[Paper Review] A coupled autoencoder approach for multi-modal analysis of cell types

Rohan Gala, Nathan W. Gouwens+7 more|arXiv (Cornell University)|Nov 6, 2019

Cell Image Analysis Techniques17 citations

TL;DR

This paper proposes a k-coupled autoencoder framework to align multi-modal single-cell data—specifically transcriptomic and electrophysiological profiles—by learning shared low-dimensional representations. It achieves ~80% accuracy in predicting transcriptomic cell types from electrophysiological data and enables discovery of cell types present in only one modality, even with sparse pairing.

ABSTRACT

Recent developments in high throughput profiling of individual neurons have spurred data driven exploration of the idea that there exist natural groupings of neurons referred to as cell types. The promise of this idea is that the immense complexity of brain circuits can be reduced, and effectively studied by means of interactions between cell types. While clustering of neuron populations based on a particular data modality can be used to define cell types, such definitions are often inconsistent across different characterization modalities. We pose this issue of cross-modal alignment as an optimization problem and develop an approach based on coupled training of autoencoders as a framework for such analyses. We apply this framework to a Patch-seq dataset consisting of transcriptomic and electrophysiological profiles for the same set of neurons to study consistency of representations across modalities, and evaluate cross-modal data prediction ability. We explore the problem where only a subset of neurons is characterized with more than one modality, and demonstrate that representations learned by coupled autoencoders can be used to identify types sampled only by a single modality.

Motivation & Objective

• To address the challenge of inconsistent cell type definitions across different single-cell modalities such as transcriptomics and electrophysiology.
• To develop a robust framework for cross-modal representation alignment that works even when only a subset of cells are measured in multiple modalities.
• To enable accurate prediction of cell types from one modality based on representations learned from another.
• To identify cell types that are uniquely present in a single modality, despite lack of direct pairing.
• To provide a scalable, differentiable method for joint analysis of multi-modal single-cell datasets with minimal assumptions on data structure.

Proposed method

• The method employs k-coupled autoencoders (k-CAE), where each modality is processed by a separate autoencoder with shared latent representations.
• A novel coupling loss is introduced based on the minimum eigenvalue of the batch covariance matrix to prevent representation collapse, avoiding issues seen with batch normalization.
• The framework is grounded in a probabilistic interpretation, where the optimization objective corresponds to maximizing a likelihood function, enabling generalization and theoretical justification.
• After training, a Gaussian mixture model is fitted to the shared latent space to enable clustering and cell type discovery.
• The approach is designed to handle partially paired datasets, where only a fraction of cells are measured in multiple modalities.
• The method allows for cross-modal prediction: predicting raw data or cell type labels in one modality from the other.

Experimental results

Research questions

• RQ1 Can cell types defined in one modality (e.g., transcriptomics) be accurately predicted from another modality (e.g., electrophysiology) using a shared representation?
• RQ2 How can representations from different modalities be aligned when only a subset of cells are measured in both?
• RQ3 What are the failure modes of coupling autoencoders, and how can they be mitigated to ensure stable, meaningful latent representations?
• RQ4 Can the framework discover cell types that are present in only one modality, even without paired data?
• RQ5 Does the proposed coupling mechanism support reliable, low-dimensional representation learning that preserves biological structure across modalities?

Key findings

• The coupled autoencoder framework achieves ~80% accuracy in predicting transcriptomic cell types from electrophysiological recordings when the transcriptomic hierarchy is resolved into 15 classes.
• Prediction accuracy drops to ~70% when the transcriptomic hierarchy is resolved into 25 classes, indicating sensitivity to class granularity.
• The method successfully identifies cell types that are uniquely present in a single modality, as shown by distinct, non-overlapping clustering in the shared latent space.
• Nearest-neighbor distance analysis confirms that unpaired types from one modality can be reliably distinguished from types in the other modality (p < 0.01 in K-S tests).
• The proposed minimum eigenvalue-based normalization effectively prevents representation collapse, outperforming batch normalization in maintaining multi-dimensional structure.
• The framework enables cross-modal reconstruction of raw data, demonstrating its utility beyond classification.