Imaging with Interactions

QC 2026-09-04

Abstract

Spatial biology has attracted growing interest because it measures two types of information at the same time: which molecules are present in a tissue and where they are located. Both are needed for understanding biology, as the organization of tissues, cells, and molecules drives biological processes in health and disease. Most leading spatial technologies recover this spatial organization from an external reference, either by imaging the tissue directly with optics or by using a substrate containing barcoded coordinates that were decoded in advance. Sequencing-based microscopy, the approach studied in this thesis, instead obtains spatial information by sequencing the interactions inside the sample. DNA-barcoded molecules interact with their neighbors, sequencing reads the corresponding interaction products,  and a reconstruction algorithm recovers coordinates from the resulting network, much like assembling a jigsaw puzzle. Thus, the spatial organization of a tissue can be recovered without optics and without a predefined spatial reference. This is particularly attractive as a natural solution to three-dimensional imaging. However, a sequencing-based microscopy experiment produces a network rather than an image, and how to reconstruct coordinates from that network is the computational problem at the center of this thesis.

Article I addresses part of this problem, which is how to recover coordinates when the interaction rule is unknown and the graph is large. We developed STRND, a graph reconstruction method that uses random walks and manifold learning to reconstruct images from proximity networks without assuming how physical distance determines edge formation. Its computational complexity is approximately linear in the number of  nodes, which allowed reconstruction of large graphs.

Article II addresses how to assess and denoise a DNA barcode network without reference ground-truth coordinates. We introduced spatial coherence, which uses three topology-based metrics to evaluate if shortest-path distances in a network follow the geometric properties expected from physical distances. The metrics are able to detect distortions in topology caused by false interactions, and also provided an objective function to guide denoising.

Article III applies these tools to an existing spatial transcriptomics method that was initially not meant for sequencing-based microscopy. This method, Slide-tags, diffuses bead-carried DNA barcodes into tissue sections and normally assigns cell positions using an optically decoded bead map. However, we found that the sequencing data also contained a cell-bead proximity network produced via diffusion during the experiment, and that this network preserved enough spatial information for a reconstruction that only uses interaction data. Applying STRND to a human tonsil dataset recovered tissue coordinates without the optical decoding step, including approximately 2,000 cells that the original analysis had excluded.

Article IV reexamines the assumption that spatial reconstruction requires networks dominated by short-range interactions. Longer-range interactions are usually treated as noise, but we found that densely connected networks could still be reconstructed accurately even after individual hop distances collapsed into a few discrete values. The reason is that each node had a distinct shortest-path profile that changed smoothly with position, thus containing spatial information. Proximity interactions are therefore sufficient for spatial reconstruction, but they are not always necessary. These results suggest that increased network connectivity can be beneficial for reconstructions under certain regimes.

Article V extends the sequencing-based microscopy logic to an application unrelated to spatial imaging, DNA data storage. Write-by-partitioning stores information in the spatial order of a DNA barcode network that forms in a hydrogel before the message is known, like a blank page that can be written on. The message is written by cutting the hydrogel into labeled sections and recovered by reconstructing the network order and detecting the partition boundaries from sequencing data. The writing chemistry is therefore independent of the message content. We showed that this strategy stored and recovered information, and spatial coherence confirmed the expected quasi-one-dimensional structure of the barcode networks.

These five articles show that DNA barcode networks contain spatial and ordinal information. The network structure can be reconstructed, evaluated, and denoised computationally without optical coordinates, and useful geometry is present in a more diverse range of interaction mechanisms and connectivity regimes than what was previously assumed. These results establish DNA barcode networks as a measurable source of spatial information and show that the same reconstruction principles can also support applications outside microscopy.

Link to DiVA