DeepProfile

Train a variational autoencoder on a tumor expression matrix twice from two random seeds and you get two different latent spaces. Both reconstruct the data about equally well, and neither tells a biologist which latent variable is a real transcriptional program and which is an accident of initialization. That instability, plus the opacity of the mapping from genes to latent variables, is why much analysis of bulk cancer expression still runs on PCA. DeepProfile treats the instability as the thing to model: it trains hundreds of VAEs per cancer type across several latent sizes, pools every latent variable they produce, clusters them, and averages each cluster into one consensus variable. What survives is the structure that recurs across restarts rather than what a single run happened to find.

The framework comes from Su-In Lee's group in the Paul G. Allen School at the University of Washington, with collaborators at Harvard Medical School, Massachusetts General Hospital and the University of Geneva, and was published in Nature Biomedical Engineering after a March 2024 preprint. Its 18 cancer-specific encoders were trained on 50,211 microarray transcriptomes from 1,098 GEO datasets, then applied unchanged to data they had never seen: 9,079 TCGA RNA-seq samples across the same cancers, and GTEx normal tissue as a malignancy contrast. It extends the group's earlier model of the same name, which learned one acute myeloid leukemia latent space to predict chemotherapy response. A different network also called DeepProfile scores coelution maps inside the Full-DIA proteomics pipeline; the two share nothing but a name.

#Key Features

• Consensus across restarts: latent variables from every VAE run are clustered and averaged, so a reported variable is one that many independently initialized models agreed on rather than one run's factorization.
• Gene and pathway attribution: integrated gradients quantify each gene's contribution to each latent variable, and Fisher's exact tests against KEGG, BioCarta, Reactome and MSigDB gene sets turn those rankings into pathway-level annotations.
• Cross-platform transfer: encoders trained entirely on microarray data embed TCGA RNA-seq cohorts, with microarray and RNA-seq embeddings of the same samples correlating closely.
• One latent space per cancer: 18 separately trained models, each tuned to the transcriptional structure of its own disease, rather than a single pan-cancer encoder.
• Normal-tissue contrast: GTEx profiles pushed through the same encoders isolate pathways that matter in malignancy but not in the matching healthy tissue.

#Technical Details

Each constituent VAE takes the top 1,000 principal components of a batch-corrected, gene-intersected expression matrix through 250-unit and 100-unit hidden layers with batch normalization and ReLU before a Gaussian latent layer, with a mirrored decoder. Six latent sizes — 5, 10, 25, 50, 75 and 100 — are each trained 100 times from different seeds, giving 600 VAEs per cancer type and 10,800 across the 18 cancers. Their latent variables are pooled and k-means clustered, and cluster averages define the final embedding; the reported models compress expression into 150 latent variables, a size selected by a g-means test of Gaussianity rather than fixed by hand, with gene attributions ensembled under the same cluster assignments. Against PCA, ICA, random projection, autoencoders, denoising autoencoders and individual VAEs, DeepProfile latent variables captured more enriched pathways in 106 of 108 test cases (proportions z-test P = 1.6 × 10⁻³⁰¹) and more oncogenic gene sets in 102 of 108.

#Applications

A group with a new tumor expression cohort can project it into the matching cancer's latent space and use the resulting variables as compact, pre-annotated features for survival, subtype or mutation-burden models — most valuable where the cohort is far too small to fit a deep model of its own. Because each variable arrives with a ranked gene list and enriched pathways, a significant association reads as a biological statement rather than a coefficient on an anonymous component. The reported annotations are a resource in their own right: 156 pathways significant in more than nine cancer types, 163 unique to one.

#Impact

The paper's own analyses show what interpretable latent spaces buy. The highest-attributed gene across all 18 cancers is IL10RA, a regulator of the balance between pro- and anti-tumor inflammation, in the top 1% of genes in 14 cancers; cell-surface receptor gene sets are enriched among top-attributed genes, which PCA components do not recover. Screening 1,077 gene sets, the authors tie MHC class II antigen presentation and DNA mismatch repair expression to patient survival and cell-cycle pathways to tumor mutation burden, and nominate tumor-associated macrophages as the source of the survival-correlated HLA-D transcripts. The design has real limits: the corpus is bulk microarray, the 18 models are separate rather than one transferable pan-cancer encoder, and the trained encoder weights are not distributed — the repository carries the training and encoding scripts and the figshare archive the preprocessed input matrices, so reproducing the published latent spaces means retraining the ensembles.