Closing the loop: Teaching single-cell foundation models to learn from perturbations.

Area of Science:

• Computational Biology
• Genomics
• Immunology

Background:

• Single-cell foundation models (scFMs) predict cellular responses to perturbations.
• Current scFMs cannot learn from experimental data to refine predictions.
• Closing the loop between prediction and experimentation is needed for iterative model improvement.

Purpose of the Study:

• To develop a closed-loop framework for scFMs that integrates experimental perturbation data.
• To enhance the predictive accuracy and utility of scFMs for biological discovery.
• To identify potential therapeutic targets for RUNX1-familial platelet disorder.

Main Methods:

• Developed a closed-loop framework extending scFMs by fine-tuning with perturbation data.
• Applied the closed-loop model to analyze T-cell activation and RUNX1-familial platelet disorder.
• Evaluated model performance by measuring prediction accuracy and positive predictive value.

Main Results:

• The closed-loop model significantly improved prediction accuracy.
• Positive predictive value for T-cell activation was increased three-fold.
• Identified two therapeutic targets (mTOR, CD74-MIF) and two novel pathways (PKC, PI3K) for RUNX1-familial platelet disorder.

Conclusions:

• Iterative incorporation of experimental data enhances foundation model predictions.
• The closed-loop framework represents a significant step toward realizing "virtual cell" models.
• This approach accelerates biomedical discovery by refining in silico predictions with experimental validation.