Speakers: Laura Guerra, Denisa Sufaj
14 Aprile 2026 | 15:45
DEIB, Sala Seminari "A. Alario" (Ed. 21)
Per maggiori informazioni: Silvia Cascianelli | silvia.cascianelli@polimi.it
Sommario
Tuesday, April 14th, 2026 at 3:45 pm a new appointment of Data Science Seminars: Bioinformatics focus will take place in DEIB "Alessandra Alario" Seminar Room (Building 21) organized by the Data Science for Bioinformatics group.These Thesis Rehearsal will be held by Laura Guerra and Denisa Sufaj on the following subjects:
"A biology-driven evaluation of cell segmentation algorithms on imaging-based spatial transcriptomics datasets" - Laura Guerra
Imaging-based Spatial Transcriptomics technologies have revolutionized our ability to map gene expression directly on the tissue at sub-cellular resolution. However, translating fluorescent signals and transcript coordinates into meaningful single-cell data relies on cell segmentation. Despite its critical role, segmentation remains one of the most significant bottlenecks in spatial data analysis pipelines. The primary objective of this research is to provide a robust and systematic evaluation of cell segmentation algorithms on spatial transcriptomics data. The key distinction between these algorithms lies in their design: while some are developed to segment cells directly from fluorescent images (image-based algorithms), others operate based on transcript distribution (point cloud-based algorithms). By only considering the computed cell shapes on the image, one might conclude that imaging-based methods perform better, as they produce more visually intuitive and regular shapes compared to transcript-based algorithms, which often generate unconventional morphologies. However, it is essential to acknowledge that an irregular shape does not necessarily imply poor performance; but it reflects a different segmentation logic driven by transcript distribution rather than pixel intensities. Consequently, to objectively compare these different approaches, this evaluation shifts the focus away from visual morphology comparison to assessing the functional integrity of the segmented units through the inspection of their internal content. We seek to determine whether the transcriptional profiles are valid and if the resulting cells represent biologically meaningful entities. To achieve this, we employ quantitative evaluation metrics specifically designed to penalize algorithms that produce biologically inconsistent results, ensuring that the segmentation reflects real cellular functions and compositions. Cell segmentation with various image-based (e.g. Cellpose), point cloud-based (e.g. Baysor, Proseg) and hybrid (e.g. RNA2Seg) algorithms was performed across 3 gold-standard datasets provided by 2 imaging spatial transcriptomics technologies (MERSCOPE and Xenium) covering 3 different tissues: Human Brain (MERSCOPE), Human Ovary (Xenium) and Human Kidney (Xenium). Results demonstrate that no single algorithm is universally superior; performance is highly dependent on tissue morphology and on the specific platform used. This work provides a critical framework for selecting the most appropriate segmentation method, highlighting that accurate boundary definition is essential to prevent the misinterpretation of downstream biological findings, ranging from cell-type assignment to the characterization of spatial microenvironments.
"Computational characterization of senescent and proliferative cell states in prostate cancer epithelium by single-cell transcriptomics"- Denisa Sufaj
Cellular senescence is increasingly recognised as a heterogeneous and context-dependent cellular state whose identification in tumour tissue remains methodologically unresolved. The absence of universal biomarkers, together with the limitations of binary classification schemes, makes it difficult to capture the phenotypic continuum occupied by senescent cells in vivo. In this thesis, we present a computational framework for the identification and characterisation of senescence-associated epithelial cell states in human prostate cancer at single-cell resolution. We integrated eight publicly available human prostate single-cell RNA-seq datasets comprising more than 136,000 epithelial cells from 107 samples and 69 patients, spanning healthy tissue, benignadjacent tissue, primary tumour, distant metastatic disease, and treated and untreated clinical conditions. Senescence- and proliferation-related transcriptional programmes were quantified using UCell scoring based on the SenMayo reference signature and curated KEGG and Reactome cell-cycle gene sets. Their joint analysis revealed that senescence- and proliferation-associated programmes partially coexist within a structured epithelial transcriptional landscape, rather than defining two sharply separated states. This observation motivated the development of a novel composite metric, the human Senescence Index Tool (hSIT), defined as the difference between the senescence score and the mean cell-cycle score. Gene-wise correlation with hSIT, together with unsupervised clustering and supervised machinelearning approaches, was used to reconstruct the senescence–proliferation axis and identify robust senescence-associated epithelial states. Epithelial cells distributed along a continuous hSIT gradient, supporting a continuum model of senescence in prostate cancer epithelium. Within this landscape, clustering analyses identified biologically interpretable compartments corresponding to senescence-enriched and proliferation-enriched, with k-means providing the most robust and coherent partition for downstream classification. Gene-ranking and feature-selection analyses further identified coordinated transcriptional programmes distinguishing senescence-like from proliferation-dominant states across multiple disease contexts. Overall, these findings establish hSIT as a biologically interpretable and analytically robust representation of epithelial senescence heterogeneity in prostate cancer and provide a transferable computational framework for studying senescence in tumour ecosystems in which senescence and proliferation coexist.