Squidiff: Predicting cellular development and responses to perturbations using a diffusion model

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Single-cell sequencing reveals cellular heterogeneity but mapping transcriptomic changes to stimuli is challenging.
• Current experimental methods for studying cellular responses to stimuli like radiation or drugs are labor-intensive.
• Elucidating disease mechanisms requires efficient tools to predict cellular behavior under various conditions.

Purpose of the Study:

• To develop a novel computational framework, Squidiff, for predicting transcriptomic alterations in diverse cell types in response to environmental stimuli.
• To demonstrate the utility of Squidiff in modeling complex biological processes and predicting cellular responses.
• To facilitate *in silico* screening of molecular landscapes for accelerated hypothesis generation and drug discovery.

Main Methods:

• Developed Squidiff, a diffusion model-based generative framework for predicting transcriptomic changes.
• Integrated continuous denoising and semantic feature learning to capture transient cell states.
• Applied the model to diverse scenarios including cell differentiation, gene perturbation, and drug response prediction.

Main Results:

• Squidiff accurately predicts transcriptomic landscapes across various cell types and conditions.
• Demonstrated robustness in modeling cell differentiation, gene perturbations, and drug responses.
• Successfully modeled blood vessel organoid development and cellular responses to neutron irradiation and growth factors.

Conclusions:

• Squidiff enables efficient *in silico* prediction of transcriptomic changes, overcoming experimental limitations.
• The framework facilitates rapid hypothesis generation and provides insights for precision medicine.
• Squidiff represents a significant advancement in computational tools for biological research and drug development.