Enhancing Biomarker Discovery in Tumor Hypoxia for Head and Neck Squamous Cell Carcinoma: Advancing Spatial Omics Data Accessibility through Convolutional Autoencoders
- Date in the past
- Monday, 3. June 2024, 09:00
- REZ, F.03.082, 3rd floor
- Verena Bitto
Address
REZ, F.03.082, 3rd floor
Organizer
Dekan
Event Type
Doctoral Examination
Spatial omics holds promise for identifying biomarkers that reflect spatial properties in tumors, such as hypoxia. However, analyzing this data is challenging due to its high dimensionality, small sample size, and significant multicollinearity – particularly when the features of interest are not the predominant signal. Also, unlike purely predictive tasks, biomarker discovery demands a high degree of explainability and involves more than merely identifying a few discriminative features. In this work, these challenges are addressed by combining convolutional autoencoders and random forests, leveraging the strengths of both algorithms to manage multicollinearity and enhance explainability. As data, mass spectrometry imaging and spatial transcriptomics data from consecutive tissue slices of head and neck squamous cell carcinoma tumor models are used. The results show that the proposed combined approach extracts consistently more biologically relevant features than random forest models alone. Furthermore, it is outlined how spatial omics can be combined to retrieve multimodal biomarkers.
Considering the increasing amount of (spatial) omics data, explainable deep learning approaches are becoming increasingly important. This work contributes to the understanding of autoencoders by demonstrating how specific characteristics in spatial omics data reflect in the latent space and how they can be tailored through hyperparameter configuration.