PertAdapt: unlocking single-cell foundation models for genetic perturbation prediction via condition-sensitive...

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Single-cell RNA sequencing (scRNA-seq) foundation models (FMs) show promise for predicting gene expression changes after genetic perturbations.
• Existing methods struggle with effective knowledge transfer and handling the imbalance between perturbation-sensitive and insensitive genes.
• This leads to limited improvements over non-pretrained baseline models.

Purpose of the Study:

• To develop a novel framework, PertAdapt, to enhance the accuracy of FMs in predicting genetic perturbation effects.
• To improve knowledge transfer from pretrained models and address the challenge of imbalanced gene sensitivity.
• To enable more precise predictions of transcriptional responses to genetic alterations.

Main Methods:

• Introduced PertAdapt, a framework integrating a plug-in perturbation adapter and an adaptive loss function.
• The adapter utilizes a gene-similarity-masked attention mechanism for encoding perturbation conditions and cell representations.
• An adaptive loss dynamically reweights perturbation-sensitive genes to better capture differential expression patterns.

Main Results:

• PertAdapt consistently outperformed non-pretrained and existing FM baselines across seven perturbation datasets (single- and double-gene settings).
• Demonstrated strong capabilities in modeling multiplexed gene interactions and generalizing in low-data scenarios.
• Showcased robustness across various backbone model sizes.

Conclusions:

• PertAdapt effectively addresses limitations in predicting genetic perturbation effects by improving knowledge transfer and gene weighting.
• The framework offers a significant advancement over existing methods for analyzing transcriptional responses.
• PertAdapt provides a robust and adaptable solution for diverse genetic perturbation prediction tasks.