LUMIC: Latent diffUsion for Multiplexed Images of Cells

Area of Science:

• Cellular biology
• Computational biology
• High-content imaging

Background:

• High-content, single-cell technologies like morphological profiling offer insights into fundamental biological processes.
• Inferring causal mechanisms in heterogeneous cell systems requires navigating complex perturbation and cell type combinations.

Purpose of the Study:

• To develop a generative model framework, LUMIC (Latent diffUsion for Multiplexed Images of Cells), for generating high-quality, high-fidelity cell images.
• To leverage generative models as priors for anticipating high-content experimental outcomes and designing more informative experiments.

Main Methods:

• LUMIC integrates diffusion models with DINO (self-Distillation with NO labels) for self-supervised feature learning and HGraph2Graph for chemical representation.
• The framework was applied to two datasets: the JUMP Pilot dataset (~27,000 images) and a newly-generated dataset (~3,000 images).
• Prediction quality was quantified using DINO embeddings, Kernel Inception Distance (KID) score, and morphological feature distribution recovery.

Main Results:

• LUMIC generated realistic, high-quality cell images across different cell lines and chemical treatments.
• The framework demonstrated generalization capabilities to unseen compounds and cell types.
• LUMIC significantly outperformed existing methods in image generation and prediction quality.

Conclusions:

• LUMIC provides a powerful tool for generating realistic cell images, advancing high-content screening and biological discovery.
• The framework facilitates the design of more efficient and informative experiments in cell biology research.
• Generative models like LUMIC are crucial for navigating high-dimensional biological data and uncovering causal mechanisms.