Representation Learning for Single-cell Perturbation Modeling: From Task-Specific Models to Foundations
Bai, Ding
Bai, Ding
Citations
Altmetric:
Author
Supervisor
Department
Machine Learning
Embargo End Date
Type
Dissertation
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Single-cell transcriptomics has opened new opportunities to understand cellular heterogeneity and predict transcriptional responses to genetic perturbations, which is a fundamental tool for uncovering gene function and identifying potential therapeutic targets. However, accurately modeling perturbation effects remains challenging, especially for multiplexed perturbations, out-of-distribution generalization, and scenarios that require accounting for gene interactions and long-range biological context. Existing methods remain limited in several important ways: task-specific models are often constrained by the scale and coverage of available perturbation datasets, while current pretrained foundation models do not directly resolve the condition-specific and out-of-distribution nature of perturbation prediction.
This dissertation studies genetic perturbation prediction from three complementary perspectives: task-specific modeling, large-scale foundation modeling, and condition-sensitive adaptation. First, we propose {AttentionPert}, an attention-based framework that models multiplexed perturbations through multi-scale global and local effects. Second, we develop {scLong}, a billion-parameter single-cell foundation model pretrained on large-scale transcriptomic data to capture long-range gene context across the full transcriptome. Third, we introduce {PertAdapt}, a framework that adapts pretrained foundation models to perturbation prediction via a plug-in perturbation adapter and an adaptive loss.
Extensive experiments across multiple single-cell perturbation benchmarks show that these methods improve the prediction of transcriptional responses under unseen conditions and provide a unified perspective on how specialized architectures, pretrained representations, and task-aware adaptation can be combined for more accurate and scalable perturbation modeling. Overall, this dissertation advances the use of artificial intelligence for single-cell analysis and contributes new methodological foundations for predictive perturbation biology."
Citation
Bai, Ding, "Representation Learning for Single-cell Perturbation Modeling: From Task-Specific Models to Foundations" PhD Dissertation, Machine Learning, MBZUAI, 2026.
Source
Conference
Keywords
Foundation Models, scRNA-seq, Genetic Perturbation