APOLLO

Application to paired scRNA-seq and scATAC-seq

APOLLO training

step 1 latent optimization

train_lord_randNoise_sharedRecon_shareseq_filter_bce_morefilter.ipynb

step 2 inference

train_lord_randNoise_sharedRecon_shareseq_morefilter_reverse_bce.ipynb

Cell type classification

train_clf_lord_shareseq_celltype_bce_morefilter.ipynb

Interpretation of partially shared latent spaces

Genes or gene ontology terms with significant changes along each principal component of the latent spaces

plot_lord_bce_pca_sampling.ipynb - identify differentially expressed genes or peaks along each principal component of the shared or modality-specific latent spaces plot_lord_bce_pca_sampling_withAnnotations_curve.ipynb - plot differentially expressed genes or peaks along each principal component of the shared or modality-specific latent spaces plot_lord_bce_pca_sampling_withAnnotations.ipynb - plot the enriched gene ontology terms of the genes or peaks represented by the shared or modality-specific latent spaces

Preprocessing of scATAC-seq data

preprocess_shareseq.ipynb

Application to paired chromatin and protein images

APOLLO training

step 1 latent optimization

train_cnnvae_splitChannels_conditional_lord_randNoise_bce.ipynb - BCE loss used for reconstruction
train_cnnvae_splitChannels_conditional_lord_randNoise.ipynb - MSE loss used for reconstruction

step 2 inference

train_cnnvae_splitChannels_conditional_lord_randNoise_reverse_bce.ipynb - inference step for the model trained with BCE loss
train_cnnvae_splitChannels_conditional_lord_randNoise_reverse.ipynb - inference step for the model trained with MSE loss

APOLLO without modality-specific latent spaces

train_cnnvae_splitChannels_conditional_lord_randNoise_fullyJoint.ipynb - step 1, latent optimization
train_cnnvae_splitChannels_conditional_lord_randNoise_reverse_fullyJoint.ipynb - step 2, inference

APOLLO without shared decoders

train_cnnvae_splitChannels_conditional_lord_randNoise_correctBCE_noSharedRecon.ipynb - step 1, latent optimization (without decoders mapping from the shared latent space to reconstruction) train_cnnvae_splitChannels_conditional_lord_randNoise_reverse_correctBCEvalLoss_noSharedRecon.ipynb - step 2, inference

APOLLO trained with one-step training as an autoencoder

train_cnnvae_splitChannels_conditional.ipynb

Phenotype classification using real images, reconstructed images from full latent space, reconstructed images from shared latent space, or protein images predicted from chromatin

plot_Clf_conditions_sampling.ipynb - plot results
train_clf_conditions_c2c_fullrecon_sampling.ipynb - train classifiers using reconstructed chromatin images from the full latent space
train_clf_conditions_c2c_sharedrecon_sampling.ipynb - train classifiers using reconstructed chromatin images from the shared latent space
train_Clf_conditions_c2p_sampling.ipynb - train classifiers using protein images predicted from chromatin
train_clf_conditions_originalImg_chromatin_sampling.ipynb - train classifiers using the original chromatin images
train_clf_conditions_originalImg_sampling.ipynb - train classifiers using the original protein images
train_Clf_conditions_p2p_fullrecon_sampling.ipynb - train classifiers using reconstructed protein images from the full latent space
train_Clf_conditions_p2p_sharedRecon_sampling.ipynb - train classifiers using reconstructed protein images from the shared latent space

Interpretation of partially shared latent spaces of paired chromatin and protein images

Manually selected chromatin and protein morphological features

getNMCO_allFeatures.ipynb - preprocess
getNMCOgroups.ipymb - group chromatin features by correlation and selecting one representative feature for each group
getNMCOgroups_protein.ipymb - group protein features by correlation and selecting one representative feature for each group

Plotting examples of chromatin or protein images along each PC

plot_examples_centerPCs_percentiles_noHeldOut.ipynb

Features with significant changes along each principal component of the latent spaces

plot_nmco_centerPCs_percentiles_chromatin_allfeatures_sampling_groupNMCOde.ipynb - identify chromatin features with significant changes along PCs of the latent spaces
plot_nmco_centerPCs_percentiles_chromatin_allfeatures_sampling_groupNMCO.ipynb - plot the significant chromatin features
plot_nmco_centerPCs_percentiles_protein_allfeatures_sampling.ipynb - identify protein features with significant changes along PCs of the latent spaces
plot_nmco_centerPCs_percentiles_protein_allfeatures_sampling_groupNMCO.ipynb - plot the significant protein features

Feature ablation test of using manually selected features to classify phenotypes

train_clf_conditions_nmco_sampling.ipynb - train phenotype classifier using all represeentative morphological features
train_clf_conditions_nmco_sampling_featureAblation.ipynb - train phenotype classifier with feature ablation
plot_clf_conditions_nmco_sampling.ipynb - plot results

Predicting gH2AX shared and modality-specific features using chromatin images

train_pred_chromatinImg2proteinFeatures.ipynb - train regression models compareImg2Features.ipynb - plot results

Image preprocessing

preprocess.ipynb

Benchmarking

benchmarking_inpainting.ipynb - compare to the previous image inpainting method for protein image prediction: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007348

Application to paired scRNAseq and protein abundance data (CITE-seq)

./citeseq directory

Application to Human Protein Atlas data

./hpa contains all three notebooks for the three models trained using each pair of chromain, ER, and microtubule markers.