scOTM: A Deep Learning Framework for Predicting Single-Cell Perturbation Responses with Large Language Models.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Accurate modeling of single-cell drug responses is crucial for healthcare.
• Current methods struggle with unpaired data and lack biological interpretability.
• Existing models often impose restrictive prior alignments, limiting their expressiveness.

Purpose of the Study:

• To develop a deep learning framework, scOTM, for predicting single-cell perturbation responses from unpaired data.
• To improve generalization to unseen cell types and enhance biological interpretability.
• To overcome limitations of existing methods in handling unpaired data and rigid prior constraints.

Main Methods:

• scOTM integrates biological knowledge from large language models into a variational autoencoder.
• Maximum mean discrepancy regularization allows flexible modeling of transcriptional shifts.
• Optimal transport establishes interpretable mappings between control and perturbed cell distributions.

Main Results:

• scOTM outperforms existing methods in predicting whole-transcriptome responses and identifying differentially expressed genes.
• The framework demonstrates superior robustness in data-limited scenarios.
• scOTM shows strong generalization capabilities across diverse cell types.

Conclusions:

• scOTM provides a powerful and flexible framework for predicting single-cell drug responses.
• The method enhances biological understanding by offering interpretable embeddings and flexible modeling.
• scOTM advances the field by effectively handling unpaired data and generalizing to new cell types.