pmVAE: Learning Interpretable Single-Cell Representations with Pathway Modules

Abstract

MotivationDeep learning techniques have yielded tremendous progress in the field of computational biology over the last decade, however many of these techniques are opaque to the user. To provide interpretable results, methods have incorporated biological priors directly into the learning task; one such biological prior is pathway structure. While pathways represent most biological processes in the cell, the high level of correlation and hierarchical structure make it complicated to determine an appropriate computational representation.ResultsHere, we present pathway module Variational Autoencoder (pmVAE). Our method encodes pathway information by restricting the structure of our VAE to mirror gene-pathway memberships. Its architecture is composed of a set of subnetworks, which we refer to as pathway modules. The subnetworks learn interpretable latent representations by factorizing the latent space according to pathway gene sets. We directly address correlation between pathways by balancing a module-specific local loss and a global reconstruction loss. Furthermore, since many pathways are by nature hierarchical and therefore the product of multiple downstream signals, we model each pathway as a multidimensional vector. Due to their factorization over pathways, the representations allow for easy and interpretable analysis of multiple downstream effects, such as cell type and biological stimulus, within the contexts of each pathway. We compare pmVAE against two other state-of-the-art methods on two single-cell RNA-seq case-control data sets, demonstrating that our pathway representations are both more discriminative and consistent in detecting pathways targeted by a perturbation.Availability and implementationhttps://github.com/ratschlab/pmvae

Machine learning for perturbational single-cell omics

“…Depending on the studied tissue, we believe that a structured approach to single-cell omic data taking advantage of spatial and temporal patterns will be more successful due to the reduced parameters and possible integration of prior knowledge (Lin et al, 2017;Fortelny and Bock, 2020;Gut et al, 2021). While we encourage the application of existing state-of-the-art models for structured data, we believe that intuitions regarding the patterns present in biological, chemical, and perturbation data will lead to cell biology-specific effective DL architectures with fewer and more interpretable input representations.…”

Biologically informed deep learning to infer gene program activity in single cells

“…We first proposed a regularized linear decoder to include domain knowledge into autoencoders for single-cell data at a conference 86 , with scalable and expressive embeddings when compared to existing factor models, such as f-scLVM 87 . Recent approaches such as VEGA 88 , scTEM 89 and pmVAE 90 also feature VAE-based architectures with linear decoders or training separate VAEs for each GP yet connected via a global loss in the case of pmVAE. In contrast, expiMap aims toward interpretable reference mapping allowing to fuse reference atlases with GPs and enabling the query of genes or GPs.…”

CanSig: discovery of shared transcriptional states across cancer patients from single-cell RNA sequencing data

“…scETM 20 computes an autoencoder-based topic model embedding that can be used to deconvolve cell expression in several signatures. pmVAE 23 learns an interpretable latent space representation where each embedding represents a separate pathway that can be used to analyze the organization of cells. Finally, ExpiMap 21 is a deep learning model that enables reference mapping through a similar organization as pmVAE, while also giving the opportunity to learn new pathways through unconstrained nodes.…”

“…To tackle the challenge of integrating scRNA sequencing from several patients, numerous methods that correct for some of these effects have been designed, including widely used scVI 15 , Harmony 17 or Scanorama 18 . Additionally, several recent methods can be harnessed to discover de novo shared transcriptional signatures [19][20][21][22][23] yet most of these methods do not address the aforementioned cancer-inherent challenges. Therefore, a computational approach for cancer-specific signature discovery needs to be developed.…”