Geometrically invariant niche representations for spatial transcriptomics
Harvard-MIT Health Sciences and Technology (HST)
Abstract: Niches shape cell states through local interactions, so comparing niches is central to understanding how tissues respond to perturbations and change with age. We present a niche representation that is invariant to translation, rotation, and reflection while encoding polar structure. The representation applies across spatial modalities, including CosMx, MERFISH, Slide-seq, STARmap, Stereo-seq, and Xenium. We demonstrate three applications: aligning tissues across samples, predicting perturbation effects in spatial Perturb-seq using niche counterfactuals, and mapping niches from tissues, including aging tissues, to annotated reference atlases.
Biography: David Goh is a PhD student in the Harvard-MIT Health Sciences and Technology (HST) program, advised by Professor Eric Sun, where he develops statistical and machine learning methods to study the spatiotemporal biology of aging. He earned his MEng in Chemical Engineering from Imperial College London, with a year abroad at MIT. His earlier work applied physics and machine learning to gene regulation, immunology, and the microbiome.