Integrative analysis of morphological and chemical signatures for understanding drug perturbations in diverse cancer cell models

High-throughput screening (HTS) and high-content imaging (HCI) have become central tools in biomedical research and drug discovery, enabling systematic interrogation of cellular responses to large numbers of chemical perturbations. However, the increasing complexity of experimental models, including co-culture systems, heterogeneous cell populations, and multi-parameter readouts, poses analytical challenges. Traditional workflows often struggle to capture context- dependent phenotypes, integrate diverse data modalities, or generalize across experimental conditions. This thesis focuses on developing and evaluating analytical strategies to improve the interpretation of HTS data generated from complex cellular assays. Using a range of multicellular in vitro models, this work demonstrates how cellular interactions can alter phenotypic responses to drug treatment, revealing effects that are not observable in simple single-cell-type assays. These findings emphasize the need for analysis pipelines that are sensitive to microenvironment- driven phenotypic variation while remaining scalable to large screening datasets. HCI is employed to capture single-cell morphological and phenotypic features across diverse treatment conditions, enabling systematic assessment of drug- induced cellular states in complex experimental contexts. To address the analytical demands of such data, this thesis develops and benchmarks computational pipelines combining classical image-based feature extraction with deep learning-based representations. The performance of these approaches is evaluated in tasks such as phenotypic profiling and mechanism of action (MOA) inference, revealing both their potential and their limitations. In particular, image-derived features are shown to be sensitive to biological context, cell type, and experimental design, which can limit robustness and generalizability across datasets. To complement morphology-based analyses, this work further investigates the use of chemistry-based descriptors for predicting drug MOA. Compared to image-based features, chemical representations provide more stable and transferable predictive performance, largely independent of cellular context. Importantly, integrating chemical and morphological information consistently improves predictive accuracy, demonstrating the value of multimodal approaches for extracting robust biological insight from high-throughput screens. Overall, this thesis contributes methodological insights and tools for the analysis of complex, high-throughput cellular assays. By highlighting both the opportunities and limitations of current phenotypic screening approaches and by demonstrating the benefits of multimodal data integration, this work supports the development of more reliable, scalable, and context-aware analytical frameworks for modern screening data. List of scientific papers 1. Gudoityte G, Sharma O, Leuenberger L, Wallin, E Fernebro J, Östling P, Bergström R, Lindberg J, Joneborg U, Kallioniemi O, Seashore- Ludlow B. Systematic profiling of cancer-fibroblast interactions reveals drug combinations in ovarian cancer. Mol Oncol. 2025 May 24. https://doi.org/10.1002/1878-0261.70051 II. Sharma O, Gudoityte G, Minozada R, Kallioniemi O, Turkki R, Paavolainen L, Seashore-Ludlow B. Evaluating feature extraction in ovarian cancer cell line co- cultures using deep neural networks. Commun Biol. 2025 Feb 25;8,303 (2025) https://doi.org/10.1038/s42003-025-07766-w III. Sharma O, Juhani K, Herren A, Kallioniemi O, Paavolainen L, Ballante F, Seashore-Ludlow B. From Chemistry to Morphology: Bi-Modal Machine Learning for MOA Prediction Across Cancer Cell Lines. [Manuscript]