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 B

Poster A2 - Computational methods for spatially-informed gene expression comparison for spatially resolved transcriptomics data across technologies

Thursday, October 24, 2024

2:45 PM – 3:45 PM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Vivien Jiang

Johns Hopkins University Undergraduate
Johns Hopkins University
Baltimore, Maryland, United States

Primary Investigator(s)

Jean Fan, PhD

Assistant Professor
Johns Hopkins University, United States

Co-Author(s)

Kalen Clifton (she/her/hers)

Graduate Student
Johns Hopkins University, United States

Introduction: High-throughput detection of targeted RNA species in fixed tissue of individual cells or small groups of cells is made possible by spatial transcriptomic technologies, which enables high-throughput gene expression profiling while maintaining the spatial positions of cells inside tissue sections. Spatial transcriptomics allows for both molecular and spatial information to be used to extract relevant biological insights, from detecting spatially variable genes (SVGs) to those that outline spatial organization and connections between distinct cell types.

There are many spatial transcriptomic technologies that have technical variations in the way that they measure transcript counts. However, quantifying variations in observed gene expression caused by differences in the method of measuring gene expression remains challenging. Quantifying technical differences between technologies will inform researchers on which technologies are best suited for their experiment and how to interpret data collected from varying spatial transcriptomic technologies. To solve this problem, we developed an approach to perform spatially-informed gene expression comparison for spatially resolved transcriptomics data across technologies through spatial structural alignment and quantifying the area of tissue where gene expression is similar across technologies relative to the area that is not. We apply this metric to compare two spatial transcriptomics platforms, multiplex error-robust FISH (MERFISH) and Xenium, for profiling a coronal section of the adult mouse brain.

Materials and

Methods: Cell positions in the MERFISH dataset were aligned to the cell positions in the Xenium dataset using STalign, to decrease variations such as tissue distortion and tears, ensuring that comparisons are made at matched locations between tissue sections. The datasets were reduced to genes shared between datasets and gene expression was normalized based on counts per million normalizations accordingly. Since the cells are not one-to-one preventing direct spatial comparisons, the datasets were rasterized using SEraster to create spatially matched values, by aggregating gene expression into spatially aligned 200um pixels that can be compared.

For each gene, we quantify its spatial similarity across the two technologies by first applying a threshold to parse out the pixels that do not have expression in either of the technologies. We then evaluate the slope (m) as the ratio of gene expression in the technologies for each pixel, such that m=1 for a pixel with the same gene expression. Log2 transformation was applied to the slopes of each pixel for all genes, which allowed the distribution to be approximately normal.

Finally, we quantify a similarity metric (S) to quantify how the expression for a given gene is similar across technologies as the proportion of pixels that are within log2(m) is half a standard deviation away from 0 compared to the total number of pixels that have expression. The mean similarity metric across all genes thus quantifies how the spatial gene expression profiles for the two tissues sampled using different technologies are similar.

Results, Conclusions, and Discussions: Results
We applied this metric to all 62 genes profiled by both MERFISH and Xenium. The mean probability of similarity for all shared genes is S=0.32. The range of probabilities S=0.004 to S=0.71. The probability of similarity is S=0.71 for Gfap (Figure 1), which is a much smaller discrepancy between gene profiles than S=0.14 for Pde11a (Figure 2).

Conclusion
From the mean probability of similarity metric of all genes, the two technologies have large discrepancies. However, the discrepancies vary on a gene-to-gene basis. Genes with small discrepancies inform researchers that there is comparable performance in the technologies. For genes with greater discrepancies, on the other hand, orthogonal spatial profiling methods or assessments of additional spatial transcriptomics datasets from other technologies would be needed to determine which technologies are providing the more accurate spatial transcriptomic characterization.

Discussion
We presented an approach to perform spatially-informed gene expression comparisons for spatially resolved transcriptomics data across technologies. Using spatial structural alignment with STalign and aggregation into common coordinate pixels with SEraster, we enabled the evaluation of gene expression at matched spatial locations. We further developed a similarity metric that quantified the spatial similarity of gene expression patterns and applied it to characterize mouse brain tissues profiled by two different spatial transcriptomics technologies. Our approach can be applied to any spatial transcriptomic dataset.

Our current approach is notably less precise at quantifying genes that have varying expression patterns between technologies, presenting potential limitations. We assume that the Log2 transformation of the slopes of each pixel is approximately Gaussian (Figure 1D). However, there can be up to three peaks in the distribution (Figure 2D) from the pixels that are only present in either of the technologies. In such a situation, half a standard deviation away from 0 may be an overestimate of the true variance, thus potentially inflating the number of pixels that are considered similar. Future applications of Gaussian Mixture Models can provide a more appropriate standard deviation on the proportion of pixels that are similar.

Overall, we anticipate this method to be useful when analyzing spatial transcriptomic datasets collected by different technologies.

Acknowledgements (Optional):