P5: Pan-Cancer Digital Pathology Analyses Predict Protein Expression and Post-Translational Modifications

Xiyue Wang*, Junhan Zhao*, Guillaume Larghero, Jiayu Zhang, Wei Yuan, Jinxi Xiang, Bao Li, Tsung-Hua Lee, Kuan-Yao Huang, Christian Engel, Jing Zhang, Eliana Marostica, Stuart Schnitt, Nancy U. Lin, Jeffrey A. Golden, MacLean P. Nasrallah, Sen Yang, and Kun-Hsing Yu

Lead Contact: Kun-Hsing Yu

Kun-Hsing_Yu AT hms DOT harvard DOT edu

Department of Biomedical Informatics, Harvard Medical School, Boston, MA

Department of Pathology, Brigham and Women’s Hospital, Boston, MA

Harvard Data Science Initiative, Harvard University, Cambridge, MA

Kempner Institute for the Study of Natural and Artificial Intelligence, Harvard University, Cambridge, MA

ABSTRACT

Characterizing the proteomic landscape is essential for understanding cancer progression and treatment response. Recent advances in mass spectrometry and protein arrays have enabled high-throughput quantification of protein expression and the identification of post-translational modifications (PTMs) linked to clinical outcomes. However, due to cost and time constraints, proteomic profiling is not routinely performed for all patients. To address this challenge, we established the Pan-Cancer Proteomics Prediction Platform via Pathology Imaging (P5), a weakly supervised machine learning framework that leverages foundation models to systematically predict proteomic profiles from whole-slide images. We analyzed 7,694 whole-slide images (WSIs) across 23 cancer types to evaluate the relationship between tissue morphology and the proteomic dysregulation of 25,158 proteins. Our AI models successfully predicted 4,913 protein markers with an area under the receiver operating characteristic curve exceeding 0.8. We validated our findings using 2,764 WSIs from 850 patients across independent study cohorts and our affiliated hospital. In addition, in-depth analysis of oncogenic pathways uncovered a direct link between tissue morphology and cell cycle regulation. We further demonstrated that P5 can expedite clinical trial enrollment by identifying patients likely to harbor the targeted proteomic profiles. Overall, P5 uncovered previously unrecognized connections between pathology imaging patterns and proteomic alterations, providing a fast and cost-effective approach to proteomic characterization that enhances cancer management and streamlines clinical trial enrollment.

Framework Overview.

Pre-requisites:

• Linux (Tested on Ubuntu 18.04)
• NVIDIA GPU (Tested on Nvidia GeForce L40s x 48GB)
• Python (Python 3.10.14),torch==2.0.0, torchvision==0.9.1+cu111, h5py==3.6.0, matplotlib==3.5.2, numpy==1.22.3, opencv-python==4.5.5.64, openslide-python==1.3.0, pandas==1.4.2, Pillow==10.0.0, scikit-image==0.21.0 scikit-learn==1.2.2,scikit-survival==0.21.0, scipy==1.8.0, tensorboardX==2.6.1, tensorboard==2.8.0.

Installation Guide for Linux (using anaconda)

1. Installation anaconda(https://www.anaconda.com/distribution/)

2. sudo apt-get install openslide-tools

git clone https://github.com/hms-dbmi/P5.git
cd P5
conda env create -f environment.yaml
conda activate P5
pip install -e .

Prepare

Step 1. prepare for data

1.Download all TCGA and CPTAC WSIs, proteomics information in here .

You need to process the WSI into the following format. The processing method can be found in CLAM

The Vichow2 feature extractor and the pretrained model can be download in

DATA_DIR
├─patch_coord
│      slide_id_1.h5
│      slide_id_2.h5
│      ...
└─patch_feature
        slide_id_1.pt
        slide_id_2.pt
        ...

The h5 file in the patch_coord folder contains the coordinates of each patch of the WSI, which can be read as

coords = h5py.File(coords_path, 'r')['coords'][:]
# coords is a array like:
# [[x1, y1], [x2, y2], ...]

The pt file in the patch_featurefolder contains the features of each patch of the WSI, which can be read as

features = torch.load(features_path, map_location=torch.device('cpu'))
# features is a tensor with dimension N*F, and if features are extracted using CTransPath, F is 768

Step 2. preparing the data set split

You need to divide the dataset into a training set validation set and a test set, and store them in the following format

SPLIT_DIR
    test_set.csv
    train_set.csv
    val_set.csv

And, the format of the csv file is as follows

slide_id	label
slide_id_1	0
slide_id_2	1
...	...

Train model

Step 1. create a config file

We have prepared two config file templates (see ./configs/) for 2560, like

General:
    seed: 7
    work_dir: WORK_DIR
    fold_num: 4

Data:
    split_dir: SPLIT_DIR
    data_dir_1: DATA_DIR_1 
    features_size: 2560
    n_classes: 2

Model:
    network: 'P5'

In the config, the correspondence between the Model.network, Train.training_method and Train.val_method is as follows

Model.network	Train.training_method	Train.val_method
P5	P5	P5

Step 2. train model

Run the following command

CUDA_VISIBLE_DEVICES=3 \
python3 train.py \
--config_path "project/configs/TCGA_protein/brca_over_sclwc.yaml" \
--set_seed \
--begin 0 \
--end 5

--begin and --end used to control repetitive experiments

Issues

• Please open new threads or address all questions to Junhan_Zhao@hms.harvard.edu, xiyue.wang.scu@gmail.com or Kun-Hsing_Yu@hms.harvard.edu

License

P5 is made available under the GPLv3 License and is available for non-commercial academic purposes.