[Paper Review] Variational autoencoders for tissue heterogeneity exploration from (almost) no preprocessed mass spectrometry imaging data

Paolo Inglese, James L. Alexander+3 more|arXiv (Cornell University)|Aug 23, 2017

Gaussian Processes and Bayesian Inference3 references3 citations

TL;DR

This paper proposes a variational autoencoder (VAE) framework for unsupervised exploration of mass spectrometry imaging (MSI) data with minimal preprocessing. By learning a non-linear, low-dimensional latent representation directly from raw spectral profiles, the VAE outperforms PCA in capturing tissue heterogeneity, enabling more accurate clustering and reconstruction of complex biochemical patterns in colon tissue samples without peak detection or alignment.

ABSTRACT

The paper presents the application of Variational Autoencoders (VAE) for data dimensionality reduction and explorative analysis of mass spectrometry imaging data (MSI). The results confirm that VAEs are capable of detecting the patterns associated with the different tissue sub-types with performance than standard approaches.

Motivation & Objective

• To address the challenge of analyzing high-dimensional, untargeted mass spectrometry imaging (MSI) data from complex tissues with unknown biochemical patterns.
• To explore whether deep generative models like VAEs can effectively capture non-linear molecular patterns in MSI data where linear methods like PCA fail.
• To evaluate VAEs for data compression, visualization, and clustering in MSI without requiring peak detection or spectral alignment.
• To compare VAE performance against standard PCA in reconstructing original spectral data and identifying biologically relevant tissue subtypes.

Proposed method

• A variational autoencoder (VAE) with a multilayer perceptron (MLP) encoder and decoder is trained end-to-end on raw MSI data to learn a continuous, low-dimensional latent space.
• The VAE uses the reparameterization trick to enable backpropagation, with the latent variable z sampled as z = μ + σε, where ε ~ N(0, I).
• The model employs a Gaussian prior p(z) = N(0, I) and a multivariate Bernoulli likelihood p(x|z) for reconstruction, with the loss function combining KL divergence and negative log-likelihood.
• The VAE is trained on desorption electrospray ionization (DESI) MSI data from 6 human colon specimens, with spectra linearly interpolated over 600–1000 m/z.
• Latent representations from the VAE and PCA are used for downstream clustering (Gaussian Mixture Models) and reconstruction, with performance evaluated via mean squared error (MSE).

Experimental results

Research questions

• RQ1 Can a VAE learn meaningful, non-linear latent representations of raw MSI data without peak detection or spectral alignment?
• RQ2 How does VAE-based dimensionality reduction compare to PCA in preserving tissue morphology and biochemical heterogeneity in MSI data?
• RQ3 Can VAE-generated latent spaces improve clustering accuracy for identifying tissue subtypes compared to PCA scores?
• RQ4 To what extent can VAEs reconstruct original MSI spectra with higher fidelity than PCA?

Key findings

• The VAE achieved significantly lower mean squared error (MSE) in spectral reconstruction than PCA, with MSE values ranging from 0.027 to 0.042 compared to PCA’s 0.065 to 0.082 across six samples.
• VAE-based latent representations enabled more accurate tissue sub-type clustering, with GMM using VAE features consistently identifying key morphological structures (e.g., central sub-type) across 2–19 cluster configurations, unlike PCA.
• The VAE model successfully captured tissue heterogeneity in less than 200 training iterations, with loss stabilizing quickly, indicating efficient training on raw MSI data.
• Visualizations of the VAE latent space revealed distinct clustering patterns (e.g., yellow cluster in sample 2) that aligned with histological features, while PCA scores showed noise-dominated variation outside tissue regions.
• The VAE model demonstrated robustness to high-dimensional, noisy MSI data, effectively compressing data while preserving biologically relevant patterns without pre-processing steps.