TY  - GEN
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/34903/
N2  - Spatial omics data shows potential to reveal novel insights into the underlying mechanisms of cancer. Yet the high-dimensional and highly correlated feature space imposes challenges on analysis. In this thesis, the implementation of convolutional autoencoders to extract explainable features for biomarker discovery is examined, exemplified on tumor hypoxia.

Mass spectrometry imaging and spatial transcriptomics experiments were performed on consecutive tissue slices of head and neck squamous cell carcinoma tumor models. To
advance accessibility of these spatial omics modalities, data was reduced by convolutional autoencoders and the resulting latent space features were ranked for association with tumor hypoxia through random forest feature importance measures. With the help of a newly proposed recovery method, the contribution of original features to a latent feature
was derived, thereby retaining biological relevant information. The derived genes and peptides were compared against the ranked genes and peptides of a random forest only model. The feature sets of the autoencoder approaches achieved consistently higher scores when evaluated using the structural similarity index measure. In contrast, the features of the random forest only models contained many more noisy hypoxia associations caused by the multicollinearity of features.

Several promising unimodal and multimodal biomarker candidates of mass spectrometry imaging and spatial transcriptomics data for tumor hypoxia were identified. Multimodal biomarkers were identified through correlation analysis of aligned serial tissue slices from
both spatial omics modalities in four samples. For a more elaborate integration, it was outlined how the molecular information of multiple spatial omics modalities may be
combined without error-prone alignment of consecutive tissue slices. Instead, the spatial omics modalities may be learned directly from the readily available microscopy images
using convolutional neural networks. Then, the learned molecular information may be predicted from microscopy images of other spatial omics modalities. Preliminary results demonstrated that the learning of the latent space features of autoencoders yielded more accurate predictions than when learning was performed on the raw and sparse spatial omics features. However, it necessitates further investigation whether also hypoxia-associated
features can be acquired accurately from microscopy images.

Overall, the findings show that convolutional autoencoders accompanied by random forest models retain more biological relevant information for biomarker discovery than without
prior feature extraction. Considering the increasing amount of available (spatial) omics data, deep learning feature extraction will become evermore important. This thesis
contributes to the overall understanding of autoencoders by showcasing how specific characteristics in spatial omics data reflect in the latent space and how they can be
addressed through hyperparameter configurations.
CY  - Heidelberg
AV  - public
TI  - Enhancing Biomarker Discovery in Tumor Hypoxia for Head and Neck Squamous Cell Carcinoma: Advancing Spatial Omics Data Accessibility through Convolutional Autoencoders
ID  - heidok34903
KW  - Autoencoder
KW  -  Random Forest
KW  -  Biomarker Discovery
KW  -  Mass Spectrometry Imaging
KW  -  Spatial Transcriptomics
A1  - Bitto, Verena
Y1  - 2024///
ER  -