Gene-Level Representation Learning via Interventional Style Transfer in Optical Pooled Screening

TL;DR

GRAPE presents a GAN-based framework that learns gene perturbation representations from Optical Pooled Screening images by interventional style transfer, using a trainable gene embedding $M \in \mathbb{R}^{107 \times 500}$ and a mapping network to generate perturbation-specific style codes. Through adversarial, cycle-consistency, and style reconstruction losses, GRAPE disentangles perturbation effects from nuisance content, enabling gene embeddings that cluster by CORUM-defined complexes and outperform engineered features and several baselines in clustering metrics, with competitive mAP. The approach demonstrates the viability of style-transfer-inspired representation learning for high-throughput gene function discovery, while noting limitations in perturbation breadth and the potential gains from expanding the perturbation set. Overall, GRAPE offers a scalable pathway to uncover latent gene relationships from OPS data, with practical implications for health and disease research.

Abstract

Optical pooled screening (OPS) combines automated microscopy and genetic perturbations to systematically study gene function in a scalable and cost-effective way. Leveraging the resulting data requires extracting biologically informative representations of cellular perturbation phenotypes from images. We employ a style-transfer approach to learn gene-level feature representations from images of genetically perturbed cells obtained via OPS. Our method outperforms widely used engineered features in clustering gene representations according to gene function, demonstrating its utility for uncovering latent biological relationships. This approach offers a promising alternative to investigate the role of genes in health and disease.

Figures (5)

• Figure 1: Overview of the GRAPE model: GRAPE builds upon StarGAN v2, and like other GAN-based models is composed of a generator and a discriminator, optimized with adversarial loss. In the GRAPE generation process, input image $x_i$ paired with its corresponding genetic perturbation $y_i$ is passed through an encoder. A target gene ($y_{target}$) is randomly selected, and its embedding is transmitted to the generator's decoder along with the encoded input, aiming to generate images reflecting the content of the input image $x_i$ and perturbation responses of $y_{target}$. The multihead discriminator takes the generated/real image $x_j$ with its corresponding generated/real perturbation $y_j$, backpropagating the loss from real vs fake classification through the respective head of $y_j$. Our primary objective in this work is to acquire effective representations for genetic perturbations from the trainable embedding layer. Finally, we assess the quality of the representations using various evaluation metrics such as mAP and clustering metrics. For a thorough investigation, we also evaluated and comapred different potential gene representations at positions . (Icons have been designed using images from www.flaticon.com.)
• Figure 2: GRAPE generates realistic perturbation phenotypes. Top: Style transfer from an input image (non-targeting) to ANAPC7 (anaphase promoting complex subunit 7) and SMARCA4 (SWI/SNF related, matrix associated, actin dependent regulator of chromatin, subfamily a, member 4) gene knockouts. Bottom: Distribution of five CellProfiler features most informative for classifying perturbed and control cells.
• Figure 3: UMAP visualization of GRAPE representations (left), and engineered representations (right). Each dot represents a single gene, and the color indicates its ground truth CORUM cluster label.
• Figure 4: Performance comparison between GRAPE representations and baselines for multiple clustering metrics (left: Normalized Mutual Information, middle: Adjusted Rand Index, right: Purity) and for different numbers of clusters $k$. The ground truth number of clusters is 14, denoted by a vertical line. The $k$-means algorithm was executed 100 times for each $k$, and the standard deviation is depicted as a shaded envelope.
• Figure 5: Comparison of mean Average Precision (mAP) for GRAPE representations, and GRAPE representation with an Exponential Moving Average (EMA) filter applied during the training process.