dpaysan/st-net

Deep learning on histopathology images.

Integrating spatial gene expression and breast tumour morphology via deep learning

ST-Net is a machine learning model for predicting spatial transcriptomics measurements from haematoxylin-and-eosin-stained pathology slides. For more details, see the acompanying paper,

Integrating spatial gene expression and breast tumour morphology via deep learning
by Bryan He, Ludvig Bergenstråhle, Linnea Stenbeck, Abubakar Abid, Alma Andersson, Åke Borg, Jonas Maaskola, Joakim Lundeberg & James Zou.
Nature Biomedical Engineering (2020).

Downloading Dataset and Configuring Paths

By default, the raw data must be downloaded from here and placed at data/hist2tscript/. The processed files will then be written to data/hist2tscript-patch/. These locations can be changed by creating a config file (the priority for the filename is stnet.cfg, .stnet.cfg, ~/stnet.cfg, ~/.stnet.cfg). An example config file is given as example.cfg.

Preparing Spatial Data

This code assumes that the raw data has been extracted into SPATIAL_RAW_ROOT as specified by the config file.

python3 -m stnet prepare spatial  # caches the counts and tumor labels into npz files
bin/create_tifs.sh                  # converts jpegs into tiled tif files

Training models

The models for the main results can be trained by running:

ngenes=250
model=densenet121
window=224
for patient in `python3 -m stnet patients`
do
    bin/cross_validate.py output/${model}_${window}/top_${ngenes}/${patient}_ 4 50 ${patient} --lr 1e-6 --window ${window} --model ${model} --pretrain --average --batch 32 --workers 7 --gene_n ${ngenes} --norm
done

To run the comparison for different window sizes:

ngenes=250
model=densenet121
for window in 128 299 512
do
    for patient in `python3 -m stnet patients`
    do
        bin/cross_validate.py output/${model}_${window}/top_${ngenes}/${patient}_ 4 50 ${patient} --lr 1e-6 --window ${window} --model ${model} --pretrain --average --batch 32 --workers 7 --gene_n ${ngenes} --norm
    done
done

To run the comparison for different magnifications:

ngenes=250
window=224
model=densenet121
for downsample in 2 4
do
    for patient in `python3 -m stnet patients`
    do
        bin/cross_validate.py output/${model}_${window}/top_${ngenes}_downsample_${downsample}/${patient}_ 4 50 ${patient} --lr 1e-6 --window ${window} --model ${model} --pretrain --average --batch 32 --workers 4 --gene_n ${ngenes} --norm --downsample ${downsample}
    done
done

To run the comparison against random initialization:

ngenes=250
model=densenet121
window=224
for patient in `python3 -m stnet patients`
do
    bin/cross_validate.py output/${model}_${window}/top_${ngenes}_rand/${patient}_ 4 50 ${patient} --lr 1e-6 --window ${window} --model ${model} --average --batch 32 --workers 7 --gene_n ${ngenes} --norm
done

To run the comparison against individual training of genes:

ngenes=250
model=densenet121
window=224
for i in `seq 10`
do
    ensg=`python3 -m stnet ensg ${i}`
    for patient in `python3 -m stnet patients`
    do
        bin/cross_validate.py output/${model}_${window}/top_${ngenes}_singletask_${i}/${patient}_ 4 50 ${patient} --lr 1e-6 --window ${window} --model ${model} --pretrain --average --batch 32 --workers 7 --gene_list ${ensg} --norm
    done
done

To run the comparison against hand-crafted features:

ngenes=250
window=224
model=rf
for patient in `python3 -m stnet patients`
do
    root=output/${model}_${window}/top_${ngenes}/${patient}_ 
    python3 -m stnet run_spatial --gene --logfile ${root}gene.log --epochs 1 --pred_root ${root} --testpatients ${patient} --window ${window} --model ${model} --batch 32 --workers 7 --gene_n ${ngenes} --norm --cpu
done

Analysis

The main results can be generated by running:

bin/generate_figs.py output/densenet121_224/top_250/ cv

The corresponding results for the comparisons are generated by running:

for i in output/densenet121_224/*; do bin/generate_figs.py $i cv; done

Generating Figures

The following blocks of code are used for generating several of the figures.

Visualization of prediction across whole slide:

bin/visualize.py output/sherlock/BC23450_cv.npz --gene FASN
bin/visualize.py output/sherlock/BC23903_cv.npz --gene FASN

UMAP Clustering:

bin/cluster.py