AI-Fun with ELLIS Seminar | Kristina Ulicna: Decoding Biology by Learning Meaningful Representations of Cell Morphologies from Single-Cell Imaging Phenomaps
| Dates: | 21 May 2025 |
|---|
| Times: | 11:00 - 12:00 |
|---|
| What is it: | Seminar |
|---|
| Organiser: | Faculty of Science and Engineering |
|---|
| Who is it for: | University staff, External researchers, Alumni, Current University students |
|---|
|
The Manchester Centre for AI Fundamentals and Manchester's ELLIS Unit are co-hosting a series of seminars featuring expert researchers working in the fundamentals of AI.
Kristina is a trained biomedical scientist turned computational biologist. After obtaining her BSc degree in Biomedical Science @ King's College London, Kristina started her PhD training with the London Interdisciplinary Doctoral (LIDo) Programme @ University College London (UCL) and completed her PhD under the guidance of Drs Alan Lowe (The Alan Turing Institute 'AI for Science' Fellow, UCL) & Guillaume Charras (London Centre for Nanotechnology, UCL).
Title: Decoding Biology by Learning Meaningful Representations of Cell Morphologies from Single-Cell Imaging Phenomaps
Abstract: High-throughput microscopy and CRISPR-based functional screening generate vast, high-dimensional datasets, offering unprecedented insights into cellular behaviour. However, extracting biologically meaningful representations from these data remains a challenge, particularly in the presence of systematic biases and limited prior knowledge of the underlying structure. In this talk, I will explore two complementary machine learning approaches for representation learning in cell imaging data. First, I will present how sparse dictionary learning—widely used in mechanistic interpretability of language models—can be adapted to extract high-level biological concepts, such as cell types and perturbation effects, from microscopy foundation models. Second, I will introduce a weakly supervised latent variable inference method that identifies and removes off-target single-cell instances in CRISPR screens, mitigating proximity bias introduced by systematic chromosome-arm correlations. Together, these approaches offer a path toward decoding biology from microscopy images with single-cell precision, helping us uncover meaningful relationships between different gene knockouts and their effects on cellular phenotypes. By learning more structured and interpretable representations, we can improve our understanding of genetic perturbations and their broader implications in functional genomics.
In person attendance is recommended but for those unable to attend in person, a you can register via the link provided to receive zoom details.
Travel and Contact Information
Find event
LT1.4
Kilburn Building
Manchester
