Revisiting Deep Archetypal Analysis for Phenotype Discovery in High Content Imaging

Mario Wieser, Daniel Siegismund, Stephan SteigeleWACV 2025

TL;DR

This paper introduces a novel deep archetypal analysis method for discovering biological phenotypes in high-content imaging (HCI) data. By relaxing the traditional simplex constraint to a hypersphere and employing an online dictionary learning approach, their method effectively identifies subtle biological differences in complex datasets, outperforming existing state-of-the-art techniques.

What Problem Does This Paper Address?

Drug discovery programs face a significant challenge in identifying promising treatment candidates for complex diseases. High-content imaging (HCI) assays generate vast amounts of multi-spectral image data from cell culture models, enabling the screening of thousands of drug candidates. However, the sheer volume of this data makes systematic human review to distinguish between disease and healthy phenotypes incredibly difficult and time-consuming. Existing automated methods, while improving, often rely on handcrafted features or require manual phenotype labeling, introducing bias. Archetypal analysis (AA) offers an unsupervised approach to uncover these phenotypes by identifying "extremal" data points. However, traditional and even recent deep AA methods struggle with high-dimensional HCI data, requiring a disproportionately large number of latent dimensions to represent a smaller number of biological phenotypes, leading to weak representations and ambiguous archetypes.

Methods

The core innovation of this paper lies in its approach to deep archetypal analysis. They propose a two-step method:

  1. Learning Hyperspherical Representations: Instead of the traditional Euclidean latent space, they leverage the Hyperspherical Variational Autoencoder (VAE) to learn a latent representation on a unit hypersphere. This is achieved by replacing the standard Gaussian distribution in the information bottleneck framework with a von Mises-Fisher distribution, which is more suitable for spherical data. This allows them to map complex, high-dimensional input images into a lower-dimensional hyperspherical latent space.
  2. Learning Approximate Archetypes: Once the hyperspherical representations are learned, they employ an online dictionary learning approach to identify the archetypes. This involves predicting mixing coefficients and learning an archetype dictionary. Crucially, they relax the simplex constraint on the archetypes, allowing for a more flexible representation. The entire process is trained end-to-end using an alternating optimization scheme.

Datasets

The authors evaluated their method on three datasets:

  • Synthetic MNIST: A standard benchmark dataset used to demonstrate the method's ability to identify distinct classes (arbitrarily chosen as digits 0, 1, and 3) as archetypes.
  • Cells-out-of-Sample (CooS) Dataset: An industry-relevant dataset containing mouse cell microscopy images. They used a subset of three distinct cell phenotypes: Endoplasmic Reticulum, Inner Mitochondrial Membrane, and Cytosol.
  • Neurite Assay Dataset: Another industry-relevant dataset focusing on neuronal development and toxicity. They used image patches categorized into three phenotype classes: healthy, damaged, and dead cells.

Key Insights and Ideas

  • Hypersphere for Phenotype Discovery: The shift from a Euclidean latent space to a hypersphere is a critical insight. Biological phenotypes often exist in complex, non-linear relationships, and a hyperspherical representation can better capture these nuances than a simple simplex.
  • Relaxing the Simplex Constraint: The traditional simplex constraint in AA is a bottleneck, especially when the number of latent dimensions needed to represent data significantly exceeds the number of desired archetypes. Relaxing this constraint allows for a more flexible and powerful model.
  • End-to-End Learning of Representations and Archetypes: By integrating representation learning (hyperspherical VAE) with archetype discovery (online dictionary learning) in an end-to-end fashion, the method ensures that the learned representations are optimized for the task of phenotype identification.
  • Online Dictionary Learning for Scalability: The use of online dictionary learning makes the method amenable to large-scale datasets, which is crucial for real-world HCI applications.

Key Results

  • MNIST: The proposed HCI-AA method achieved a Mean Squared Error (MSE) of 0.002 on the MNIST dataset, significantly outperforming clustering (GMM, MSE 0.19) and linear AA (MSE 0.27) methods, and competitive with other deep learning approaches like DeepAT (MSE 0.171).
  • HCI Datasets (CooS and Neurite): For both industry-relevant HCI datasets, HCI-AA demonstrated superior performance in terms of image reconstruction error (MSE), which serves as a proxy for the quality of learned representations.
  • CooS: HCI-AA achieved an MSE of 0.0072, compared to 0.0284 for DeepAT and 0.0265 for AAnet.
  • Neurite: HCI-AA achieved an MSE of 0.0001, significantly better than 0.0014 for both DeepAT and AAnet.
  • Qualitative Results: Visually, HCI-AA was able to uncover all three biologically expected phenotypes in both HCI experiments, whereas baseline methods often failed to identify all phenotypes, detecting only one or two.

Limitations and Future Work

The authors acknowledge the following limitations and suggest future directions:

  • Approximate Archetypes: Similar to other online learning methods, their approach finds approximate archetypes, unlike methods that consider the entire dataset.
  • Prior Knowledge of Archetype Count: The method currently requires prior knowledge of the number of archetypes. Future work aims to develop data-driven methods for automatically selecting the number of archetypes.
  • Latent Space Interpretability: In some cases, interpreting the latent archetypal space can be challenging due to the high flexibility of neural networks. Future work will focus on incorporating additional invariances to improve interpretability.
  • Sophisticated Representation Learning: They plan to investigate more advanced representation learning techniques for future improvements.

Summary updated

BibTeX
@inproceedings{wieser2025revisiting,
  author = {Mario Wieser and Daniel Siegismund and Stephan Steigele},
  title = {{Revisiting Deep Archetypal Analysis for Phenotype Discovery in High Content Imaging}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure