pertTF: context-aware AI modeling for genome-scale and cross-system perturbation prediction.

Area of Science:

• Genomics
• Computational Biology
• Cell Biology

Background:

• Predicting single-cell responses to genetic perturbations is crucial for understanding cell states and diseases.
• Current models struggle to predict phenotypic outcomes beyond gene expression and generalize across various perturbations and biological contexts.

Purpose of the Study:

• Introduce pertTF, a transformer-based AI model for predicting single-cell genetic perturbation responses.
• Enhance prediction accuracy for gene expression changes and infer cell identity shifts.
• Enable predictions in complex biological systems like primary human islets using transfer learning.

Main Methods:

• Trained pertTF on a unique dataset of single-cell expression profiles from 30 gene knockouts across 14 human cell types during pancreatic development.
• Utilized a transformer architecture to model complex gene interactions and cellular responses.
• Applied transfer learning to adapt the model to new cellular contexts and perturbation types.

Main Results:

• pertTF surpasses existing methods in predicting expression changes for unseen genes and cellular contexts.
• The model successfully infers perturbation-induced shifts in cell identity and population composition.
• Demonstrated generalizability through in silico CRISPR screens, identifying key regulators in stem cell and pancreatic development.

Conclusions:

• pertTF provides a powerful framework for integrating large-scale single-cell perturbation data with AI.
• The model advances the prediction of genetic perturbation effects across diverse cellular systems and disease contexts.
• Enables predictions in challenging experimental settings, facilitating research in developmental biology and disease modeling.