[Paper Review] Deep Autoencoders for Dimensionality Reduction of High-Content Screening Data

Lee Zamparo, Zhaolei Zhang|arXiv (Cornell University)|Jan 7, 2015

Cell Image Analysis Techniques14 references16 citations

TL;DR

This paper proposes stacked denoising autoencoders (SdA) for dimensionality reduction in high-content screening data, leveraging deep neural networks to capture non-linear relationships in millions of high-dimensional cell images. SdA outperformed PCA, LLE, Kernel PCA, and Isomap in preserving biologically meaningful phenotypic clusters, as measured by higher homogeneity in Gaussian Mixture Model clustering after reduction to 10–50 dimensions.

ABSTRACT

High-content screening uses large collections of unlabeled cell image data to reason about genetics or cell biology. Two important tasks are to identify those cells which bear interesting phenotypes, and to identify sub-populations enriched for these phenotypes. This exploratory data analysis usually involves dimensionality reduction followed by clustering, in the hope that clusters represent a phenotype. We propose the use of stacked de-noising auto-encoders to perform dimensionality reduction for high-content screening. We demonstrate the superior performance of our approach over PCA, Local Linear Embedding, Kernel PCA and Isomap.

Motivation & Objective

• To address the challenge of clustering millions of high-dimensional cell image features in high-content screening.
• To improve phenotype discovery by enabling non-linear dimensionality reduction that preserves biologically relevant structure.
• To develop a scalable, unsupervised method suitable for large-scale, unlabeled screening data with limited labeled examples.
• To evaluate whether deep autoencoders can outperform classical linear and manifold learning methods in downstream clustering performance.
• To demonstrate that SdA produces embeddings that are more amenable to clustering into biologically meaningful phenotypic subpopulations.

Proposed method

• Stacked denoising autoencoders (SdA) were trained using mini-batch stochastic gradient descent on scaled, normalized high-content screening data with 916 features per cell.
• Each denoising autoencoder layer corrupts input features by randomly setting a proportion of values to zero, then learns to reconstruct the original input through a non-linear mapping.
• The encoder maps high-dimensional inputs to a lower-dimensional latent space, while the decoder reconstructs the input from the latent representation, minimizing reconstruction loss.
• Model architecture and hyperparameters (learning rate, noise rate, momentum, weight decay) were tuned via grid search on a GPU-accelerated cluster using Theano.
• The final SdA model reduced data from 916 to 10–50 dimensions, followed by Gaussian Mixture Model (GMM) clustering with cross-validated initialization.
• Performance was evaluated using homogeneity scores averaged over five independent runs, with standard deviation reported.

Experimental results

Research questions

• RQ1 Can stacked denoising autoencoders achieve better dimensionality reduction than classical methods like PCA and Isomap in high-content screening data?
• RQ2 Does the non-linear modeling capacity of SdA improve the separation of biologically distinct phenotypic subpopulations in reduced space?
• RQ3 How does SdA compare to kernel PCA and LLE in preserving inter- and intra-cluster distances relevant to phenotype classification?
• RQ4 Can SdA scale effectively to millions of data points while maintaining high performance on unsupervised clustering tasks?
• RQ5 Is the learned low-dimensional representation from SdA more suitable for identifying phenotypically coherent clusters than those from linear or manifold-based methods?

Key findings

• SdA consistently outperformed PCA, LLE, Isomap, and Kernel PCA in clustering homogeneity across all reduced dimensions (10 to 50).
• The three- and four-layer SdA models achieved the highest average homogeneity, with the best-performing SdA model exceeding Kernel PCA by a significant margin.
• SdA produced wider mean inter-class distances compared to Kernel PCA, indicating better separation between distinct phenotypic groups.
• Although SdA did not always yield the tightest intra-class clustering, it maximized inter-class separation, which enhances clustering reliability.
• The SdA model achieved superior performance with minimal hyperparameter tuning, particularly when the base learning rate was optimized.
• The use of mini-batch stochastic gradient descent enabled efficient training on large-scale data, supporting scalability to millions of cells.