C-T.19: Foundation Model-Guided Cell Type Annotation for Imaging-Based Spatial Transcriptomics
Authors
Nico Kraus
Goethe University Hospital
Cristina Ortiz
Goethe University Hospital
Christoph Welsch
Goethe University Hospital
Marcel H. Schulz
Goethe University Hospital
Keywords
Spatial Transcriptomics, Imaging-Based, Foundation Model, Cell Type Annotation
[sponser-meet-now-chat][/sponser-meet-now-chat]
Spatial transcriptomics (ST) enables the linking of gene expression profiles with spatial coordinates. In particular, imaging-based methods such as CosMx or MERSCOPE offer subcellular resolution, thereby enabling the investigation of a wide range of biological questions. However, these methods are inherently limited to targeted marker panels, leaving crucial markers for downstream analyses such as cell type annotation unmeasured.
Various approaches have been developed to close this information gap, broadly falling into two categories: computational prediction of missing gene expression on the one hand, and reference-based mapping of single-cell RNA sequencing (scRNA-seq) data onto ST datasets on the other. However, both strategies have shown only limited success, as the inherent technical differences between scRNA-seq and ST assays restrict reliable information transfer. So-called foundation models, pre-trained on large amounts of biological data, offer a promising alternative. Unlike the previously mentioned approaches, foundation models learn robust, generalizable transcriptomic representations. In the field of single-cell research, such models have already demonstrated their effectiveness. Recently, models trained on combined SC- and ST-data have emerged (Nicheformer, scGPT-spatial, sCT), whose broad transcriptomic representations effectively compensate for the restricted marker panels of imaging-based ST assays.
We demonstrate our approach on a complex dataset spanning multiple conditions and various tissue samples, where previously established methods have shown limited success. By combining a foundation model with targeted fine-tuning via Low-Rank Adaptation on a small subset of the original data, we extract enriched embeddings that serve as base for an informed clustering, yielding more structured clusters and an improved downstream analysis.
Co-authors: Nico Kraus, Cristina Ortiz, Christoph Welsch, Marcel H. Schulz
Contact Attendee
Warning: Attempt to read property "user_email" on string in /home/1276969.cloudwaysapps.com/ydbgzhdjeq/public_html/wp-content/plugins/my-conference-now/functions.php on line 2826