial gene expression heatmaps from spatial transcriptomics slides using a sliding-window inference strategy. Predicted expression values are assigned to the center patch of each window, and optional Gaussian smoothing can be applied to both ground-truth and predicted expression maps for visualization.

An example script to run the visualization:

python visualization.py \
  --base_dir heatmap \
  --count_file filtered_feature_bc_matrix.h5 \
  --h5_path features/wsi_features.h5 \
  --model_path checkpoints/model_best.pth \
  --gene_names_pt data/dataset.pt \
  --genes_csv data/common_genes.csv \
  --out_dir results \
  --offsets_json wsi_crop.json \
  --gene ACTA2

Note: --gene_names_pt should be the training dataset (dataset.pt), which stores the gene ordering used during model training.

Reference

Please cite our publication: Han et al., "Predicting gene expression from whole slide images in prostate cancer using deep learning". doi: xxx.