Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session A

Poster A5 - Alignment of Spatial Transcriptomics Data with Intrinsic Coordinate Systems discovered by Machine Learning

Thursday, October 24, 2024
10:00 AM - 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  Spatial transcriptomics (ST) is a high-resolution and high-throughput technology that characterizes the gene expression level of cells in a tissue sample along with the spatial locations of these expressions. An important application of this technology is the ability to systematically compare the spatial expression profiles of multiple tissue samples that may represent varying biological or medical conditions. The potential for this comparative analysis drives the need for a technique that aligns heterogeneous tissue samples to each other to better identify meaningful similarities and differences among aligned regions of tissue. However, the cell-cell alignment of tissue samples using ST profiles is an unsolved computational problem. The best-performing of the few methods available today use a combination of expression and spatial information to align samples, but operate under the assumption that all genes in the transcriptome are important for capturing spatial information. We develop a new alignment algorithm that first learns how spatial expression patterns inherently capture features of the tissue’s spatial organization, and then uses this encoded information to compare cells for sample alignment. We show that our new method, “SEAlign”, creates more accurate alignments than existing tools, both in terms of matching biologically analogous tissue regions and in minimizing the spatial distortion required for alignment. We support this claim through systematic evaluations on ST datasets obtained through different technologies, as well as on synthetic data sets with known ground truth alignments.

Materials and

Methods:  SEAlign takes inputs in the form of spatial transcriptomics data on two biologically related tissue samples, and produces a mapping between pairs of cells from either tissue as an alignment. It first trains a neural network model called “SpaceExpress”, under parallel development in our group, to learn how tissue organization is reflected in the expression patterns of genes that are learned to be spatially-informative. The result of this step is a low-dimensional representation of each cell in both samples that captures the expression of these spatially-informative genes, inherently encoding spatial information. SEAlign then evaluates the similarity of the learnt low-dimensional representations for each cell across both samples. The most similar cells across samples are mapped to one another, and the collection of the mapped pairs forms a sample-sample alignment.

Results, Conclusions, and Discussions: We performed extensive quantitative assessments of SEAlign to demonstrate its improved accuracy compared to three existing alignment methods: PASTE, PASTE2, and SLAT. Our first evaluations were performed on synthetic ST data. This data’s expression profile included a subset of genes whose expression levels were informative about spatial location in a tissue, and they included a controlled degree of technical distortion between the two synthesized samples. The next evaluations were performed on ST data of consecutive pairs of tissue slices from a Drosophila embryo, obtained using Stereo-seq ST technology. The tissue slices in this dataset displayed relatively little biological distortion from one another. In contrast to this benchmark, our third evaluation was performed on more heterogeneous samples of the human dorsolateral prefrontal cortex from different donors, obtained using 10x Visium spatial transcriptomics technology. In each of these evaluations, we found SEAlign to be better than or competitive with the three current alignment methods at matching biological domains and preserving spatial configuration after alignment. These results indicate that SEAlign is a promising new computational algorithm for aligning spatial transcriptomics samples, even among heterogeneous samples.

Acknowledgements (Optional):