Abstract: FR-PO033
Multimodal Data Analysis with Spatial Transcriptomics
Session Information
- AI, Digital Health, Data Science - II
November 03, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Border, Samuel, University of Florida Department of Biomedical Engineering, Gainesville, Florida, United States
- Lucarelli, Nicholas, University of Florida Department of Biomedical Engineering, Gainesville, Florida, United States
- Naglah, Ahmed, University of Florida Department of Medicine - Quantitative Health Section, Gainesville, Florida, United States
- Mimar, Sayat, University of Florida Department of Medicine - Quantitative Health Section, Gainesville, Florida, United States
- El-Achkar, Tarek M., Indiana University School of Medicine Department of Medicine Division of Nephrology, Gainesville, Florida, United States
- Jain, Sanjay, University of Washington St Louis Department of Medicine Division of Nephrology, St. Louis, Missouri, United States
- Melo Ferreira, Ricardo, Indiana University School of Medicine Department of Medicine Division of Nephrology, Gainesville, Florida, United States
- Eadon, Michael T., Indiana University School of Medicine Department of Medicine Division of Nephrology, Gainesville, Florida, United States
- Sarder, Pinaki, University of Florida Department of Medicine - Quantitative Health Section, Gainesville, Florida, United States
Background
Spatial transcriptomics (ST) methods have provided researchers the ability to link morphological observations in histology with molecular mechanisms. This expanded view has great potential in a clinical setting towards improving patient stratification and treatment design. However, it is difficult for a human to handle all of these data. Machine Learning (ML) methods are specially equipped to handle dense datasets, exhibiting high performance on image-related tasks. ML can be used to provide details about cell populations and their cell states in non-specially prepared tissues.
Methods
We first built a visualization tool, FUSION, that allows users to examine the distribution of cell types and their states across functional tissue units. Users can also select regions of interest over which to aggregate ST data. We then made a prediction model (SpotNet) to characterize cellular information from histology tissues. Our model uses a ResNet50 encoder and a set of branched layers to predict cell types and states. Two models were trained using glomerulus and tubule-specific cells. There were more tubules than glomeruli. A dynamic patching approach greatly increased the diversity of our training data without substantial augmentation. Predictions made on a patch-level were converted into heatmaps for visualization of cell type localization.
Results
FUSION is deployed on the web for public access. SpotNet achieved a Mean Absolute Error (MAE) of 6.4±3.2%, 10.1±1.8%, 1.2±1.16%, 11.9±4.2% on prediction of 6 cell types and 4 cell states for glomeruli and 12 cell types and 4 cell states for tubules respectively. Predicted cell localizations overlapped well with ground truth.
Conclusion
We present a method for interactive viewing of ST data and a novel ML model for prediction and interpretation from images. Integration of this model into our visualization platform will enable highly granular analysis of histology.
Funding
- Other NIH Support