Efficient Self-Supervised Domain Adaptation of Vision Transformers for Single-Cell Morphological Profiling

Sammanfattning

Morphological profiling via the Cell Painting assay is an important tool in modern toxicological screening and drug discovery. While classical engineered feature extraction pipelines, such as CellProfiler, remain the industry standard, deep learning foundation models offer unprecedented capabilities for capturing complex phenotypic representations. However, full-parameter fine-tuning of massive Vision Transformers (ViTs) to specialized multi-channel data is computationally prohibitive. This thesis investigates the efficacy of Efficient Self-Supervised Adaptation (ESSA), utilizing Low-Rank Adaptation (LoRA) and a Bag of Channels (BoC) encoding strategy, to adapt the DINOv3 foundation model to single-cell microscopy data of iPSC-derived hepatocytes. The adapted models were evaluated across diverse high-content screening tasks, including targeted single-cell anomaly detection (multi-nucleolar and multi-nuclei classification), macroscopic dose-response modeling, and linear probing against classical engineered feature spaces. The results demonstrate that parameter-efficient, self-supervised adaptation significantly outperforms the unadapted DINOv3 baseline and competes robustly with domain-specific pre-trained models (DINO BoC).Crucially, the inference extraction strategy dictated downstream performance. Extracting representations using explicitly masked cell crops (Masked CLS) prevented spatial context leakage from dense cellular microenvironments, yielding the highest accuracy in localized biological tasks without the need to train the models with masked crops. Furthermore, linear probing confirmed that perfect spatial alignment allows the self-supervised ViT to implicitly encode classical biology, successfully capturing over half of the mathematical variance present across 3,876 standard CellProfiler features. The evaluations additionally highlighted that sphering normalization actively suppresses essential biological outliers required for anomaly detection. Finally, scalability was confirmed via a Multi-plate ESSA model that maintained generalized robustness while trained on independent cell-painting plates with a higher total variance. Ultimately, this work proves that parameter-efficient, self-supervised adaptation is a highly viable and computationally accessible framework for translating massive vision foundation models into sensitive, domain-specific tools for biomedical research.